Behavior Analysis Of Object For Video Surveillance

Visual Surveillance is becoming more and more important in our society. The traditional surveillance systems need a lot of interactions with humans while lacking of intelligence module that can analyze behavior of objects appearing in the video automatically. Such kind of system is not only leads to great waste of human resources, lay too much burden on the human operators, but also may get unreliable results. So the objective of this research is develop some intelligent functions relying theory of pattern recognition and machine learning for surveillance systems, so that these systems can be easier to use and more reliable.This paper is mainly on tracking methods of non-rigid object,object undergoing affine motion and multiple objects through occlusion. For tracking of non-rigid object, we propose a local patch-based tracking decomposition framework. Our algorithm is based on a visual tracking decomposition scheme for the ef?cient design of basic trackers. The combination of basic trackers can solve the problem of abrupt motions. Furthermore, the fusion of multiple features makes the tracker more robust to illumination. For tracking of object undergoing affine motion, the 2-D af?ne motion of a given object template is estimated in a video sequence by means of coordinate-invariant particle ?ltering on the 2-D af?ne group instead of local coordinate-based approach, which gets more accurate estimation for parameters of affine motion. For tracking of multiple objects through occlusion, we propose a novel approach for real-time tracking multiple objects, which combines feature correspondence with a probabilistic appearance model. The locations of multiple objects through occlusion can be estimated accurately via probabilistic voting. All the proposed schemes and algorithms are tested with video sequence, and achieve their expected performances.Recognizing human action is also of great significance for visual surveillance system. A new spatio-temporal interest point detector using 2D Gabor ?lters is presented to extract features of human action accurately, which is robust to occlusion, lighting changes and camera zooming. Furthermore, centering on the detected spatio-temporal interest point, a polyhedron with eighty faces model-based spatio-temporal gradient descriptor is created to illustrate the spatio-temporal visual features of human action. The low-level weight histogram and high-level semantic attributes are fused together and the latent Support Vector Machine (SVM) combined with coordinate descent is adopted to find the local optimum of the prediction model instead of the single SVM classifier. Experiments using some kinds of typical datasets demonstrate that our approach achieves a higher recognition rate compared to existing methods.