Adversarial learning-based skeleton synthesis with spatial-channel attention for robust gait recognition
Ying Chen,Shixiong Xia,Jiaqi Zhao,Yong Zhou,Qiang Niu,Rui Yao,Dongjun Zhu,Hao Chen
DOI: https://doi.org/10.1007/s11042-022-12665-x
IF: 2.577
2022-04-19
Multimedia Tools and Applications
Abstract:Person re-identification (ReID) aims to identify the same person across multiple cameras. Gait recognition is the person ReID using human gait to identify a walking person, which is an effective identification technology with many advantages, such as remote identification and without invasion. State-of-the-art solutions solve the problem of extensive annotation skeleton information and labeling process by employing encoder-decoders to reconstruct skeleton data, which encodes gait feature with a fixed-length vector limiting the performance of this architecture. In this paper, we propose an end-to-end pipeline dubbed as SCA-GAN, which assembles spatial-channel attention with GAN-like framework to synthesize skeleton sequences reversely without labeled skeleton data. A disadvantage of traditional encoder-decoder architecture is that, because of the fixed-length latent vector encoded from the encoder, the decoder fails to learn the reasonable standard to generate imperfect samples. Therefore, we design a GAN-like framework for discriminative gait feature extraction via adversarial learning. In addition, for learning the rich global information of skeleton data, the information of skeleton is extracted via convolutional block embedding locality-aware attention mechanism. Specifically, a contrastive feature loss is constructed between the gait encoder and the gait decoder to minimize their pixel-wise distance explicitly. The proof-of-principle experiments and ablation study on several benchmarks prove that the proposed method significantly outperforms gait recognition counterparts in precision.
computer science, information systems, theory & methods,engineering, electrical & electronic, software engineering