Experiments on gait analysis by exploiting nonstationarity in the distribution of feature relationships

We consider the use of nonstationarity in the distribution of feature relationships over time for walking gait-based recognition. We statistically model the features of a person by computing the distribution of the relations among the features, rather than the features themselves. These relational distributions of feature relations are represented as points in a space of probability functions. Our database presently consists of twenty subjects walking outdoors along three different paths at 0° (frontal-parallel), 22° and 45° with respect to the image plane and walking in both directions, left to right and right to left. We performed statistical tests to demonstrate that variations between persons are statistically more significant than the variations due to walking angles and walking directions. We also present identification results on people walking at different directions and different angles.