GaitCloud: Leveraging Spatial-Temporal Information for LiDAR-Base Gait Recognition with A True-3D Gait Representation

Shaoxiong Zhang, Hiromitsu Awano, Takashi SatoWACV 2025

TL;DR

This paper introduces GaitCloud, a novel 3D gait representation for LiDAR-based gait recognition that preserves spatial integrity better than existing depth-image-based methods. By leveraging voxelization and introducing specialized network modules (LE and HCP), GaitCloud achieves state-of-the-art performance, demonstrating the potential of true 3D representations for robust human identification.

What Problem Does This Paper Address?

Gait recognition, identifying individuals by their walking patterns, has seen significant progress using camera-based methods. However, these methods often struggle with variations in viewpoints and clothing, as 2D image projections lose crucial 3D spatial information. While LiDAR sensors offer precise 3D data, existing LiDAR-based gait recognition methods still project point clouds into 2D depth images, inheriting the limitations of camera-based approaches. This paper addresses the challenge of effectively utilizing the rich 3D spatial-temporal information from LiDAR point clouds to achieve more robust and accurate gait recognition, overcoming the compromises made by 2D projections.

Methods

The core innovation is GaitCloud, a true 3D gait representation built upon voxelization. Instead of projecting point clouds into 2D depth images, GaitCloud converts each frame of a LiDAR gait sequence into a 3D voxel grid. These voxelized frames are then temporally fused by summing their voxel-wise values to create a compact, fixed-size 3D representation.

To process this 3D representation, the authors introduce two novel network modules:

  • Layer Encoder (LE): Addresses the inherent disparity in vertical and horizontal resolutions of LiDAR point clouds. It extracts features within "layers" (defined by laser channels) to mitigate information imbalances.
  • Horizontal Convolutional Pooling (HCP): An adaptation of Horizontal Pyramid Pooling (HPP), designed to extract positional features from the 3D feature maps generated by the encoder, enhancing interpretability.

The overall network architecture uses a 3D ResNet backbone, modified to incorporate the LE as an input layer and HCP for feature extraction.

Datasets

The primary dataset used is SUSTech1K, an open-access LiDAR-based gait dataset. The authors follow the dataset's established protocol for training and testing. The dataset contains gait data from 1050 identities captured from 12 viewpoints, with various attributes like "Normal," "Bag," "Clothing," "Carrying," "Occlusion," "Night," "Uniform," and "Umbrella" to simulate real-world variations.

Key Insights and Ideas

  • 3D is Better than Projected 2D: The fundamental insight is that preserving the full 3D spatial information through voxelization, rather than projecting it into 2D, significantly improves gait recognition performance, especially under challenging conditions.
  • LiDAR's Unique Properties Require Tailored Modules: The inherent resolution differences in LiDAR data (vertical vs. horizontal) necessitate specialized modules like the Layer Encoder (LE) to effectively process the information.
  • Voxelization as a Compact Representation: GaitCloud's voxelization approach effectively reduces data volume while preserving spatial integrity, making it a viable alternative to dense depth images.
  • Robustness to Viewpoint Variations: The proposed method, particularly with gait rotation, shows remarkable robustness to different viewing angles, a common challenge in gait recognition.

Key Results

GaitCloud achieved state-of-the-art performance on the SUSTech1K dataset:

  • Variance Experiments: Achieved accuracies of 93.1% across various conditions (Bag, Clothing, Carrying, Occlusion, Night, Uniform, Umbrella). It particularly excelled in the "Umbrella" subset, outperforming previous methods.
  • Cross-View Experiments: Achieved 89.2% accuracy, demonstrating strong performance even when the probe and gallery sets are from different viewpoints.
  • Comparison with LidarGait: GaitCloud consistently outperformed the LidarGait baseline, especially in cross-view scenarios, highlighting the benefits of the true 3D representation.
  • Module Effectiveness: The combination of LE and HCP modules led to significant accuracy improvements, particularly in cross-view experiments without gait rotation.

Limitations and Future Work

  • Computational Complexity: While GaitCloud is more efficient than some 3D representations, extending the temporal dimension can significantly increase computational cost, as noted with the 3D-LidarGait comparison.
  • LiDAR Sensor Limitations: The performance of LiDAR-based methods can degrade with lower-resolution sensors or at longer distances, similar to challenges faced by low-resolution cameras.
  • Imbalanced Data: The study noted that imbalanced training data distribution could lead to slightly compromised accuracy for minor attributes.
  • Future Work: The paper suggests further exploration of the unique properties of LiDAR point clouds and the development of more efficient network architectures.

Summary updated

BibTeX
@inproceedings{zhang2025gaitcloud,
  author = {Shaoxiong Zhang and Hiromitsu Awano and Takashi Sato},
  title = {{GaitCloud: Leveraging Spatial-Temporal Information for LiDAR-Base Gait Recognition with A True-3D Gait Representation}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
}
Figure