Automatic gait recognition using area-based metrics

Abstract

A novel technique for analysing moving shapes is presented in an example application to automatic gait recognition. The technique uses masking functions to measure area as a time varying signal from a sequence of silhouettes of a walking subject. Essentially, this combines the simplicity of a baseline area measure with the specificity of the selected (masked) area. The dynamic temporal signal is used as a signature for automatic gait recognition. The approach is tested on the largest extant gait database, consisting of 114 subjects (filmed under laboratory conditions). Though individual masks have limited discriminatory ability, a correct classification rate of over 75% was achieved by combining information from different area masks. Knowledge of the leg with which the subject starts a gait cycle is shown to improve the recognition rate from individual masks, but has little influence on the recognition rate achieved from combining masks. Finally, this technique is used to attempt to discriminate between male and female subjects. The technique is presented in basic form: future work can improve implementation factors such as using better data fusion and classifiers with potential to increase discriminatory capability.