Bounded Graph Conditioning for LiDAR 3D Object Detection Under Sensor Degradation.

Area of Science:

• Computer Vision
• Robotics
• Sensor Fusion

Background:

• Light Detection and Ranging (LiDAR) 3D object detection performance degrades significantly with sensor noise, outliers, and calibration errors.
• Existing detector evaluations often use clean benchmarks, not reflecting real-world performance under adverse conditions.

Purpose of the Study:

• To improve the robustness of existing LiDAR 3D object detectors against various sensor degradations.
• To introduce a novel front-end processing method that enhances sensor data quality before detection.

Main Methods:

• Introduced Bounded Graph Conditioning (BGC), a deterministic pre-voxelization front-end using k-nearest-neighbor (kNN) averaging with bounded residual correction.
• Developed a reproducible sensor-degradation stress protocol and risk-constrained operating-boundary analysis for evaluation.
• Tested BGC with established detectors (PointPillars, SECOND, Voxel R-CNN) on KITTI and nuScenes datasets.

Main Results:

• BGC significantly improved detection quality and feasible operating coverage under strong noise and outlier conditions.
• Performance gains were backbone-dependent and smaller for other degradation types.
• A range-adaptive BGC variant further enhanced performance on nuScenes under strong noise (mAP/NDS: 0.2687/0.4493 vs. baseline 0.2471/0.3846).

Conclusions:

• Bounded Graph Conditioning (BGC) offers a practical, sensor-side enhancement for LiDAR 3D object detection robustness against specific degradations.
• BGC's effectiveness is conditional, depending on the backbone, fault type, and operating conditions.
• Severe sensor drift, like translation drift, remains a challenge beyond BGC's scope, highlighting explicit sensing boundaries.