Research On Multi-Target Tracking Based On Hough Forest In Complex Scenes

With the rapid development of computer hardware and video surveillance technology,target detection and tracking have gradually become the focus of computer vision research and always attract wide attention of researchers.Especially,after the concept of artificial intelligence is put forward,the video intelligence analysis technology is pushed to a new high.As its important branch,multi-target tracking is mainly to confirm the identity of each target in the video sequence,locate the target,and further form the trajectory of the specific target.At present,the research of multi-target tracking has made technological progress continuously due to long-term research and development,but there are still some points for further study,such as complex tracking scene,the change in the appearance of the target,occlusion and deformation,real-time tracking and so on.This dissertation focuses on multi-target tracking methods in complex scene,which deal with some problems of similar targets,target occlusion and deformation effectively.The main innovative research and achievements of this dissertation are as follows:Firstly,a target detection algorithm based on DPM is proposed for the different moving targets in complex environment.The HOG feature is improved and enhanced on the basis of traditional HOG feature,which can describe the targets better and reduce the computational complexity effectively.This method utilizes the combination of the root filter and multiple part filters to locate the outline of the target and detect the key parts of the target accurately,and then to determine the real goal according to the comprehensive response and further improve the precision of target detection.Secondly,on the basis of target detection,we propose a multi-target tracking algorithm based on Hough Forest.According to the primary association strategy,the algorithm generates reliable tracklets,extracts training samples online and constructs the target feature model with color histogram of multi-part fusion,HOG feature and optical flow information to further ensure the better multi-dimensional description of the target.By the discriminative learning of the Hough Forest,we summarize the statistics stored in the leaf nodes which are appearance and motion features from samples to estimate the linking probability between tracklets.Moreover,the reliable tracklets are hierarchically associated into the long trajectories of the targets,which effectively solves the interference caused by the similar targets and partial occlusion in the tracking.Finally,in order to deal with the errors in primary association and the inherent accumulation of errors in Hough Forest,a multi-target tracking algorithm based on the multi-feature fusion and Hough Forest is proposed.On the basis of the original method,the method adds double-threshold association strategy,which effectively reduces the mismatch in primary association and ensures the reliability of tracklets and samples.According to the results of the first level output from Hough Forest,we propose a trajectory matching method with multi-feature fusion,which combines similarity measure in color histogram and feature points matching based on multi-scale Gabor filter to build the matching probability model with the weighted factor.In this way,the long trajectories that Hough Forest output are hierarchically linked to the complete trajectories of the targets by matching the long trajectories of the targets respectively.Meanwhile,the algorithm can effectively improve the ability of solving the problems of target mutual occlusion and deformation in the video sequences,and realize the robust tracking of multiple targets.