Research On Methods Of Visual Object Tracking

The visual object tracking is a key issue in many vision-based applications, such as visual surveillance, visual navigation of robots, human-computer interaction, medical diagnose and military guidance. Along with the rapid growth of the information techniques in the last tens of years, the object tracking has attracted many researchers'attention and has become a very popular research topic. Although many effective visual object tracking methods has been proposed, there are still a lot of difficulties in designing a robust tracking algorithm due to the challenging complex scenarios such as significant illumination changes in environment, pose variations of the object and non-linear deformations of shapes, and noise and dense clutters in complex background, etc.To deal with the difficult problems in tracking object, the research is focused on the two major components of tracker: filtering and object appearance model in this dissertation, and several new effective methods have been proposed. The main contributions of this dissertation are summarized as follows:1. A tracking algorithm based on the color distribution and adaptive Kalman filter is proposed. The object model based on the distributions of the major colors of the object is built so that the object can be represented with a few of the model parameters, and its performance is superior to the histogram used widely in object tracking. A Kalman filter whose parameters can be adapted by occlusion is introduced into the tracking algorithm to improve the robustness of the tracking algorithm. The proposed tracking algorithm can track objects robustly in complex scenarios such as appearance change and occlusions.2. For the disadvantage of the color distributions that it can not accurately track scale of the object due to the fact that it omits spatial information, a probabilistic tracking algorithm based on the Spatially Constrained Color Model (SCCM) is proposed. The algorithm builds the object model that includes the color distributions and the spatial configuration of the object to improve the discriminative power of the object model. Meanwhile, the robustness of the tracking is strengthened by integrating the SCCM into particle filter. The experiments show that the proposed tracking algorithm is robust and effective to scale change of the object, pose variations of the object, partial occlusions, etc.3. To deal with the change of object appearance, a novel adaptive model update algorithm is proposed in the framework of particle filter. The algorithm estimates tracking reliability from the distribution of the posteriori probability density and similarity between observation image and object model, and the object model is updated only during the reliable tracking. The proposed algorithm effectively avoids the model drift caused by update procedure and improves the robustness of tracking.4. A tracking algorithm based on the blob model is proposed. The algorithm partitions the image into a few of blob features, and achieves the tracking by matching the blobs. Since the blob model includes color, spatial information and shape of the object, the discriminative power and robustness of the model are improved. The experiments on real scenarios demonstrate that the proposed algorithm can effectively track object in complex circumstance.5. For the problem of the low signal to noise ratio and luminance contrast, a robust algorithm based on adaptive feature selection is proposed for tracking objects in forward looking infrared (FLIR) imagery. In this method, the gray features which can distinguish the object from its surrounding background are selected adaptively to build the object model. The tracking is achieved by the Mean Shift algorithm. The experiments on real FLIR imagery have shown that the proposed algorithm is robust for the infrared object.