Robust Visual Object Tracking Algorithm Based On Particle Filter

Visual object tracking is a comprehensive technique for locating and tracking object by analyzing and understanding video information captured by visual sensor. It has an important status in the field of computer vision. In the research and application relating to visual object tracking, robustness is the most basic and important problem. With rapid development of computer hardware technique, real-time ability of visual object tracking algorithms relys more and more on hardware mechanism. The pivotal issue of visual object tracking is to improve the robustness of dealing with all kinds of disturbances in all kinds of environments.Recently, among various visual tracking algorithms, particle filter tracking is a robust one that is able to solve the popular problems of un-linear object state and un-Gaussian noise distribution, track various object states simultaneously, and adapts to not only stationary visual platform but also moving visual platform such that it has drawn considerable attention both in the research and applications on visual object tracking. However, after in-depth study of particle filter tracking, it is found that particle filter algorithm, object feature model, and similarity measurement of feature model are the three key aspects of particle filter tracking, and they all have some problems to be addressed. Furthermore, like other visual object tracking algorithms, particle filter tracking also lacks of sufficient intelligence so that it cannot handle as human a variety of complex environment changing and choose a suitable scheme in real-time. Therefore, it is of great significance to deal with the problems lying in particle filter tracking and explore external mechanism of human vision to endow particle filter tracking with some intelligence in order to enhance the robustness of particle filter tracking.After the survey and analysis of the current research work, we present in this thesis our research on the problems of visual object tracking based on particle filter with the aim to promote the robustness of visual object tracking algorithms by analyzing the three aspects—particle filter algorithm, model similarity measurement, and object feature model—of particle filter tracking that influence its robustness and investigating the external mechanism and physiological characteristics of human visual system so as to improve the intelligence of particle filter tracking. In this paper, a visual tracking algorithm called grey model particle filter (GMPF), an adaptive control model for adjusting noise distribution, a modified version of Bhattacharyya coefficient for similarity measure between feature models, an object feature model named elliptical region covariance descriptor which enables fusion of various spatial-temporal features, a visual tracking framework that simulates human visual intelligence with the corresponding algorithms, and a computation model of the top-down visual attention mechanism are proposed. The main contributions of this thesis are as follows.(1) Both of proposal distribution selection and particle sampling range regulation concerned in particle filter algorithm that affect the performance of particle filter tracking is investigated.For proposal distribution selection, in order to make proposal distribution approach to posterior density of object state well, a method for generating proposal distribution is proposed by modeling grey prediction model of object state based on grey system theory and yielding proposal distribution around object state prediction. This proposal distribution is combined with particle filter tracking, called grey particle filter tracking (GMPF). With qualitative and quantitative comparison experiments of GMPF with particle filter tracking, kalman particle filter tracking, and unscented particle filter tracking, the outperformance of GMPF is demonstrated. Also, the effect of variation of the number of particles and the length of grey prediction sequence is studied.For particle sampling range regulation, we analyzed the relationship between the process noise factor of transition model and particle sampling range, and interpreted that the peak of observation probabilities of particles is applicable to act as evaluation criteria of particle sampling range. Thus, an adaptive regulation model of process noise factor is proposed by using the peak of observation probabilities of particles as variable. Experiment results illustrated the applicability of the proposed adaptive regulation model of the process noise factor, compared it with fixed process noise factor and the adaptive regulated process noise factor controlled by prediction error, and showed the outperformance of our proposed method in difficulty cases of partial and short-time fully occlusion. (2) The model similarity measurement, Bhattacharyya Coefficient (BC), that is widely used in particle filter tracking using histogram feature model is studied. We first analyzed the effect of BC on particle filter tracking, and presented the reason why particle filter tracking with BC only adapts to object shrinking while not to object dilation such that tracking accuracy is difficult to ensure. To tackle this problem, a modified Bhattacharyya coefficient is proposed, termed MBC for simplicity. It is proved in theory that MBC is endowed with single peak attribute, hence by using MBC, particle filter tracking can adapt not only to object shrinking but also to object dilation. Moreover, the effect of the two parts of MBC are analyzed. Experimental results evaluated the performance of particle filter tracking with MBC, and demonstrated the validity of theoretical analysis of MBC.(3) Object feature model of particle filter tracking is investigated. Since object feature model constructed by fusion of multiple features can augment the robustness of particle filter tracking, we first analyzed the drawbacks of state of the art, and proposed a multiple feature fusion method, terms as elliptical region covariance descriptor. This descriptor mainly has three advantages. First, this descriptor enables fusion of a variety of spatial-temporal features into an unified feature model. It can reflect both self-correlation of the features and cross correlation among them. Second, this descriptor has three parameters of major axis, minor axis, and rotation angle such that it has the capability to represent scale and angle variation simultaneously. Third, the dimension of this descriptor is determined by the number of the used features and independent of the dimensions of the used features so as to reduce the computational cost and ensure the real-time applicability. Experimental results test the robustness of particle filter tracking based on the proposed elliptical region covariance descriptor against the difficult cases of object scale and angle changing, object fast moving, partial and short-time fully occlusion, illumination variation, and strong noise.(4) With the motivation to cope with the issue of automatically recovering object tracking, how to improve the intelligence of particle filter tracking is deliberated. In order to mimic the function of human visual tracking, a visual tracking framework is proposed according to the way human’s eye tracks object. With particle filter tracking and visual attention mechanism of human visual system (HVS), two particle filter tracking algorithms are proposed in our proposed visual tracking framework. The first particle filter tracking algorithm employs visual attention mechanism to imitate the global object searching process of HVS and particle filter tracking to simulate the local object tracking process of HVS. The global object searching and local object tracking are dynamically integrated to perform robust object tracking. The second particle filter tracking algorithm terms the original proposal distribution of particle filter as local proposal distribution and defines visual saliency map of visual attention as global proposal distribution. This algorithm extends the traditional particle filter tracking by dynamically combines the local and global proposal distributions. Through comparing our proposed algorithms to state of the art particle filter tracking algorithms, experimental results confirm the robustness and outperformance of our proposed algorithms. Additionally, when we studied the algorithms to mimic human visual tracking using visual attention mechanism, we also take a research on visual attention mechanism. A computation model of the top-down visual attention is proposed in terms of the mechanism of generating visual attention, and is used to compute the visual saliency map. After that, a method for detection of visual attention focus transition area is proposed by analyzing in multi-scale the Shannon entropy density of the visual saliency map. This method has been applied in our proposed first particle filter tracking algorithm.