Object Detection,Visual Tracking And Image Fusion Algorithms For Intelligent Video Surveillance

Intelligence video surveillance(IVS) is a novel technology because it draws the achievements of computer vision into traditional video surveillance. IVS can realize image understanding automatically and keep alert to abnormally individual or group behaviors. Since an IVS system often detects object, tracks moving target and achieves video analysis with the help of multi-camera. Hence, the corresponding researches on this domain focus on image fusion, object detecting, target classification, visual tracking and video scene understanding etc.The main works of this dissertation include moving object detection, visual tracking and multi-image fusion. The major challenges in the research of object detection, tracking and image fusion are three-fold: first, how to eliminate the climate or light changes, the background disturbance, the movement shadow and the shaking of camera etc in object detection. Second, how to solve the occlusion problem in single-target or multi-target visual tracking. Third, how to realize fast image fusion. This dissertation focuses on above-mentioned challenges and investigates robust and fast object detection, visual tracking and image fusion algorithms for intelligence video surveillance. The main contributions are summarized as follows:(1) The common works for detecting moving object are background subtraction(BS). The BS algorithms operate in the spatial domain and require a large amount of training sequences to estimate a background model. The training process always imposes high computational complexity. In order to overcome the limitation of BS algorithms, a compressed sensing based algorithm is proposed for the detection of moving object in video sequences. The proposed algorithm does not require any training sequences for estimating background model and uses the compressed measurement to achieve robust object detecting. In our algorithm, we first use a three dimensional sampling method to yield compressed measurements. Then, we propose a compressed sensing and matrix decomposition based object detection model to simultaneously reconstruct the foreground, background and video sequence. Finally, a confidence map is estimated through using the reconstructed video sequence. The foreground reconstruction result in the object detection model can be improved by utilizing the confidence map. Experimental results show that the proposed object detection algorithm is robust to the movement turbulence and sudden illumination changes.(2) In the real-world video surveillance, object detection is difficult because it not only suffers the movement turbulence and the illumination changes, but also faces with the video noise and camera shaking. To enhance the detection robustness in video noise and camera shaking, another compressed sensing based algorithm is proposed for object detection. This algorithm uses a compressed sensing and graph cut based object detection model to yield robust foreground support and background reconstruction results simultaneously. Since graph cut is robust to image noise, hence, different from the first one, the second object detection model can gives an exact foreground reconstruction results in video noise and camera shaking without any post-processing. Moreover, the proposed model can detect the aperiodic moving tendency such as smoke etc.(3) Recently, sparse representation based object tracking algorithms have become a hot topic. The main challenge in this domain is how to enhance the robustness and reduce the computational complexity of tracking algorithm. To solve this problem, a multi-task sparse representation based algorithm is proposed for robust visual tracking, in which, the best target is obtained by estimating the sparse representation of compressed particle observations jointly. To enhance the robustness of visual tracking performance, the proposed algorithm adopts a nonlocal regularizer to simultaneously exploit both local and nonlocal relationship among different sparse representation results. Experimental results show that the proposed visual tracking algorithm can achieve a better tracking performance than the state-of-the-art tracking methods.(4) Only using one camera to achieve object detection and tracking in severe occlusion(such as fog) is very difficult. Instead, making TV camera and infrared camera work together can solve this problem. The multi-image fusion becomes the key point in aforementioned case. In this dissertation, a compressed sensing based framework is proposed for fast image fusion. The advantages of the proposed framework are three-fold: first, the fusion framework processes one image block at a time to economize the storage space and enhance the transmission capability. Second, the dual channel Pulse Coupled Neural Network model is used as an important weighting-factor in the fusion scheme, which can give a compensation for the loss in image compression. Third, image reconstruction is accomplished by using self-adaptively modified Landwebber(SAML) algorithm, which has fast convergence and is robust to video noise. Experimental results show that the proposed image fusion framework can highlight moving target feature for object detecting and tracking.