Research On Video Abnormal Event Detection And Its Application Using High Order Feature And Saliency

In recent years, with the rapid development of computer and Internet technologies, and the increasing social security needs of many segments of society, intelligent video analysis has been booming. Intelligent video analysis technology analyses, understands video content, detects the abnormal events, and make early warning, by using artificial intelligence, machine learning, computer and other multi-disciplinary theories and methods. It is an important research field of intelligent analysis of big data and has important theoretical significance and wide application prospect. On the other hand, with the rapid growth of video data, the protection of video data has become a hot research topic.In this thesis, researches on abnormal event detection and video authentication are carried out deeply, which aiming to improve the performance of intelligent video analysis system. By mining deeply the motion attributes in videos, expanding the application of human visual perception mechanisms in the field of video analysis, finding more effective methods for event representing, it can improve the performance of abnormal evemt detection. The main work is summarized as follows:1. Taking into account a sudden change may cause abnormal movements, and the conventional video description method based on optical flow statistics can only describe the velocity, but not the change of velocity, a video event representation method based on high-order feature is proposed, to characterize the motion content of video more adequately and comprehensively. To avoid tracking algorithm in dealing with complex or crowded scenes, propose a short-time tracking method based on optical flow to obtain higher-order motion feature. In order to adapt the method to more video scenes, it integrates the traditional low-order features, and determines the optimal integration of low and high order feature using an online way. A large number of experimental results demonstrate the effectiveness of high-order features.2. In normal cases, the abnormal video content is often manifested as salient, while the visual attention mechanism is the powerful tool for processing high-dimensional and redundant data, so an abnormal event detection method based on spatial-temporal saliency is proposed. According to the spatial-temporal saliency map, non-salient regions are removed, and abnormality detection is carried on further. Due to the removal of the redundant content, buildding a normal event model of the entire scene efficiently becomes possible. Therefore, region-level models but not the block-level models are used. This not only greatly reduces model construction time, but also overcomes the problem existing in block-level models, and improves the detection performance.When constructing the spatial-temporal saliency map, two saliency detecton methods for image are proposed and be used to produce the saliency map for video by integrating the temporal saliency map. The two image saliency detection methods are as following:(1) An image saliency detection method based on extended region contrast and supervised locality preserving projection(ERC-SLPP) is presented. Two saliency maps based on region contrast and machine learning are constructed respectively, and be fused to form the final saliency map, realizing the complementary. In the process of constructing regional contrast based saliency map, image boundary extension is introduced by making full use of the prior that ’image boundary is most background. This operation can increase the contrast of salient regions with other areas in the image, thereby reach the goal that highlight the salient object and inhibit the background. When constructing the saliency map based on machine learning, SLPP algorithm is utilized for dimensionality reduction on the high-dimensional low-level visual features, and support vector machine is used for classification. The experimental results have demonstrated the excellent performance of ERC-SLPP in image saliency detection.(2) By using the boundary prior, an improved multi-manifold ranking(IMMR) based multi-view image saliency detection is proposed. Unlike most saliency detection methods that based on the contrast of salient region with other regions, the proposed method focus on the non-salient regions of image, considering the saliency detection as a multi-manifold ranking problem. In this method, image elements on boundaries are regarded as the seeds, and all elements will be assigned a ranking score according to the similar degree matrix constructed using multi-view features. Here the image element is described using multi-view features, and the correlations between the features are considered using IMMR algorithm, which realize the feature fusion in the process of saliency inference. Extensive experiments on several databases have shown its superiority.In saliency based abnormal events detection, video saliency map is constructed by combining with the two image sliency maps respectively. A large number of experimental results demonstrate the excellent performance of saliency based anomaly detection.3. In order to achieve authentication of video, an abnormal detection based video authentication and self-healing method is proposed. Because the abnormal area is often an important and sensitive areas and in view of "important objects should be protected important ly", a multi-levels recovery scheme is proposed, to recover the abnormal area loseelessly and improved recovery ability at the same time. In order to ensure the same abnormal location between embedding and extracting end, a synthetic frame is used to detect abnormality instead of the original video frames. In dual watermark embedding process, it selects the embedding position and embedding schemes adaptively based on the abnormal area to achieve blind extraction and blind recovery. Experimental results show that the proposed method can accurately locate the spatial, temporal and spatial-temporal tampering. It can restrore the tampered areas under a certain tampering rate, and restore the abnormal area losslessly. In all, it can protect the videos effectively.