Complex Scenes Video Object Detection And Tracking Algorithm

Video object tracking and segmentation in image sequences are an important topic in the field of computer vision. It is a critical technology in wide applications of computer vision field such as surveillance, video retrieval and human-machine interaction etc. And it is of a wide development prospect as well. Its results will affect high-level video processing directly. However, visual target segmentation and tracking often become very difficult due to complex background condition and the random target motion. At present, moving target detection and tracking are not well-considered, many problems and difficulties in theory research and in applications are still unsolved. Visual target detection and tracking often becomes very difficult in real environments due to many factors such as illumination variation, dynamic background, object rotation, scale variation, occlusion, target complicated and random motion. Therefore, the researching on robust target detection and tracking under complicated scenarios have the great significance in theory and practical application. The thesis conducts research of the target detection and tracking algorithms under fixed monocular vision. In this dissertation, the research is focused on the key technologies of movement target detection and segmentation, cast shadows elimination, target describing and tracking. The main contributions are summarized as follows:(1) To the visual target detection in complex environments, an improved Adaptive-K Gaussian Mixture Model (AKGMM) method for background modeling was proposed in this dissertation. The AKGMM algorithm altered the dimension of the parameter space at each pixel based on the changing frequency of pixel value in this method. The number of GMM reflected the complexity of pattern at the pixel. An improved learning method was proposed for Gaussian Mixture Model in this dissertation. An adaptive learning rate was calculated for each Gaussian at every frame for speeding up the convergence without compromising model stability. Compared with the conventional GMM method, the method proposed in this dissertation gets a faster convergence while maintaining good robustness against complex environment.(2) Shadow detection and suppression technologies in a system for moving visual object detection and tracking were introduced in this dissertation. A shadow model was established by the analyzing optical properties of cast shadow. In this dissertation, a set of moving pixels were firstly obtained based on AKGMM algorithm, and then a novel method was proposed to discriminate shadow region from the set of moving pixels by using the properties of brightness, chromaticity and texture in sequence.(3) Degeneracy and sample impoverishment problem are inevitable results of sequential importance re-sampling particle filter. A mass of degenerated particles seriously affects the tracking ability of particle filter. In order to achieve precision and robust tracking, a large number of samples for particle filter are required. This dissertation investigated particle filter algorithms based on intelligent optimization. To optimize the distribution of predicted particle in the state space, ant colony optimization (ACO) and particle swarm optimization (PSO) were embedded into particle filter framework respectively. The intelligent optimization algorithms drove all particles towards high likelihood regions. Both ant colony algorithm and particle swarm algorithm effectively eliminated particle degeneration and sample impoverishment. A novel adaptive method of particle filter-based tracking was also introduced in this dissertation. Adaptation was applied to the standard deviations of the noise in the state vector based on the tracking accuracy. The above mentiond measures greatly increased the tracking robustness and accuracy in complex environments.(4) A tracking approach based on multi-cue adaptive fusion was proposed to resolve the visual target tracking problem in complex environments. It was observed that different discrimination was obtained by different characteristics. This proposed method fused multiple cues according to its individual reliability of the tracking system. Different cues were reweighed based on the predefined discrimination. The problem of instable tracking by using single measurement source was solved with this proposed measure. The method also enhanced the ability of object discrimination and enabled the tracking algorithm more accurate and stable under complex environments.