Research On Image Appearance Analysis And Application Method Under Multi-target Tracking

Multi-target tracking technology is a research hotspot in the field of computer vision and pattern recognition,and has a wide range of application scenarios.In the prior art,appearance models and motion models are often combined,different targets are compared and matched,and finally the matching results are obtained by association.For multi-target tracking,the appearance model is crucial and indispensable.This article mainly studies the appearance of images under multi-target tracking,and extracts the appearance characteristics of targets through neural networks.At the same time,it also takes into account the effects of unstable target detection,frequent occlusion of targets,and changes in background lighting.Improve the similarity of appearance features of the same target between frames,and reduce the similarity of appearance features of different targets.Firstly,in order to solve the problem that the target detector gives unstable original detection before tracking operation,this paper combines a human key point extraction and side-by-side height method,and proposes a new side-by-side height and human body.The key frame detection frame correction method can correct the original detection frame.Secondly,in order to accurately measure the pedestrian occlusion relationship and the visible area of pedestrians in the target detection frame,this paper proposes the concept of occlusion relationship and visible area measurement method,which can quantify the situation in the detection frame.Thirdly,by referring to the siamese network framework in pedestrian re-identification and combining the proposed pedestrian visible area measurement method,a new image appearance matching framework method based on pedestrian visible area is proposed.The appearance features extracted through the neural network can reduce the appearance feature distance of the same target and widen the appearance feature distance of different targets,which can well match the same target between frames,including the occluded target.Including the trajectory segment generation method proposed later in this paper,the association of multi-view trajectory segments and the coupling of multi-view trajectory segments are all combined with this method to design the powerful matching ability of the same target.Fourthly,build a complete multi-target tracking system.The trajectory segment is generated using the trajectory segment generation method based on image appearance matching,and the cross-view image appearance similarity method is used to associate the trajectory segments of multiple perspectives,and the multi-perspective trajectory segment is coupled by combining the appearance similarity and the motion similarity,and the trajectory segment Coupled as nodes of the network flow,a minimum cost flow multi-target tracking system is established,the optimal solution is obtained by using the continuous shortest path algorithm,and the complete trajectory is finally output.The tracking system of this paper is tested on the PETS2009 dataset,and compared with traditional methods and other leading tracking methods at home and abroad,the advantages of the multi-target tracking method designed in this paper are highlighted,and problems that need to be improved are identified.