Object Detection (OD) stands as a fundamental task in the area of autonomous driving environment perception. This study introduces a 3D OD method grounded in deep learning and an improved Euclidean clustering algorithm, aiming to improve the accuracy and efficiency of point cloud segmentation and OD. The core methodological innovations include: (1) the integration of the Cloth Simulation Filter (CSF) for accurate ground and non-ground point separation, combined with a K-Dimensional Tree (KD-Tree) structure and an adaptive parameter mechanism to enhance clustering robustness and efficiency; and (2) an enhanced PointNet architecture incorporating multi-scale grouping (MSG), multi-resolution grouping (MRG), and skip connections to improve local feature extraction and multi-level feature fusion. This method is differentiated from prior works by its holistic integration of density-aware segmentation and hierarchical feature aggregation, addressing key bottlenecks in handling sparse and uneven LiDAR data. The proposed method is rigorously evaluated on the KITTI and NuScenes benchmarks. It achieves segmentation accuracies of 94.96% and 93.12%, with single-frame processing times of 15.63 ms and 17.24 ms, respectively, demonstrating a superior balance of speed and precision compared to traditional Euclidean clustering and other baseline methods. For the 3D OD task, the model attains average detection accuracies of 94.36% and 92.68% on the respective datasets, representing statistically significant improvements ( p < 0.001) over the standard PointNet. The detection speed reaches 34 fps and 31 fps, meeting real-time requirements while outperforming existing frameworks in challenging scenarios involving occluded and multi-scale objects. The findings confirm that the proposed framework provides a robust, efficient, and generalizable solution for 3D environmental perception in autonomous driving systems.
Zhang et al. (Mon,) studied this question.