Research On Multi-target Tracking Across Camera

The task of multi-target multi-camera tracking aims to find all the trajectories of pedestrians in different cameras.The challenge of this task is that on the one hand,processing multiple targets of multiple videos means a large amount of computation,on the other hand,in actual applications,long-term occlusion caused by non-overlapping placement of different cameras,as well as pedestrian attitude,illumination,viewpoints,etc.,cause a large difference in pedestrian appearance under different cameras,which adversely affects pedestrian trajectory matching.To deal with these problems,this thesis applies a multi-target tracking algorithm based on a hierarchical inference structure to reduce the computational complexity,and designs better appearance features to enhance the appearance correlation association between observations to achieve the purpose of matching trajectories under different cameras.In this thesis,a single camera multi-target tracking algorithm based on graph partition model is used to track.The algorithm includes a two-layer inference structure to reduce the complexity of the algorithm.Each level is performed in a sliding window to achieve real-time purposes.The sliding window for the second layer may cause missed detections near the junction.We propose a method of overlapping sliding windows to reduce the missed detection rate at the junction.To solve the problem of identity switches that may occur in the case of mutual occlusion between targets,we propose a method of trajectory constraint to reduce the occurrence of identity switches.To solve the problem of trajectory appearance changing caused by viewpoints under different cameras,this thesis proposes a multi-target multi-camera tracking algorithm combining pedestrian directions and block HSV histogram features as trajectories' appearance feature,which constrains the different pedestrian trajectories in the same direction,and the weight of the same pedestrian trajectory in the same direction is increased,finally the influence of the pedestrian direction change on the appearance characteristics of the trajectory is easied,and the tracking accuracy is improved.Aiming at the problem that the feature generalization ability of manual design is not strong,and the multi-target tracking across the camera is a heavy recognition problem,this thesis proposes a cross-camera multi-target based on the fusion learning direction and the overall feature based on the previous algorithm framework.The tracking algorithm,which designs a triplet loss function that fuses the direction information and the overall appearance characteristics.The loss function can suppress the influence of the pedestrian direction change on the pedestrian appearance characteristics,and has the overall appearance feature with strong discriminative ability.lighting,etc.are robust.Finally,the proposed algorithm further improves the tracking effect.