| Risky driving behavior is an important cause of traffic accidents.It is important to monitor the driver’s behavior and give early warning before the occurrence of risky driving.The existing computing capability and data acquisition equipment have ability to realize real-time data collection and analysis,making the advance warning possible.When the driver loses control of his emotions while driving the vehicle,the driver’s driving performance will be impaired.Existing studies mainly focus on driving characteristics under different emotions,but few studies take emotions as a feature of behavior prediction.Therefore,in this paper,driver’s driving behavior is predicted by integrating driver’s facial expression data and vehicle kinematics data.In order to obtain the most realistic data,this study carried out the naturalistic driving experiment,and cleaned the naturalistic data through a variety of pre-processing methods,and then used the machine learning algorithm to predict driving behavior.Specific research contents are as follows:In order to obtain large-scale real driving data,online ride-hailing drivers were taken as subjects to carry out naturalistic driving tests.Millions of vehicle motion data and video data of vehicles are collected through experiments.The video was analyzed by professional software to obtain emotional data and micro-expression data.Raw data was preprocessed by methods including time synchronization,time splicing,invalid value processing and data smoothing.The physical model considering road linearity and road friction and the acceleration threshold were used to identify the three kinds of risky driving behaviors,including abrupt lane change,abrupt acceleration and abrupt deceleration,and the data of risky driving behaviors were labeled.Five representative machine learning models Decision Tree,Random Forest,Gradient Boosting Decison Tree,Support Vector Machine,Bayesian Network were selected.The predictive ability of the model was tested on different feature sets.Formula 1 with comprehensive consideration of precision and recall was selected as the evaluation index of the model,and the optimal model was selected for subsequent research.The results show that the performance of the Gradient Boosting Decison Tree(GBDT)is better than other models in each feature set,and when the feature set of vehicle kinematics data and emotion data was used,the model can obtain the best prediction performance,F1 score is 92.66%.The above optimal model GBDT was selected to study the combination of time window and feature set with optimal prediction ability.Different feature sets and time intervals were used to build GBDT,and the prediction ability of each model was compared.The results show that the best prediction accuracy can be obtained by using the emotion and vehicle kinematics data of 0-2s before the behavior(F1=93.97%).In the case that the driver is given a 1s reaction time,the optimal prediction performance can be obtained by using vehicle kinematics data and facial expression data from 1-13 s before the behavior(F1=85.88%).To sum up,based on large-scale naturalistic driving data,this paper integrates multi-source data such as facial expression data and vehicle kinematics data to predict driving behavior.The actual prediction performance of five machine learning models was compared to select the most appropriate prediction model.Through the grid search strategy,the optimal special solicitation and time interval were found.The proposed driving behavior prediction scheme can be directly applied to real-time risky driving behavior prediction and provide technical support for the development of human-machine interaction of intelligent vehicles. |
