To address the challenges in autonomous driving environment perception — namely, that pure vision-based schemes are prone to misjudgment, existing 3D point cloud object detection algorithms suffer from insufficient feature representation due to the sparsity and inhomogeneity of point clouds, low detection accuracy in complex scenarios, and the fact that most algorithms only validate single-category objects, failing to meet the multi-category perception requirements of national standards — a feature-enhanced PointRCNN network (FE-PointRCNN) is proposed. To overcome the limitations of traditional MLP structures, which possess strong global feature extraction capabilities but weak local geometric feature capture, along with the high computational overhead of attention mechanisms, Enhanced Multi-Layer Perceptrons (E-MLPs) and Residual Channel Attention (RC-Attention) modules are designed. These modules are integrated to construct an improved E-PointNet++ feature extraction architecture that enables parallel multi-scale feature extraction and adaptive weighted enhancement of key channel features, thereby alleviating the gradient vanishing problem and enhancing feature discriminability. Comparative and ablation experiments are conducted on the KITTI and ONCE datasets. The results demonstrate that the proposed FE-PointRCNN achieves accuracy improvements over the baseline PointRCNN* for the Car, Pedestrian, and Cyclist categories and exhibits strong robustness on the complex-scene ONCE dataset. The proposed algorithm effectively improves point cloud feature extraction capability and multi-category object detection performance, providing valuable technical support for LiDAR-based perception systems in autonomous driving. 

针对自动驾驶环境感知中纯视觉方案易误判、现有三维点云目标检测算法受点云稀疏不均特性约束导致特征表达不足、复杂场景检测精度偏低、多数算法因仅验证单一类别目标因而难以满足国标多类别感知要求等问题, 提出一种特征增强的PointRCNN网络(FE-PointRCNN)。为克服传统MLP结构全局特征提取能力强而局部几何特征捕获不足、注意力机制计算开销大等缺陷, 设计增强多层感知机(E-MLPs)和残差通道注意力(RC-Attention)模块, 构建改进的E-PointNet++特征提取架构, 实现多尺度特征并行提取与关键通道特征自适应加权增强, 缓解梯度消失并提升特征判别性能。在KITTI与ONCE数据集上的对比实验和消融实验结果表明, 与基线PointRCNN*相比, FE-PointRCNN在Car, Pedestrian和Cyclist类别上均能实现精度的提升, 且在ONCE复杂场景数据集上表现出良好的鲁棒性。所提算法能够有效地改善点云特征提取能力和多类别目标检测性能, 可为自动驾驶激光雷达感知系统提供技术支撑。