Research On Abnormal Behavior Detection In Surveillance Videos

This paper focuses on the application of intelligent video surveillance analysis technology in monitoring city appearance violation behavior. The primary contents are as follows:1、Foreground detection is one of the key step for the follow identification work in intelligent video surveillance, three effective algorithms were proposed to detect the foreground(including stationary foreground, such as abandoned object, long standing pedestrians, vehicles). Firstly, combining the characteristics of the moving average method and approximate median method, a double background modeling method was proposed to detect the stationary target by the difference of the fast and the slow background with the different refresh rate. This algorithm is robust with less computation, and does not need to track the targets. Then, two backgrounds were constructed to detect static foregrounds, an evidence aggregate image and a permitted occlusion time parameter were introduced with the purpose to reduce the false alarms and handle occlusions necessarily occurred. This method runs average around50frames per second for a sequence with an image of352×288in Matlab7.1for real-time surveillance tasks. Finally, an effective static objects detection method by improving the updating approach of Gaussian mixture model (GMM) was proposed to detect the stationary targets. This method retains the advantages of GMM and can handle occlusions well. The experimental results under different conditions demonstrate that the proposed method achieves a better performance under complex background condition.2. Two classification algorithms were proposed to distinguish the abandoned and removed objects (stolen things, ghost, etc.). Illumination/shadow filter is designed for the detected static foreground, long standing pedestrians and vehicles were excluded to further improve the robustness of the stationary target detection. Then, an identification method was proposed based on a weighted feature of rich colors and edges, the experimental results show that the classification results were greatly improved comparing to the existing algorithm edge-based method and color-based method. To further improve the timeliness and the classification results, a real-time abandoned and theft discrimination method was presented based on a boundary spatial color contrast. Compared to other existing methods, this method greatly reduces the computational time-consuming and classification accuracy rate is higher. This algorithm provides a solid foundation for the subsequent tracking, classification, processing and has a strong practical value for its good performance and real-time detection.3、A head-reference human contour model (HHCM) was proposed for the purpose to improve the human motion recognition performance with human contour as the main feature. Compared with the existing approach to human behavior recognition with human silhouette as the main feature, this method has the following characteristics:1) quickly extract human continuous closed contour by employing the Level Set method without re-initialization;2) Since the head was positioned only on the human body which is unaffected by background, the head positioning is simple and fast;3) human contour information and motion characteristics from the typical key frame were enough to identify human behavior;4) human contour model based-on the head can effectively improve the clustering performance of feature vectors. In order to detect abnormal behavior in human interaction under various environments, a human abnormal behavior detection based on the optical flow feature was proposed. The author defines the motion entropy based on the amplitude-based weighted orientation histogram to measure the anomaly of human activity, and an anomaly judge coefficient was proposed with the weighted sum of motion entropy and motion energy to recognize abnormal events. The test results under different behaviors dataset demonstrate the effectiveness of this method.