A Method for Estimating Motions of Contours with an Application to Gait Recognition

Abstract

In this thesis, I propose a novel method for estimating motions between image contours. The method is fast and can handle both small and large displacements. It uses the distance transform and its gradients to estimate correspondences between points. The distance transform of a binary contour represents the distance of each pixel to the nearest contour pixel. The gradient of the distance transform points in the direction normal to the contour. By combining unit vectors obtained from the gradient with the original distance transform, the method produces vectors that correspond to the normal displacement between pairs of contours. This method can then be extended to compute true motion near corners as well as parameterized motion models. Experiments on various shape contours show the method's efficacy in computing normal displacements. Cases that do not correspond to normal motion are analyzed and corrected through the use of contour normals and motion models. The method is also applied to gait recognition, the goal of which is to identify people in videos based on their unique walking pattern. Many gait recognition methods operate on sequences of binary silhouettes. Since contours are easily obtainable from silhouettes, the proposed method is well-suited to the task. I describe two representations, the Histogram of Motion and the Edge Motion Vector, that allow for the comparison of contour motion between frames and sequences. These representations are tested on a large gait recognition database and achieve rank 1 and rank 5 performance that is comparable to the state of the art. The success of the method on gait recognition shows it is useful as a standalone representation, but it can also be used to improve other techniques and for other applications - this is left to future work.