Abnormal Behavior Detection And Analysis Based On Videos

With the popularization of digital video cameras, smart phones and monitoring probes, large amounts of video data are produced every day. Consequently, people pay more and more attention to the storage and analysis of the massive video data. Abnormal behavior detection and analysis is one of the research hotspots, which has broad application prospects in intelligent surveillance, video retrieval and video rating.In this paper, we attempt to study both of the specific and nonspecific abnormal behavior detection problems. Some effective approaches are proposed in the aspects of feature extraction and abnormal behavior recognition.We investigated the challenging task of detecting violence in videos and propose a new video feature extraction framework based on sparse coding. Specifically, local spatio-temporal features are extracted from the video clip. To eliminate the feature noise and improve the efficiency, Kernel Density Estimation is exploited for feature selection on the original local features. In order to obtain the highly discriminative video feature, sparse coding scheme and max pooling procedure are employed to further process the low-level feature. Encouraging experimental results are obtained based on two challenging datasets which record both crowded scenes and non-crowded scenes.To detect the abnormal limb movement caused by neurological diseases, we proposed a detection method based on motion trajectories. Some markers pasted on the limbs are detected first. Constant velocity model and Kalman filter based method are exploited to track these markers. The final decision of abnormality is based on the average measurement residual in current trajectories. The experiments are carried out on videos provided by doctors and high accuracies have been achieved.The nonspecific abnormal behavior detection problems make the implicit assumption that normal instances are far more frequent than anomalies in the test data. The abnormal behavior can be identified as irregular behavior from normal ones. Based on this assumption, an unsupervised anomaly detection method is proposed. To describe the motion information, multi-scale histogram of optical flow orientation and overlap sliding window strategy are employed to extract the frame feature. Then sparse coding is adopted to transform the feature into sparse code vector over the normal dictionary. Normal behavior is likely to generate sparse vector with a small reconstruction cost, while abnormal behavior is dissimilar to any of the normal basis, thus generates a dense representation with a large reconstruction cost. Therefore the sparse reconstruction cost is computed to measure the abnormality of the testing sample. The effectiveness of the proposed method is validated on public datasets.