Research On Pedestrian Multi-object Tracking Under Single-source Camera

Multi-object tracking is an intermediate task in the field of computer vision,and the task has very important value in academic research and industrial applications.In particular,recent years have seen the greater importance in the research on multi-object tracking of the application and development of technologies such as automatic driving,intelligent security in public places,military drones,and fighter jets.The multi-object tracking of pedestrians is the cornerstone of multi-object tracking research,and most of the multi-object tracking research is developed on the basis of pedestrian multi-object tracking research.Therefore,the academic research and industrial application of multiobject tracking technology for pedestrians is still worthy of research.Although this technology has been developed and applied for a long time,it inclines to appear the following situation when dealing with the mutual occlusion of the object to be tracked and the disappearance of the object,the sudden change of speed and shape of the object to be tracked,the change of illumination and the complex background environment.(1)The detection performance of the detector using the detection-based tracking technology route is insufficient;(2)The apparent feature extraction ability of the multi-object tracking algorithm is poor and needs to be improved.The above shortcomings have made it difficult to meet the needs of more complex real-world situations.At the same time,in practical applications,certain requirements are put forward for the lightweight,real time and robustness of the algorithm,and the reduction of the number of identity switching of the object to be tracked.This paper takes pedestrian multi-object tracking as the main task,committing to reducing the number of identity switching of the object to be tracked as much as possible while ensuring the tracking accuracy,and meeting the needs of real time and robustness.In order to achieve the above goals,this paper is mainly divided into two major aspects to improve,one is to improve the detection ability of the object to be tracked,and the other is to improve the ability to extract the apparent features of the object to be tracked.Finally,the improved object detection network and the improved appearance feature extraction network are combined into the multi-object tracking algorithm DeepSORT to form a more advanced multi-object tracking algorithm.In the detection part of this paper,mainly used the following ways:(1)Adding NAM attention module to improve YOLOv5m;(2)Improving the detection head of YOLOv5 m to a decoupled detection head;(3)Improving the bounding box loss calculation method of YOLOv5 m to WiseIoU method;(4)Encapsulate super large convolution kernel,structural reparameterization,depthwise convolution,etc.into Rep LKDe Xt module and combine the module with YOLOv5 m to improve the YOLOv5 m object detector network.In the tracking part of this paper,mainly used the following ways:(1)Replacing the original apparent feature extraction network with a lightweight network architecture OSNet;(2)Using the Relation-aware global attention RGA to improve OSNet to comprehensively improve the apparent feature extraction ability.After improving the object detection ability and the apparent feature extraction ability of the tracking object under the above method,the final network model is evaluated on the MOT17 data set.HOTA has been increased by 3.86% in total;IDF1 has been increased by 3.65% in total,and MOTA has been increased by 3.91% %;IDSW has been reduced by 93 times,which is 16% lower than the original model;and FPS has been increased by 6 frames.The improved tracking model not only improves the tracking performance but also reduces the number of identity switching of the object to be tracked.There are two main innovations in this paper:(1)Aiming at the problem that the object detection ability of the object detector in the DeepSORT tracking algorithm is not strong enough,this paper mainly forms a more advanced object detection network YOLOv5m＿De＿NAM＿Wise＿Rep LK through a series of improvement measures on the YOLOv5 m network;(2)Aiming at the problem of insufficient apparent feature extraction ability in the original DeepSORT tracking algorithm,this paper finally forms a more advanced apparent feature extraction network OSNet-RGA through a series of improvements to DeepSORT’s apparent feature extraction network.Finally,a more advanced multi-object tracking algorithm YOLOv5m*-DeepSORT-OSNet* was formed by combining the above two new improved networks on DeepSORT.This paper adopts the technical scheme of detection-based tracking based on deep learning,mainly by improving the YOLOv5 m object detection network and DeepSORT’s apparent feature extraction network and combining the improved two new networks into the DeepSORT algorithm framework to form a more advanced multi-object tracking algorithm.The resulting new multi-object tracking algorithm not only improves the tracking performance but also reduces the number of object identity switches to be tracked.At the same time,the actual scene evaluation shows that the improved tracking algorithm has a better pedestrian tracking effect under the single-source camera.