Multiple Temporal Aggregation Embedding for Gait Recognition in the Wild

Gait recognition in the wild is a cutting-edge topic in biometrics and computer vision. Since people is less cooperative in the wild scenario, view angles, walking direction and pace cannot be controlled. It leads to high variance of effective sequence length and bad spatial alignment of adjacent frames, which degrades current temporal modeling method in gait recognition. To address the aforementioned issue, we propose a multi-level and multi-time span aggregation (MTA) approach for comprehensive spatio-temporal gait feature learning. With embedded MTA modules, a novel gait recognition architecture is proposed. Results of extensive experiments on three large public gait datasets suggest that our method achieves an excellent improvement on gait recognition performance, especially on the task of gait recognition in the wild.