Research On Multi-target Tracking Method Based On Detections

In recent years,with the continuous development and popularity of digital technology,it's easier to acquire,store,and process various video data to extract valuable information to achieve high-level comprehension of data.For maneuvering target tracking in visible sequence images,the feasibility,universality and rea l-time performance of tracking algorithm are all key factors to be considered.However,many existing methods show limitation in tracking accuracy and computing efficiency.In this paper,the characteristic,the deficiency,and the development tendency of many classical tracking algorithms are discussed and summarized.Accordingly,research is made focusing on the balance of tracking accuracy and efficiency.The main achievements is as follows:1.Based on the reading of extensive target tracking literatures,the paper reviewed the common target detection and tracking methods in sequence images.Focusing on the core idea and operation steps of two excellent multi-target tracking algorithms,the inadequacy in the tracking field was clear to set the basement for multi-target tracking algorithms with higher performance.2.A short interval algorithm based on multi-model and a long interval algorithm based on Kalman Filter in multi-target tracking is proposed to relieve the tracking difficulties caused by changing appearance and occlusion of the target.Firstly,the similarity between two targets in adjacent frames is calculated combining the color,position and size features according to the detection results from the HOG classifier,and the paper sets multiple criteria to get the reliable short tracklets.Secondly,the appearance discrimination model,the motion model,the time model,the detection model and the scene structure model are contributed and combined to gain the similarity matrix of the short tracklets.Then the Hungarian algorithm is used to calculate the optimal solution for those short tracklets to form longer tracklets.More complete trajectories are acquired by repeating the above operations.Finally,focusing on the targets that are occluded for a long time but with higher appearance similarities,the paper uses the Kalman Filter to get further association for the trajectories.Experimental results show that the algorithm is effective in improving the continuity of the target trajectories,and the algorithm can obviously relieve the interference of target occlusion.3.A hierarchical multi-target tracking algorithm based on bi-directional motion estimation is proposed in this paper.First of all,in order to improve computational efficiency of the above multi-model-based tracking method,this algorithm utilizes only the motion model of the target.Then the theory of bi-directional motion estimation is introduced to enhance the reliability of association method based on this single model,which means to calculate the backward and forward estimated target simultaneously from two tracklets and then measure the association degree of these two tracklets using the similarity of those estimated targets.In the final phase the optimal association of the short tracklets are formed by using multiple threshold conditions to gain complete target trajectories.From the experimental verification,the algorithm is not only proved to be superior in handling occlusions and imperfect detections but also shows advantage in real-time tracking..