| In modern society, the security issue attracts more and more interests and incurs the idea of intelligent surveillance system. A complete such system includes varies kinds of video analysis modules, which requires large amounts of computations. Thus, the dissertation proposes a solution of surveillance platform based on the CELL multi-core processor and a concept of Feature Library which is used to manage all kinds of sub-modules. The key area of the dissertation, video object tracking, is one of the features in the library.Video object tracking is a fundamental research area in the field of intelligent video analysis. The precise tracking result is critical to many high level applications. After more than ten years of research, although scientists have provided several practical tracking algorithms, the current fruits are far from perfect. The dissertation has focused on the CAMSHIFT algorithm and particle filter algorithm. After comparing their strengths and weakness, particle filter is chosen to be used as the framework of multi-target tracking in the complex environment. By modifying the program of single target tracking, the new algorithm fixes the problem of occlusion. The test result shows that the modified algorithm of the dissertation works properly when the target is partially or even completely occluded.One issue of particle filter is its computational complexity. To meet the real time requirement, it has to be optimized according to the architecture of CELL BE. The optimization is implemented in two different ways: one is to utilize SIMD instructions to accelerate functions such as calculating color histograms; the other is to divide the computations into several parts and deliver them into different synergic cores. The test result shows that the speedup after optimization has something to do with target number and target size. In general, the speedup equals two times the target number (for example, 16 targets tracking can achieve 32 times speedup). |
PDF Full Text Request