Huang Jinzhe, Xie Yiyuan, Chen Zhuang, Su Ye
College of Computer and Information Science, Chongqing Normal University, Chongqing 401331, China.
College of Electronics and Information Engineering, Southwest University, Chongqing 400715, China.
Sensors (Basel). 2025 Mar 9;25(6):1697. doi: 10.3390/s25061697.
The pursuit of robust 3D object detection has emerged as a critical focus within the realm of computer vision. This paper presents a curriculum-guided adversarial learning (CGAL) framework, which significantly enhances the adversarial robustness and detection accuracy of the LiDAR-based 3D object detector PointPillars. By employing adversarial learning with prior curriculum expertise, this framework effectively resists adversarial perturbations generated by a novel attack method, P-FGSM, on 3D point clouds. By masterfully constructing a nonlinear enhancement block (NEB) based on the radial basis function network for PointPillars to adapt to the CGAL, a novel 3D object detector named Pillar-RBFN was developed; it exhibits intrinsic adversarial robustness without undergoing adversarial training. In order to tackle the class imbalance issue within the KITTI dataset, a data augmentation technique has been designed that singly samples the point cloud with additional ground truth objects frame by frame (SFGTS), resulting in the creation of an adversarial version of the original KITTI dataset named Adv-KITTI. Moreover, to further alleviate this issue, an adaptive variant of focal loss was formulated, effectively directing the model's attention to challenging objects during the training process. Extensive experiments demonstrate that the proposed CGAL achieves an improvement of 0.8∼2.5 percentage points in mean average precision (mAP) compared to conventional training methods, and the models trained with Adv-KITTI have shown an enhancement of at least 15 percentage points in mAP, compellingly testifying to the effectiveness of our method.
对强大的三维目标检测的追求已成为计算机视觉领域的一个关键焦点。本文提出了一种课程引导的对抗学习（CGAL）框架，该框架显著提高了基于激光雷达的三维目标检测器PointPillars的对抗鲁棒性和检测精度。通过运用具有先验课程专业知识的对抗学习，该框架有效抵抗了一种新型攻击方法P-FGSM对三维点云产生的对抗性扰动。通过巧妙地基于径向基函数网络为PointPillars构建一个非线性增强块（NEB）以适应CGAL，开发了一种名为Pillarillar Pill-RBFN的新型三维目标检测器；它在不经过对抗训练的情况下展现出内在的对抗鲁棒性。为了解决KITTI数据集中的类别不平衡问题，设计了一种数据增强技术，即逐帧单独对带有额外真实物体的点云进行采样（SFGTS），从而创建了一个名为Adv-KITTI的原始KITTI数据集的对抗版本。此外，为了进一步缓解这个问题，制定了一种自适应的焦点损失变体，有效地在训练过程中引导模型关注具有挑战性的物体。大量实验表明，与传统训练方法相比，所提出的CGAL在平均精度均值（mAP）上提高了0.8至2.5个百分点，并且使用Adv-KITTI训练的模型在mAP上至少提高了15个百分点，有力地证明了我们方法的有效性。
