Enhancing 3D Object Detection by Using Neural Network with Self-adaptive Thresholding

TL;DR

This work tackles false positives in 3D LiDAR-based object detection for urban field robotics under dynamic and adverse weather conditions. It introduces a post-processing Neural Network with self-adaptive, distance-aware thresholds, deriving a distance-based adaptive threshold via a quadratic fit to confidence as a function of distance and enforcing a decision rule $\mathcal{Q}_Δ$ based on $m_{xy}$, $\sigma_{xy}$, and a distance-dependent function $f(x,y)$, enabling seamless integration with detectors like PointPillars. The method demonstrates improved robustness on OS2 LiDAR data, including fog and rain scenarios, and reduces false positives from clutter such as foliage while preserving near and far object detections. The approach offers a practical path to safer autonomous navigation in real-world driving and suggests potential for real-time, learning-based extension in diverse driving conditions.

Abstract

Robust 3D object detection remains a pivotal concern in the domain of autonomous field robotics. Despite notable enhancements in detection accuracy across standard datasets, real-world urban environments, characterized by their unstructured and dynamic nature, frequently precipitate an elevated incidence of false positives, thereby undermining the reliability of existing detection paradigms. In this context, our study introduces an advanced post-processing algorithm that modulates detection thresholds dynamically relative to the distance from the ego object. Traditional perception systems typically utilize a uniform threshold, which often leads to decreased efficacy in detecting distant objects. In contrast, our proposed methodology employs a Neural Network with a self-adaptive thresholding mechanism that significantly attenuates false negatives while concurrently diminishing false positives, particularly in complex urban settings. Empirical results substantiate that our algorithm not only augments the performance of 3D object detection models in diverse urban and adverse weather scenarios but also establishes a new benchmark for adaptive thresholding techniques in field robotics.

Figures (2)

• Figure 1: The Nerual-Network framework of the proposed algorithm. Traditional 3D detection models typically employ a static, binary-threshold approach for post-processing. In contrast, our method utilizes Neural Network thresholding, which dynamically adjusts thresholds based on specific environmental inputs and detection contexts, thereby enhancing detection accuracy and reliability.
• Figure 2: Results for comparison of our method and traditional method. Our algorithm significantly enhances the efficacy of 3D object detection models in complex urban environments characterized by adverse weather conditions such as fog and rain. By effectively minimizing false positives and accurately discerning objects from misleading point clouds generated by inclement weather, the algorithm improves the precision of object detection. This advancement fosters safer navigation for autonomous objects by ensuring more reliable and accurate perception under challenging conditions.