On automated model-based extraction and analysis of gait

We develop a new model-based extraction process guided by biomechanical analysis for walking people, and analyse its data for recognition capability. Hierarchies of shape and motion yield relatively modest computational demands, while anatomical data is used to generate shape models consistent with normal human body proportions. Mean gait data is used to create prototype gait motion models, which are adapted to fit individual subjects. Our approach is evaluated on a large gait database, comprising 4824 sequences from 115 subjects, demonstrating gait extraction and description capability in laboratory and real-world capture conditions. Recognition capability is illustrated by an 84% CCR in laboratory conditions, which is reduced for real-world (outdoor) data. Preliminary results from a statistical analysis of the extracted gait parameters, suggest that recognition capability is primarily gained from cadence and from static shape parameters, although gait is the cue by which these are derived.