Driving Style Assessment Using Maneuver Transition Probabilities And Driver Operation Aggressiveness

Risky maneuvers(e.g.,tailgating,frequently changing lanes)are engrained in Chinese traffic culture.With the rapid development of telematics,it is of high priority to analyze driving style for the improvement of driving safety in China.Driving style analysis can be utilized to detect risky driving style,and provide drivers with offline feedback report for education or online risk alert when necessary.However,the current researches concerning driving style are not comprehensive and only focus on a few aspects of driving aggressiveness evaluation.To solve these two problems,this paper aims to analyze driving style from the perspectives of frequency and intensity of risks occurring on road.To achieve this goal,an effective hierarchical structure for maneuver recognition was proposed.Based on recognized maneuvers,the transition probabilities between them can be obtained to evaluate driving style from the perspective of frequency,and the aggressiveness of each maneuver can be calculated to evaluate driving style from the perspective of intensity.Thus,the overall driving style can be estimated from these two aspects,frequency and intensity.For the recognition of longitudinal driving maneuvers on highways,deceleration,time headway and the change rate of time headway were adopted.For the recognition of lateral maneuvers,four variables were adopted,including Shannon entropy of steering wheel angle in a 2s time window,root mean square of lateral acceleration in a 2s time window,standard deviation of yaw rate in a 5s time window and Shannon entropy of speed in a 4s time window.These four variables were used as the input feature vector of a random forest classifier,and the output of the classifier was used for lateral maneuvers recognition.Validation results showed that the correct recognition rates ranged from 86% to 98%,proving the effectiveness of the proposed methodology in maneuver recognition.To evaluate driving style from the aspect of risk frequency,five maneuver transition probabilities were selected using a feature selection method,including transition from near following to constraint right lane change,from constraint right lane change to constraint left lane change,from constraint left lane change to approaching,from approaching to constraint right lane change,and from constraint left lane change to free driving.These five maneuver transition probabilities were used as the input feature vector of a random forest classifier,and its output was used for the recognition of driving style from the aspect of risk frequency.Cross validation results showed that the correct recognition rate achieved 93%,18% higher than the correct recognition rate when using a traditional method.To evaluate driving style from the aspect of risk intensity,the exponent of longitudinal,lateral and equivalent acceleration,namely aggressiveness index,were proposed to indicate the risk a driver felt while driving.The aggressiveness index in five aspects(including acceleration,baking,headway time control,lane change control,and turning control)were weighted to obtain an overall aggressiveness score,based on which drivers were classified into three risk groups(low-risk,moderate-risk,high-risk).The validation tests based on videos and on-site experiments showed that the correct recognition rates were 85% and 92%,respectively.Based on the driving style evaluation results from the aspect of risk frequency and intensity,a decision tree model was proposed for overall driving style estimation.Cross validation results showed that the decision tree model could achieve a correct recognition rate of 89%.