Silhouette based Gait Recognition: Research Resource and Limits

As is seen from the work on gait recognition, there is a de-facto consensus about the silhouette of a person being the low-level representation of choice. It has been hypothesized that the performance degradation that is observed when one compares sequences taken on different surfaces, hence against different backgrounds, or when one considers outdoor sequences is due to the low silhouette quality and its variation. If only one can get better silhouettes the perfomance of gait recognition would be high. This thesis challenges that hypothesis. In the context of the HumanID Gait Challenge problem, we constructed a set of ground truth silhouttes over one gait cycles for 71 subjects, to test recognition across two conditions, shoe and surface. Using these, we show that the performance with ground truth silhouette is as good as that obtained by those obtained by a basic background subtraction algorithm. Therefore further research into ways to enhance silhouette extraction does not appear to be the most productive way to advance gait recognition. We also show, using the manually specified part level silhouettes, that most of the gait recognition power lies in the legs and the arms. The recognition power in various static gait recognition factors as extracted from a single view image, such as gait period, cadence, body size, height, leg size, and torso length, does not seem to be adequate. Using cummulative silhouette error images, we also suggest that gait actually changes when one changes walking surface; in particular the swing phase of the gait gets effected the most