Kernel-based Object Tracking

Visual tracking is one of the most popular research fields at present, and it is gradually applied in many areas, such as visual surveillance, human computer interface, robotics, autonomous drive, etc. Many algorithms, such as Optical Flow, Mean Shift, Partial Filter, etc, have been proposed for this task. Mean shift tracking algorithm, as one of the most successful algorithms among these tracking algorithms, however, is still facing great challenges to make a robust, real-time, visual tracking system under the conditions of light varies, partial occlusion, changes in object scale and appearance.One of the basic disciplines of artificial intelligence is to gain intelligence via imitating the behavior and the neuronal architecture of human beings. By following this basic discipline, my dissertation is to imitate the way and process of the visual system of human beings and other primates, who have a remarkable ability to interpret complex scenes in real time, despite the limited speed of the neuronal hardware available for such tasks. This dissertation proposes a novel target description based on visual attention-based selection mechanism after studying the classical target description and tracking process. We also proposed horizontal projection and vertical projection to combine spatial information to the target description. Then, we implemented the mean shift optimizing procedure to the spatial projection instead of kernel histogram. I divided my dissertation into the following several parts:1. Introduce the three steps of classical mean shift tracking algorithm and its limitations.2. Introduce visual attention selection mechanism and its application on target representation.3. Propose the horizontal projection and vertical projection on target representation and mean shift optimizing procedure based on spatial projection.4. Introduce the novel tracking procedure of visual attention based spatial projection tracking algorithm and its experiment results.Based on the previous work, the major contribution of this paper are like the following:1. We proposed a novel mechanism, named Visual Attention, to get a simple and robust target model by highlighting the importance of the salient pixels of the target while suppressing the background and less salient pixels. We do not go with the classical target description method– to describe the target precisely by feature extracting (color, texture, corner, or shape information, etc). The proposed visual attention based model is a model that based on the differences between the target and the background. Hence, the visual attention based target description model may not be disturbed by light varies, partial occlusion, changes in object scale and appearance. That is to say the proposed target model could be more robust than classical target model.2. We employ the horizontal and vertical projection to the target description and we apply the mean shift optimizing procedure based on the spatial projection instead of color histogram. Various experiments show that the proposed tracking algorithm based on spatial projection can localize the target more precisely than classical mean shift tracking algorithm.