Novel algorithms for video object tracking

Abstract

Visual tracking is a way to trace the position of an object in a video sequence. It has been widely used in film industry, interactive gaming, surveillance and auto-robotics. In this thesis, we will initially introduce some existing approaches for target representation, such as templates, color histograms and orientation histograms. These models are then used to examine the similarity between a target and possible candidates. Let us recall some more tracking techniques. The template matching approach takes a template (a sub-image) for searching within a pixel region with the least distance to it. The mean shift is a gradient-decent method which indicates that the candidate with the highest similarity stays at the zero-gradient position. The particle filter follows the sequential Monte-Carlo method to form a discrete expression of posterior pdf (probability distribution function of possible states given the known observations), where the expectation is designated as estimate. Experiment results illustrate that the particle filter is much robust than other two approaches. However, the computation is much more complex subsequently. In the research work, we propose a new sampling strategy for particle filtering in object tracking. A particle filter formulation is not always able to generate effective samples at the step of importance sampling. It requires heavy computation in order to achieve acceptable tracking results. We propose a multi-step recursive sampling method to replace the direct importance sampling. This relies on the feedback derived from the resampling procedure. New particles are sampled recursively from the existing particles with high weights. After several iterations, particles become densely populated, and this new sampling policy gives significant contributions to achieve accurate position estimation. Besides, in order to strengthen the density around the probable area, we have to generate a good sampling pattern for the proposed candidates. Usually, the particle transition vector is simulated by element-wise Gaussian samplings in the literature, i.e. the covariance is a diagonal matrix. However, it cannot react properly due to the changing target motion. Therefore, we may apply a 2D predictive transition vector to update the pattern of the multivariate Gaussian sampling. Before this, several predictive models are analyzed for comparison. The adaptive averaging model wins over others which include the AR model and Taylor series expansion. Then it has been decomposed into rotational angle and magnitude, which help update the sampling pattern. Finally, the adaptively updating sampling pattern is able to establish a more appropriate searching region which can reach the real state. Experimental results indicate that the proposed method reduces computation substantially and it also preserves good tracking results comparable to other algorithms in the literature. Also the static histogram-based representation sometimes may not be robust enough in observing really good candidates and serving the long tracking process. So we have adopted the structural histogram set to settle the problem of low-confident similarity map. During tracking, every partial histogram in the set is assigned a dynamic weight according to its performance. The weight dominates its contribution to the final summed similarity result. Meanwhile, the referenced target model is dynamic via selecting one in a candidate set. It helps ensure the model to follow the real changing appearance of target. Furthermore, the elements in the set can also be replaced by a real adorable estimate in a fixed interval. Experimental results illustrate that the new model can effectively track objects with clutter background. Acceptable tracking results are achieved for long videos.