基于改进的 PointPillars 小目标检测算法

针对传统点云检测算法中小目标漏检、误检等问题, 文章提出一种改进的 PointPillars 小目标检测算法。首先, 在主干网络中引入特征金字塔模块, 融合高层次语义信息与低层次定位信息来优化对小目标的检测; 其次, 加入残差模块和卷积块注意力模块 (convolutional block attention module, CBAM), 结合通道信息与空间信息来完成对深层特征的提取; 最后, 通过三检测头输出不同尺寸的特征图, 以实现对不同类别目标的精确识别。在公开数据集 KITTI 上进行的可行性分析结果表明, 相较于传统 PointPillars 算法, 该文算法对小目标识别的平均精度均值提升了 1.74%。实车试验结果表明, 该文算法对道路目标检测结果准确, 检测频率达 65 Hz, 满足实时性需求。

In view of the missing and false detection of small and medium targets in the traditional point cloud detection algorithm, this paper proposes a small target detection algorithm based on improved PointPillars. Firstly, feature pyramid module is added into the backbone network to optimize the detection of small targets by combining high-level semantic information and low-level positioning information. Secondly, by adding the residual module and the convolutional block attention module (CBAM), the channel information and spatial information are combined to complete the extraction of deep features. Finally, three detection heads output feature maps of different sizes to achieve accurate recognition of different types of targets. The feasibility analysis of the proposed algorithm is carried out on KITTI data set. Compared with traditional PointPillars algorithm, the proposed algorithm improves the mean average precision (mAP) for small target recognition by 1.74%. The results of real vehicle test show that the algorithm is accurate for road target detection, and the detection frequency of 65 Hz meets the real-time requirement.