Open and closed contours tracking based on shape priors and training

Abstract

This dissertation presents a new open and closed contours tracking algorithm using shape prior and its training based on a Bayesian framework, where the contour is a part (open contour) or the whole (closed contour) of the object's boundary. The shape of an object is a very important feature for many vision tasks such as object recognition and tracking. Specifically, the tracking performance can be increased if the target is determined and the tracker utilizes its shape information. The proposed method provides a new state space model for the representation of contours, which can reflect the shape information to the contour and handle rigid and non-rigid motions of contours independently. This model enables us to focus on the non-rigid motion during the tracking, and the model works for challenging rigid motion scenarios. In addition, for the robust tracking of contours, a measurement function that considers the contrast on object boundaries, target appearance, and temporal coherence is proposed. The proposed method is tested for various cases of contours such as open contour, closed contour and multi-contours. The state space model and measurement functions are modified a little bit in consideration of each contour model. First, an open contour is modeled and tracked by the proposed method, which has received little attention during several decades compared with the closed contour or bounding box shape tracking. The proposed state space model can represent an open contour that is moved by the dynamic model where rigid and non-rigid motions are absolutely separated. The measurement is designed with contrast, local track and appearance terms that indicate the proper position of the target and make the tracking more robust. The proposed method is applied to two examples of open contours targets (human omega shape and a cheetah profile), and experimental results show that the proposed method achieves superior performance to the conventional contour tracking methods. The proposed method is also compared with recent bounding box tracking methods for the object tracking purposes, and the comparison shows that the proposed method works robustly to fast motions and yields more accurate estimate of object's location than the conventional bounding box tracking methods. Second, the proposed method is tested for the closed contour tracking which is usually carried out by segmentation algorithms or level set methods. A closed contour is described by the proposed model and deformed by the dynamic model. Measurement function is the same to the case of open contour tracking except the local track term, which is calculated with partial object appearances that are denoted by some local patches and their relative positions. As an application example, automobiles in blackbox video sequences are tracked by the proposed method. Experimental results show that the proposed method accomplishes higher performance than conventional tracking methods where some of them presents the target as a bounding box and others extract the object boundary using segmentation methods. Moreover, the document capture and tracking algorithm is also proposed, which is suitable for applying the proposed method because the shape of document is well known (a quadrilateral) and its boundary can be estimated by the proposed method. This system is based on building quadrilaterals as document proposals using line segment detector and tests all proposals to find the best one with measurement terms. The proposed algorithm makes good marks at 2015 ICDAR competition. Finally, multi-contours tracking algorithm is devised based on the contour tracking method. It is assumed that targets belong to the same category and their appearances, colors and shapes are similar to each other. Thus, the proposed method trains only one shape model to track multi-contours. The state space vector is amended such that all contours can be represented by one state vector. In order to consider interactions between targets, the interaction term is attached to the existing dynamic model. As an example, human legs are tracked by the proposed method which may help to recognize the gaits. Experimental results show that conventional algorithms have troubles in tracking and distinguishing between the two legs, whereas all targets are well estimated accurately by the proposed method.