基于语义分割引导的三维目标检测  

3D object detection based on semantic segmentation guidance

作　　者：崔振东 李宗民[1,2] 杨树林 刘玉杰 李华[3,4] CUI Zhen-dong;LI Zong-min;YANG Shu-lin;LIU Yu-jie;LI Hua(College of Computer Science and Technology,China University of Petroleum,Qingdao Shandong 266580,China;College of Big Data and Basic Science,Shandong Institute of Petroleum and Chemical Technology,Dongying Shandong 257061,China;Key Laboratory of Intelligent Information Processing,Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190,China;School of Computer Science and Technology,University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国石油大学(华东)计算机科学与技术学院,山东青岛266580 [2]山东石油化工学院大数据与基础科学学院,山东东营257061 [3]中国科学院计算技术研究所智能信息处理重点实验室,北京100190 [4]中国科学院大学计算机科学与技术学院,北京100049

出　　处：《图学学报》2022年第6期1134-1142,共9页Journal of Graphics

基　　金：国家重点研发计划项目(2019YFF0301800);国家自然科学基金青年基金项目(61806199);国家自然科学基金项目(61379106);山东省自然科学基金项目(ZR2013FM036,ZR2015FM011);中国石油大学(华东)研究生创新基金项目(22CX04037A)。

摘　　要：三维目标检测是计算机视觉领域的热门研究内容之一。在自动驾驶系统中,三维目标检测技术通过捕获周围的点云信息与RGB图像信息,对周围物体进行检测,从而为车辆规划下一步的行进路线。因此,通过三维目标检测实现对周边环境的精准检测与感知是十分重要的。针对三维目标检测技术中随机采样算法导致前景点丢失的问题,首先提出了基于语义分割的随机采样算法,通过预测的语义特征指导采样过程,提升了前景点的采样比重,进而提高了三维目标检测精度;其次,针对三维目标检测定位置信度与分类置信度不一致的问题,提出了CL联合损失,使得网络倾向于选择定位置信度与分类置信度都高的3D候选框,避免了传统的NMS仅考虑分类置信度所带来的歧义问题。在KITTI三维目标检测数据集进行了实验,结果表明,该方法能够在简单、中等、困难3个难度下均获得精度的提升,从而验证了其在三维目标检测任务中的有效性。3D object detection is one of the most popular research fields in computer vision. In the self-driving system,the 3D object detection technology detects the surrounding objects by capturing the surrounding point cloud information and RGB image information, thereby planning the upcoming route for the vehicle. Therefore, it is of great importance to attain the accurate detection and perception of the surrounding environment. To address the loss of foreground points incurred by random sampling in the field of 3D object detection, a random sampling algorithm based on semantic segmentation was proposed, which guided the sampling process through the predicted semantic features, so as to increase the sampling proportion of foreground points and heighten the precision of 3D object detection. Secondly, to address the inconsistency between the location confidence of 3D object detection and the classification confidence, the CL joint loss was proposed, leading the network to select the 3D bounding box with high location confidence and classification confidence, so as to prevent the ambiguity caused by the traditional NMS only considering the classification confidence. Experiments on KITTI 3D object detection datasets show that the proposed method can improve the precision at the three levels of difficulties: easy, moderate, and hard, which verifies the effectiveness of the method in 3D object detection task.

关 键 词：深度学习 三维目标检测 点云语义分割 采样算法 定位置信度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]