Cross-view and multi-view gait recognitions based on view transformation model using multi-layer perceptron

Gait has been shown to be an efficient biometric feature for human identification at a distance from a camera. However, performance of gait recognition can be affected by various problems. One of the serious problems is view change which can be caused by change of walking direction and/or change of camera viewpoint. This leads to a consequent difficulty of across-view gait recognition where probe and gallery gaits are captured from different views. In this study, a novel method is proposed to solve the above difficulty based on View Transformation Model (VTM) under an uncalibrated single camera system. VTM is constructed based on regression processes by adopting Multi-Layer Perceptron (MLP) as a regression tool. VTM smoothly estimates gait feature from one view using motion information in a well selected Region of Interest (ROI) on gait feature from another view. Thus, pre-trained VTMs can normalize gait features from across views into a same view before gait similarity measurement is carried out. Comprehensively, the proposed method is logically extended for multi-view gait recognition where gallery gaits from multiple views are used to recognize probe gait from single view. This is addressed using multi-view to one-view transformation where VTM is now employed to estimate gait feature from single view using motion information in well selected ROIs on gait features from multiple views. The proposed method is tested on a large benchmark gait database which contains 124 subjects from 11 views. Extensive experimental results demonstrate that our method significantly outperforms other baseline methods in literature for both across-view and multi-view gait recognitions. In our experiments, particularly, average accuracies of 99%, 98% and 93% are achieved under multiple view gait recognition by using 5 cameras, 4 cameras and 3 cameras respectively.