Enhancement of Automobile Safety using Machine Learning

Transcription

1 Enhancement of Automobile Safety using Machine Learning Prof. Govindaraj M 1, Nischal S 2 1,2 Department of MCA, New Horizon College of Engineering Abstract The Goal of Machine Learning is to improve the performance of a System based on previous experience, given that the electronic safety systems that make Automobiles safer and more reliable, the performance of these systems remains constant irrespective of the various external conditions that impact them. The performance of these systems can be optimized to better suit the external conditions by constantly analyzing the data collected by the external factors that act on these systems. There are numerous electronic safety systems that rely on Electronic Control Unit for the input and output of the system, the Electronic Control Unit can be programmed to constantly store the input data and analyze the same using machine learning techniques such as Artificial Neural Networks, KNN algorithm etc. to produce an output that is more optimal to the given situation. In this paper we implement machine learning techniques to few of the Electronic Safety Systems and discuss how these changes will affect the performance of these systems and what advantages are gained over the traditional system. Keywords Antilocking Braking System, Electronic Brake distribution, Machine Learning, Automated Testing, Artificial Neural Network, Electronic Control Unit, Training Data, Corner Case. I. INTRODUCTION Significant development in Machine Learning (ML) techniques like Artificial Neural Networks (ANNs) over the last decade has enabled the development of safety-critical ML systems like autonomous cars that adapt their behavior based on their environment as measured by different sensors (e.g., camera, Infrared obstacle detector, etc.). Several major car manufacturers including GM, Ford, Tesla, and BMW are building and actively testing Machine Learning applications. Recent results show that Machine Learning has become very efficient in practice and already driven millions of miles on Autonomous cars without any human intervention. Twenty US states including California, Texas, and New York have recently passed legislation to enable testing and deployment of autonomous vehicles. However the pace at which the development of these products are going on is very slow and seems unlikely to be available for consumer use until the second half of the 21st century, even when these driverless cars arrive on the roads it will take many more decades for them to completely replace human drivers, in the meantime the same technologies that give a hope that driverless cars are possible can be applied to existing diver aid electronics to slowly began the transition of making the cars think, process, and react. These applications ones proven to be useful, can contribute hugely to the development of autonomous vehicles in the long run. At a conceptual level, the erroneous corner-case behaviors in ML based software are analogous to logic bugs in traditional software. Similar to the bug detection and patching cycle in traditional software development, the erroneous behaviors of ANNs and other MI techniques, once detected, can be fixed by adding the error-inducing inputs to the training data set and also by possibly changing the model structure/parameters [1][2], unlike traditional software where the program logic is manually written by the software developers, Machine Learning software automatically learns its logic from a large amount of data with minimal human guidance. Moreover, the logic of a traditional program is expressed in terms of control flow statements while Machine Learning uses previous experience in similar input cases to predict the output. These differences make machine learning based applications DOI: /IJRTER GENTZ 454

2 prone to lot of errors until it has access to sufficient information to process and predict the correct output. Large software companies like Google that have already deployed machine learning techniques in several production-scale systems including speech recognition, image search, etc. have noted that these applications improve their efficiency over time as users get used to the behavior of the machine and the outputs gets more saturated. Challenges that we may face during the implementation of machine learning into automobile safety. 1. Dealing with Corner Case Behaviors 2. Behavior of the system when the training data is very small. 3. Intrusiveness of the system may affect driving style of an individual in the initial stages. II. RELATED WORKS As stated in the introduction many major Automobile manufacturers are in the race to develop driverless vehicles for the future and these projects greatly rely on machine learning algorithms. Machine learning is specifically applied in safety systems of these vehicles because it is difficult to produce traditional flow based software as these system act on varying real-world conditions. These systems gather data from the various sensors that supply information to the Electronic Control Unit (ECU) that processes the data in real time speeds and predict the necessary output to handle the situation. Machine Learning System use existing large amount of data called as training data, to study the required output under various circumstances and create a virtual map of the input output pairs to predict the output when the algorithm is under use. This virtual map gets more and more accurate as the amount of training data increases. As Stated above there are many challenges that we may face when using these approach and experts have relied on traditional software to monitor the corner case behaviors, and let the machine learning come into play only when the amount of training data available is sufficient to predict a reasonable output. In the initial stages the System is completely handled by Traditional software and control is passed to machine learning only when the amount of training data is sufficient for the algorithms to study. III. SYSTEM ANALYSIS 3.1. Problem Statement: To come up with Automobile Safety Systems that are based on traditional Software Systems and to replace the software with Machine Learning based algorithms that improve efficiency and behavior of these system based on the Driver s style and other external factors which act on these systems Goal and motivation: To Create 3 critical Safety Systems using Machine Learning algorithms such that they improve their output over time based on the Driving style of the user and varying vehicle conditions and external factors. 1. ABS (Antilocking Braking System) 2. EBD (Electronic Brake Distribution) 3. Driver Fatigue Detection. IV. HARDWARE SPECIFICATION 4.1. Antilocking Braking System Uses Wheel Speed sensors on the wheels. The Brake Master Cylinder is connected to the Brake Force Modulator which receives commands from the ECU to decrease the brake force in case the sensors detect that one of the wheels have All Rights Reserved 455

3 4.2. Electronic Brake distributor uses traction sensors on the wheels to calculate which wheel has maximum traction to the road surface and transfers maximum brake force to that wheel, it also uses the Brake force modulator to achieve this [3] Fatigue Detection system are of many types, however the most efficient ones uses infrared cameras to monitor the driver s eyes and calculates the time intervals between blinks to predict All Rights Reserved 456

4 In addition to the above hardware, to achieve our goal of improving the efficiency of these systems over time we need additional hardware to monitor other aspects that are not considered by traditional Systems. We need Tire pressure monitors, Tire Wear indicators, Tire and Road Temperature Sensors. To process the data collected by this hardware the machine learning algorithms need as much data as they can access and we need high speed storage system that is least prone to failure to hold the training data and provide it when the algorithm needs to access this data [5]. V. SOFTWARE SPECIFICATION To achieve our goal of enhancing the safety system of automobiles, we need efficient software to make best use of the available hardware. But as stated earlier there are many challenges that we face by directly applying machine learning to these Safety systems, and hence we need traditional flow based software to monitor the corner case behaviors and for handling outliers in the training data, and also to take control of the system when the training data is not sufficient to predict efficient results. We need the software to gather information from all the sensors and store the data in a particular format to make it available as the training data for the various machine learning algorithms implemented. The machine learning algorithms that we use must be very robust and less prone to corner case behaviors even though we have traditional software to back handle these situations. The algorithms must periodically change the mapping of the output prediction by analyzing the training data and produce sufficient feedback to the driver. We can also use cloud services to gather the feedback provided by these algorithms and use the data to improve the overall hardware and software design which will play a major role in achieving the ultimate goal of robust driverless vehicles [4]. VI. PROPOSED WORK 6.1. Antilocking Braking System The Antilocking Braking System was a major breakthrough in the automobile safety, and many countries have made it mandatory for certain class of vehicles. However, it did come with many drawbacks as well, the system is highly intrusive for certain driving styles, and it does not consider the external factors that may lead to proper functioning of this system. The system solely depends on wheel speed sensors to and predicts if a wheel is locked if it is rotating at a different speed relative to the other wheels. When the system intrudes the normal functioning of the brakes it makes sure that the wheels do not lock up and the vehicle does not spin out of control. But this intrusiveness would increase the braking distance that may cause fatal collisions when the driver does panic braking [6]. To reduce the intrusiveness of the system in such cases the Machine learning algorithm uses additional data such as reaction time, Brake pressure applied by the driver, Line of travel of the vehicle body and the G forces acting on it. By Storing these information as the training data, the system can improve the effectiveness over time based on the drivers style of using the brakes, such that the ABS intrudes only when the driver All Rights Reserved 457

5 6.2. Electronic Brake Distribution There is a misconception among some drivers that EBD and ABS are one and the same, but there is significant difference on how these system works and what it does. ABS make sure that the wheels of the vehicle do not lock, whereas the EBD system makes sure the vehicle stays in a straight line and does not roll out of the path the driver does not intend it to. Regardless of the people s understanding, ABS and EBD works best when both the systems are present on a vehicle. The Electronic Brake distributor monitors the traction of each wheel and makes sure the brake force is distributed based on the amount of traction available at that wheel. EBD is necessary because at any given moment all the wheels of a vehicle may not be rolling on the same kind of surface, and all the wheels may not have the same amount of wear and hence traction at these wheels vary a lot and sending the same amount of brake force to all the wheels would make the vehicle pull towards the wheel which has maximum traction and it rolls out of line the driver intends it to be. The traditional EBD system works purely based on the calculated traction available at each wheel, but there are factors that may help this system to behave more in a way the driver intends to. The System must also take into account the total weight of the vehicle and also consider at which part of the vehicle the weight is located. For example, if the driver has stored heavy objects in boot of the vehicle. The rear wheel will initially have higher traction but when brakes are applied the weight is shifted to the front wheel and the traction at the rear wheels will reduce. This case may also involve differently weighted passengers seated at different parts of the vehicle in a small car. Considering these factors, the system can learn what kind of distribution would be more efficient under the current situation Driver Fatigue Detection The fatigue Detection system uses infrared cameras to monitor the driver s eyes and calculates the time between blinks to predict if the driver could fall asleep. When fatigue is detected, the system raises an alarm and asks the driver to take a break before continuing driving. And also in some cases the system restricts the vehicle speed till the driver takes the breaks for a set period of time. Using the data collected, the machine learning algorithm can analyze at what time interval usually the driver experiences fatigue and raise an alarm before the fatigue hits the driver, such that the slim chances of driver not noticing the warning is eliminated [7]. VI. DRAWBACKS As discussed various times before in this paper, Machine Learning algorithms still need backing of traditional software for this particular application, Machine learning can only be used to enhance the behavior of the existing Safety system that uses traditional software by considering various additional factors, but machine learning cannot take over the place of Traditional software altogether. The Training data collected by the system may include data that represents an individual s driving style, but the system must also be able to detect if the vehicle is being driven by a different driver that usual. This can be achieved by proving guest mode on the vehicle, where the users manually tell the vehicle that it s been driven by a different driver, but that would again put extra effort on the user. As the amount of Training data increases in volume the speed of access of the data and mapping of input-output pairs will become slow, and hence the system must also ensure older data that is not very useful must be All Rights Reserved 458

6 There is no automated testing methods to test Machine Learning algorithms and hence we need to use manual testing methods to achieve perfection and that would consume a lot of time and would eventually lead to high price of the systems. VIII. CONCLUSION Recent advancement in technology and proven efficiency of Machine learning algorithms to adapt itself and improvise the outputs have made it necessary for replacement of traditional software by Machine learning algorithms wherever possible. Although not totally perfect, these systems would take automobile safety one step further by making it adaptive and improvising based on the user s behavior, the vehicles condition and other external factors. From the data collected by these systems overall advancement of automobile safety would drastically improve and could be used on driverless cars which are believed to be the future of transportation. REFERENCES I. Yuchi tian, Deep Test: Automated testing of DeepNeural-Networks driven automated cars, arxiv: v2 [cs.se] 20 Mar 20. II. Suman Jana, Deep Test: Automated testing of DeepNeural-Networks driven automated cars, arxiv: v2 [cs.se] 20 Mar 20. III. Dr.Porag Kalita, Study on Vehicle Computersied Wheel alignment Volume 3, issue 2, February2016 IV. Ahmed ElShafee, Mahmoud EIMenshawi, and Mena Saeed, Integrating Social Network Services with Vehicle Tracking Technologies, International, Journal of Advanced Computer Science and Applications, Vol. 4, No. 6, 2013 V. Amita Mundhe, Jayashree Otari, Automatic Number Plate Recognition Using Smart Phones,International journal of Engineering Research & Technology, Vol. 2 Issues 4, April VI. Ayman A. Aly1, 2, El-Shafei Zeidan1, 3, Ahmed Hamed1, 3, Farhan Salem1 1 Department of Mechanical Engineering, Faculty of Engineering, Taif University, AlHaweiah, Saudi Arabia 2 Department of Mechanical Engineering, Faculty of Engineering, Assiut University, Assiut, Egypt 3 Department of Mechanical Power Engineering, Faculty of Engineering, Mansoura University, Mansoura, Egypt ayman_aly@yahoo.com Received April 23, 2011; revised May 18, 2011; accepted May 25, VII. Zutao Zhang,Sichuan, Driver Fatigue Detection Based Intelligent Vehicle Control,Key Lab of Signal and Information Processing Southwest Jiaotong University Chengdu, , All Rights Reserved 459