OG-PCL: Efficient Sparse Point Cloud Processing for Human Activity Recognition

The introduction of the Occupancy-Gated Parallel-CNN Bi-LSTM (OG-PCL) network marks a significant advancement in human activity recognition (HAR) using mmWave radar technology. With an impressive accuracy of 91.75% on the RadHAR dataset and a lightweight parameter size of only 0.83M, OG-PCL outperforms existing methods such as 2D CNN, PointNet, and 3D CNN. This efficiency is further enhanced by the innovative Occupancy-Gated Convolution (OGConv) block, which addresses the challenges of processing sparse point clouds. The tri-view parallel structure of OG-PCL has been validated through ablation studies, demonstrating its effectiveness in preserving spatial information across three dimensions. As privacy concerns grow with traditional camera and wearable-based HAR methods, the OG-PCL offers a robust and efficient alternative for real-time applications, paving the way for future developments in radar-based activity recognition.