Pedestrian Detection And Behavior Recognition Based On Vehicle Vision

As the number of accidents in traffic accidents increases year by year,and the performance of computer hardware continues to improve and Ai's ability to solve in specific scenarios shows amazing performance beyond humans,pedestrian detection and behavior recognition technology for analyzing vehicle driving systems has considerable application prospects..Most of the current applications of behavior recognition technology are used to monitor scenes,and few studies have applied behavior recognition to vehicle driving systems.This paper examines how pedestrian detection techniques and pedestrian behavior recognition techniques can be applied to vehicle driving.The main contents of this paper are as follows:(1)Firstly introduce the basic composition and execution principle of convolutional neural networks,then introduce the pedestrian detection technology and behavior recognition technology based on deep learning,and analyze and compare the different technologies.Then the TOLOVT3-TINY pedestrian detection algorithm is analyzed in detail.Finally,the YOLOV3-TINY is optimized for the specific application scenarios of the vehicle vision,including the calculation of pedestrians' anchors,the retraining of the pedestrian class,the optimization of the YOLOV3-TINY feature extraction network,and the addition of YOLOV3.-TINY predictive scale,improving pedestrian detection accuracy through target tracking.(2)Research on behavior recognition based on continuous multi-frame images.This part firstly uses C3 D model to extract multi-frame picture features and then classify them.Then,the optimization strategy is proposed for the network structure from the activation function and the network depth,and the performance before and after the optimization is compared.Finally,the classic design idea in CNN design is introduced to design the 3D-RDBC module,which is compared with the 3d residual network.,greatly improving the detection performance.(3)Research on picture-based behavior recognition,using CNN to extract and classify pedestrian characteristics,then use the BBOX of the pedestrians returned by YOLO as the input of CNN to get the specific behavior,and then design the CNN network model of their own design.Some classic network models do the accuracy comparison,and the multi-target tracking corrects the behavior,and the accuracy isobviously improved.(4)Combining the above pedestrian detection and behavior recognition results,the realization of the early warning area is designed.The warning area is drawn in front of the vehicle,and the early warning strategy is designed according to the different behaviors of pedestrians in different areas.