Study On Algorithm About Multiple Targets Tracking Via Multiple Cameras

The concept target tracking stated in this paper, which exactly means visual target tracking, has the main task to study how to analyze the interesting target's movement status from continuous images or videos. Target tracking is one of the important and hot issues in computer vision. It is an active research topic which concerns pattern recognition, image processing, probability theory and statistical inference, random process, and system state estimation.With the development of target tracking technology, there have been many successful target tracking systems applied to all aspects of real life. Meanwhile, due to the actual needs, how to track multiple targets is an urgent need, so the concept multi-targets tracking (MTT) comes into our vision. Multi-targets tracking has already become one of the hot andifficult issues in visual target tracking research.In most cases, due to the limitation of camera's vision, the supervision system can not observe the entire interesting region through a single camera. In order to monitor a vaster region, what we need is a distributed supervision system using multiple cameras.Obviously, multi-targets tracking via multi-cameras technology is significant not only in academic research, but also in engineering applications .This paper studies three key technologies of the task of multi-target tracking via multi-cameras: target detecting, multi-targets tracking and multi-cameras coordination. My work can be summed up as follows:(1) Target detecting. First, I described the current status of target detecting technology, and focused on analyzing the advantages and disadvantages of the gaussian mixture model (GMM) and kernel density estimation(KDE) algorithm. Then I proposed a target detecting algorithm based on time-space and movement features, which has the advantages of strong anti-noise and speed.(2) Multi-targets tracking. First, I analyzed the difficult aspects of multi-targets tracking compared with single-target tracking, and proposed a particle filter algorithm based joint state; Then I introduced how to use Markov chain Monte Carlo (MCMC) in the particle filter framework to ensure real-time tracking, as well as how to use Markov Random Field (MRF) model and the appearance model to handle occlusions.(3) Multi-cameras coordination. First, I Introduced the method for mapping the targets' position between multiple cameras, and proposed an algorithm for detecting the calibration points automatically by using SIFT and virtual camera. Then, I introduced how to match the same targets between different cameras. At last, an algorithm for coordinating and assigning cameras was proposed.