Interim Report: Phase 1 Development of a New Methodology to Characterize Truck Body Types along California Freeways Contract Number: 11-316 Principal Investigator: Stephen G. Ritchie, Ph.D. Prepared for: California Air Resources Board Prepared by: Stephen G. Ritchie, Ph.D. Director, Institute of Transportation Studies Professor, Department of Civil and Environmental Engineering University of California, Irvine, 92697-3600 March 20, 2013
Disclaimer: The statements and conclusions in this report are those of the authors and not necessarily those of the California Air Resources Board. The mention of commercial products, their source, or their use in connection with material reported herein is not to be construed as actual or implied endorsement of such products. Acknowledgement: This Interim Report was submitted in partial fulfillment of ARB Contract Number 11-316, Development of a New Methodology to Characterize Truck Body Types along California Freeways by the Institute of Transportation Studies at the University of California, Irvine, under the sponsorship of the California Air Resources Board. This Interim Report was completed in March 2013. i
Table of Contents Abstract... vi Executive Summary... 1 Background... 1 Objectives... 1 Methods... 1 Results... 2 Conclusions... 2 Phase 1: Proof-of-Concept Truck Body Classification Models for Inductive Loop Signature and Weigh-in- Motion Data... 3 Introduction... 3 Background... 3 Objectives... 5 Modification to Proposed Tasks... 5 Data Collection and Processing... 5 Study Sites... 5 Groundtruth System Description and Progress... 7 Body Classification Scheme... 9 Task 1.1 Body Type Classification Model using Inductive Loop Signature Data... 12 Objective... 12 Methods... 13 Results... 14 Discussion... 22 Task 1.2 Body Type Classification Model using Weigh-in-Motion and Inductive Loop Signature Data. 22 Objective... 22 Methods... 22 Results... 25 Discussion... 29 Task 1.3 Hardware Integration of WIM Controllers with Inductive Loop Signature Data... 30 Objectives... 30 Overview of WIM Controllers... 30 Preliminary Investigation... 31 Proposed Integration Design for 1060 WIM Controllers... 32 ii
Comparison of Inductive Signatures between 1060-IST and isinc controllers... 36 Summary and Conclusions... 40 Recommendations... 41 References... 42 Glossary of Terms, Abbreviations, and Symbols... 43 Appendix... 44 iii
List of Figures Figure 1 Project Phases and Tasks Flow Chart... 4 Figure 2 Data Collection Sites... 6 Figure 3 Data Collection System Architecture... 7 Figure 4 Data Collection Setup at Fresno WIM station... 8 Figure 5 Data Groundtruth System User Interface... 9 Figure 6 Examples of Drive Unit Body Types: (a) Single Unit Truck and (b) Semi-Tractor Drive Unit... 12 Figure 7 Examples of Trailer Unit Body Classes: (a) Enclosed Van and (b) 40ft Box Container... 12 Figure 8 Loop signature Classification Model Feature Selection Example... 13 Figure 9 Loop signature Classification Model structure... 14 Figure 10 Confusion Matrix for the Single vs. Multi-Unit Classification Model for Loop Signatures... 14 Figure 11 Confusion Matrix for the Single-Unit Truck Body Classification Collapsed Class Model for Inductive Loop Signatures... 16 Figure 12 Examples of Trailer Axle Configurations: (a) Single Trailers (S), (b) Semi-Trailers (ST), and (c) Multi-Trailers (MT)... 17 Figure 13 Confusion Matrix for the Multi-Unit Truck Classification Collapsed Class Model for Trailer Body Type for Inductive Loop Signatures... 19 Figure 14 Multi-Unit Truck Classification Model for Trailer Body Class... 21 Figure 15 Data Fusion Approach for Combined WIM and Inductive Loop Signature Model... 24 Figure 16 Model Structure for the Combined Inductive Loop Signature and WIM Body Classification Model... 25 Figure 17 Confusion Matrix for FHWA Class 5 Body Classification Model Collapsed Classes for Inductive Loop Signature and WIM... 27 Figure 18 Confusion Matrix for FHWA Class 9 Trailer Body Classification Model Collapsed Classes for Inductive Loop Signature and WIM... 29 Figure 19 Types of WIM controllers deployed in California: (a) 1060 series and (b) isinc Lite... 31 Figure 20 1060 WIM Controller Loop Sensor Module (Left) and IST-222 Loop Sensor Module Adapter (right)... 34 Figure 21 Modified 222 Input File for IST-222 LSM Adapter... 34 Figure 22 Field Processing Unit used to Log Inductive Signature Data from IST-222 Detector Cards... 35 Figure 23 Field Setup of Prototype IST-222 LSM Adapter, Modified 222 Input File and Field Processing Unit at WIM Data Collection Site... 35 Figure 24 Comparison of Hardware Setup for Standalone 1060 WIM Controller (top) and 1060 WIM Controller Integrated with IST-222 Detector Cards for Inductive Signature Data Logging... 36 Figure 25 Experimental Setup: I-5 Yale Referred to as IST - IST (Left), and I-405 Westminster Referred to as IST - isinc (Right)... 37 Figure 26 Signature Transformation: Normalization (a), Shift (b), and Stretch Step (c)... 38 Figure 27 IST-iSinc Inductive Loop signatures Before (left) and After (right) Transformation... 38 iv
List of Tables Table 1 Summary of Data Collection Sites (listed from north to south)... 6 Table 2 Summary of Data Groundtruth Progress as of March 20 th, 2013... 9 Table 3 Classification Scheme for Drive Units... 10 Table 4 Classification Scheme for Trailer Units... 10 Table 5 Single-Unit Truck Body Classification Model for Inductive Loop Signatures CCR Results by Class 15 Table 6 Multi-Unit Truck Body Classification Model for Trailer Body Type for Inductive Loop Signatures CCR Results by Class... 18 Table 7 Available Variables from WIM and Inductive Loop Signatures for Model Development... 23 Table 8 Summary of Feature Extraction/Selection for WIM and Inductive Loop Signature Model... 24 Table 9 FHWA Class 5 Truck Body Classification Model Results for Inductive Loop Signatures and WIM Model... 26 Table 10 FHWA Class 9 Trailer Body Classification Model Results for Inductive Loop Signatures and WIM Model... 28 Table 11 Proposed Pin-out Cross-assignment Between 1060 Loop Sensor Module and IST-222 Inductive Loop Detector Cards... 33 Table 12 Median and 85 th Percentile Errors for Signature Transformation... 39 Table 13 Summary of Modeling Results for the Inductive Loop Signature Only and WIM and Inductive Loop Signature Models... 40 Table 14 Literature Review for Classification Studies... 40 v
Abstract The purpose of this project is to develop a new methodology to characterize truck body types along California Freeways to be later incorporated into ARB s California Vehicle Activity Database (CalVAD) to provide a comprehensive reference for users that integrates a variety of sensor technologies and data sources within a single tool. By combining existing technologies, namely inductive loop signature and weigh-in-motion (WIM) detectors, high resolution truck data can be derived and applied to freight and emissions models. The focus of this interim report is on the methods and deliverables for Phase 1 of the Project, which covers three main tasks: development of an inductive loop signature body classification model (Task 1.1), development of a combined inductive loop signature and WIM data based body classification model (Task 1.2), and investigation of a hardware interface configuration to obtain inductive signatures at WIM stations (Task 1.3). Proof-of-concept body classification models were developed for Tasks 1.1 and 1.2 in order to proceed to model enhancement in Phase 2 of the Project. Data was collected at five WIM sites and one VDS site across California and comprise approximately 34,000 truck samples. The truck body classification scheme, which was initially based on the Vehicle Inventory and Use Survey (VIUS), was re-defined to include expanded classes totaling 29 drive/single unit truck body types and 26 trailer body types reflecting various body types found in the collected data. A data architecture and specialized software user interface was established to link photo, inductive loop signatures, and WIM data collected at the six sites. The resulting processed data was used for model development and validation. A Feed Forward Neural Network model was employed for the proof-ofconcept body classification models. Results of the proof-of-concept models were promising; the inductive loop signature-only model (Task 1.1) had correct classification rates (CCR) of 77% for single unit trucks with 27 classes and 82% for multi-unit trucks with 28 classes. The combined inductive loop signature and WIM model (Task 1.2) had a CCR of 69% for FHWA class 5 trucks with 33 body classes and 85% for the trailer configuration of FHWA class 9 trucks with 21 trailer body classes. Model enhancements such as alternate modeling structures, selection of inductive loop signature features, and additional data collection were identified from the proof-of-concept models of Phase 1 and will be carried out during Phase 2. Lastly, investigation of a hardware interface between inductive signature equipment and WIM controllers yielded a working configuration by which inductive signatures could be collected from a WIM controller. Additionally, comparison of inductive signatures between two versions of the WIM controller hardware units (1060 and isinc) confirmed the compatibility of the signatures and therefore the potential ability of the classification model to produce accurate classification results from both types of controllers. Keywords Truck body classification, inductive loop signature, Weigh-in-motion (WIM), Feed Forward Neural Network, 1060 WIM Controller, isinc WIM Controller vi
Executive Summary Background The purpose of this project is to develop a new methodology to characterize truck body types along California Freeways and will be integrated into ARB s California Vehicle Activity Database (CalVAD) so that a comprehensive overview of commercial and non-commercial vehicle activity can be provided for policy analysis and research. CalVAD currently provides activity estimates of heavy-heavy duty trucks (HHDT) in the state of California by combining data from existing weigh-in-motion (WIM) and vehicle detector station (VDS). For this project, the use of inductive signature technology will be used to develop and deploy advanced vehicle classification models at selected locations for two types of facilities: existing VDS equipped with inductive loop detectors and existing WIM stations. Through this approach, it is possible to obtain higher resolution truck data which can enable more accurate estimates of GHG and other truck emissions, allow for decision makers to make more informed decisions for pavement management across a wider range of locations, and provide insight into the spatial distribution of body types for freight forecasting applications. The project will enhance CalVAD by incorporating a higher level of detail of commercial vehicle body classes, thus expanding the estimation of HHDT activity by CalVAD. The project has been divided into three phases with the following general goals: Phase 1 will develop proof-of-concept body classification models; Phase 2 will enhance the proofof-concept models and create techniques for propagating WIM classifications to VDS locations; and Phase 3 will deploy the developed classification models to selected WIM and VDS sites throughout the State. This report marks the completion of Phase 1 of the project. Objectives The research carried out under this project represents a completely new method for obtaining high resolution truck data and therefore provision of prototype models was deemed necessary for continuation of later project phases. The objectives of Phase 1 were to: 1) establish a proof-of-concept truck body type classification model using inductive loop signature data, 2) establish a proof-of-concept truck body type classification model using fused WIM and inductive loop signature data, and 3) to establish the hardware interface connections between the WIM controllers and the inductive signature detector cards. The proof-of-concept models will serve as a baseline for further model development in Phase 2. Methods Data was collected at seven WIM sites and one VDS site across California. Collected data including inductive loop signatures, WIM controller outputs, and photos, were pre-processed, loaded into a relational database, and processed using a specially developed software user interface to integrate each collected sample s data (signature, WIM data, and photo) and classify the record according to the body classification scheme. The truck body type classification scheme was originally derived from the body classes defined in the 2002 Vehicle Inventory and Use Survey (VIUS), and was further refined to reflect the variety of body types found in the collected data. The processed data records, which follow the adapted body classification scheme, were used for model development and validation. An inductive loop signature feature processing approach and a WIM and inductive loop signature fusion methodology were developed under Phase 1. The proof-of-concept model for loop based classification included data from a single VDS site while the proof-of-concept model for loop and WIM combined data from two sites but focused only on the two most common truck classes: the FHWA class 5 single-unit two-axle 1
trucks and FHWA class 9 five-axle semi-tractor-trailers. Both classification models were based on the Feed Forward Neural Network architecture. Results Deliverables of Phase 1 include a proof-of-concept classification model for body type classifications using signature data (Task 1.1), a proof-of-concept body classification model using WIM and signature data (Task 1.2), as well as a hardware interface specification for implementing signature collection equipment at WIM stations (Task 1.3). For the loop only model, correct classification rates (CCR) were 77% for the single-unit model with 27 classes and 82% for the multi-unit model with 28 classes. For the combined loop and WIM model, the CCRs were 69% for FHWA class 5 trucks with 33 body classes and 85% for FHWA class 9 tractor-trailers with 21 body classes. In addition to the hardware specification, hardware components were also developed and shown to produce a stable data connection between the WIM controller and the inductive signature detector cards at all six of the study sites. Conclusions The proof-of-concept models show promising results for body type classification using inductive loop signatures alone, as well as for inductive loop signatures integrated with WIM data. The proof-ofconcept models are able to correctly classify vehicles into over 20 body configurations which is well beyond the level of detail found in existing studies, which at best contain around 10 body classes with most distinguishing between axle classes only. The models will serve as a baseline for further model development in Phase 2, which will focus on model enhancement. Several model enhancements have been identified as a result of the proof-of-concept modeling including further refinement of heterogeneous body classes during the groundtruthing process, development of alternate methods for signature normalization, pre-processing, feature selection, and fusion approaches, and organization of model outputs by weight class categories established by CARB. Additionally, data collection and further groundtruthing are required for several body classes which currently have a low number of samples. A hardware specification was developed to integrate inductive signature detector cards with 1060 WIM controller systems. This would allow the combined inductive signature and WIM data classification model that will be refined in Phase 2 to be deployed at most of the existing WIM sites in California. A signature comparison test between inductive signature data collected from 1060 WIM controllers retrofitted with inductive signature cards and newer isinc WIM controllers was also performed. Results show that the inductive signatures from both controllers are similar, and indicated that the models developed using data obtained from the 1060 WIM controllers can also be deployed on isinc controllers. Lastly, the hardware specifications for interfacing with newer WIM controller units, i.e. isinc controllers, will be examined for feasibility in Phases 2 and 3. 2
Phase 1: Proof-of-Concept Truck Body Classification Models for Inductive Loop Signature and Weigh-in-Motion Data Introduction Background A significant proportion of goods movement is transported by trucks, and the value and tonnage of goods are expected to grow over time. Trucks have a significant impact on pavement infrastructure, traffic congestion, pollution and quality of life. To provide a better understanding of the behavior of freight-related truck movements, it is necessary to obtain an abundant high resolution truck data. However, current traffic detection infrastructure in the state of California is not designed to collect sufficiently detailed truck data to address these concerns. The prevailing traffic detection infrastructure managed by the California Department of Transportation (Caltrans) comprise of Weigh-In-Motion (WIM) and Vehicle Detector Stations (VDS). There are about 160 WIM sites (by location and direction) within the state that are used primarily for continuous collection of axle weight measures and axle-based classification of trucks 1. These stations are located along freeways and arterials at corridors that experience heavy truck usage. Although there are nearly 8,000 VDS currently being deployed in California, these were not originally designed to provide detailed truck data. The purpose of this project is to develop a new methodology to characterize truck body types along California Freeways. To do so, inductive signature technology will be installed at existing VDS and WIM data collection sites. Through this approach, it is possible to obtain higher resolution truck data which can enable more accurate estimates of GHG and other truck emissions, allow for decision makers to make informed decisions for pavement management, and provide insight into the spatial distribution of body types for freight forecasting applications. The project has been divided into three phases with the following general goals: Phase 1 developed proof-of-concept body classification models; Phase 2 will enhance the proof-of-concept models and create techniques for propagating WIM classifications to VDS locations; Phase 3 will deploy inductive signature capabilities to selected WIM and VDS sites throughout the State. This report marks the completion of Phase 1 of the project. The flow chart shown in Figure 1 depicts the overall scope of the project. At the conclusion of this project, the body classification estimates output from the developed models will be incorporated into ARB s California Vehicle Activity Database (CalVAD). CalVAD [integrates] disparate data sources to develop a comprehensive view of Vehicle Miles Traveled (VMT) and speeds for different vehicle classes in the state of California (CalVAD, 2013). Expanding CalVAD to include inductive signature based vehicle classification models such as those to be developed in this project, will provide researchers and analysts with a single, comprehensive tool by which to monitor and study commercial and non-commercial vehicle traffic at the widest range of locations in California at the highest level of detail, compared to what any of the data sources is able to provide individually. The current project will leverage the database structure, user oriented web interface, established spatial relationships between WIM and VDS stations, and data imputation techniques by which to expand the 1 There are approximately 50 additional WIM sites in California which are designated as PrePass stations used to screen overweight trucks for further assessment at Commercial Vehicle Enforcement Facilities, and not typically used for WIM data collection. 3
classification details from WIM to VDS, while contributing a more refined and detailed body classification scheme at specific locations as an enhancement to CalVAD. Figure 1 Project Phases and Tasks Flow Chart 4
Objectives The research carried out under this project represents a revolutionary method for obtaining high resolution truck data and therefore provision of prototype models was deemed necessary for continuation of later project phases. The focus of this interim report is on the final methods and deliverables for Phase 1 which covers three main tasks: development of an inductive loop signature body classification model (Task 1.1), development of an inductive loop signature and WIM body classification model (Task 1.2), and investigation of a hardware interface configuration to obtain inductive signatures at WIM stations (Task 1.3). The proof-of-concept models will serve as a baseline for further model development in Phase 2. Modification to Proposed Tasks Under the original proposal, Phase 1 prototype models were to be developed using data previously collected by the UCI research team. The existing data consisted of inductive signature records and video recordings of commercial trucks at the San Onofre Weigh and Inspection Facility, but did not contain WIM data. The original proposed plan was to derive axle count from the video to serve as synthetic WIM data for input to the prototype models. This was proposed because the hardware interface configuration between the WIM controller and the inductive signature detector cards was expected to require a longer development time. However, the hardware integration interface was completed earlier than expected and used to collect new data. Therefore, derivation of axle counts was not necessary and instead, actual WIM and signature data collected concurrently was available. Accordingly, the prototype models of Phase 1 are based on actual WIM and signature data. Because the models now depend on actual WIM and signature data, there was a slight delay incurred in the development of these models defined under Tasks 1.1 and 1.2 due to the additional effort required to collect and process the new data. Fortunately, the modification to Phase 1 results in progress on Phase 2, Task 2.1 in which data collection was to be carried out at WIM stations. Hence, there was no expected delay to the overall progress of the project. Data Collection and Processing Study Sites Data collection was performed at seven sites across the State of California. Figure 2 shows the locations of the seven data collection sites. The site name, data type, date of collection, and number of collected truck samples are summarized in Table 1. The WIM site along southbound (SB) I-5 in Irvine, was used for testing and development of the hardware interface configuration between the WIM controller and the inductive signature detector cards. Sites at Redding, Willows, and Fresno were selected to increase the diversity of the truck body types for model development. For example, sites in northern and central California were expected to have a greater presence of logging trucks and agricultural trucks that would not typically be found at WIM stations closer to UCI such as Irvine and San Onofre. Sites at San Onofre and Leucadia were selected for later use in Phase 2 to develop the propagation methodology by which detailed classification based on WIM data can be propagated to VDS equipped with inductive loop signature capabilities. All sites other than the Saigon station contained 1060 series WIM controller units which require use of the configured hardware interface. The Saigon site along I-405 uses a newer WIM controller system called isinc Lite which has the capability to collect inductive signature data without additional detector hardware. Data was collected at this station to establish a comparison between inductive loop signatures produced by various controller types, i.e. inductive signature detector cards at 1060 controllers and isinc controllers. 5
Figure 2 Data Collection Sites Table 1 Summary of Data Collection Sites (listed from north to south) Site Location Site/Controller Type Dates Number of Truck Samples Collected SB I-5 at Redding 1060 WIM Dec. 10-12, 2012 5,110 NB I-5 at Willows 1060 WIM Dec. 10-12, 2012 6,908 SB SR-99 at Fresno 1060 WIM Nov. 7-8, 2012 9,718 SB I-405 at Saigon* isinc WIM and VDS Oct. 9, 2012 97 SB I-5 at Irvine 1060 WIM and VDS Sept. 21, Oct. 2-3, 2012 6,963 SB I-5 at San Onofre PrePass 1060 WIM Jan. 9-10, 2013 1,906 SB I-5 at Leucadia 1060 WIM Jan. 9-10, 2013 4,017 Total 7 sites 34,719 samples * For signature comparison between different controller systems 6
Groundtruth System Description and Progress In addition to inductive loop signature and WIM data, still image data was collected for each passing vehicle. A customized data collection system was developed which consisted of the hardware interface between the WIM controller and inductive signature detector cards as well as a digital SLR camera which captured a series of images of the passing vehicle which was triggered by the inductive signature detector card. The architecture of the customized data collection system is shown in Figure 3. An example of the data collection setup at the Fresno WIM station is shown in Figure 4. Figure 3 Data Collection System Architecture 7
Figure 4 Data Collection Setup at Fresno WIM station A database was designed and developed in PostgreSQL to store and integrate inductive loop signature, WIM record, and photo data. A specially developed software user interface was developed in Visual Basic to efficiently process and classify collected vehicle samples. The user interface was designed to communicate with the database and allow the user to pan through photos and select the vehicle class parameters while also linking inductive loop signature records and WIM data records to the appropriate vehicle record. Figure 5 shows the groundtruth system user interface with an example at the I-5 Irvine site. The photo and left signature image are obtained at the VDS while the center FHWA image and right signature image are from the WIM controller. The user recorded the vehicle class information in five parts: truck axle configuration, trailer axle configuration, truck body configuration, and trailer body configuration, and total number of axles. 8
Figure 5 Data Groundtruth System User Interface Of the almost 35,000 collected vehicle samples, a total of 7,909 records have been fully processed, i.e. photo, vehicle body configuration, inductive loop signature, and WIM record have been assigned to the vehicle record. A summary of the data groundtruth progress as of March 20 th, 2013 is shown in Table 2. For the proof-of-concept modeling in Phase 1, a sub-set of the completed records were used. The subset includes records from the I-5 Irvine and SR-99 Fresno sites. Table 2 Summary of Data Groundtruth Progress as of March 20 th, 2013 Site Location Number of Truck Number of Groundtruth Samples Collected Records Completed SB I-5 at Redding 5,110 411 NB I-5 at Willows 6,908 608 SB SR-99 at Fresno* 9,718 2,931 SB I-405 at Saigon 97 Not used for classification SB I-5 at Irvine* 6,963 3,746 SB I-5 at San Onofre 1,906 0 SB I-5 at Leucadia 4,017 213 Total 34,719 collected 7,909 completed *Sites used for model development in Phase 1 Body Classification Scheme The truck body type classification scheme outlined in the Project Proposal and based on the Vehicle Inventory and Use Survey (VIUS) defined body classes was further refined to reflect the variety of body 9
types found in the collected data. The refined scheme expands on the originally defined 28 VIUS body classes into 29 drive/single-unit truck body types and 26 trailer body types, as shown in Table 3 and Table 4, respectively. Examples of drive/single-unit trucks are shown in Figure 6 and correspond to FHWA classes 4 through 7. Examples of trailer units are shown in Figure 7 and correspond to FHWA classes 8 through 12. Table 3 Classification Scheme for Drive Units Body Category Body Type Ambulance, Street Sweeper, Fire Truck Wrecker Winch and Crane Trucks Garbage Dump Service Trucks Bottom Dump Dumpster Transport Flatbed Tow Truck Concrete Mixer Utility Truck Recreational Vehicle/Coach Busses 30 ft Bus (City or School) 20 ft Bus (City of School) Multi-Stop of Step Enclosed Vans Drop Frame Curtain-side Open Top Low Boy/Drop Frame Platforms Basic Tank Tanks Pneumatic Livestock Specialty Logging Beverage Conventional Cab Semi-Truck Drive Extended Sleeper Cab Units Cab-Over Engine Cab Automobile Transport Cab Table 4 Classification Scheme for Trailer Units Body Category Body Type Enclosed Vans Drop Frame Curtain-side 10
Platforms Tanks Specialty Dump Containers Small Trailers Open Top Agricultural Basic Low Boy Platform with devices Chemical/Dry Bulk Pneumatic Hopper Beverage Pole/logging/pipe Automobile Transport Livestock Belly Dump End Dump Bottom Dump Container Chassis 40ft Box Container 20ft Box Container 20ft Box Container on 40ft Chassis 53ft Box Container Recreational Vehicle or 5 th Wheel Passenger Vehicle Small trailer or dolly 11
(a) (b) Figure 6 Examples of Drive Unit Body Types: (a) Single Unit Truck and (b) Semi-Tractor Drive Unit (a) (b) Figure 7 Examples of Trailer Unit Body Classes: (a) Enclosed Van and (b) 40ft Box Container Task 1.1 Body Type Classification Model using Inductive Loop Signature Data Objective The purpose of Task 1.1 was to demonstrate the ability of inductive signatures to distinguish truck body types. Past research has shown that conventional inductive loop signatures are capable of producing axle-based classification according to the FHWA scheme s 13 classes (Jeng and Ritchie, 2008) as well as by five broad body classes for emissions models (Liu et al, 2011). This project represents the first attempt to produce a highly detailed commercial vehicle body classification model based on conventional inductive loop signatures with a goal of distinguishing VIUS-based body classes. For this reason, a prototype model was developed as part of Phase 1, prior to continuing to Phases 2 and 3 of this project. The prototype, or proof-of-concept, model developed in Phase 1 will be enhanced as part of Phase 2. The final model will be useable at VDS locations which are equipped with inductive signature detector cards. Also, as part of Phase 2, the enhanced model will be linked to the model developed for WIM stations through a propagation approach, thus possibly allowing for a higher level of detailed class and weight information to be estimated at a VDS site. 12
Methods The methodology for the inductive loop signature body classification model was divided into two parts: (1) Feature selection and (2) Model Development. Each is described in this section. Feature Selection A feature set of 29 magnitude differences from the duration and maximum magnitude normalized inductive loop signature were selected as model inputs. Figure 8 shows the process of feature extraction from a normalized signature for single and multi-unit trucks. Each feature represents the difference between two of 30 equally spaced points along the normalized inductive loop signature. Figure 8 Loop signature Classification Model Feature Selection Example Model Development All models are implemented as Feed Forward Neural Networks with a single hidden layer comprised on 15 neurons. The number of input features and output classes vary by model. A neural network approach was chosen based on the previous success of this method with classifying inductive signatures (Tok and Ritchie 2011; Liu et al, 2011) and the ease of implementation through Matlab. The simplicity and effectiveness of the model implementation makes it ideal as a baseline reference for future models developed in Phase 2. Each dataset was proportionally split by vehicle body class into a training set (60% of the total samples), validation set (20%), and testing set (20%). The model structure for the inductive loop signature-based model is presented in Figure 9. The first tier of the model, label as Single vs. Multi-Unit, assigns a vehicle to the appropriate sub-model based on the first three and last three features of the feature set which were found to best distinguish between single and multi-unit truck classes. The second tier of the classification is comprised of two sub-models: (1) a single unit body classification sub-model and (2) and multi-unit trailer body class sub-model. The third tier of the model distinguishes the tractor body class for each trailer type. Each trailer body class output by the multi-unit sub-model was then further broken down into tractor unit body class. In other words, for each trailer body class, 13
there exists a separate sub-model to estimate the tractor unit body class where the number of output tractor classes varies by trailer type. Figure 9 Loop signature Classification Model structure Results Model results are presented for the test dataset in terms of the overall correct classification rate (CCR) which is calculated as the number of correctly classified samples divided by the total number of samples, as well as the individual class CCRs which are calculated as the number of correctly classified samples for the class divided by the total number of samples for that class. While overall and class CCR are provided for all models, the collapsed class model results (which condense common classes into a collapsed class set) are given as confusion matrices in which the columns represent the output classes of the model and the rows represent the target, or true, classes. Thus elements along the diagonal of the matrix represent correct classifications and off-diagonal elements indicate misclassifications. Single vs. Multi-Unit Truck Model The Single vs. Multi-Unit Truck Model consisted of six input features (the tail ends of the inductive loop signature) and two output classes. The overall CCR was 98%. The confusion matrix for the test data is shown in Figure 10. Test Dataset Targets Outputs Single Unit Multi Unit 14 Actual Count Single Unit 451 5 456 99% Multi Unit 14 328 342 96% Predicted Count 465 333 798 98% Figure 10 Confusion Matrix for the Single vs. Multi-Unit Classification Model for Loop Signatures CCR
Single-Unit Truck Body Classification Model The Single-Unit Truck Body Classification Model consists of 29 input features obtained from the inductive loop signature and 27 outputs corresponding to body type classes. The overall CCR was 77%. Table 5 summarizes the CCR by vehicle class for single-unit truck body types. For some classes with a small number of samples, all samples were partitioned into the training dataset and therefore none were available for the test dataset. Figure 11 shows the confusion matrix for the collapsed set of 15 single-unit truck body classes with an overall CCR of 80%. Classes were collapsed based on body type group and misclassification distribution. Table 5 Single-Unit Truck Body Classification Model for Inductive Loop Signatures CCR Results by Class Target Classes No. CCR Samples (%) Sedan or Coupe 0 - SUV 1 0 Minivan 0 - Pick-up (4 tire) 159 92 Pick-up (6 tire) 1 0 12 Passenger Van 71 94 RV (Coach) 2 0 30ft Bus 13 85 20ft Bus 13 69 Conventional Cab 1 0 Extended Cab 2 50 Passenger Vehicles Busses Semi- Tractors Service Trucks Vans Automobile Transport Cab 0 - Basic Platform 34 53 Beverage Truck 2 0 Ambulance, Street Sweeper, Fire truck 4 0 Wrecker 6 17 Winch or Crane Truck 4 0 Garbage Truck 22 100 Dump Truck 12 42 Dumpster Transport 10 50 Flatbed Tow Truck 5 60 Concrete Mixer 3 100 Utility Truck 21 43 Multi-Stop or Step Van 8 38 Enclosed Van 68 79 Curtain-side Van 0 - Chemical/Dry Bulk Tank Truck 3 0 15
Figure 11 Confusion Matrix for the Single-Unit Truck Body Classification Collapsed Class Model for Inductive Loop Signatures Multi-Unit Truck Classification Model for Trailer Body Type The Multi-Unit Truck Classification Model for Trailer Body Type consists of 29 input features from the inductive loop signature and 28 output classes. The output classes are distinguished by both body type and axle configuration. Where applicable, body type configurations were further sub-categorized by axle configuration into single trailers (S), semi-trailer (ST), and multi-trailers (MT). Single trailers generally correspond to FHWA-CA classes 5 and 6 with small trailers and class 14. Semi-Trailers correspond to FHWA classes 8, 9, and 10. Multi-Trailers correspond to FHWA classes 11, 12, and 13. Examples of each axle type are shown in Figure 12. The overall CCR for the Multi-Unit Truck Classification Model for Trailer Body Type was 82% with 28 classes. Individual class CCR are shown in Table 6. The confusion matrix of the collapsed class model is shown in Figure 13 with an overall CCR of 83% for 25 classes. 16
(a) (b) (c) Figure 12 Examples of Trailer Axle Configurations: (a) Single Trailers (S), (b) Semi-Trailers (ST), and (c) Multi-Trailers (MT) 17
Table 6 Multi-Unit Truck Body Classification Model for Trailer Body Type for Inductive Loop Signatures CCR Results by Class Targets No. CCR Samples (%) Basic S 0 - Basic ST 13 38 Basic MT 0 - Low Boy Platform 9 89 Platform with Devices 2 0 Enclosed ST 122 91 Enclosed MT 2 0 Drop Frame 4 50 Open Top 30 93 Curtain-side 1 0 40ft (2 TEU) Box Container 2 0 20ft (1 TEU) Box Container 0-53ft Box Container 8 13 Container Chassis 1 0 Pneumatic Tank ST 0 - Platform Vans Containers Tanks Specialty Dump Non- Semi Pneumatic Tank MT 1 0 Chemical Tank S 1 100 Chemical Tank ST 7 71 Auto. Transp. S 3 0 Auto Transp. ST 4 50 Beverage ST 1 0 Beverage MT 0 - Bottom Dump MT 26 100 End Dump S 10 80 End Dump ST 5 100 RV Trailer/5th Wheel 5 40 Passenger / Small Vehicle 3 0 Small Trailer / Dolly 54 96 18
Figure 13 Confusion Matrix for the Multi-Unit Truck Classification Collapsed Class Model for Trailer Body Type for Inductive Loop Signatures 19
Multi-Unit Truck Classification Model for Tractor Body Type The Multi-Unit Truck Classification Model for Tractor Body Type was based on the observed tractortrailer combinations in the dataset. Each trailer body type has a distinct set of possible tractor types. For example, box container trailers are only pulled by semi-tractor configuration drive units whereas small trailers can be pulled by a larger variety of drive units ranging from passenger vehicles to service trucks. For this reason, the tractor body type was determined separately for each trailer body type. Therefore, the number of output tractor classes varies across trailer body types. The set of 29 input features from the inductive loop signature were used in the model. Figure 14 shows the observed combinations of tractor-trailers along with the resulting CCR by trailer class. Note that for several trailer classes, there was only one observed tractor body type, so no model was needed. Also, note that when there are only a small number of observations for a given trailer class, no model was developed. 20
* Not enough samples, **Only one drive unit body type observed Figure 14 Multi-Unit Truck Classification Model for Trailer Body Class 21
Discussion The results of the body classification model for inductive loop signature data are promising. The CCRs were 77% for the single-unit truck model with 27 single-unit truck body types, and 82% for the multiunit truck model with 28 trailer body types. For single unit classes with large sample sizes, individual class CCRs ranged from 40 to 100%. Low performing classes with more than 10 samples (less than 80% CCR) include 20 ft buses, basic platforms, dump trucks, dumpster transports, and utility trucks. High performing classes include pick-ups, 30ft buses, garbage trucks, concrete mixers, and enclosed vans. Low performance can be traced to expected misclassifications such as utility trucks and pick-ups, and platforms and enclosed vans. For multi-unit trailer classes with large sample sizes, individual class CCRs ranged between 38 and 100%. Basic platform semi-trailers had a low CCR and were commonly misclassified as enclosed van semi-trailers. 53ft Box Containers also had a low CCR and were commonly misclassified as enclosed van semi-trailer. On the other hand, several classes showed high CCRs including low boy platform semi-trailers, enclosed van semi-trailers, open top van semi-trailers, bottom dump multi-trailers, end dump semi-trailers, and small trailers. Task 1.2 Body Type Classification Model using Weigh-in-Motion and Inductive Loop Signature Data Objective Task 1.2 sought to establish proof-of-concept for producing high resolution truck body type data by fusing WIM and inductive loop signature data. By adding length, spacing, and weight information measured at WIM stations into the classification model, it is possible to obtain a higher level of detail than can be achieved by using only inductive loop signature data. The purpose of the two model approach, i.e. an inductive loop signature only model and a combined loop signature and WIM model, is to provide state-wide coverage through separate models useable at VDS sites or at WIM sites and to expand the level of detail known at WIM sites to VDS sites. The proof-of-concept models developed under this task focused on FHWA classes 5 (single-unit trucks) and 9 (semi tractor-trailers) since these represent the most commonly observed truck classes. As part of Phase 2, body classification models will be developed for all FHWA classes. Methods The methodology was divided into two parts: (1) Feature extraction through data fusion and (2) Model development. Feature Extraction The variables available from the WIM controller and inductive loop detector (ILD) card for model development are summarized in Table 7. At the most primitive level, if inductive loop signature data is used to distinguish body type, then weight data can be included to differentiate ARBs weight classes (medium-heavy, heavy, heavy-heavy, etc.) by body type. A more enhanced method, that was adopted in Task 1.2, was to use the WIM axle spacing and vehicle length data to parse the inductive loop signature corresponding to components of multi-unit vehicles such that features are only extracted for the portion of the signature pertaining to the unit being distinguished. For example, given a signature from a semi-tractor-trailer combination, the WIM axle spacing was used to break the signature into the tractor and trailer portions, which would then be feed separately into their corresponding classification models. 22
Table 7 Available Variables from WIM and Inductive Loop Signatures for Model Development Data Type Variables WIM Gross Vehicle Weight (GVW) Inductive Signature Vehicle Length Right and Left Axle Weights Axle Spacing FHWA Class Speed Change in inductive magnitude over time Since FHWA class 5 comprises mainly of single-unit trucks, the body type characteristics are contained in the entire signature, so the signature does not need to be parsed. Instead, the feature set of 30 interpolated magnitude differences was enhanced by adding the vehicle length and spacing of between axles 1 and 2, 2 and 3, and 3 and 4 as separate features. For FHWA class 9 semi tractor-trailers, an inductive loop signature parsing method was developed. Since there are many possible combinations of tractors and trailers, separating the full inductive signature into their tractor and trailer components can yield a higher classification performance, as the variability in one component such as the tractor unit features would not adversely influence the classification of the trailer component. Figure 15 shows an example of how the WIM axle spacing data is used to parse the tractor and trailer portions of an inductive loop signature. The WIM controller axle spacing measurement between the steering and drive axles is applied to the length normalized inductive loop signature to extract the tractor portion and trailer portion. 10 and 20 interpolated magnitude points are then calculated from the tractor and trailer components of separated signature, respectively. Lastly, the feature set is derived as the change in magnitude between each pair of equally spaced interpolated magnitudes. A summary of the feature selection for FHWA classes 5 and 9 is shown in Table 8. 23
Figure 15 Data Fusion Approach for Combined WIM and Inductive Loop Signature Model Table 8 Summary of Feature Extraction/Selection for WIM and Inductive Loop Signature Model Vehicle Type Approach Feature Set FHWA Class 5 30 interpolated magnitude differences from the length normalized inductive signature 30 Signature Features, Axle 1-2 spacing, Axle 2-3 spacing, Axle 3-4 spacing, Vehicle Length FHWA Class 9 Parsed inductive signature based on spacing of 1 st and 2 nd axle measurements from which the interpolated magnitude differences are computed separately for tractor and trailer units 10 Signature Features for Truck, Axle 1-2 spacing, Axle 2-3 spacing 20 Signature Features for Trailer, Axle 3-4 spacing, Axle 4-5 spacing, Vehicle Length Model Development The structure of the combined inductive loop signature and WIM body classification model is shown in Figure 16. The first tier of the model, labeled as FHWA Class from WIM, is obtained as a direct output from the WIM controller. The WIM controller classifies a vehicle into one of 15 FHWA-CA axle classes based on the gross vehicle weight, axle count, and spacing between each axle through a look-up table (Lu et al 2002; see Appendix). Note that, Class 15 consists of vehicles that cannot be categorized into classes 1 through 14 based on spacing and weight or due to system error. The second tier of the model consisted of two sub-models; one for FHWA class 5 trucks and one for FHWA class 9 trailers. The third tier contained a separate body classification model for each trailer class. 24
Classification models for the Class 5 and Class 9 sub-models are based on the Feed Forward Neural Network architecture with a single hidden layer which comprises of 15 neurons. The number of input features and output classes vary by model based on the parsing approach for the inductive loop signatures and the number of body classes to be distinguished. Each dataset was proportionally sampled by vehicle body class into a training set (60% of the total samples), validation set (20%), and testing set (20%). Figure 16 Model Structure for the Combined Inductive Loop Signature and WIM Body Classification Model Results Model results are presented for the test dataset in terms of the overall CCR, which was calculated as the number of correctly classified samples divided by the total number of samples, as well as the individual class CCRs which were calculated as the number of correctly classified samples for the class divided by the total number of samples for that class. While overall and class CCR are provided for all models, the collapsed class model results (which condense common classes into a collapsed class set) are given as confusion matrices, where columns represent the output classes of the model and rows represent the target, or true, classes. Thus elements along the diagonal of the matrix represent correct classifications and off-diagonal elements indicate misclassifications. FHWA Class 5 Body Classification Model The FHWA Class 5 Truck Body Classification Model consists of 34 input features from the inductive loop signature and WIM controller and 33 output body classes. Input features include 30 interpolated magnitude differences, vehicle length, spacing between axles 1 and 2, axles 2 and 3, and axles 3 and 4. The output classes are a combination of single-unit truck types, e.g. Class 5 with no trailers, in addition to single-unit trucks with single trailers (S). Additionally, the body types include semi-trailer (ST) configuration which have been misclassified by the WIM controller as FHWA class 5. 25
The overall CCR was 69%. Table 9 summarizes the CCR by vehicle class for FHWA class 5 trucks. For some classes with a small number of samples, all samples were partitioned into the training dataset during the sampling procedure, and therefore none are available in the test dataset. Figure 17 shows the confusion matrix for the collapsed set of 13 body classes which had an overall CCR of 72%. Classes were collapsed based on body type group and misclassification distribution. Table 9 FHWA Class 5 Truck Body Classification Model Results for Inductive Loop Signatures and WIM Model Targets Samples CCR (%) 4 Tire Pick-up 6 67 4 Tire Pick up S 4 75 6 Tire Pick-up 0-6 Tire Pick-up S 3 33 12 Pass. Van 11 100 12 Pass. Van S 0 - Recreational Vehicle (Coach) 3 0 Recreational Vehicle (Coach) S 1 0 30ft Bus 4 50 20ft Bus 10 70 Passenger Vehicles Busses Semi- Tractors Conventional Cab 1 0 Conventional Cab ST 0 - Platforms Basic Platform 21 38 Basic Platform S 1 100 Beverage Truck 1 0 Ambulance, Street Sweeper, Fire Trucks 3 0 Ambulance, Street Sweeper, Fire Trucks S 2 0 Wrecker 5 100 Wrecker S 1 0 Winch or Crane Truck 2 0 Winch or Crane Truck S 0 - Dump Truck 1 0 Flatbed Tow Truck 4 75 Flatbed Tow Truck S 0 - Utility Truck 15 87 Utility Truck S 0 - Multi-Stop or Step Van 7 43 Enclosed Van 53 94 Enclosed Van S 1 0 Curtain-side Van 0 - Open Top Van S 0 - Chemical/Dry Bulk Tank Truck 1 0 S = Single Trailer, ST = Semi-Trailer Service Trucks Vans 26
Figure 17 Confusion Matrix for FHWA Class 5 Body Classification Model Collapsed Classes for Inductive Loop Signature and WIM FHWA Class 9 Trailer Body Classification Model The FHWA Class 9 Trailer Body Classification Model consists of 21 input features from the inductive loop signature and WIM controller and 21 output trailer body classes. Input features include 20 interpolated magnitude differences of the parsed inductive loop signature that correspond to the trailer, and vehicle length. The output classes are a combination of semi tractor-trailer configurations, e.g. FHWA Class 9, and non-semi configured single unit trucks with small trailers (S). The latter category represents vehicles that were misclassified by the WIM controller. Hence, one important benefit of the body classification model developed in this Task is its ability to identify vehicles that were misclassified by the WIM controller as Class 9 tractor-trailers. The overall CCR for the Class 9 trailer body classification model was 85%. Table 10 summarizes the CCR by vehicle class for Class 9 tractor-trailers and non-semi-trucks. Due to the small number of samples in some classes, all samples were partitioned into the training dataset, hence none are available in the test dataset. Figure 18 shows the confusion matrix for the collapsed set of 17 body classes with an overall CCR of 86%. Classes were collapsed based on body type group and misclassification distribution. 27
Table 10 FHWA Class 9 Trailer Body Classification Model Results for Inductive Loop Signatures and WIM Model Targets # Samples CCR (%) Basic 19 89 Low Boy Platform 10 90 Platform with Devices 2 0 Enclosed 153 98 Drop Frame 3 0 Open Top 27 70 Curtain-side 2 0 40ft Box Container 4 75 20ft Box Container 0-20ft on 40ft Chassis 0-53ft Box Container 8 50 Container Chassis 1 0 Pneumatic Tank 2 0 Chemical Tank 18 67 Auto Transp. 4 75 Agricultural Van 0 - Livestock 1 100 End Dump 4 50 Platform Vans Containers Tanks Specialty Non- Semi Auto. Transp. S 1 100 Passenger / Small Vehicle S 0 - Small Trailer / Dolly S 0-28
Figure 18 Confusion Matrix for FHWA Class 9 Trailer Body Classification Model Collapsed Classes for Inductive Loop Signature and WIM Discussion The results of the proof-of-concept body classification model for combined inductive loop signature and WIM data are promising. The CCRs were 69% for the FHWA Class 5 body classification model with 33 body types and 85% for the FHWA Class 9 body classification model with 21 trailer body types. For FHWA Class 5 single-unit truck body classes, individual class CCRs ranged from 38 to 100% for classes with at least five samples. Low performing classes include 30 ft buses, basic platforms, and multi-stop vans. High performing classes include pick-ups, 20ft buses, wreckers, utility trucks, and enclosed vans. From the results shown in the confusion matrix, low performance can be traced to misclassifications between platforms and enclosed vans and in other cases to a low number of test samples. For FHWA Class 9 trailer body classes, individual class CCRs ranged between 50 and 100%, for classes with more than five samples in the test dataset. Chemical tank semi-trailers had a low CCR and were commonly misclassified as platform semi-trailers. 53ft Box Containers also had a low CCR and were commonly misclassified as enclosed van semi-trailers. On the other hand, several classes showed high CCRs including basic and low boy platform semi-trailers, enclosed van semi-trailers, open top van semitrailers, 40ft box containers, automobile transport semi-trailers, livestock semi-trailers, and non-semitrailers such as single trailer automobile transports. 29
Task 1.3 Hardware Integration of WIM Controllers with Inductive Loop Signature Data Objectives One of the main purposes of this study is to develop a truck classification model that is based on the integration of WIM and inductive signature data. Hence, the main purpose of this Task 1.3 was to investigate the possibilities of integrating inductive signature data with WIM data for either or both the earlier 1060 series and the latest isinc WIM controllers. The deliverable of this task as defined in the proposal was the hardware integration specification of at least one of these WIM controllers. Another objective of Task 1.3 was to determine if the classification models developed from this study could be applied to both controllers. Since the data for model development was obtained solely from 1060 controllers, there was a need to demonstrate that the signature data obtained from the 1060 controllers are similar to the isinc controllers to ensure that the models are cross-platform compatible. Overview of WIM Controllers Two main types of WIM controllers are currently deployed in the State of California: the earlier DOSbased 1060 series controllers and the current Linux based isinc family of controllers, which include the isinc WCU-II and isinc WCU-3 Lite. Pictures of the 1060 and isinc Lite controllers are shown in Figure 19. The main distinction between the controllers for the purpose of this study is in their built-in ability to log inductive signature data. The loop sensor module (LSM) of the 1060 WIM controllers is designed only to obtain conventional bivalent inductive loop data. On the other hand, the LSM of the isinc Lite controller has the ability to obtain inductive signature data. The caveat for the isinc controller however, is that inductive signature data is currently designed only for diagnostic and troubleshooting purposes. Hence, the inductive signature data can only be manually logged when the system is in diagnostic mode, and is not currently available as an operational feature within the system. 30
(a) (b) Figure 19 Types of WIM controllers deployed in California: (a) 1060 series and (b) isinc Lite Preliminary Investigation Two methods of signature integration for the 1060 series WIM controller were initially considered. The first and preferred option was to develop an adapter interface to replace the existing bivalent module with IST-222 (IST) inductive loop signature detector cards. The second less desirable option was to splice into the inductive loop sensor leads such that the inductive loop leads would directly connect to both the 1060 LSMs as well as IST-222 detector cards. However, the concern was that the splice would adversely affect the inductive loop measurement accuracy for both detector cards. The initial approach to investigating the potential of inductive signature operations of the isinc controller was to perform a trial inductive signature data collection to investigate the quality and compatibility of the inductive loop signature data obtained from isinc controllers. This would determine if the inductive signature data obtained from isinc controllers are suitable for vehicle classification applications. However, since the inductive signature function of the isinc controller is a proprietary feature, further development of the inductive signature feature for operations would depend on International Road Dynamics (IRD), the developers and vendors of isinc, which according to them would require developmental time and costs. Because of the perceived challenges to get isinc controllers operational with inductive loop signature data within the proposed timeline of this study, the initial hardware integration effort was focused on the 1060 series controller. Furthermore, 1060 series controllers are currently deployed at about 80 percent of current WIM sites within the State of California. Hence, despite their age, a hardware integration solution with the 1060 series controllers would be applicable to a much larger number of candidate sites currently available for deployment consideration. Nevertheless, an investigation was still performed to compare the signatures between IST-222 detector cards collected at WIM controller locations and signature data directly logged from isinc controllers to ensure that the vehicle 31
classification models developed from this study using the data obtained from 1060 series WIM controllers would be implementable on the isinc deployed WIM locations when the inductive loop signature feature concerns of the isinc controller have been addressed. Proposed Integration Design for 1060 WIM Controllers From the pin-out specifications of the 1060 LSM obtained from IRD as well as the IST-222, it was determined that the IST-222 was a viable candidate for adapting to the 1060 LSM. However, the 1060 LSM is a four channel detector, while the IST-222 only possesses two channels. Hence, the adapter would have to replace an existing 1060 LSM with two IST-222 detector cards. Table 11 shows the proposed cross-assignment between the 1060 WIM controller LSM and two IST-222 ILD cards. A prototype IST-222-to-LSM adapter was subsequently designed following this proposed assignment scheme to adapt two IST-222 detector cards to replace a single 1060 WIM LSM. The prototype adapter was developed as shown in Figure 20, which comprises of a 64-pin connector mounted on a PCB sharing an identical form factor of the 1060 WIM LSM and connected to two 15-pin VGA-style female connectors, with each connector designed to connect to an IST-222 detector card via a 44-pin edge connector interface. 32
Table 11 Proposed Pin-out Cross-assignment Between 1060 Loop Sensor Module and IST-222 Inductive Loop Detector Cards 1060 Pin IST Card No./Pin 1060 Pin IST Card No./Pin 1 33 2 34 #1/H 3 #1/F 35 4 36 5 #1/D 37 6 38 #1/E 7 #1/7, #1/20, #2/7, #2/20 39 8 40 9 #1/X 41 10 42 #1/W 11 #1/A, #2/A 43 12 44 #1/J 13 #1/K 45 14 46 #1/L, #2/L 15 47 16 48 #2/H 17 #2/F 49 18 50 19 #2/D 51 20 52 #2/E 21 53 22 54 23 #2/X 55 24 56 #2/W 25 57 26 58 #2/J 27 #2/K 59 28 60 29 #1/C, #2/C 61 30 #1/B 62 #2/B 31 63 32 64 33
Figure 20 1060 WIM Controller Loop Sensor Module (Left) and IST-222 Loop Sensor Module Adapter (right) A modified 222 input file was also fabricated to interface with the IST-222 LSM adapter as shown in Figure 21. The back panel of the input file was removed and replaced with individual 44-pin edge connectors for each IST-222 detector card. Each edge connector was wired to a male VGA style connector for the purpose of interfacing with the LSM adapter. Because detector cards plugged into the modified 222 input file draw power directly from the 1060 WIM controller, the modified input file was not connected to a power supply. Inductive loop signature data was logged into a field processing unit (shown in Figure 22) via the USB port located on the front panel of each IST-222 detector card. Figure 23 shows the prototype LSM adapted and modified 222 input file deployed at 1060 WIM controller sites for data collection efforts in Tasks 1.1 and 1.2. Schematic layouts showing a comparison of the hardware setup for a standalone 1060 WIM controller and the proposed integration with IST-222 detector cards is shown in Figure 24. Figure 21 Modified 222 Input File for IST-222 LSM Adapter 34
Figure 22 Field Processing Unit used to Log Inductive Signature Data from IST-222 Detector Cards Figure 23 Field Setup of Prototype IST-222 LSM Adapter, Modified 222 Input File and Field Processing Unit at WIM Data Collection Site 35
Figure 24 Comparison of Hardware Setup for Standalone 1060 WIM Controller (top) and 1060 WIM Controller Integrated with IST-222 Detector Cards for Inductive Signature Data Logging Comparison of Inductive Signatures between 1060-IST and isinc controllers The main objective of this sub-task under Task 1.3 was to determine the similarity of inductive loop signature data collected at two types of WIM controllers, namely, 1060 and isinc, which are deployed in the State of California. As described in the previous sections, the 1060 controllers will need to be equipped with IST-222 ILD cards to enable inductive signature data logging, while isinc controllers have the built-in capability for sampling and logging inductive signature data, although this feature is currently available only in diagnostic mode. Because the inductive signature data logging hardware differs and performs under different sampling rates for these two controllers, there may be some differences in the quality of the inductive signatures obtained. This evaluation task was performed to determine if the models, which were developed from data collected from the 1060 series WIM controllers, are compatible with the isinc controllers, thus eliminating the need to develop different classification models for each controller type. The ideal comparison would require both types of controllers, i.e. a 1060 and an isinc to be located in close proximity to each other. This would allow samples of inductive loop signatures of the vehicles to be captured by both controllers. Hence each vehicle that traverses the study site would generate two signatures one from each controller type for direct comparison. Unfortunately, such a configuration is not available in the State of California. Hence, an alternative experimental setup was designed using two independent sites one for each controller type with an adjacent VDS station equipped with IST-222 ILD cards. Unlike the inductive loop sensors at the WIM sites which have a 6 foot square loop configuration, the inductive loop sensors 36
at the adjacent VDS sites have a 6 foot round loop configuration. Consequently, inductive signature features obtained between the WIM and VDS sites will inherently possess some differences due to the geometric differences in the loop configuration. However, it can be concluded that inductive signatures obtained from the two WIM controller types are compatible, if the difference in inductive signature pairs obtained between each WIM controller and their adjacent VDS location are similar, indicating that the differences are attributed only to the loop geometry configuration between WIM and VDS locations, and not in the controller hardware itself. To compare inductive signatures between the two hardware systems, a statistical test which removed the effects of differing loop configurations (round vs. square) was performed. Experimental Setup The experimental setup for collecting inductive loop signatures from the 1060 WIM controller equipped with IST detector cards and the isinc WIM controller at Yale (I-5 SB) and Westminster (I-405 SB) sites is shown in Figure 25. At the Yale site, the VDS and adjacent WIM stations were both equipped with IST inductive loop detector cards. This site is referred to as IST-IST. At the Westminster site, the VDS station was equipped with IST cards while the WIM station was equipped with isinc LSMs. Hence, the site is referred to as IST-iSinc. Since the separation between the VDS and WIM stations at both sites was less than 100 feet, each of these sites were ideal for obtaining samples of inductive signatures from the same truck across the WIM and adjacent VDS station. 400 and 97 vehicle records were used in the comparative analysis at the Yale and Westminster locations, respectively. The analysis was performed in two steps: (1) Signature Transformation and (2) Statistical Comparison. Figure 25 Experimental Setup: I-5 Yale Referred to as IST - IST (Left), and I-405 Westminster Referred to as IST - isinc (Right) Signature Transformation In order to remove effects from the geometric loop configurations, scaling of the signatures by the controller/detector cards, speed differences, and lateral positioning over the loop between the two sites, a signature transformation step preceded the statistical comparison. Signature normalization was performed as the first step to remove the effects of different detection thresholds and sensitivities 37
between the two systems. To normalize each signature, the sampled magnitudes and durations were divided by their maximum values. The second step, referred to as shift and stretch, involved using the WIM signature as a reference and horizontally shifting and stretching the VDS signature to achieve the best fit more specifically, the minimum sum of vertical differences between the two signatures. (a) (b) (c) Figure 26 Signature Transformation: Normalization (a), Shift (b), and Stretch Step (c) At the aggregate level, to compare the effect of the transformation, the average median error was used in conjunction with visual examination to determine the ability of the shift and stretch approach to improve signature similarity. The average median error between the IST-iSinc and the IST-IST was defined as the average of the 50th percentile errors from the 97 vehicle records at Westminster and the 400 vehicle records at Yale, respectively. As shown in Figure 27, the magnitude difference between IST and isinc signatures were significantly reduced after the transformation step. The effects of the signature transformation are numerically summarized in Table 12. Both the median error and 85 th percentile (85 th %) are provided to measure the effects of the transformation. For the IST-IST case, the average median error before the transformation was 0.030 which decreased to 0.021 after the transformation. For the IST-iSinc case, the average median error was 0.048 before the transformation, which reduced to 0.021 after the transformation. Therefore, the transformation approach had positive effects. Figure 27 IST-iSinc Inductive Loop signatures Before (left) and After (right) Transformation 38
Table 12 Median and 85 th Percentile Errors for Signature Transformation Average Standard Average Std Median Error Deviation (Std) 85 th % Error IST-IST Before 0.030 0.025 0.072 0.046 After 0.021 0.012 0.052 0.029 IST-iSinc Before 0.048 0.019 0.164 0.062 After 0.021 0.013 0.056 0.032 Statistical Comparison A two tailed t-test of the log transformation of the median and 85 th percentile errors from the IST-IST and IST-iSinc was used to statistically compare the transformed inductive loop signatures. A log transformation was applied to normalize the median and 85 th percentile error to satisfy the normality requirements of the t-test. The null and alternate hypothesis for the t-test is shown in Equation 1. H µ µ 0 : e _ IST = e _ isinc H 1 : µ e _ IST µ e _ isinc Equation 1 Critical values for the t-test were 0.5532 (p= 0.5820) for the median error and -1.3505 (p=0.1791) for the 85 th percentile statistical comparisons at the 5% level of significance. Thus, the statistical tests confirm that the inductive signatures from the 1060 WIM controller equipped with IST-222 inductive loop detector cards and isinc LSMs were similar. Specifically, at the 5% significance level, inductive signatures from the 1060 and isinc were not statistically different. Therefore, this result concludes that the classification model developed from IST-222 inductive loop detector cards at the 1060 WIM sites is expected to be applicable to WIM sites equipped with isinc controllers. 39
Summary and Conclusions The proof-of-concept models developed under Phase 1 (Tasks 1.1 and 1.2) show promising results for body classification using inductive loop signature only and inductive loop signatures combined with WIM data. A summary of the proof-of-concept model results in terms of the correct classification rate (CCR) and number of body classes is provided in Table 13. It should be noted that direct comparison between models is not relevant due to the different datasets used for model development as well as the level of variability of body types for the data used in each model. Although the overall CCRs of the proof-ofconcept models are high, individual class CCRs were rather low for several key classes such as box containers and tanks. This is possibly due to the inability of the feature set to distinguish between similar body types or heterogeneity of body configurations with the classification scheme established for this project. Table 13 Summary of Modeling Results for the Inductive Loop Signature Only and WIM and Inductive Loop Signature Models Model Loop Only Model (Task 1.1) WIM + Loop Model (Task 1.2) Structure Single Units Multi-Units FHWA 5 FHWA 9 # Classes 27 28 33 21 CCR 77% 82% 69% 85% In comparison to the previous classification methods developed and summarized in Table 14, the body classification models developed under Tasks 1.1 and 1.2 were able to distinguish between many more classes with higher CCR. For example, even the most detailed model by Tok and Ritchie (2011) contains only 10 trailer unit types, compared to the proof-of-concept model of Task 1.1 with 28 multi-unit truck trailer body classes and Task 1.2 with 21 multi-unit truck trailer body classes. Table 14 Literature Review for Classification Studies Author (year) Classifications CCR Ki and Baik (2006) 5 Classes 91.5% Sun and Ritchie (2000) 7 Classes 82-87% Jeng and Ritchie (2008) 13 Classes 93% Liu et al. (2011) 5 MOVES Classes 97.6% Tok and Ritchie (2011) 9 Drive and 10 Trailer Unit 81.8% The integration of IST-222 inductive loop detector cards with 1060 series WIM controllers was accomplished under Task 1.3. This enables 1060 WIM controllers to collect inductive signature data simultaneously with WIM data. A comparison test of inductive signatures from a 1060 WIM controller and an isinc controller also showed that the signatures were similar. This shows that the truck classification models developed in this study would be applicable to both controller types which are deployed in the State of California. In conclusion, in addition to the first of their kind detailed body classification models developed in this study, the data collected and processed present a valuable source for truck information which can be leveraged beyond the classification model development. 40
Recommendations Based on the performance and misclassifications resulting from the proof-of-concept models as well as comments from ARB, the following modeling enhancement approaches have been identified: 1. Inclusion of passenger vehicles in the dataset. To facilitate the initial groundtruthing efforts, passenger vehicle photos were removed from the dataset and therefore modeling datasets did not include a significant number of passenger car records. Passenger vehicle samples wil be added to the data set in Phase 2 of this study to ensure that the developed models can distinguish passenger vehicles from trucks in general traffic. 2. Organization of body classes by weight categories defined by ARB for emissions modeling. WIM data provides gross vehicle weight (GVW) for each vehicle. This means that body class estimates resulting from the body classification models developed under this project can be further divided between weight categories such as those defined by the ARB for emissions estimation (heavy, heavy-heavy, etc.). In Phase 2, body classification results will be presented in this manner. Additionally, for the Phase 2 propagation task, weight classes will be propagated from WIM to VDS locations based on the distributions of weight within each body class. 3. Separation of two and three axle tractor body classes. Based on the suggestions of ARB, the tractor classes shown in the inductive loop signature only multi-unit truck classification model and the combined inductive loop signature and WIM model will be further refined to include axle configuration during Phase 2 of this study. This should allow for distinction between two and three axle tractors which have different emissions rates. 4. Further refinement of heterogeneous classes within single and multi-unit classes. This will be carried out in Phase 2 of this study by adding body type sub-classes to what we expect are largely varying body types such as platforms and enclosed vans and then groundtruthing according to the expanded scheme. 5. Development of alternate methods of classification. Currently, the proof-of-concept models were implemented as feed forward neural networks with a single hidden layer of 15 neurons. While this is sufficient for proof-of-concept, improved models will be developed in Phase 2. Probabilistic neural network algorithms will be explored in addition to the current feed forward neural network models used in Phase 1.. 6. Exploration of alternate signature normalization and parsing approaches. Signatures in the proof-of-concept models were normalized by their individual peak magnitudes and durations. This approach removes some of the class specific features from the signature. As a model enhancement, an adaptive normalization approach will be explored in Phase 2, which uses station level data or class specific data to normalize the signature. Also, the proof-of-concept model used the same feature set for all sub-models. As an enhancement additional features may be added or the feature set will be reduced to only the features necessary for distinguishing classes within the sub-model. Finally, the parsing approach can be further refined for FHWA class 9 vehicles and adapted for all remaining FHWA classes not included in the proof-of-concept models. 7. Additional data collection efforts to boost number of samples for underrepresented truck types. As shown in the summary tables in the modeling results section, several classes had only a few test samples. To expand the model, it is necessary to have a representative number of samples for each body class. This will be carried out by further groundtruthing efforts at the six stations listed in the Data Description, as well as possible additional data collection efforts at new locations, if required, during Phase 2 of this study. 41
References The California Vehicle Activity Database (CalVAD) Viewer, California Traffic Management Laboratories (CTMLabs), http://calvad.ctmlabs.net/ Accessed September 10, 2013. Jeng, S.T. and Ritchie, S.G., Real-time vehicle classification using inductive loop signature data, Transportation Research Record No. 2086, 2008, pp 8-22. Ki, Y.K., Baik, D.K., 2006. Vehicle-Classification Algorithm for Single-Loop Detectors Using Neural Networks, IEEE Transactions on Vehicular Technology 55(6). Liu, H., Tok, Y.C.A., and Ritchie, S.G., Development of a Real-Time On-Road Emissions Estimation and Monitoring System, IEEE ITSC 2011 Conference Paper, October 5-7, 2011, Washington, D.C. Lu, Q., Harvey, J., Le, J., Quinley, R., Redo, D., and Avis, J. (2002), Truck Traffic Analysis using Weigh-In- Motion (WIM) Data in California, Pavement Research Center, Institute of Transportation Studies, University of California, Berkeley, CA. Sun C. and Ritchie, S.G., 2003. Inductive Classifying Artificial Network for Vehicle Type Categorization, Computer-Aided Civil and Infrastructure Engineering 18, 161-172. Tok, A., Ritchie, S.G., 2010. Vector Classification of Commercial Vehicles using a High Fidelity Inductive Loop Detection System, In proceedings of the 89h Annual Meeting of the Transportation Research Board (DVD), Washington D.C. Vehicle Inventory and Use Survey (Discontinued), U.S. Census Bureau, 2002 Economic Census. 42
Glossary of Terms, Abbreviations, and Symbols Abbreviation Definition 1060 1060 Weigh-in-motion controller hardware unit CCR Correct Classification Rate FHWA Federal Highway Administration IRD International Road Dynamics isinc/isinc Lite isinc Weigh-in-motion controller hardware unit IST-222/IST Inductive Signature Technologies 222 detector card LSM Loop Signature Module of the Weigh-in-motion controller MT Multi Trailer Axle Configuration S Single Trailer Axle Configuration SB/NB Southbound/Northbound ST Semi-Trailer Axle Configuration UCI University of California, Irvine USB Universal Serial Bus VDS Vehicle Detection Station WCU Weigh-in-motion Controller Unit WIM Weigh-in-motion 43
Appendix FHWA-CA Classification Scheme for Commercial Vehicle Classes 4 through 14 (from Lu et al 2002) 44