The Estimation About The 3D Location Of Targets In Image And Its Applications In Multi-Target Tracking

In recent years,multi-target tracking has become a research hotspot in the field of computer vision,especially in such fields as intelligent monitoring applications are very wide.Usually,the computer can be used to detect and track the target,which can largely replace the work of human beings and reduce the consumption of various resources.In the early stage of its research,multi-objective tracking is basically carried out in a single perspective.So far,many scholars have proposed many excellent algorithms.However,there are still limitations in the single angle of view.The most prominent problem is that it is impossible to avoid occlusion in the single angle of view.At the same time,if the target appearance information in the image obtained from the single angle of view is fuzzy,it will have a great impact on tracking.In this regard,many people consider to do multi-objective tracking in multi-perspective environment,so that the redundant complementary information between different perspectives can be used to solve some problems in single-perspective tracking.Compared with the single-perspective multi-objective tracking research,the data correlation between perspectives(reconstruction)should be considered when multi-objective tracking is conducted in the multi-perspective environment,which is also a research focus in recent years.Generally,the 3D coordinates of the same target in different perspectives are not changed under ideal conditions,which is also the key to solve the problem of data correlation between perspectives.This paper aims to realize the fusion of redundant complementary information based on the consistency of 3D coordinates of the same tracking target in different perspectives,so as to improve the overall tracking performance.The main work of this thesis is as follows:(1)In this paper,an improved camera parameter estimation method based on pedestrian path information is proposed to help transform the tracking of a target in a two-dimensional image into a three-dimensional space.Compared with the original camera parameter estimation method,this method can obtain more accurate camera parameters.(2)In this paper,a new 3D coordinate reconstruction method-"MI method”is proposed for multi-object tracking.In this method,the idea of image similarity measurement is introduced.By matching the position information of the same tracking target from different angles,the optimal ground height is found and the 3D coordinates of the tracking target are reconstructed.Compared with the original 3D coordinate reconstruction methods,the MI method" improves the accuracy of the reconstructed 3D coordinates and has good anti-noise performance.(3)At the end of this paper,a multi-perspective multi-objective tracking system based on the minimum cost flow is built and the performance of the system in the front-end data processing stage is optimized by using the "MI method".At the same time,the performance of the system was tested on the video data set of PETS2009 and compared with other tracking systems.The results show that the tracking system built in this paper has achieved good tracking effect.