HEGait: Optimizing heatmaps with an end-to-end framework for higher-accuracy gait recognition.

Abstract

Pose-based gait recognition methods can effectively exclude covariates unrelated to gait (e.g. background, carriers, and clothing), making them a prominent research direction. Presently, the SOTA pose-based methods rely on heatmaps derived from pretrained pose estimators. However, these heatmaps primarily encode static keypoint positions and fail to capture motion patterns, limiting their suitability for gait recognition. To address this limitation, we propose HEGait, a heatmap-based end-to-end gait recognition framework. Unlike existing methods, HEGait jointly optimizes the pose estimator and gait recognition module under the supervision of gait-specific losses, enabling the generated heatmaps to implicitly encode gait-related motion features. Furthermore, we design a heatmap refinement module (HRM) that integrates keypoint-level and pixel-level attention mechanisms to explore the contributions of different keypoints and heatmap positions. This work provides a novel perspective for gait recognition research: optimizing intermediate representations. Extensive experiments on CASIA-B, CCPG, and SUSTech1K datasets show that HEGait achieves SOTA results among pose-based methods, validating its effectiveness.