Frontal-view human gait recognition based on Kinect features and deterministic learning

Abstract

A frontal-view gait recognition approach based on Kinect features and deterministic learning is presented in this paper. The utility of Kinect sensor eliminates the adverse interference of background for gait feature extraction. Different kinds of Kinect features are extracted and gait dynamics underlying different individual's time-varying Kinect features are captured by using radial basis function (RBF) neural networks (NNs) through deterministic learning (DL). The obtained knowledge of gait system dynamics is stored in constant RBF networks. A bank of estimators is constructed to represent the training gait patterns. By comparing the set of estimators with the test gait pattern, a set of recognition errors are generated. The average Li norms of the errors are taken as the similarity measure between the training gait patterns and the test gait pattern. The test gait pattern can be recognized rapidly based on smallest error principle and dynamic pattern recognition. Finally, experiments are conducted on the self-constructed Kinect gait database, and three main conclusions are obtained: (1) lower limb possesses richer gait information than upper limb does; (2) joint angle features does not perform as well as relative distance features; (3) best performance and comprehensive feature group is generated by combining joint angle, relative distance and anthropometric features together.