Towards LiDAR point cloud geometry compression using rate-distortion optimization and adaptive quantization for human-machine vision

Abstract

Due to rapid advances in 3-Dimensional (3D) sensing and rendering technologies, point clouds have become increasingly widespread, bring significant challenges for transmission and storage. Existing LiDAR Point Cloud Compression (PCC) methods primarily focus on enhancing compression efficiency and maintaining high signal fidelity, with insufficient considering human and machine joint perception. This paper proposes Rate Distortion Optimization (RDO) and Adaptive Quantization (AQ) for LiDAR Point Cloud Geometry Compression (PCGC) to balance human–machine vision performance. Specifically, we first propose Hybrid Distortion RDO (HDRDO) using hybrid distortion and Lagrange multiplier, where the optimal weights are determined by Differential Evolution (DE) algorithm. Furthermore, by comprehensively analyzing the impacts of point clouds on a Gaussian-based classification method on overall quality, we propose a HDRDO-based AQ method to adaptively quantify important and non-important points by optimal Quantization Parameter (QP) selection. We implement on Geometry-based Point Cloud Compression (G-PCC) Test Model Category 1 and 3 (TMC13), called the anchor method. Compared with the anchor method, the proposed algorithm achieves consistent PSNR for human vision tasks and improves by 2.66% and 21.18% on accuracy at low bitrates for detection and segmentation, respectively. Notably, the proposed overall method performs better than the existing method.