Learning view?invariant features using stacked autoencoder for skeleton?based gait recognition

Human gait recognition in a multicamera environment is challenging task biometrics because of the presence large pose and variations illumination among different views. In this work, to address problem view, we present novel stacked autoencoder for learning discriminant view-invariant representations. Our can efficiently progressively translate skeleton joint coordinates from any arbitrary view common canonical without requiring prior estimation angle or covariate type losing temporal information. Then, construct discriminative feature vector by fusing encoded features with two other spatiotemporal feed into main recurrent neural network. Experimental evaluations CASIA A B datasets demonstrate that proposed approach outperformed state-of-the-art methods on single-view recognition. particular, method achieved 46.31% 33.86% average correct class probe set ProbeBG ProbeCL, respectively, dataset while considering variation; 0.3% 30.68% higher than previous best-performing methods. Furthermore, cross-view recognition, our shows better results over when view-angle variation 36°.