Anomaly Event Detection-based Traffic Surveillance Video Summarization

With the increase of the number of cameras in traffic scene, surveillance video data is growing rapidly while most of the events that take place in the scene are not the focus of attention. We name the event that people pay attention to as abnormal event in this paper and propose a method to get the video summary with anomaly event which shows the main content that we focus on.The method proposed in this paper consists of object detection and tracking, abnormal event detection and video summarization. We extract the feature of the frames in the given video and localize the objects frame-by-frame with Latent-SVM. Then track the objects to obtain the trajectories which are representative of motion events. After analyzing the characteristics of anomaly events we propose a new feature vector and build Random Forests with Bootstrapping as classification model. Then the model is used to detect the anomaly event. Finally, we propose two types of video summarization method while one is implemented based on key areas and another is based on key frames. In the first method, key areas are reorganized into new frames to generate a new video. The second is implemented by filtering the frames which contain the detection area of the anomaly event to generate a new video.Experiments confirm the effectiveness, accuracy and simplicity of the novel method.