An Integrated Study Of Privacy-enabled Video Coding And Analysis Based On CS

Compressive sensing(CS) is an emerging research that provides a framework for image reconstruction using sub-Nyquist sampling rates. The core idea of the CS theory is:if the signal is sparse in some linear transform domain, by virtue of the sub.Nyquist sampling rates can there be a perfect reconstruction of the original signal, or a robust approximation in the presence of noise. For computer vision applications, it is known that natural images have a sparse representation in the wavelet domain. According to the CS theory, by taking random projections of a scene onto a test function set that is incoherent with the wavelet basis, it is able to reconstruct the image by convex optimization algorithm.The CS theory became the focus of the image circles at home and abroad as soon as it was generated. It’s stated that the CS theory can easily capture the key information in the image using relatively few projections obtained by extraordinary measurements. Naturally, the video composed by a series of static images was included within the range of CS applications. The main research contributions of this thesis are as follows:1. A CS-based tracking algorithm with security features is proposed in this thesis, achieving target tracking and privacy protection synchronously. This algorithm integrates CS theory and the classical background subtraction algorithm, which can make the extensive placement of video surveillance systems increase the capability of individual protection. Here, the confidentiality is mainly reflected in two aspects:First, the inherent confidentiality of CS measurement. The measurement matrix and the projection process are regarded as a key and an encryption of sensitive areas respectively. The decoder can not decrypt when there is no key in hand, then it protects the information security. In other words, it is impossible to reconstruct the original video content from the encoded random projections alone, so the security of data information is protected. Second, this algorithm doesn’t need to reconstruct the video frame but be able to track objects of interest, so it has the advantage of privacy protection which other general background subtraction algorithms don’t have. In addition, according to the distance of the target from the camera, the paper designs a background update method in projection domain. Simulation results show that the algorithm proposed here can effectively reconstruct real image from the incomplete measurements with the right measurement matrix, and can also have the advantage of computational secure which can be more useful in some secrecy applications.2. The thesis studies a joint CS theory and video analysis mutual feedback and iterative optimization system. It uses the predicted positions and sizes of the bounding boxes obtained from the video analysis as a priori information or constraints for the CS reconstruction or decoding module to guide the foreground reconstruction, improving the reconstruction quality. The improved quality is in turn used to optimize the results of target tracking. When the signal analysis tasks are embedded in the compressive sensing theory, which measures and reconstructs a certain type of signals, and by iterative constraints and optimization there will be potential coding gain and decoding quality improvement. Compared with the classical algorithm, the algorithm proposed here achieves bit-rate decrease to some extent, meanwhile, obtains a potentially lager gains in performance.