Andrei Sharf

A persons manner of walking or their gait is an important feature in human recognition and classification tasks. Gait serves as an unobtrusive biometric modality which yields high quality results. In comparison with other biometric modalities, its main strength is its performance even in data that are captured at distance or at low resolution. In this paper, we present an algorithm for classification of gait disorders arising from neuro-degenerative diseases such as Parkinson and Hemiplegia. We focus on motion anomalies such as tremor, partial paralysis, gestural rigidity and postural instability. The analysis and classification of such motions are challenging since they consist of a multiplicity of subtle formations while lacking a regular pattern or major cycle. We introduce a gait representation which is invariant to the walking cycle and yields an efficient similarity metric. Our method performs on the joints’ motion trajectories of a 3D human skeleton captured by a Kinect sensor. The algorithm is robust, in that it does not require calibration, synchronization or a careful capturing setup. We demonstrate its efficiency by classifying different degenerative cases with high accuracy even in the presence of noise and low-resolution acquisition.