Research On Object Description Based On Motion Perception And Anomaly Attention

The object of anomaly detection and analysis in video is to detect and locate the anomaly in scene, and has important academic research value and bright future of applications. The difficult point is that the definition of anomaly has diversity and complexity in different applied scenes, and current researches mainly focus on the recognition of limited types of simple behaviors, or anomaly behavior detection in a specific scene.Human visual attention mechanism is rarely considered of being employed as the key criteria of anomaly recognition system. Motion perception is the process of inferring the speed and direction of elements in a scene based on visual. Human Visual System has high priority of motion perception feature over other types of low-level features. Visual stimuli in the scope of perception will be paid attention to, and the stimuli out of the scope will be ignored. Modeling motion perception provides the strongest clue to detecting abnormal activities.In general, anomaly detection method builds prior template or statistical parameter model of normal behavior firstly, and anomaly can be classified by computing difference between test samples and prior template (or model). Surprise computation model based on Bayesian theory measures the differences between prior distribution and posterior distribution.Their key to successful modeling, however, lies in the extraction of highly descriptive and discriminative features for a minimal compact representation of scene activity. It is shown in recent researches that Human Visual System has the ability of capturing sparsely critical information in natural scene. Sparse coding model enables activities in the scene can be represented with compact and few basis vectors.For above issues, the research on object description method of anomaly discovery based on dynamic perception model, surprise computation model and sparse coding model is conducted in the doctoral dissertation.The main work and innovation are as follows.1) Combining with the motion attention mechanism of Human Visual System, a novel abnormal object discovery method based on motion perception model is proposed. Motion attention model based on DCT blocks classification is exploited to model dynamic perception region, obtaining the collection of motion attention blocks, the features of HNF in the motion attention blocks are extracted as feature samples to input into the sparse coding model to produce a dictionary learned. The reconstruction error is employed as objective function to judge if there is existing abnormal object discovered in the scene. Experiment shows that the proposed method is effective, practical and implemented easily.2) For the distinctive differences between prior distribution and posterior distribution caused by the appearance of anomaly in video, a method of crowd abnormal event discovery is presented. Block matching motion estimation is used firstly for video sequences, obtaining motion vector map in every frame of the scene. The number of low-speed motion blocks and high-speed motion blocks are computed statistically in multiple directions bins, and multi-scale motion histogram feature is extracted. Surprise computation model is exploited to acquire the surprise value of every frame, which is the discrimination rule of the existence of crowd abnormal events in current frame. The experiment results show that the proposed method can discover specific type of crowd events and the comparison with ground-truth shows the efficiency.3) An individual anomaly discovery method based on Bayesian surprise computation model is proposed. Gaussian Background Model is employed for background modeling, and the bounding box of foreground objects and the ratio of width and height. Optical flow field estimation is conducted meanwhile, obtaining optical flow motion field in every frame, by which the average velocity of foreground objects is computed. Bivariant surprise model is used for detecting the abnormal object with abrupt feature changes in temporal. Experimental results demonstrate that the presented method can discover two types of individual anomaly, running abruptly and falling down.4) Combining with the ability of capturing the key information of nature scenes sparsely in Human Visual System. An anomaly crowd detection algorithm for video based on spare coding model is proposed. The feature of multi-scale motion histogram is extracted, and fast sparse coding algorithm is exploited to learn the dictionary. If the representation error of the features in the scenes based on the learned dictionary is beyond a predefined threshold, it is discriminated as anomaly. Two standard behavior recognition dataset are used in the work and the results demonstrate the practical applicability and efficiency of the method.5) An individual anomaly detection algorithm for video based on spare coding model is proposed. The HOG/HOF of space and time interest points is extracted as high-dimension feature, and fast sparse coding algorithm is exploited to learn the dictionary. If the representation error of the features in the scenes based on the learned dictionary is beyond a predefined threshold, it is discriminated as anomaly. Experimental results demonstrate that the presented method can discover effectively two types of individual anomaly, running abruptly and falling down.