The Parallel Design And Implementation Of MCMC Multi-Object Tracking Algorithm

With the rapid development of computer vision technology,the video object tracking technology has been made extensive research by domestic and foreign scholars in many fields,such as Intelligent Video Monitor and Alarm System,human-computer interaction,Visual Navigation and so on.According to the number of moving objects in the video,the video object tracking can be divided into Single target tracking and Multi-target tracking.In Multi-target tracking,the system needs to handle more than one target's prediction of the position and correction of the trajectory in the process of tracking,so it produces higher computational complexity and has more challenge.In recent years,many algorithms have been proposed,but it is still a hot and difficult problem to design a tracking algorithm with both good stability and real time implementation.Nowadays,the general parallel computing graphics processor has become the basic configuration of the mainstream computer.It is an inevitable trend to improve the computational efficiency of the algorithm by using the computing power of GPU.The traditional multi-target tracking algorithm is usually serial computing,GPU is not taken into account in the design,and the algorithm does not make full use of the computing resources.In this thesis,we do our research in the parallelization of the MCMC algorithm,using the static background video sequences as the research object,based on the CPU-GPU heterogeneous computer system.We first introduce the basic architecture of GPU and discuss the basic concept of the CUDA instruction system,and then describe the design process and the method of GPU based parallel programing.After the discussion of basic realization methods of Bayesian tracking model,the thesis gives an Extended Bayesian tracking model in the multi-target tracking problem.Then the thesis describes the MCMC multi-target tracking algorithm and analyzes the key problems and technical difficulties in the implementation of the algorithm.In the section of parallel target detection,the fundamental principles of the moving object detection in static background video are discussed and the parallel Gaussian Mixture Model based background subtraction method is analyzed.In order to meet the requirements of the sampling of multi-target state in the tracking part,we propose an optimization method based on the confidence map.We add the confidence map generation part in the parallel Gaussian Mixture Model based background subtraction algorithm;calculate the pixel confidence value within the target region using the moving target's intensity and the strength spread function,so it can provide more information for the tracker in the estimation of the target state.In the section of the parallelization of the multi-target tracking,this thesis describes the basic principle of IP-MCMC and the parallel tempering.For the time-consuming problem of MCMC algorithm in target state sampling,we propose an optimization algorithm based on the multi-chain Monte Carlo algorithm.In this algorithm,multiple sampling chains are constructed in the thread blocks on GPU,firstly it gets the initial stable distribution sampling points using the parallel tempering method,after that it obtains the sampling results through the implementation of Multi-Chain MCMC algorithm,and then it obtains the posterior estimation of the multi-target state.The experimental results shows that the proposed method can make full use of the parallel computing power of GPU,can reduce the sampling time of multi target state and improve the computational efficiency of the algorithm.The algorithm proposed in this thesis has certain theoretical value to the parallel computing method in the research of real-time video multi-target tracking.Due to the parallel reconstruction introduced in the target sampling stage,the robustness of the algorithm is subject to some restrictions,so it still cannot meet the need of practical engineering application.How to continue to improve the accuracy and robustness of the multi-target tracking algorithm is the problem to be solved in the future work.