Pedestrian Detection And Tracking In Tunnel Environment Based On Deep Learning

The correct detection and tracking of pedestrian targets in highway tunnels is of great significance to highway safety.Due to lighting in the tunnel,camera installation,etc.,the surveillance video image is blurred,and the pedestrian target is relatively small in size and easily blocked,making traditional methods unable to effectively detect and track pedestrian targets.Deep learning model have achieved good results in the field of pedestrian detection and tracking due to their good feature extraction and expression capabilities,but it is difficult to exert their good feature extraction capabilities for complex tunnel environments.Therefore,research on pedestrian target detection and tracking algorithms based on deep learning in tunnel environment has important theoretical and practical significance.In view of the high detection accuracy of the two-stage target detection algorithm,the thesis is based on the Faster R-CNN target detection algorithm,and proposes two optimization schemes for the poor detection effect of the Faster R-CNN in tunnel environment.One scheme is to Faster R-CNN feature extraction layer low-level information is introduced into high-level information to increase high-level feature information;another solution is to draw on the idea of pooling pyramid module in PSPNet algorithm,and add attention mechanism to enhance Faster R-CNN algorithm Context information.Aiming at the problem of occlusion during pedestrian tracking,a method combining Kalman filtering and pedestrian re-identification algorithm is proposed.The main work and contributions of this article are as follows:(1)In response to the problem of poor detection of low-resolution pedestrian targets in the tunnel environment of Faster R-CNN algorithm,the low-level network information of the feature extraction layer in Faster R-CNN network is added to the high-level network to increase the feature information in the high-level network,Improve the accuracy of pedestrian target detection in Faster R-CNN algorithm.(2)Aiming at the problem that the Faster R-CNN algorithm has insufficient ability to extract pedestrian target features,a dilated pyramid structure is designed to increase the context information in the network and improve the network’s ability to extract pedestrian target features.The characteristic information is scattered,increasing the attention mechanism.By adding the dilated pyramid module and the attention mechanism module to the Faster R-CNN network,the network feature extraction capability is enhanced,and the detection accuracy of pedestrian targets is improved.(3)Aiming at Kalman filtering,the target tracking fails due to occlusion in the process of pedestrian target tracking.In this paper,a pedestrian re-identification algorithm with pedestrian target feature extraction and feature information matching is introduced,and a method of combining Kalman prediction and pedestrian re-recognition is proposed.Experiments show that this method can effectively improve the pedestrian tracking effect in the tunnel environment.Based on the above,a method suitable for the detection and tracking of pedestrians in the tunnel environment is formed.Collect the Chongqing Expressway Tunnel Surveillance Video to make dataset,train and test the improved model.Experimental results show that compared with the original Faster R-CNN algorithm and Kalman filter tracking,the improved method in this paper has improved detection accuracy and anti-occlusion ability,and has also achieved good results in the actual highway tunnel environment.