Research On Online Multi-Target Pedestrian Tracking

With the development of intelligent hardware,the field of computer vision has entered a period of rapid development.In this context,online multi-target tracking under video sequences has great scene application value.For example,in the video surveillance flow analysis,emerging artificial intelligence application scenarios such as smart city traffic and smart visual navigation play a very high role.Therefore,this paper establishes a set of algorithm models for common pedestrians with specific tracking objects,which provides a certain reference for the study of related issues.At present,the mainstream framework of multi-target tracking adopts the mode of detection-assisted tracking.This kind of strategy first obtains the region of interest through the detection model,that is,the target detection frame is extracted using the target detection network.Then a tracker is established to predict the next position of each tracking target;the target detected in the current frame is measured and matched with the tracked target.Innovating in the pedestrian target detection phase and data correlation phase of the tracking model can improve the overall performance of the tracking model.The main work of this article is as follows:1.In the pedestrian target detection phase of the tracking model,for the existing multi-target pedestrian detection methods,there are problems such as small target miss detection and low robustness in crowded scenes.This article innovates from three aspects: first,in target detection In the network design stage,a pedestrian target detection network with multi-scale feature fusion is designed.Second,in the candidate frame selection stage,the size of the candidate frame is re-clustered according to the characteristics of the slender pedestrian frame,so that the detection network is more effective when classifying the target It is targeted;third,in the case of mutual occlusion of pedestrians,the exclusion penalty weight is added to the original non-maximum suppression algorithm to improve the problem of overlapping detection frames being easily eliminated.2.In the data association phase of the tracking model,for the problem that pedestrian multi-target tracking is prone to mismatch in the case of object occlusion and so on,this article innovatively designs from two aspects: first,the pedestrians in the area are roughly classified based on blocks,Limiting the area of the tracking area during rematching,making the global matching problem optimized to the local domain matching problem,and improving the efficiency of the tracking algorithm;second,introducing a priori information about pedestrian groups and constructing a set of pedestrian groups The criterion of the identification makes pedestrians with group labels more robust in the tracking algorithm.From the perspective of the overall model design,the advantages of this article are various aspects of optimization and innovation.The details are optimized for the problem in different aspects.From the pedestrian target detection stage to the pedestrian target data correlation stage,they rely on each other to achieve a Satisfactory results.Finally,according to the content of the algorithm,a simple and easy-to-use online multi-target pedestrian tracking system is designed to provide an auxiliary reference for the embedded platform system.