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

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Pose-based gait recognition methods effectively isolate gait from other factors.
• Current state-of-the-art (SOTA) methods use heatmaps from pretrained pose estimators, but these lack dynamic motion information.
• This limitation hinders their performance in gait recognition tasks.

Purpose of the Study:

• To propose HEGait, a novel heatmap-based, end-to-end framework for gait recognition.
• To enhance gait recognition by optimizing intermediate pose representations.
• To develop a method that implicitly encodes gait-related motion features within heatmaps.

Main Methods:

• HEGait jointly optimizes the pose estimator and gait recognition module using gait-specific losses.
• A heatmap refinement module (HRM) is introduced, incorporating keypoint-level and pixel-level attention.
• This approach enables exploration of keypoint and heatmap position contributions.

Main Results:

• HEGait achieves state-of-the-art (SOTA) performance among pose-based methods on CASIA-B, CCPG, and SUSTech1K datasets.
• The framework demonstrates effective encoding of gait-related motion features.
• Joint optimization and heatmap refinement significantly improve recognition accuracy.

Conclusions:

• Optimizing intermediate representations, like heatmaps, offers a new direction for gait recognition research.
• HEGait provides a robust and effective solution for accurate gait recognition.
• The proposed method validates the effectiveness of integrating pose estimation and gait analysis.