Reaserch On Kernel Based Target Tracking Algorithm

Target tracking is a classical computer vision problem with many important applications areas such as robotics, surveillance and driver assistance. The task is to follow a target or multi targets in a video sequence and the target can be any object of interes. Though significant progress has been made in the last few years in target tracking, the interference of the target shape changes, occlusion, background and other factors lead that target tracking remains a difficult and open research topic. Studies and improvements are made for the kernel based mean shift(MS) tracking in this paper; and a novel kernel based expanded multi-channel correlation filter(EMCCF) tracking algorithm is proposed. The main research work and contributions of this dissertation are summarized as follows:(1) Aiming at solving the special problem that tracking nonlinear motion target under occlusion, the cross-bin color histogram based MS with globle research scheme is proposed. First, the cross-bin color histogram is used instead of traditional bin-bin color histogram describing the target to reduce the influence of background and improve the tracking accuracy; when the target is lost after severe occlusion, a scale change adjustment mechanism is used for searching target position in global scope to improve MS anti-occlusion capability.(2) Aiming at solving the general problem that tracking linear motion target under short time occlusion, the method that can take two layers of kalman filter(KF) framework into MS with corrected background-weighted histogram(CBWH) is proposed. First, the mathematical model of first layer is established through kinematics equation; second, the relationship among the Bhattacharyya coefficients, the filter noise and the tracking results is used to adjust the tracking results self-adaptively, getting target position, reducing the occlusion effect on the tracking results; last, in the second layer, all nonzero elements of the target template histogram are filtered through dynamic filter residual and Bhattacharyya coefficient, getting filtered target template by adjusting and updating parameters of the filter real-time, reducing effect on tracking results due to changes in target feature and achieving update synchronization of target template with CBWH.(3) A novel expanded multi-channel correlation filter(EMCCF) target tracking algorithm is proposed. On the basis of multi-channel correlation filter(MCCF), the optimal solution of proposed EMCCF is got by constructing EMCCF model from time domain under kernel ridge regression, which uses the relationship between ridge regression and correlation filter(CF). Because MCCF only exists a time-cost linear optimal solution, so proposed EMCCF sovles the disadvantage that MCCF is only suitable for offline target detection task but not suitable for online target tracking.