基于GCR-PointPillars的点云三维目标检测  

3D object detection in point cloud based on GCR-PointPillars

作　　者：伍新月 惠飞[1,2] 金鑫 WU Xinyue;HUI Fei;JIN Xin(School of Information Engineering,Chang’an University,Xi’an 710018,China;School of Electronics and Control Engineering,Chang’an University,Xi’an 710018,China)

机构地区：[1]长安大学信息工程学院,陕西西安710018 [2]长安大学电子与控制工程学院,陕西西安710018

出　　处：《现代电子技术》2024年第11期168-174,共7页Modern Electronics Technique

基　　金：国家自然科学基金面上项目(52172380);陕西省重点研发计划(2021ZDLGY04-06)。

摘　　要：针对PointPillars算法中存在识别与定位不准确的问题,提出一种GCR-PointPillars三维目标检测模型,该模型首先在Pillar特征网络中引入全局注意力机制,学习点云特征之间的相关性,增强伪图特征的全局信息交互能力;其次,基于ConvNeXt V2重新构建特征提取网络,提取更加丰富的语义信息,从而有效提升网络的学习能力;最后引入RDIoU来联合引导分类和回归任务,有效缓解分类和回归不一致的问题。文中模型在KITTI数据集中与基准网络相比,汽车类别在简单、中等、困难三种难度级别下分别提高了2.69%、4.29%、4.84%,并且推理速度达到25.8 f/s。实验结果表明,文中模型在保持实时性速度的同时,检测效果也有明显提升。In view of the inaccurate recognition and localization in PointPillars algorithm,a 3D object detection model based on GCR-PointPillars is proposed.In this model,a global attention mechanism(GAM)is introduced in the Pillar feature network to learn the correlation between the point cloud features,so as to enhance the global information interaction ability of the pseudo-map features.The feature extraction network is reconstructed based on ConvNeXt V2 to extract richer semantic information,which improves the learning ability of the network effectively.The RDIoU is introduced to jointly guide the classification and regression tasks,which effectively alleviates the inconsistency of classification and regression.In the KITTI dataset,in comparison with the benchmark network,this model improves the car category detection by 2.69%,4.29%and 4.84%at the three levels of simple,moderate and difficult,and its inference speed reaches 25.8 f/s.The experimental results show that the detection effect of the proposed model is improved significantly while maintaining real-time speed.

关 键 词：三维目标检测 注意力机制 ConvNeXt V2 损失函数 激光雷达点云 自动驾驶 

分 类 号：TN958.98-34[电子电信—信号与信息处理] TP391[电子电信—信息与通信工程]