基于特征融合的改进型PointPillar点云目标检测  

Improved PointPillar point cloud object detection based on feature fusion

作　　者：张勇[1] 石志广[1] 沈奇 张焱[1] 张宇[1] ZHANG Yong;SHI Zhiguang;SHEN Qi;ZHANG Yan;ZHANG Yu(National Key Laboratory of Science and Technology on Automatic Target Recognition,College of Electronic Science and Technology,National University of Defense Technology,Changsha 410073,China)

机构地区：[1]国防科技大学电子科学学院ATR重点实验室,湖南长沙410073

出　　处：《光学精密工程》2023年第19期2910-2920,共11页Optics and Precision Engineering

基　　金：国家自然科学基金资助项目(No.62075239)。

摘　　要：针对PointPillar在自动驾驶道路场景下对点云稀疏小目标检测效果差的问题,通过引入一种多尺度特征融合策略和注意力机制,提出一种点云目标检测网络Pillar-FFNet。针对网络中的特征提取问题,设计了一种基于残差结构的主干网络;针对馈入检测头的特征图没有充分利用高层特征的语义信息和低层特征的空间信息的问题,设计了一种简单有效的多尺度特征融合策略;针对主干网络提取的特征图中信息冗余的问题,提出了一种卷积注意力机制。为验证所提算法的性能,在KITTI和DAIR-V2X-I数据集上进行实验。实验结果表明,所提出的算法在KITTI数据集上与PointPillar相比,汽车、行人和骑行者的平均精度最大提高分别为0.84%,2.13%和4.02%;在DAIR-V2X-I数据集上与PointPillar相比,汽车、行人和骑行者的平均精度最大提高分别为0.33%,2.09%和4.71%,由此证明了所提方法对点云稀疏小目标检测的有效性。A point cloud object detection network,Pillar-FFNet,is proposed by introducing a multiscale feature fusion strategy and an attention mechanism to address the ineffectiveness of PointPillar in detecting small sparse objects in point clouds in autonomous driving road scenarios.First,a backbone network based on a residual structure is designed for feature extraction in the network.Second,a simple and effec⁃tive multiscale feature fusion strategy is designed to address the problem that the feature maps fed into the detection head do not make full use of the semantic information of high-level features and the spatial infor⁃mation of low-level features.Finally,a convolutional attention mechanism is proposed to treat information redundancy in the feature maps extracted using the backbone network.To validate the performance of the proposed algorithm,experiments are conducted on the KITTI and DAIR-V2X-I datasets.The results show that the proposed algorithm achieves maximum average accuracy improvements of 0.84%,2.13%,and 4.02% for cars,pedestrians,and cyclists,respectively,on the KITTI dataset and maximum average accuracy improvements of 0.33%,2.09%,and 4.71% for cars,pedestrians,and cyclists,respectively,on the DAIR-V2X-I dataset compared with the PointPillar results.Experimental results demonstrate the effectiveness of the proposed method for the detection of sparse small objects in point clouds.

关 键 词：小目标检测 点云稀疏 PointPillar 残差结构 多尺度特征融合 卷积注意力 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]