Research On Fusion Of Multiple Features In Intelligent Video Surveillance

Nowsday public safety and traffic safety have been paid more attention on. Traditional video surveillance system can not afford of the demand for it need more human resource and people can't concentrate these attentions on work in long time. Intelligent video surveillance system become a new research hot spot in computer vision. Intelligent video surveillance is the application of computer vision and pattern recognition in filed of video surveillance. A key issue in intelligent video surveillance is how to understand the context of image effectively. For complex of scene and confine of feature, it is difficult that have good performance using a single feature in pattern classification. In many applications, it can have good performance using fusion of multiple features. The work in this paper is mainly how to classification of foreground and background effectively using fusion of multiple features in the moving object detection and tracking.The main works and invanations of this paper as follows:(1) A new recursive background learning model based on edge feature is proposed in this paper and this method is robust for illumination change .This method can eliminate the effect of local illumination change and the tiny swing of video captures , but the shortcoming of this method only can have the sparse represention of the moving object. The background models basing on color feature are compared and quantitative analyzed in this paper. The background model based on color feature can the dense represention of the moving object, but this method is sensitive for illumination change. When object stay long time in background begins to move and local illumination change, the flase object will be detected. A new background model basing on fusion of color and edge features is proposed in this paper and this method can elimate the flase object results for the above reasons effectively.(2) For cast shadow shares the same characteristices in some features, cast shadow is usually mistook to the moving object. It is difficult of the classification of moving object and cast shadow using a single feature .A new feature selection framework for classification of foreground and shadow is proposed in this paper. By learning using the give samples, a linear classifier of fusion of multiple features is trained. Experierment demonstrates this model has better performance than the model using a single feature.(3) A new layered background model on fusion of region feature and local feature is proposed in this paper. The background is layered by region feature. The scene is divided to dynamic background layer and static background layer. Different background model is adopted for different layer. A Gaussian Mixture Model based on color feature is adopted in static layer. A background model basing on fusion of motion feature and color feature is adopted in dynamic layer. Experierment demonstrates this model has better performance and less run time than the previous model.(4) Scale change, rotate change and occluding of moving object can cause to the failure of object tracking. The problem of object tracking is a binary label problem. First the probabilities of pixies belong to forground and background is calcaculted. Cosidering the space coherence of pixels, a markov random filed of foreground segment is modeled. Experierment demonstrates this model is robust for scale change, rotate change and occluding.