Research On Key Techniques Of Multi-target Tracking In Multi-view Environment

Multi-target tracking is popular in the field of computer vision in recent years for its great significance in the field of intelligent monitoring.Detecting and tracking the target of interest by computer instead of by people can greatly reduce human resource consumption.The original multi-target tracking is based on single-view environment.So far,there have been a lot of excellent single-view multi-target tracking algorithms.However,they still can't solve the occlusion problem well.The use of redundant complementary information from multiple views provides the possibility to solve the occlusion problem.Compared with single-view multi-target tracking,multi-view multi-target tracking should solve both temporal data association and data association between cameras.The data association between cameras is generally based on the consistency of 3D positions,which belong to the same target under different views.This thesis aims to solve information fusion under different views based on 3D position,and make full use of the complementary information of each view to improve the overall tracking performance.The main work of this thesis is as follows:(1)This thesis proposes a new 3D coordinate reconstruction method---"Z0 search method",which obtains the optimal solution by quickly changing horizon's location.Compared with the original monocular 3D coordinate reconstruction method,the method is more accurate.Compared with the method of computing the minimum distance of two skew lines,the method is easier to solve.(2)Based on the "Z0 search method" and the monocular 3D coordinate reconstruction method,this thesis proposes a "double probability" pairing method based on detection.Its idea is to select at most two candidate detections in another view for each detection.The reconstruction error of the candidate pairing is transformed into the corresponding probability by the Gaussian mapping function,and the probability is then as the pairing weight for the subsequent tracklet association process.The tracklet association method based on "dual probability pairing" exerts the compensation advantage of the reliable tracklet to the 3D reconstruction process.And the tracklet association method forms a reliable coupling based on tracklets.(3)Finally,the coupling is established as a node in the network flow graph to establish a minimum cost flow tracking model.The tracking system of this thesis is tested on experiment data set of PETS2009.And MOT matrix scores are compared with methods' in other references,which verifies the advantages of the multi-view multi-target tracking system designed in this thesis.