Research On Video Abnormal Event Detection And Authentication Methods

Recently, with the rapid developments of computer hardwares and Internet, the video surveillance system has been widely applied in various fields of human life. The vast surveillance videos also require more intelligent processing techniques. How to intelligently analyze the large amount of high-dimensional video data and effectively find emergency and abnormal events in them is a crucial task for social stability and property safety. It is also one of the most important research topics in the age of big data. At the same time, due to the insufficient of Internet transmission and management statute, the transmission of video data is likely to be malicious attacked and manipulated, which will hamper the data security. Therefore, how to protect the sensitive information in the video data has also become a hot research subject in the video authentication field.Since the video data is always with the characters of redundant, high-dimensional, complex, fuzzy, fickle and noisy, the abnormal events detection in the videos becomes a tough task. Simultaneously, the various attack techniques in the Internet also make it more difficult to prevent the videos from manipulating.In this thesis, the key problems in the video analysis, i.e., abnormal event detection and authentication are researched. The main works and contributions of this study are summarized as follows:1. In the video data representation aspect, we proposed a video abnormal events detection algorithm, which can effectively reduce the computation complexity caused by the high-dimensional video data. In this algorithm, the features are multiple hierarchically extracted to represent the input video based on the local spatio-temporal neighborhood information. Then, a feature selection technique is proposed based on the correlations between each feature and the class information to remove redundant features. At last, the efficient of our algorithm is evaluated on the UCSD database. Through comparing with other methods, it can be found that the proposed algorithm achieves better detection performance.2. In order to improve the effective and accuracy of video abnormal events detection, a joint algorithm which combines dimensionality reduction and sparse representation is proposed. In this algorithm, we firstly extract the motion features of the video by considering both the direction and energy of the motion objects to refine the optical flow histogram(OFH). Then, the dimensional reduction technique and sparse representation are combined for abnormal events detection and a effective strategy is also presented to optimize the objective function of our algorithm. By this strategy, the local neighborhood structure and main information of the original high-dimensional data can be well preserved.Furthermore, in order to deal with the video in which the object movements is slight, a sample selection strategy is also proposed based on the clustering. Through the sample selection, we can get the training data which best reflects the information of the events. Finally, the experimental results on UMN database demonstrate the efficient of our algorithm.3. For the sake of protect the sensitive information in the video data, we propose a video authentication algorithm based on visual hierarchy and sparse representation. The algorithm first exploits the correlation between the cover data and hiding information by sparse representation technique to reduce the amount of data to be embedded. Then, in order to further improve the invisibility and guarantee the safety of sensitive information during transmission, we propose a visual hierarchal strategy for video protection. This strategy utilizes the visual attention mechanism to partition the cover data into various saliency levels and embeds the hiding information hierarchically. In our study, we adopt the iris images as the hiding information due to their safety and stability. The sensitive information of video can be authenticated by either checking the integrality of iris images or iris image recognition. The experimental results on UCSD and UMN show that the proposed method exhibits good security, invisibility, and high capacity.