A Survey of Robust 3D Object Detection Methods in Point Clouds

Zimmer, Walter; Erçelik, Emeç; Zhou, Xingcheng; Ortiz, Xavier Jair Diaz; Knoll, Alois · 2022 · OpenAlex-citations

DOI: 10.48550/arxiv.2204.00106

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This survey paper reviews the state-of-the-art in LiDAR-based 3D object detection for autonomous driving, addressing the critical need for robust perception systems capable of handling sensor noise, occlusions, and adverse environmental conditions such as rain, snow, and fog. The authors aim to provide a comprehensive overview of novel detection architectures, data augmentation techniques, sampling strategies, activation functions, attention mechanisms, regularization methods, normalization techniques, learning rate schedules, and loss functions. The study is motivated by the superior performance of LiDAR over camera-only systems, particularly in low-light conditions and at ranges of 50–60 meters, making it vital for safe autonomous navigation. The authors categorize 3D object detection methods into point-based, voxel-based, range-view-based, and multi-view-based approaches. They analyze specific architectures, distinguishing between one-stage detectors (e.g., PointPillars, 3DSSD, SA-SSD, CIA-SSD, SE-SSD) and two-stage detectors (e.g., PV-RCNN, PV-RCNN++, CenterPoint). The review details technical innovations within these models, such as PointPillars’ efficient 2D voxelization, PV-RCNN’s integration of voxel and point-based features, and CenterPoint’s anchor-free center estimation. Additionally, the paper evaluates ten major autonomous driving datasets, including KITTI, nuScenes, Waymo, and ArgoVerse, comparing their modalities, class distributions, and environmental coverage. Experimental evaluations were conducted on the KITTI, nuScenes, and Waymo datasets to assess mean average precision (mAP) and inference speed. The authors report that LiDAR-based detectors significantly outperform camera-only methods, with the best LiDAR algorithm achieving 68.63% higher accuracy on the KITTI 3D car class. Specific performance metrics include PointPillars running at 62 Hz using TensorRT, SA-SSD achieving 25 Hz while matching two-stage accuracy, and PV-RCNN++ improving mAP by 4% on Waymo. The survey also highlights the impact of advanced techniques like AutoAugment for data augmentation, Farthest-Point-Sampling for efficient key point selection, and Focal Loss for addressing class imbalance. The significance of this work lies in its systematic comparison of robustness, speed, and accuracy across diverse detection frameworks and datasets. The authors identify current challenges, such as point cloud sparsity and occlusion, and propose multi-frame 3D object detection as a promising direction for future research. By aggregating features from multiple frames through implicit processing, scene-level aggregation, or attention mechanisms, future systems can better handle dynamic traffic scenarios. This survey serves as a foundational reference for researchers aiming to develop more reliable and efficient perception systems for autonomous vehicles.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• distraction detection algorithms