Driving Behavior Analysis Based on Vehicle OBD Information and AdaBoost Algorithms

Transcription

1 , March 18-20, 2015, Hong Kong Driving Behavior Analysis Based on Vehicle OBD Inforation and AdaBoost Algoriths Shi-Huang Chen, Jeng-Shyang Pan, and Kaixuan Lu Abstract This paper proposes a novel driving behavior analysis ethod based on the vehicle on board diagnostic (OBD) inforation and AdaBoost algoriths. The proposed ethod collects vehicle operation inforation, including vehicle speed, engine RPM, throttle position, and calculated engine load, via OBD interface. Then the proposed ethod akes use of AdaBoost algoriths to create a driving behavior classification odel, and finally could deterine whether the current driving behavior belongs to safe driving or not. Experiental results show the correctness of the proposed driving behavior analysis ethod can achieve average of 99.8% accuracy rate in various driving siulations. The proposed ethod has the potential of applying to real-world driver assistance syste. Index Ters Driving behavior analysis, driver assistance syste, AdaBoost algorith, on board diagnostic (OBD) C I. ITRODUCTIO URRETLY, with the econoy developing, the aount of the vehicles increases every year. As the sae tie, the aount of non-professional drivers increases rapidly. Since ost novice drivers are unskilled, unfailiar with the vehicle condition and in weak awareness of traffic safety, the drivers' personal factors have becoe the ain reasons of traffic accidents. The driving auxiliary equipent is urgently needed to reind drivers the vehicle inforation in tie and correct iproper driving behavior. Fig. 1. Distribution of Traffic Accidents Statistics [1]-[2]. Fig. 1 shows the distribution of traffic accidents statistics Manuscript received Jan. 8, 2015, revised Feb. 10, This work was supported in part by the Ministry of Science and Technology, Taiwan under Grant SC E MY3. Shi-Huang Chen is with the Departent of Coputer Science and Inforation Engineering, Shu-Te University, Kaohsiung City, 824, Taiwan. (phone: ext. 4607; Fax: ; e-ail: shchen@stu.edu.tw). Jeng-Shyang Pan is with the Fujian University of Technology and the Harbin Institute of Technology Shenzhen Graduate School, China (e-ail: jengshyangpan@gail.co ) Kaixuan Lu was with the Harbin Institute of Technology Shenzhen Graduate School, China (e-ail: lukaixuan203@sina.co ) [1]-[2]. According to the statistical analysis data shown in Fig. 1, the ost accidents are caused by the personal factors of driver. They include iproper operation, incorrect judgent, and etc [1]-[2]. To further reduce the accident and to protect the driver/pedestrian safety, it should not only continuously iprove the vehicle safety equipent as well as road conditions, but also should pay attention to the driver as the research object. Various literatures have introduced driving behavior analysis ethods based on coputer vision iage processing, ulti-sensor fusion, and etc. Yoshifui Kishioto and Koji Oguri proposed a HMM (hidden Markov odel)-based ethod to estiate the possibility of the driver braking fro past driving period [3]. Antonio Prez, M. Isabel Garca, Manuel ieto, and other scholars realized a syste called Argos. The Argos is an advanced in-vehicle data recorder which can help other researchers to better study the driver behavior analysis [4]. Reza Haghighi Osgouei and Seungoon Choi copleted the driving behavior odel for different drivers based on objective principles. Their objective principle is deterined by a rando distance of HMM odel trained by each driver [5]. Cuong Tran, Anup Doshi, and Mohan Manubhai Trivedi ake use of video optical changes and HMM to coplete the driver's foot oveent estiated algorith [6]-[7]. The scholars of University of California extract the video iages fro outside of the vehicle to analyze the driving behavior by the positional relationship with other vehicle, and then judge the vehicle is in safe state or danger state [8]. Shan Bao and Linda g Boyle copleted intersection driving experients in different age drivers and established a different age driver rearview irror observation odel [9]. Toshikazu Akita and others proposed the driver behavior odeling ethod based on the fusion syste, which is ainly recognition and analysis the car following state, accurate recognition rate is higher [10]. According to the above literatures, current research ethods of driving behavior analysis include the driving data collection, driving odeling algoriths, and applications. Driving data collection includes autootive video capture, car-ounted sensors, and the on board diagnostic (OBD). In ters of driving behavior odeling algoriths, there are HMM, support vector achine (SVM), decision trees, and other principles. The ain application of the driving behavior analysis is the identification of driving lethargy or the driver's actions forecast. In this paper, a novel driving behavior analysis ethod based on the vehicle OBD and AdaBoost algoriths is proposed. This proposed ethod collects the vehicle operation inforation, including vehicle speed, engine RPM, throttle position, and calculated engine load,

2 , March 18-20, 2015, Hong Kong fro the OBD interface. Then the proposed ethod akes use of AdaBoost algoriths to create a driving behavior classification odel, and finally could deterine whether the current driving behavior belongs to safe driving or not. Experiental results show the correctness of the proposed driving behavior analysis ethod can achieve average of 99.8% accuracy rate in various driving siulations. The reaining sections of this paper are organized as follows. Section II brief introduces the AdaBoost theore. The detail of the proposed driving behavior analysis ethod is presented in Section III. Section IV shows experiental results. Finally, Section V concludes this paper. II. THE BASIC PRICIPLE OF ADABOOST AdaBoost is a classical classification achine learning algorith. The basic principle of AdaBoost algorith is to use a large nuber of weak classifiers cobined together by a certain ethod for a strong classifier. The strong classifier has a strong ability of classification [11]-[13]. Strong classifier generated as follows: Assuing given a two-classification training data set: T { x1, y1, x2, y2,, x, y } (1) where each saple coposed by the instance and flags. n Instance xi R, flags yi Y { 1, 1}, X is the instance space, Y is flags set. AdaBoost use the follow algorith to generate the strong classifier. A. Basic AdaBoost Algorith T { x, y, x, y,, x, y }, Input: training data set where xi n R, yi Y { 1, 1} ; weak learn algorith. Output: Strong classifier G(x). (1) Initialization of the weight value distribution of the training data 1 D1 ( w11,, w1 i,, w1), w1 i, i 1, 2,, (2) (2) 1, 2,, M ( is the ties of train.) (a) Using training data set has the weight distribution D to learn, get the basic classification G ( x): X { 1, 1} (3) (b) The classification error rate of G (x) is calculated on the training data e P( G( x) yi) wii( G( x) yi) (4) (c) Calculation the coefficient of G (x), the logarith is the natural logarith 1 1 e log (5) 2 e (d) Update the weight value distribution of the training data D ( w 1,1,, w 1, i,, w 1, ) (6) wi w 1, i exp( yg i ( xi)) (7) Z where, Z is noralization factor which could ake the D becoe a probability distributions. Z wi exp( yg i ( xi)) (8) i 1 i 1 (3) Build a linear cobination of basic classifiers M f ( x) G ( x) (9) 1 The final classification can be expressed as M G( x) sign( f ( x)) sign G( x) (10) 1 III. DRIVIG BEHAVIOR AALYSIS METHOD The proposed driving behavior analysis ethod consists of driving operation acquisition odule, data preprocessing odule, and AdaBoost classification odules. Driving behavior analysis data will be divided into training set and test set. Preprocessing and feature extraction are siultaneously applied to both sets. Classification akes the test saples into the driving odel based on AdaBoost algoriths to classify and deterine the test saple category. The nuber of rightly or wrongly classified saples divided by the nuber of the test set saples is the classification correct rate or error rate. In the data acquisition odule, at first, a good driving data and bad driving data should be collected as training set. Then collecting another data set includes good driving behavior data and bad driving behavior data as a test set using the sae procedure. After data preprocess step, each tie slice saples can be regarded as the rate of change the driving operation. This paper uses the training set to establish the driving classification odel by the AdaBoost algoriths, and then the proposed ethod could use the test set to judgent the accuracy of the odel. Finally, the proposed ethod judges the erits of driving behavior. Fig. 2 shows the flowchart of the entire proposed ethod. Fig. 2. The flowchart of the proposed driving behavior analysis ethod. A. OBD Syste Current OBD syste, or called OBD-II, was proposed in 1996 to replace the deficient forer syste, naely OBD-I. The specifications of OBD-II were recognized by the Environental Protection Agency (EPA), US, and California Air Resources Board for air-pollution control. Since 1996, all vehicles are required to be equipped with OBD-II under EPA regulation in the USA. When vehicle exhausts higher level of air-pollution contents, OBD-II will generate Diagnostic Trouble Code (DTC) essages and Check Engine light will display. Soe sybols of Check Engine light are shown in

3 , March 18-20, 2015, Hong Kong Fig. 3. Meanwhile, OBD-II will save this DTC in the eory inside ECU. Thus the DTC essages can be retrievable through an OBD-II scan tool [14]-[16]. Fig.3. Variety of icons of Check Engine light. The ain features of OBD-II are (1) unified J pin socket and data link connector (DLC) (as shown in Fig. 4); (2) unified DTC and eanings; (3) storage and display DTC; (4) contains vehicle record capability; and (5) auto-clear or reset function for the DTC [18]. Hence, the advantage of OBD-II is its standardization. In other words, just one set of OBD-II scan tool is able to perfor the diagnosis and can scan against variety of vehicles which equipped with OBD-II syste. (a) (b) Fig. 4. (a) J1962 OBD-II 16-pin socket, (b) OBD-II DLC. The OBD-II socket has 16 pins and is usually installed below the driving panel. Aong these 16 pins, nine of the have fixed functions and the rest pins are left to the discretion of the vehicle anufacturer [16]. B. OBD Signal Generator Fig. 5. The control panel of OBD-II signal generator. This paper uses a custoized OBD-II signal generator to coplete and siulate the OBD-II vehicle inforation data collection. The control panel of the OBD-II signal generator is shown in Fig. 5. The left and right sides of the control panel are OBD data input control terinal and output control terinal, respectively. The input control terinal has knob input and jack input ethods. The knob input is anual adjustent ode. One can control ultiple data changes on the OBD-II signal generator, including the coonly 7 kinds of OBD data. Jack input is the external analog signal input. The outputs control terinal has indicator lights and jack output ethod. There are 15 kinds of engine running data. Indicator lights will show the operation status of engine coponents, e.g., fuel injections, ignition coils, and etc. The jack output can achieve data output displayed on other devices, such as an oscilloscope. The iddle of the control panel contains interlocking switches and OBD-II data link connector. The interlocking switches is designed to control the synchronization between the throttle value and the engine RPM. When it is in the open state, the accelerator pedal is available to control the throttle and engine RPM at the sae tie. When it is in the close state, one can control the throttle and engine RPM individually. The OBD-II data link connector achieves the OBD data output. Because the OBD-II interface is not a standard equipent of general coputer, an OBD-II to Bluetooth adapter was fabricated to accoodate with the PC to acquire the vehicle status. This paper selects EZ-SCA5 as the OBD-II to Bluetooth adapter. EZ-SCA5 supports ost of OBD-II counication protocols including SAE J1850 PWM, SAE J1850 VPW, ISO , ISO KWP, and ISO CA. The OBD-II essages will be decoded and transitted to OBD-II diagnosis encoder via EZ-SCA5. C. The Decode of OBD Signal The OBD-II vehicle inforation data collected in this paper include the vehicle speed, engine speed (RPM), throttle position, and engine load. These data could reflect the results of the driving behavior paraeter data. One can decode these OBD data via OBD Mode01 operation with appropriate PIDs [18]. The PID stands for Paraeter ID. The PIDs of vehicle speed, engine speed (RPM), throttle position, and engine load are defined in SAE J1979 [17] and given in Table 1. Table 1. PIDs definitions of vehicle speed, engine speed (RPM), throttle value, and engine load PID Data Bytes Description Units Forula# OD 1 Vehicle speed k/h A OC 2 Engine speed rp ((A*256)+B)/ Throttle position % A*100/ Engine load value % A*100/255 #A and B are represented the returned data bytes. Engine speed is one of the ost iportant paraeter in the engine output data. The best driving operation is not just the high vehicle speed and low engine speed, also the driver s operation can ake the car through the low efficiency area of the engine in the shortest tie and keep the vehicle speed and engine speed running in sooth. Vehicle speed is a direct response to the driver's operating. It can directly reflect the current driving state whether speeding or aintain safe. The throttle position is adjusted to send the nuber of fuel into engine. Suitable throttle opening in order to ensure the efficient operation of the engine. Too high or too low value of throttle will cause incoplete cobustion of fuel and air pollution. The engine load is an indicator of a driver operating environent. The clibing or running overload will cause the engine load to too large. It is easy to daage the engine and therefore is not conducive to safe driving. This paper develops two criteria for data collection based on the engine characteristic curve. (1) The noral vehicle condition data: The relative ratio of the vehicle speed and engine speed is aintained at between 0.9 and 1.3 (test in the sae gear); the relative ratio of the engine speed and throttle valve is aintained at between 0.9 and 1.3; the engine load is aintained between 20% and 50%.

4 , March 18-20, 2015, Hong Kong (2) The bad vehicle condition data: The relative ratio of the vehicle speed and engine speed is aintained ore than 1.3 or less than 0.9; the relative ratio of the engine speed and throttle valve is aintained ore than 1.3 or less than 0.9; the engine load is aintained greater than 50% or less than 20%. According to the above data collection procedure, a total of 1000 sets of vehicle inforation siulation data, including noral and bad driver condition, are collected in this paper. Soe collected vehicle data are listed in the Table 2. Table 2. The vehicle OBD-II inforation data (partial) used in this paper. The vehicle speed Engine speed Throttle Engine load k/h r/in % % D. Vehicle OBD Inforation Data Preprocessing This paper uses three characteristics, i.e., the relative ratio of the vehicle speed and engine speed, the relative ratio of throttle position and engine speed, and the engine load to analyze whether the current state of the driving behavior is safe or dangerous. First, the proposed ethod needs to copute the of vehicle speed, engine speed and throttle position by the Equation (11), where t 2 - t 1 = 1. data( t2) data( t1) Dt () t2 t1 A part of results are given in Table 3. Table 3. The vehicle OBD-II data Vehicle Speed Engine speed Throttle (11) Engine load % Because there are larger difference between the values of vehicle speed, engine speed, and throttle opening. This paper should copute the relative ratio of the vehicle speed and engine speed, the relative ratio of throttle position and engine speed. Calculation ethod shows in Equations (12) and (13): cs() t zs() t Rcz () t (12) where R cz (t) is the relative ratio of the vehicle speed and engine speed, cs(t) is the vehicle speed at tie t with axiu speed of 220, zs(t) is the engine speed at tie t with axiu engine speed of jq () t zs () t Rjz () t (13) ax( jq ) ax( zs ) where R jz (t) is the relative ratio of throttle position and engine speed, jq (t) is the of throttle position at tie t, zs (t) is the of engine speed at tie t, ax(jq ) and ax(zs ) are the axiu values of throttle position and engine speed, respectively. Finally, the proposed ethod cobines these three feature values which are the relative ratio of the vehicle speed and engine speed, the relative ratio of throttle position and engine speed, and engine load to deterine whether the vehicle data is noral or bad driving condition. The proposed ethod adds positive and negative labels to each data saple. OBD inforation data saple is stored as following forat: S OBD (t) = {the vehicle speed and engine speed, the relative ratio of throttle position and engine speed, engine load, the label} IV. EXPERIMETAL RESULTS This paper uses the toolkit-gml-adaboost-atlab [19] of MATLAB to coplete the data preprocessing and driving behavior odeling. After pretreatent, a set of experiental data saple is shown in Table 4. It follows fro Table 4 that there are three characteristics and one label. The nuber of iterations of AdaBoost algorith is obtained by the three different ways. They are Gentle AdaBoost, Modest AdaBoost, and Real AdaBoostas. The results of Real AdaBoost are better than those of Modest AdaBoost and Gentle AdaBoost. The accuracy rates of driving behavior analysis based on these three AdaBoost algoriths are given in Table 5. Table 4. A set of the vehicle OBD inforation data saples The relative The relative ratio of the ratio of Engine load vehicle speed throttle valve % and engine and engine label speed speed In contrast experients, this paper also copletes training the identification function based on SVM algorith. Under the optial paraeters c = 0.25, g = 4 condition, the

5 , March 18-20, 2015, Hong Kong recognition accuracy rate is %. The result is poorer than the results of AdaBoost algoriths. Table 5. AdaBoost algorith accuracy rates under different iterations Iteration tie Gentle AdaBoost 99.39% 100% 100% 100% Modest AdaBoost 92.12% 97.98% 98.99% 100% Real AdaBoost 100% 100% 100% 99.8% [15] E/E Diagnostic Test Modes Equivalent to ISO/DIS , SAE Standard J1979, [16] Diagnostic Connector Equivalent to ISO/DIS , SAE Standard J1962, [17] SAE J1979/ISO , E/E Diagnostic Test Modes, SAE International. Revised: [18] SAE International. On-Board Diagnostics for Light and Mediu Duty Vehicles Standards Manual. Pennsylvania, 2003 [19] (GML AdaBoost Matlab Toolbox Manual) V. COCLUSIO A novel driving behavior analysis ethod using vehicle on board diagnostic (OBD) inforation and AdaBoost algoriths is described in this paper. The proposed driving behavior analysis ethod utilizes OBD interface to collect a nuber of critical driving operation data, i.e., vehicle speed, engine speed (RPM), throttle position, and calculated engine load. Then it uses AdaBoost algoriths to coprehensive the analysis of driving behavior. Experiental results show the correctness of the proposed driving behavior analysis ethod can achieve average of 99.8% accuracy rate in various driving siulations. Using the real AdaBoost algorith, the proposed ethod could obtain 100% accuracy rate under 15 iterations. The proposed ethod has the potential of applying to real-world driver assistance syste. REFERECES [1] Miyaji, M., Danno, M., Oguri, K.: Analysis of driver behavior based on traffic incidents for driver onitor systes. In: Intelligent Vehicles Syposiu, IEEE(2008) pp [2] Liang, J., Cheng, X., Chen, X.: The research of car rear-end warning odel based on as and behavior. In: Power Electronics and Intelligent Transportation Syste, IEEE (2008), pp [3] Kishioto Y, Oguri K. A odeling ethod for predicting driving behavior concerning with driver s past oveents, IEEE International Conference on Vehicular Electronics and Safety (IEEE ICVES 2008), pp: [4] Perez A, Garcia M I, ieto M, et al. Argos: an advanced in-vehicle data recorder on a assively sensorized vehicle for car driver behavior experientation, IEEE Transactions on Intelligent Transportation Systes, 2010, 11(2): pp [5] Osgouei R H, Choi S. Evaluation of driving skills using an h-based distance easure IEEE International Workshop on Haptic Audio Visual Environents and Gaes (HAVE 2012), pp [6] Tran C, Doshi A, Trivedi M M, Modeling and prediction of driver behavior by foot gesture analysis. Coputer Vision and Iage Understanding, 2012, 116(3), pp [7] Tran C, Doshi A, Trivedi M M, Pedal error prediction by driver foot gesture analysis: A vision-based inquiry IEEE Intelligent Vehicles Syposiu, 2011, pp [8] Sivaraan S, Trivedi M M. Active learning based robust onocular vehicle detection for on-road safety systes IEEE Intelligent Vehicles Syposiu, 2009, pp [9] Bao S, Boyle L. Age-related differences in visual scanning at edian-divided highway intersections in rural areas. Accident Analysis & Prevention, 2009, 41(1), pp [10] Akita T, Suzuki T, Hayakawa S, et al. Analysis and synthesis of driving behavior based on ode segentation, 2008 IEEE International Conference on Control, Autoation and Systes, pp [11] Wu B, Ai H, Huang C, et al. Fast rotation invariant ulti-view face detection based on real AdaBoost., 2004 the Sixth IEEE International Conference on Autoatic Face and Gesture Recognition, pp [12] Hang Li, Statistical Learning Method. Tsinghua University Press, [13] Parviz, M., Moin, M. S.: Boosting approach for score level fusion in ultiodal bioetrics based on AUC axiization, Journal of Inforation Hiding and Multiedia Signal Processing, 2(1), 2011, pp [14] Diagnostic Trouble Code Definitions Equivalent to ISO/DIS , SAE Standard J2012, 2002.