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 from decreasing point cloud density and simultaneous human-sensor motion.
• Existing methods struggle with data sparsity and complex dynamics inherent in long-range outdoor HAR.

Purpose of the Study:

• To develop a novel framework, the Pose-Enhanced Point Cloud Network (PePNet), specifically designed for distance-aware, long-range LiDAR-based HAR.
• To introduce a new dataset, Momo, to facilitate research and evaluation in challenging long-range outdoor HAR scenarios.

Main Methods:

• The core Pose-Enhanced Point Cloud Block (PeP Block) integrates Dynamic Enhancement, Pose Prompter, and Adaptive Point Selection modules to address point cloud sparsity and motion complexities.
• A Spatiotemporal Tube Embedding (ST-Tube) combined with the Mamba state space model is employed to capture long-range dependencies and intricate motion dynamics.
• A large-scale dataset, Momo, focusing on long-range (2-30 m) outdoor HAR with sparse point clouds and simultaneous motion was created.

Main Results:

• PePNet demonstrates consistent performance improvements over existing methods on the newly introduced Momo dataset.
• The PeP Block module shows effectiveness as a plug-and-play component, enhancing other point cloud-based action recognition frameworks for long-range outdoor settings.

Conclusions:

• PePNet offers a robust solution for LiDAR-based HAR in challenging long-range outdoor environments by effectively managing sparse data and complex motion.
• The developed Momo dataset provides a crucial benchmark for advancing research in this specialized domain.