Research On Pedestrian Tracking Algorithm Based On Deep Learning

Pedestrian tracking is an important task in the field of computer vision.It is the basis for other advanced vision tasks such as human pose estimation,motion recognition,and behavior analysis.It is widely used in emerging fields such as autonomous driving,intelligent security and service robots.Pedestrian tracking faces many influencing factors,including environmental lighting changes,pedestrian occlusion,the randomness of the appearance and disappearance of new and old targets,and camera movement.Therefore,improving the accuracy of the pedestrian tracking algorithm is extremely valuable and challenging.In recent years,due to the significant improvement in the performance of target detection,tracking-bydetection framework has gradually become the mainstream.The tracking process is usually divided into four steps: pedestrian detection,feature extraction,similarity calculation,and data association.This paper uses tracking-by-detection framework to carry out related research to improve the accuracy of pedestrian tracking algorithms.The main research contents are as follows:(1)In a scene where pedestrians are more densely tracked,it is difficult for the algorithm to stably correlate with the previous trajectory when pedestrians reappear after being temporarily occluded,resulting in pedestrian identity switching and trajectory discontinuity.In response to this problem,this paper proposes a pedestrian re-identification method combining context and multi-path feature fusion.This method can extract the apparent features of pedestrians with strong discriminativeness,so that the distance between the same target features is close,and the distance between different target features is far,which helps to improve the accuracy of data association.Experimental results show that this method performs better on the Market1501 and Duke MTMC-re ID data sets.(2)The tracking-by-detection framework relies heavily on pedestrian detection,and excellent detection algorithms can significantly improve tracking performance.Aiming at the problem of low accuracy of pedestrian detection in scenes with small targets,occlusion,and crowding,this paper designs a joint attention module,and combines this module with the YOLO-v4 target detection model to design a pedestrian detection algorithm based on joint attention,and Experiments are carried out on the Wide Person pedestrian detection data set,and the experimental results show the effectiveness of the algorithm.At the same time,a lightweight model of the algorithm is designed with Mobile Net-v2 as the feature extraction network.The detection accuracy is slightly reduced,but the speed is improved.(3)In view of the inaccuracy of using Kalman filter to predict pedestrian motion state in the case of camera motion,this paper proposes a pedestrian motion estimation model fused with camera displacement,which compensates error for the motion estimation caused by camera displacement by introducing image alignment.When calculating the apparent similarity,this paper designs a feature selection method based on appearance confidence to select highquality features to calculate the similarity and improve the accuracy of data association.The experimental results on the MOT16 dataset show that the algorithm in this paper can track pedestrians stably in complex scenarios.