Moving Target Detection And Classification Based On Saliency For Outdoor Video Surveillance

Saliency detection is a classic subject in machine vision, which simulates the human visual processing to emphasis the region of rapid change in color and intensity. Because of not considering movement, some static areas may be highlighted for the reason of rapid change in gray. Moving target detection emphasizes on the areas different from the whole background, but some interference elements, such as camera movement and snaking branches, can affect the real small moving targets. So combining the saliency detection with motion detection, we can overcome the interference and get the significant moving target.Target classification is a significant procedure in behavior understanding and can provide information on target tracking and behavior analysis. Outdoor video surveillance is likely to be effected by many factors, such as light, weather, shelter and so on. In order to increase the speed of classification and improve accuracy, typical characteristics must be selected to meet that.The main jobs and innovations in this paper are as following:(1) By combining the global and regional characters, this paper proposes a modified FT algorithm based on multi-resolution, which can get a target object with clear boundary and uniform saliency.(2) A motion estimation algorithm based on block gray projection is proposed here. By comparing the gray projection of current frame and the reference frame in horizontal and vertical directions, this algorithm calculates movement vector and get the motion saliency map.(3) With a dynamic fusion technology and the dynamic preferred thoughts, the target can be detected by fusing the static saliency map and motion saliency map.(4) In targets classification, the classification features, such as aspect ratio, dispersion and duty cycle are selected, and the classifier is selected to use the SVM which is suitable for small sample study and has good generalization ability. Vehicles, pedestrians, cyclists and other targets are classified.(5) The system uses Matlab to test videos in the database from http://ftp.pes.rdg.ac.uk/.Comparing the inter-frame difference algorithm, time average method and Gaussian mixture model, the method in this paper has higher accuracy and robustness in target detection. In classification a grid optimization and5-fold cross-validation method is used to find the optimal classification kernel function RBF kernel and optimal parameters, and finally get the classification result with high recognition.