关键词: 3D 点云检测；PointPillars 算法；INT8 量化；后训练量化；RMSNorm；自动驾驶

Abstract: Aiming at the real-time requirement of 3D point cloud detection algorithms in autonomous driving scenarios, INT8 quantization is a mainstream lightweight solution, but traditional quantization methods in the image field suffer from severe accuracy loss when applied to point cloud algorithms. Taking the PointPillars algorithm as the research object, through quantitative sensitive layer analysis, it is found that the pillar feature net (PFN) is the core bottleneck of quantization accuracy loss, and its input activation values show the distribution characteristics of large dynamic range difference and many outliers. Drawing on the distribution optimization idea of large models, this paper proposes to introduce an root mean square layer normalization (RMSNorm ) at the input end of the PFN layer and perform fine-tuning to compress the feature dynamic range into an INT8 quantization-friendly interval. Experiments show that this method reduces the mean average precision (mAP) loss of the PointPillars model after INT8 quantization from 13.7% to 1.4% on the NuScenes dataset, increases the inference speed by 3.1 times, and compresses the model volume by 75%. Moreover, this scheme is also effective on the CenterPoint algorithm, controlling the mAP loss within 1.8%. The research provides a feasible path for efficient quantization of 3D point cloud detection models and lays a foundation for ultra-low bit quantization.

Key words: 3D point cloud detection; PointPillars algorithm; INT8 quantization; post-training quantization; RMSNorm; autonomous driving