Research On Driver Intention Recognition Algorithm For Intelligent Driving

The driver is not only the most important factor in the"human-vehicle-road"closed-loop system,but also the most unstable factor.From the perspective of safety,up to 94%of traffic accidents are caused by drivers’mistakes.If drivers’driving thoughts can be predicted in advance,the number of traffic accidents can be effectively reduced.From the perspective of intelligent vehicle development,full autonomous driving is difficult to be realized in the short term.It must first enter the stage of man-machine co-driving,in which the driver and the vehicle share the control right.At this time,the decision-making layer of the system needs to identify the driver’s intention and realize the synchronization of driving decisions.Therefore,in the development process of intelligent driving,the driver factor,especially the identification of the driver’s intention,must not be ignored.Based on the Brain4Cars data set,this paper carries out research on driver intention recognition before driving operation in different time windows,and achieved certain results that the optimal model reached 82.6%in F₁.The main work contents are as follows:（1）Using 3D convolution to extract abstract features of cockpit video and carry out research on driver intention recognition with it.CNN can only extract features from spatial dimensions and cannot consider the correlation existing in continuous frames of moving video.Therefore,the ability of 3D convolution to extract spatial-temporal features is introduced in this paper,and 3D Res Next101 is used to extract abstract features of video in the cockpit to identify the driver’s intention.Due to the small number of samples,this paper takes the model pretrained on Kinetic as the basic model through transfer learning,carries out training with fixed partial weights,and uses various methods to avoid overfitting.The results show that 3D convolution can effectively extract abstract features of the cockpit video.The trained model can identify the driver’s intention with an accuracy rate of 76.3%and F₁ of 75.3%in the test set.（2）Study on driver intention identification by using CNN-LSTM to predict future frame motion feature of traffic conditions.In this paper,the optical flow diagram containing motion information is used as the model input,CNN is used for feature extraction and LSTM is used to predict the motion features of future frames to identify the driver’s intention.This paper compares the effects of using Lucas-Kanade algorithm,Farnback algorithm and Flownet2.0 network to generate optical flow.Considering that the scene studied in this paper has obvious light change and the speed of moving target is fast,Flownet2.0 network is finally selected to process the frame image to obtain optical flow.Because CNN-LSTM network can only predict the motion feature of a certain frame in the future,the overall recognition performance of the model is relatively low,with accuracy of 64.7%and F₁ of 68.2%.（3）Research on driver intention identification by using feature fusion of upper and lower branches.The abstract features of cockpit video extracted by 3D convolution were effectively combined with the future frame motion feature predicted by CNN-LSTM as the recognition features.The effect of the model on the test set shows that the abstract features of cockpit video can complement the video motion information of traffic conditions effectively,which is helpful to the identification of driver intention.In addition,compared with other methods,the results show that the F₁ of the proposed method in this paper is better when the time window T=5s.At the same time,in order to improve inference speed of the model,this paper optimized the model based on Open Vino.