, pp.151-155 http://dx.doi.org/10.14257/astl.2018.150.35 Pothole Detection using Machine Learning Hyunwoo Song, Kihoon Baek and Yungcheol Byun Dept. of Computer Engineering, Jeju National University, Korea chzhqk1994@gmail.com, masinogns@gmail.com, ycb@jejunu.ac.kr Abstract. Pothole detection is import to decrease accidents across the world. Many researches have been done but they require some specific devices or tools to acquire sensor data. In this paper, we propose an hany way to implement pothole detection using a smartphone, and classification is performed using Transfer Learning. The experimental result shows that the proposed approach provides us efficiency from the view point of implementation and performance. Keywords: Pothole detection, Transfer Learning, Inception V3 1 Introduction and Related Works The number of vehicles drastically increases every year, and the number of accidents proportionally does too. The condition of road surface affects directly on our safety. The American Automobile Association estimated in the five years prior to 2016 that 16 million drivers in the United States had suffered damage from potholes to their vehicle with a cost of 3 billion USD a year[1]. In India, 3,000 people per year are killed in accidents involving potholes. Britain has estimated that the cost of fixing all roads with potholes in the country would cost 12 billion EURO[2]. According to the World Health Organization, road traffic injuries caused an estimated 1.25 million deaths worldwide in the year 2010. That is, one person is killed every 25 seconds. Only 28 countries, representing 449 million people (seven percent of the world's population), have adequate laws that address all five risk factors (speed, drunk driving, helmets, seat-belts and child restraints)[3]. By the way, there is close relationship between the accident and road condition including a pothole. According to Austroads[4], road accidents occur as the result of one, or more than one of the following factors: human factors, vehicle factors, road and environment factors. Vogel and Bester[5] introduced risk factors (human, vehicle and environment factors) for 14 accident types that can be used as a reference point to determine the likely cause of an accident of a specific type. A research had been done from a little bit different point of view, where the researchers proposed a costeffective solution to identify the potholes and humps on roads and provide timely alerts to drivers to avoid accidents or vehicle damages. Ultrasonic sensors are used to identify the potholes and humps[6]. A low cost model for analyzing 3D pavement images was proposed, which utilizes a low cost Kinect sensor which gives the direct depth measurements, thereby reducing computing costs[7]. Lin and Liu have proposed a method for pothole detection based ISSN: 2287-1233 ASTL Copyright 2018 SERSC
on SVM (Support Vector Machine). This method distinguishes potholes from other defects such as cracks. The images are segmented by using partial differential equations. In order to detect potholes, the method trains the SVM with a set of pavement images. However, the training model fails to detect the pavement defects if the images are not properly illuminated[6][8]. In the previous researches, some specific devices and tools are needed to detect the status of roads, which causes some extra costs and inconveniences. In this researches, we introduce an efficient way to detect road distress using a mobile devices. Anybody can install an app into his/her mobile to detect the status of roads. 2 Proposed Approach Every movements should make different sensor values if we use some sensors including gyroscope and accelerator. Fortunately, almost all of the recent mobile smart phones have the two sensors, which is easy to handle and has advantages in the cost and efficiency. Therefore, we utilize a smart phone as a sensor to acquire movement information, and the sensed data will be fed into a classifier to detect the status of a road. In this point, a classifier is one of the key component to implement a successful system. Fig. 1. Inception V3 and Transfer Learning Can we utilize Inception V3 to a specific domain problem, which is one of the well-known classifiers for classifying existing general items and objects including TVs, vehicles, refrigerators, airplanes, and etc.? We tried to find the answer for this question in this research. We use the existing knowledge in Inception V3 except the final fully connected layer in it. They say it is Transfer Learning which gives us some advantages in computation time and efficiency for recognition. Figure 1 shows that what Transfer Learning means and which part of Inception V3 will be retrained. The overall process for our approach includes (1) acquiring road status information using gyroscope and accelerate sensors which is called logging, (2) data preprocessing for Convolutional Neural Networks(CNN), that is, Inception V3. There 152 Copyright 2018 SERSC
are two steps: learning and testing. In learning step, the hyper parameters in fullyconnected layer is tuned using some portion of data, and test is performed using the rest data to verify our proposed approach. Fig. 2. An example of sensor data for a pothole (1 try) Figure 2 shows an example actual data acquired for sensors which is captured while passing over a pothole one time. We can see some number of oscillation in the center area having high frequency for a pothole. Relatively low frequency on both sides means a vehicle is passing over a flat path. 4 Experimental Results To verify our proposed approach, we located some bumps and potholes we could easily find around us to collect some sensor data. An Android app was implemented to capture the data using Gyroscope and Accelerator sensors. Table 1. Data description Heading level Count Description Normal 100 Pothole 100 Acquired from a Bump 100 vehicle Total 300 3 classes (types) of road status (normal, pothole, and bump) were considered and 5 real instances on a road were found. We collected 20 times of sensor data for each instance, which makes 100 count of sensor data for each class. A total number of 300 number of data were collected finally as showed in Table 1. Copyright 2018 SERSC 153
Fig. 3. The change of loss values during learning To train our Convolutional Neural Networks, 70 percent of data (70 instances) for each class separated. After 18,000 learning epochs, the loss value converged toward almost zero, which means the learning process performed successfully. Finally, we tested using the rest data, which is 30 percent of the original data (30 instances for each class). All of the instances were recognized correctly showing 100% of classification rate. 5 Conclusion and Future Works In this paper, we proposed an efficient method to recognize a pothole on a road from the viewpoint of cost and implementation. This is handy way because sensor data is acquired using a smartphone everybody has nowadays. To make the implementation easier, we utilized Inception V3 and Transfer Learning which gives us a very flexible way for application. Interesting thing about this research is that general knowledge works for a specific domain problem. That is, the knowledge acquired in Inception V3 to recognize common objects around us can be transferred to recognize a totally difference signal patterns as we showed in previous session. Meanwhile, the success of Transfer Learning is depend on the variety of data which might change according to sort of vehicles, the shape of bump and pothole, and etc. Many types and shapes means difficulty of learning, but big data will be helpful to solve it. Also, not a general Inception V3 but a domain specific Convolutional Neural Networks should be considered to handle the problem well. Acknowledgments. This research was financially supported by The Project Management Center Cultivating Smart Grid & Clean Energy Manpower (CK-1), JNU. References 1. Pothole Damage Costs U.S. Drivers $3 Billion Annually, https://www. oregon.aaa.com/ 2016/02/pothole-damage-costs-u-s-drivers-3-billion-annually/. (2017) 154 Copyright 2018 SERSC
2. Bad roads killed over 10k people in 2015; 3,416 deaths due to potholes. https://timesofindia. indiatimes.com/india/bad-roads-killed-over-10k-people-in-2015-3416-deaths-due-topotholes/articleshow/53482615.cms. (2015) 3. List of countries by traffic-related death rate. Wikipedia, https://en.wikipedia.org/wiki/ List_of_countries_by_traffic-related_death_rate. (2018) 4. AUSTROADS, 1994. Road Safety Audit. Sydney (2006) 5. L. Vogel, C. J. Bester : A Relationship between accident types and causes. (1999) 6. Rajeshwari Madli, Santosh Hebbar, Praveenraj Pattar, and Varaprasad Golla : Automatic Detection and Notification of Potholes and Humps on Roads to Aid Drivers. IEEE Sensor Journal. VOL. 15, pp. 4313--4318. (2015) 7. I. Moazzam, K. Kamal, S. Mathavan, S. Usman, M. Rahman: Metrology and visualization of potholes using the Microsoft Kinect sensor. Proc. 16th Int. IEEE Conf. Intell. Transp. Syst, pp. 1284 1291. (2013) 8. J. Lin, Y. Liu : Potholes detection based on SVM in the pavement distress image. Proc. 9th Int. Symp. Distrib. Comput. Appl. Bus. Eng. Sci., pp. 544 547. (2010) Copyright 2018 SERSC 155