Research On Pedestrian Detection And Tracking Based On Deep Learning

With the rapid development of computer vision in recent years,the multi-target detection and tracking technology for pedestrians has emerged as a popular research direction,playing an essential role in intelligent monitoring,intelligent transportation,human-computer interaction,and automatic driving,among other fields.But the multi-target detection and tracking technology for pedestrians are in the face of a challenge posed by complex environment and characteristics of the pedestrian.Therefore,this paper makes an in-depth study on the multi-target detection and tracking technology for pedestrians in complex scenes.The specific research contents are as follows:(1)In response to the problem that there are not enough scenes in the public pedestrian dataset,this paper combines the Crowd Human database and self-collected data to build the pedestrian dataset needed for research.It has 19,375 19375 pictures,including complex scenes such as streets,shopping malls,squares,etc.(2)To deal with the problems in the pedestrian multi-target detection process,such as low detection rate,interference from the complex environment,occlusion,and high missed detection rate of small-scale pedestrian targets,this paper proposes three improvements based on the YOLOv4 detection algorithm as follows: first,by simplifying the backbone feature extraction network,a lightweight CSPDarknet-41 is proposed,which reduces the depth and parameters of the detection network model and effectively improves the real-time performance of the detection algorithm;secondly,lightweight,efficient channel attention for convolutional neural networks is introduced into the backbone feature extraction network to enhance the detection network’s ability to extract the local features of pedestrian targets and improve the detection accuracy of occluded targets;finally,the multi-scale output layer is reconstructed.Based on the multi-scale feature fusion,the multi-scale shallow feature fusion is carried out,and finally,a feature output layer is added for the detection of small-scale pedestrian targets,which effectively reduces the missed detection rate of small-scale pedestrian targets.In addition,tests are carried out on the self-collected pedestrian data set and compared with other common target detection algorithms.Experiment results suggest that the improved YOLOv4 detection algorithm proposed in this paper can achieve the best detection rate and detection accuracy.(3)In response to the problems of high walking speed of pedestrian target,different trajectories,and sudden change of apparent characteristics in the process of pedestrian multitarget tracking,on the basis of the Deep Sort tracking algorithm,this paper effectively improves the prediction of nonlinear-moving pedestrian target’s position and the matching ability of pedestrian target by integrating EKF filtering algorithm and DIOU matching.This paper uses the MOT20 multi-pedestrian dataset to test the performance of the improved Deep Sort tracking algorithm and compare it with other mainstream tracking algorithms in the test.The test results suggest that the improved Deep Sort tracking algorithm proposed in this paper is significantly better than other tracking algorithms in all performance indexes.What’s more,it can also better deal with the problems in the process of pedestrian multi-target tracking.