Gait Recognition in Large-scale Free Environment via Single LiDAR

TL;DR

This work tackles in-the-wild gait recognition using a single LiDAR by introducing HMRNet, a hierarchical, multi-representation network that fuses dense range-view cues with rich 3D motion from raw point clouds. It pairs a hierarchical adaptive cross-representation mapping with motion-aware embedding and gait-saliency enhancement to capture both static body structure and dynamic gait cues, achieving state-of-the-art results on SUSTech1K and the new FreeGait dataset. FreeGait provides long-range, unconstrained, multi-modal gait data in real-world scenes with occlusions and crowds, enabling more realistic evaluation and fostering robust method development. The approach demonstrates strong cross-view and low-light robustness, highlighting the practical potential of LiDAR-based gait recognition for security, healthcare, and smart environments.

Abstract

Human gait recognition is crucial in multimedia, enabling identification through walking patterns without direct interaction, enhancing the integration across various media forms in real-world applications like smart homes, healthcare and non-intrusive security. LiDAR's ability to capture depth makes it pivotal for robotic perception and holds promise for real-world gait recognition. In this paper, based on a single LiDAR, we present the Hierarchical Multi-representation Feature Interaction Network (HMRNet) for robust gait recognition. Prevailing LiDAR-based gait datasets primarily derive from controlled settings with predefined trajectory, remaining a gap with real-world scenarios. To facilitate LiDAR-based gait recognition research, we introduce FreeGait, a comprehensive gait dataset from large-scale, unconstrained settings, enriched with multi-modal and varied 2D/3D data. Notably, our approach achieves state-of-the-art performance on prior dataset (SUSTech1K) and on FreeGait.

Figures (6)

• Figure 1: Examples of diverse visual modalities and 2D/3D representations in FreeGait.
• Figure 2: Examples of predefined walking path in constrained environments on SUSTech1K shen2022lidar, contravenes the free gait patterns observed in real-world scenarios.
• Figure 3: The pipeline of our method. We extract dense body structure information from range views, and undistorted geometric and motion features via motion-aware feature embedding (MAFE) from point clouds. Then, adaptive cross-representation mapping module (ACM) is applied to fuse two-representation features at different levels hierarchically. Lastly, the gait-saliency feature enhancement (GSFE) module is leveraged to highlight more gait-informative features for final identification.
• Figure 4: The procedure of point-wise flow to learn motion relation between two adjacent frames.
• Figure 5: The exemplar range views on SUSTech1K shen2022lidar in four different attributes with serious appearance variance between gallery set and probe set.