Research On Object Tracking Using Prior Oriented Active Contour

Object tracking refers to the task of finding the location of the specific object of interest in each image frame of a video sequence, so as to generate the trajectories of the moving objects in the sequence of images. Object tracking is a challenging research topic in the field of computer vision and has been widely used in many vision based applications such as surveillance. Given an initial target area, the tracking procedure firstly establishes a target model based on it. Then according to this target model, the tracker separates the specific object from the complex background in the new coming image frame. Finally, the tracking result of this frame is used to update the target model, serving for the tracking task of next image frame. Therefore, extracting sample data from exact target area to establish discriminative target model is of significant importance for robust tracker. However, most currently proposed tracking algorithms use simple geometric shapes, such as rectangular or oval, to present the tracking results. This kind of rough presentation on the one hand could not provide well description for the real world objects with complicated shapes, and on the other hand will inevitably introduce a large amount of background pixels in the target area. These background pixels will affect the target model?s establishment and updating. This will weaken the discrimination of the target model and ultimately lead to failure of the trackin g task. Further, the inaccurate target presentation in tracking results will also limit its effectiveness in the upper applications. For example, the shape based human action recognition and event detection applications usually have a higher requirement on the accuracy of the tracking results.Based on the above problem description, I focus my efforts on the research of object contour tracking based on prior oriented segmentation theory. Active contour model is an energy based segmentation approach, whose main idea is to evolve a contour denfined in the image domain by an energy function and make it a ct from the inintial positon toward the expected region boundary. This thesis studies how to make use of the target prior knowledge in the context of object tracking to supervise the active contour evolution in image segmentation theories, finally realize object contour tracking technique which outputs the accurate target contour in each frame. Particularly, depending on the tracking environment and the different types of target prior knowledge, the current research mainly includes the following four aspects:This research firstly proposes a level set based mean shift tracker to realize contour tracking of objects. The proposed algorithm introduces the binary level set active contour model into the mean shift sampling framework. By conducting curve evolution on each sample and using the final curve energy to weight the sample, the proposed method realizes accurate contour tracking of the target object. Compared with the conventional trackers, the contributions of the proposed method are: first, compared to the previous mean shift algorithms that use rectangular or oval to present the tracked target, the proposed method introduces the active contour model to achieve contour tracking of the target, improving the accuracy of the output tracking results and reducing the drift; second, compared to the traditional level set methods, the proposed method adopts binary level set model, which uses a binary level set function to replace the signed distance function in traditional level set model, thus avoiding the re-initialization of the level set function and greatly improving the computational efficiency. The experimental results indicate that this method can obtain more accurate object region and increase robustness when tracking objects with complex shape.Although we introduce level set active contour model in above method to achieve contour tracking of the tracked target, this contour tracking mechanism mainly relies on the original segmentation characteristics of the level set model, without considering the target priori knowledge in the context of tracking, which makes the model more applicable for segmenting uni-featured foreground object from uni-featured background. However, real world objects and tracking circumstances often contain multiple feature distributions, so the original segmentation mechanism of the level set model is not adequate for these complex cases. Based on this problem description, we propose an object contour tracking method based on a supervised level set model. The proposed method fully consi ders the object tracking context, uses the boosting manner to learn the knowledge of the tracked target and refines the curve evolution of the level set model, which can ensure a more accurate convergence to the exact target we want to track. Finally, accurate target region qualifies the samples fed to the target appearance modeling procedure as well as the target model prepared for the next time step. The experimental results indicate the effectiveness of the method in complex scene.To address the problem that traditional kernel based trackers such as mean shift methods adopt rectangular or elliptical kernel, we propose an adaptive data-driven kernel for object contour tracking. Since an ideal kernel is expected to have the shape of the actual tracked object, how to select the suitable kernel to best fit the object shape is an important issue facing to kernel trackers. This method introduces the concept of supervised level set model into the sample space of the mean shift, making the level set curve distinguish between the two classes of samples in the sample space and simultaneously obtain the kernel that ideally match the object shape. Therefore, the proposed data-driven kernel can fit the object shape and be evolved to adapt to target change, thus give a better estimation bias. The experimental results indicate that the proposed method outperforms the previous methods using fixed-shape kernels.To deal with the cases where the target and background in real world circumstances may have various and unstable appearance but relatively stable shape characteristics, we propose an object contour tracking algorithm based on shape prior. The algorithm uses graph cut model and evaluates the segmentation by simultaneously considering its global shape and local edge consistencies with the object shape priors to implement shape-preferred contour evolution. By iteratively implementing the optimization, the proposed method can achieve joint estimation of the optimal segmentation and the most likely object shape encoded by the shape priors, and eventually converges to the candidate result with maximum consistency between these two estimations. Finally, we employ an ellipses segmentation application which contains various challenges for regional segmentation to evaluate the proposed method. The experimental results validate the effectiveness of the proposed method of dealing with contour tracking task on objects with particular shape.Based on the above, this research improves the accuracy of tracking results expression on targets with complex shape, and achieves directional convergence of the active contour towards a particular interested tracking target. It obtains a kernel tracker where the kernal can fit the target shape and adapt to target changes, and a shape oriented tracker which is able to fully tap the specific target shape prior to accurately estimate the target contour, making the contour tracking of objects with complex appearance more accurate and efficient.