PePNet: Pose-Enhanced Point Cloud Network for LiDAR-Based Human Action Recognition in Outdoor Long-Range Scenarios

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Sensor Data Analysis

Background:

• LiDAR-based human action recognition (HAR) in outdoor, long-range scenarios faces challenges due to decreasing point cloud density with distance and simultaneous human-sensor motion.
• Existing methods struggle to maintain accuracy under these conditions, necessitating novel approaches for reliable performance.

Purpose of the Study:

• To develop a distance-aware framework, the Pose-Enhanced Point Cloud Network (PePNet), for improved long-range LiDAR-based HAR.
• To address point cloud sparsity and motion complexities through specialized modules and advanced modeling techniques.

Main Methods:

• Proposed the Pose-Enhanced Point Cloud Network (PePNet) featuring a core Pose-Enhanced Point Cloud Block (PeP Block) with Dynamic Enhancement, Pose Prompter, and Adaptive Point Selection modules.
• Integrated Spatiotemporal Tube Embedding (ST-Tube) with the Mamba state space model to capture long-range dependencies and complex motion dynamics.
• Introduced Momo, a large-scale LiDAR-based HAR dataset specifically for long-range (2-30 m) outdoor scenarios.

Main Results:

• PePNet demonstrated consistent performance improvements over existing methods on the newly introduced Momo dataset.
• The PeP Block proved effective in mitigating point cloud sparsity and handling motion artifacts in long-range HAR.
• The proposed framework shows significant potential for applications in robotics and autonomous vehicles.

Conclusions:

• PePNet offers a robust solution for long-range LiDAR-based HAR in challenging outdoor environments.
• The PeP Block serves as a versatile, plug-and-play module capable of enhancing other point cloud-based action recognition systems.
• The Momo dataset provides a crucial benchmark for advancing research in long-range outdoor HAR.