Toward Robust LiDAR based 3D Object Detection via Density-Aware Adaptive Thresholding

TL;DR

Robust LiDAR-based 3D object detection in real-world urban environments is challenged by false positives and distance-dependent detection quality. The paper introduces a post-processing distance-aware adaptive thresholding framework that modulates the detector confidence threshold using a quadratic distance model, with a cap beyond distance δ. Demonstrated across multiple LiDAR detectors (e.g., PointPillars, SECOND, PointRCNN, PV-RCNN), the method preserves or improves mAP while enhancing precision for near objects and recall at longer ranges, using learned parameters α, β, γ, δ, and k. Experiments on KITTI and a custom urban dataset, including adverse weather scenarios (fog, rain), show reduced false positives (e.g., misidentified bushes) and improved robustness, supporting safer autonomous navigation.

Abstract

Robust 3D object detection is a core challenge for autonomous mobile systems in field robotics. To tackle this issue, many researchers have demonstrated improvements in 3D object detection performance in datasets. However, real-world urban scenarios with unstructured and dynamic situations can still lead to numerous false positives, posing a challenge for robust 3D object detection models. This paper presents a post-processing algorithm that dynamically adjusts object detection thresholds based on the distance from the ego-vehicle. 3D object detection models usually perform well in detecting nearby objects but may exhibit suboptimal performance for distant ones. While conventional perception algorithms typically employ a single threshold in post-processing, the proposed algorithm addresses this issue by employing adaptive thresholds based on the distance from the ego-vehicle, minimizing false negatives and reducing false positives in urban scenarios. The results show performance enhancements in 3D object detection models across a range of scenarios, not only in dynamic urban road conditions but also in scenarios involving adverse weather conditions.

Figures (4)

• Figure 1: The dynamic environment involves moving vehicles, pedestrians, and obstacles such as bushes and road signs (RGB image). 3D object detection using a single threshold misclassifies the bushes as vehicles, which results in false positives (single threshold). Our method significantly improves the robustness of 3D object detection in complex urban scenarios by effectively minimizing false detections, thus enhancing autonomous driving systems' overall performance and safety (ours).
• Figure 2: Framework of the proposed algorithm. 3D Detectors zhou2018voxelnetren2015fasteryan2018secondlang2019pointpillarsshi2019pointrcnnshi2020pvduan2019centernet normally apply the post processing using the single threshold. In contrast, the adaptive thresholding.
• Figure 3: Confidence score tendency with a single threshold (=0.5). The red dots represent the score mean values at each distance (10 m), and the green-shaded area indicates the standard deviation.
• Figure 4: Results of our algorithm. In challenging urban road scenarios, such as fog and rain, our algorithm enhances the performance of 3D object detection models by reducing false positives and accurately distinguishing vehicles from point clouds caused by adverse weather conditions. This leads to improved overall precision of object detection, ensuring safer driving for autonomous vehicles.