Spatiotemporal training‐free LiDAR denoising and hybrid pedestrian detection on airport aprons: Focus on low‐speed and stationary targets

Abstract LiDAR sensors in apron environments enable detailed monitoring of aircraft, vehicles, and pedestrians, but their clouds points often contain irregular noise from reflections, clutter, and dynamic interference. This study proposes the density and temporally aware filter (DTF‐Filter), which combines local density‐adaptive thresholding with temporal consistency across consecutive frames. A multi‐stage pedestrian recognition pipeline integrates adaptive ground segmentation, density‐based clustering, and a hybrid classifier of traditional machine learning and a lightweight deep learning model. On real‐world apron data (15 000 frames) and KITTI, DTF‐Filter achieves a mean outlier reduction of 78.0% ± 4.2% (max: 95.2%) versus 48.3% ± 6.1% for statistical outlier removal (SOR) ( p < 0.001) and improves the pedestrian detection ‐score from 0.891 ± 0.023 to 0.952 ± 0.012 ( p < 0.001), with consistent performance across weather conditions (clear: 0.961 ± 0.009, fog: 0.943 ± 0.015, rain: 0.937 ± 0.018). The method provides a computationally efficient (11 ± 1.2 ms/frame on CPU), training‐free solution for real‐time LiDAR perception of stationary and low‐speed targets in safety‐critical apron environments, intended to complement higher velocity perception modules.