The Research On Object Detection And Tracking Of Smart Visual Surveillance

Smart visual surveillance has attracted increasing attention in recent years and has become a hot research topic in computer vision for visualization, intelligence, networking and rich information. Smart visual surveillance has many applications in different fields such as in traffic, military, security, human computer interaction and medical areas. The technologies of moving object detection and tracking are the fundamental technologies in smart visual surveillance and are the basis of the smart visual surveillance. They are critical to the post-processing of smart visual surveillance. The dissertation focuses on the key technologies of object detection and tracking of smart visual surveillance for improving object detection and tracking greatly in real-time, robustness and accuracy.First, we study the background modeling method, review the state-of-the-art algorithms in background modeling method, focus on the codebook algorithm, analysis the limits of the original codebook algorithm. On this basis, we proposed a box-based code algorithm. The algorithm builds the model of the background by constructing a box-based measurement to match the pixel value of the image, classifies the input pixel as background or moving object by the background subtraction. Background model of the visual scene is constructed by the background modeling method, and then moving objects are detected by background subtraction. The algorithm considers illumination and shadow, can handle multiple mode background, and reduces the computational complexity. During the training period, it allows moving object in the training sequence, and it is suitable for moving objects under the static camera scenarios.Second, we explore the feature point-based tracking algorithms, analyze the drawbacks of the feature point-based tracking algorithms. A SIFT feature-based tracking algorithm is proposed by combination of the codebook background modeling method and the feature point-based tracking method. The algorithm obtain the area of the moving object by the background subtraction method, and then matches the objects between the SIFT features of the object in the current frame and the SIFT features of the object in the next frame to realize the object tracking. The algorithm extracts the SIFT features of the moving object region only from the region of the object, so that there is no the feature grouping problem. And since the SIFT feature extracting only from the object area instead of the full image, the computation is reduced. The algorithm can handle partial occlusion problem.Third, we research on active contour-based algorithms, present a active contour-based multiple object tracking algorithm on analyzing the existing active contour-based object detecting and tracking algorithms. The algorithm detects and tracks multiple objects by using a global minimization of active contour model, will not fall into local minimum, using a fast numerical implementations, the algorithm reduces the computational complexity, and can detect and track multiple objects under noise.Finally, we study on particle filter in object contour tracking to tracking defoming objects, and analyze the algorithm of DPF-MT(particle filter with mode tracker for tracking nonaffine contour deformations). A tracking algorithm which combines the Kalman filter algorithm and the algorithm of DPF-MT is proposed for tracking deformable fast moving object under complex background to improve the algorithm of DPF-MT. The algorithm firstly predicts the global translation velocity of the deforming object, then predicts the deformation of the object using particle filter, splits the state space into the effective basis space(small number of dimensions) and residual basis space, performs importance sampling on effective basis and mode tracking on residual space. The algorithm can track the deforming moving object under complex background.