Range-Aware Sparsification and Feature Pyramid Consistency for Pedestrian Detection in Sparse 3D LiDAR Point Clouds

3-D object detection in open space is a critical component of autonomous driving systems. Owing to the small physical size of pedestrians and the inherent sparsity of LiDAR points, current detectors exhibit low recall for pedestrian instances. This paper presents an end-to-end joint-optimization framework that simultaneously augments training data and strengthens the model’s capacity to recognize pedestrians. Specifically, (i) we propose a Density-Adaptive Sampling (DAS) strategy that dynamically adjusts the "ground-truth pasting" policy according to local point-cloud density; (ii) we introduce Physics-Aware Range Projection (PARP), which emulates the beam-divergence characteristics of a spinning LiDAR to synthesize high-fidelity, long-range hard examples whose physical attributes match real acquisitions; and (iii) we design a Consistency-Enhanced Feature Pyramid Network (CE-FPN) that replaces transposed convolutions with interpolated convolutions to eliminate checkerboard artifacts, and further embeds a self-supervised consistency regularization loss to enhance the discriminative power of sparse features. Extensive experiments on the KITTI benchmark show that the proposed framework improves the 3-D Average Precision and Bird’s-Eye-View Average Precision of pedestrian detection by 2.36% and 2.29% over the baseline, respectively, demonstrating significant gains for pedestrian localization.