Research On Multiple Target Tracking Using Visual Information

Multi-target tracking can be achieved by continuous position estimation for multi-target locations in specific application scenarios. Multi-target tracking technology has been widely used in many fields, such as video conference, surveillance system, transportation network system, robot operation, military, etc. Multi-target tracking can extract kinds of information, and the usage of video information contained therein can obtain various parameters of target easily and intuitively. With the development of science and technology, it's very easy to obtain video material. But lacking manpower to analyze video content makes it very urgent to research and design a feasible scheme for video analysis by means of machine intelligence.The most critical technical problem in multiple target tracking is the research of distinction and association of multiple targets. It's valuable to adopt the methods of joint probability data and nearest neighbor when the number of targets is small. But the effect of these methods would be limited with multiple targets and the explosive growth of amount of computation of data association technology. Nowadays, multi-target tracking method, based on Random Finite Sets, has become a hot research topic. It can reduce the computation amount and shorten the operation time. More importantly, this technology avoids complex data association and makes it less sensitive to target quantity.The purpose of this thesis is to achieve multi-target tracking by virtue of video information and to analyze the tracking effects based on Multi-Bernoulli Filter Method and Random Finite Sets Theory. The main works are as follows:(1) To study the theories of video information collection, processing, feature extraction, etc. To research the similarity of targets by means of matching HSV color-histogram of colorful images with SIFT features, both of which, with adoption of fuzzy logic fusion, can successfully solve these problems, such as target deformation, blocking, image noise, etc.(2) To complete the simulation of video tracking algorithm combined with HSV color histograms and video contour information based on particle filter algorithm of nonlinear objective tracking of Bayesian Filtering Theory.(3) To study Random Finite Sets Theories and analyze the algorithm of PHD, CPHD generated by Random Finite Sets and the algorithm of Multi-Bernoulli Filter Method. To make a comparison among the three simulation algorithms above and to select Multi-Bernoulli filter in order to achieve multiple video-targets tracking on the basis of the simulation results.(4) To make target feature information database. To complete some processes of video-object tracking via Multi-Bernoulli filter with adoption of feature fusion method. To achieve a relatively smaller-ratio target tracking and analyze the tracking results. To modify the related parameter factors according to the video feature and get a better tracking robustness with adoption of feature fusion method.