Research On Kernel Based Visual Tracking Algorithm For Varying Scale Target

Video target tracking is also called visual tracking which is a kind of technique that keeps tracking the interest region effectively in video sequence. How to realize stable video target tracking is always a classic problem in the field of computer vision. When distance between observing platform and moving target is varying, both the background and target region become dynamic. So the target's scale variability should be considered in tracking algorithm.In this dissertation, the research is focused on kernel based visual tracking algorithm for varying scale target. Because the video target is a kind of plane object, visual tracking algorithm should not only estimate the target's position, but also should estimate the target's scale. On the other hand, to estimate the target's position accurately, the visual tracking algorithm has to extract the target's features in a correct scale bound to realize features' effective matching. So how to estimate target's scale is one of the fundamental problems for a stable visual tracking algorithm. However, traditional kernel based video target tracking algorithm lacks tracking window's scale update mechanism, so it's tracking result is not complete and stable. To deal with the difficulty, five tracking methods have been proposed. The five methods are named respectively as: kernel based visual tracking with scale invariant features; kernel based visual tracking with variant spatial resolution model; target tracking based on adaptive target model; kernel based visual tracking with reasoning about adaptive distribution image; and probabilistic motion switch tracking method based on mean shift and double model filters.The main contributions of this dissertation are summarized as follows:1. For boosting up the feature's discriminating ability, both scale invariant features and kernel based color distribution features are used as descriptors of tracked object. The proposed algorithm can keep tracking object of varying scales even when the surrounding background being similar to the object's appearance.2. To complement the ability of KBT, both Cartesian and Log-polar coordinates are used as a joint tracking coordinate. By using target's space-variant resolution model, the proposed algorithm can estimate the target's scale and rotation parameters within the framework of KBT.3. Kernel based visual tracking with continuous adaptive distribution is proposed. By integrating the CAMSHFIT method into KBT, the proposed algorithm can update the tracking window's scale. And the target model can also be updated within the framework of mean shift based on target's continuously adaptive distributions.4. Based on the kernel based visual tracking with continuous adaptive distribution, a reasoning engine which can deals with fundamental constraints on the spatial-temporal continuity of target's motion is proposed. Based on voting reasoning about the target's adaptive distributions image and target's motion, the proposed algorithm can update the tracking window's scale and target model more stably.5. Within the framework of multiple model filters, an efficient image tracking method that integrated mean shift and double model filters is proposed. Two motion models can switch each other by using a probabilistic likelihood. Experiment results show the method can successfully keep tracking target, no matter the target's velocity is large or small, changing or constant, with modest requirement of computation resource.In short, to deal with the problem of varying scale target tracking, SIFT features matching, space variant resolution model, target model adaptive technique and interactive multiple model filters are integrated into the framework of kernel based visual tracking. All these methods can improve the performance of traditional kernel based visual tracking method, and solve the problem of varying scale target tracking.