Research On Tracking And Classification Of Vehicles Based On Compressive Sensing

The tracking and classification of vehicles in video is a very important part in the field of automatic driving.At present,there are many achievements on the tracking and classification of vehicles in video,but most of them are based on fixed cameras and scenes,which are not suitable for automatic driving and intelligent assistant.Aiming at solving the problems of low real-time performance and scene interference in the method of vehicle tracking and classification in mobile background video sequence,a system of tracking and classification of vehicles in video is designed based on compressive tracking algorithm.The main work and achievements of this paper are as follows:(1)Research on the area detection of moving targets.Aiming at the detection of the moving area under the moving background,a detection method combining the frame subtraction method and the background modeling method is proposed,which can effectively improve the accuracy of detection result.(2)Research on vehicles extraction from video sequences.Aiming at the problem of balancing the real-time performance and accuracy of existing vehicles detection algorithms,a vehicles detection model named Attention-YOLOv4 based on video sequences is proposed,and attention mechanism is introduced to further improve the detection speed and accuracy of the model.(3)Research on fine-grained recognition of vehicle types.A lightweight residual network(Lightweight Res Net,LRN)based on knowledge distillation is proposed,and the teacher-student system is used to train the LRN model.The results show that LRN has a classification ability similar to the teacher network,and effectively reduces the computational complexity of the model.(4)Research on vehicle tracking based on compressive sensing.Aiming at the real-time performance of vehicle tracking and the sensitivity to factors such as illumination and occlusion,a vehicles tracking algorithm based on an improved measurement matrix is proposed with compressive tracking as the basic framework.The algorithm has good real-time performance and is resistant to occlusion.Factors such as illumination changes have certain robustness.This paper designs a tracking and classification method of vehicles under a mobile background,which has been significantly improved in terms of real-time performance and accuracy.In the follow-up research,the finegrained recognition of non-motor vehicles' types and the tracking method of night-driving vehicles will be discussed.