Segmentation-Guided Denoising and Gradient-Aware Depth Completion for Enhanced LiDAR-Based Human Activity Recognition

Human Activity Recognition (HAR) using 3D-LiDAR sensors presents significant challenges due to occlusions, noise, and sparse depth data. While traditional filtering techniques and deep learning-based methods exist, they typically focus on autonomous driving applications and require difficult-to-obtain ground truth LiDAR datasets. In this work, we propose a novel segmentation-guided denoising and depth completion approach that enhances LiDAR depth images for HAR. Our method utilizes a U-Net-based human segmentation model to remove irrelevant background noise and a gradient-guided depth completion algorithm to interpolate missing depth values. We first calibrate RGB and depth data to generate accurately aligned training datasets. Our segmentation model, trained on these aligned images, achieves high accuracy (Dice Score: 0.958 on seen subjects, 0.923 on unseen subjects) in detecting human shapes. The depth completion algorithm then iteratively reconstructs missing values using a weighted average approach guided by local depth gradients, ensuring structural consistency in completed depth maps. Unlike previous works, our approach does not rely on ground-truth LiDAR datasets but instead leverages calibrated RGB-depth alignment for segmentation training. Experimental evaluations demonstrate that our method significantly improves depth image quality while maintaining computational efficiency, making it suitable for real-time HAR applications. To the best of our knowledge, this is the first study to combine segmentation-based denoising with gradient-aware depth completion for LiDAR-based HAR. Future work will focus on further improving segmentation accuracy and refining depth completion techniques for more complex environments.