| The term gait recognition is typically used to signify the identification of people in image sequences ‘by the way they walk’. Gait is determined by the physical characteristics of each individual, and so is believed to be as unique to the person as a fingerprint is. Gait is also one of the few biometrics that can be measured at a distance, which makes it useful in surveillance applications as well. In this thesis, our objective is to develop robust methods for extracting discriminant gait features automatically and passively from low-resolution video. To this end, we explore two different gait recognition techniques: a non-parametric approach that derives its classification feature from the image self-similarity plot of a walking sequence, and a parametric approach that estimates explicit parameters of gait, viz. cadence, stride length and apparent height and uses them for classification. Both methods derive their robustness from the fact that they do not rely on image feature tracking or correspondence. The methods are extensively tested on a variety of datasets, and achieved up to 70% with the first method and 50% with the second method in classification accuracy on reasonably large datasets. Finally, because carried loads are gait-altering, we also describe a method to determine whether or not a walking person is carrying an object prior to applying gait recognition on the sequence. The method exploits known shape and periodicity cues of human walking figure. It achieves 85% detection rate and 12% false alarm rate when tested on 41 realistic outdoor sequences. |
