Research On Robustness Of Object Tracking System

With the development and popularization of computer science, electronic technology and artificial intelligence, object tracking technology has been widely applied to both civilian and military, such as visual surveillance, autonomous navigation, intellginet traffic monitor system, human-computer interaction system, video compression. In recent years, a large number of rearchers have studied the object tracking extensively and in-depth. Many effective video tracking algorithms are proposed for a variety of application environments. However, the object tracking system is quite complex.. Therefore, it is still one of the hot research fields of computer vision to develop a strong robustness and good applicability object tracking algorithm.To deal with the difficult problem in tracking object, the research is focused on three major components of the object tracking system:feature selection, target presentation model, similarity measure and target localization algorithm in this dissertation. And several new effective methods have been proposed. The main contents and contribtuons of this dissertation are summarized as follows:1. When the background and object have the similar color distribution, CAMSHIFT tracking algorithm can not successfully track the object. Thus, the robust distribution extraction algorithm based on adaptive Gaussian Mixture Model (GMM) is proposed. First, a fast and efficient adaptive GMM modeling method is used to model the target and background. The separability of each Gaussian component of object model is evaluated by the improved Bhattacharyya distance. The Gaussian components of object model that have a high separability are selected to generate the robust distribution. Experimental results show our algorithm can successfully extract the robust distribution and find the stable part of the object. When the robust distribution is applied to CAMSHIFT tracking algorithm, the accurate and robust tracking results are obtained and the tracking performance is improved.2. The existing feature selection methods based on distance measure are mostly only suitable to measure the distance between the two unimodal distributions. However, in fact, the background and object distributions are often multimodal distribution. Therefore, we propose a new feature selection based on improved Bhattacharyya distance. For each feature, the object and background is modeled with a GMM. The improved Bhattacharyya distance is used to evaluate the discriminability of each Gaussian component of object model, which is added together to represent the separability of this feature. Feature selection is done according to the separability of each feature. The subjective experimental results on the static images show that our measure can effectively evaluate the distance between two multi-peak distributions, and select the most discriminative features to track the object, which helps to improve the tracking performance.3. The traditional EM-based GMM modeling method is more time-consuming. A fast and effective GMM modeling method in a joint gray-spatial space is proposed. And the integral image is used to improve the computational efficiency of the candidate model parameters. Meanwhile, the approximate symmetric KL divergence between two GMMs is proposed to compute the similarity between the object GMM and the candidate GMM. The expremental results show the proposed modeling method can significantly reduce the target modeling time and the computational time of the parameters of the candidate model, and the proposed measure is robust and has a strong discriminability. Moreover, the tracking performance is significantly improved.4. The existing spatiogram similarity measures are either not robust, or not discriminative enough. Two spatiogram similarity measures are proposed:one is based on the symmetric KL divergence, and another is based on improved Jensen-Shannon Divergence (JSD). The former takes advantage of the strong discriminability of the symmetric KL divergence. The latter makes full use of the weight of the Gaussian component to strengthen the discriminability of JSD. Theorectical and experimental results show that our proposed two measures are better than existing measures, and the tracking performace is greatly improved.