针对现有基于伪点云的3D目标检测算法精度远低于基于真实激光雷达(Light Detection and ranging,Li-Dar)点云的3D目标检测,本文研究伪点云重构,并提出适合伪点云的3D目标检测网络.考虑到由图像深度转换得到的伪点云稠密且随深度增大逐渐稀疏,本文提出深度相关伪点云稀疏化方法,在减少后续计算量的同时保留中远距离更多的有效伪点云,实现伪点云重构.本文提出LiDar点云指导下特征分布趋同与语义关联的3D目标检测网络,在网络训练时引入LiDar点云分支来指导伪点云目标特征的生成,使生成的伪点云特征分布趋同于LiDar点云特征分布,从而降低数据源不一致造成的检测性能损失;针对RPN(Region Proposal Network)网络获取的3D候选框内的伪点云间语义关联不足的问题,设计注意力感知模块,在伪点云特征表示中通过注意力机制嵌入点间的语义关联关系,提升3D目标检测精度.在KITTI 3D目标检测数据集上的实验结果表明:现有的3D目标检测网络采用重构后的伪点云,检测精度提升了2.61%;提出的特征分布趋同与语义关联的3D目标检测网络,将基于伪点云的3D目标检测精度再提升0.57%,相比其他优秀的3D目标检测方法在检测精度上也有提升.
3D Object Detection Based on Feature Distribution Convergence Guided by LiDar Point Cloud and Semantic Association
In view of the accuracy of existing 3D object detection algorithms based on Pseudo-LiDar is far lower than that based on real LiDAR(Light Detection and ranging),this paper studies the reconstruction of Pseudo-LiDar and proposes a 3D object detection algorithm suitable for Pseudo-LiDar.Considering that the Pseudo-LiDAR obtained by image depth is dense and gradually sparse along the increase of depth,a depth related Pseudo-LiDAR sparsification method is proposed to reduce the subsequent calculation amount while retaining more useful Pseudo-LiDAR in the middle and long distance,so as to realize the reconstruction of Pseudo-LiDAR.Furthermore,a 3D object detection algorithm based on object feature distri-bution convergence under the guidance of LiDar point cloud and semantic association is proposed.During network train-ing,a laser point cloud branch is introduced to guide the generation of Pseudo-LiDAR object features,so that the generated Pseudo-LiDar object feature distribution converges to the feature distribution of laser point cloud object,thereby correcting the detection error caused by the difference between the two data sources.Aiming at the insufficient semantic association between Pseudo-LiDar in the 3D candidate bounding-box obtained by RPN(Region Proposal Network)network,an atten-tion perception module is designed to embed the semantic association between points through the attention mechanism in the feature representation of Pseudo-LiDar,so as to improve the accuracy of 3D object detection.The experimental results on KITTI 3D object detection dataset show when the existing 3D object detection network adopts the reconstructed Pseudo-LiDar,the detection accuracy is improved by 2.61%.Furthermore,the proposed 3D object detection network with the fea-ture distribution convergence and semantic association improves the accuracy by 0.57%.Compared with other excellent methods,it also improves the detection accuracy.
3D object detectionPseudo-LiDarsemantic associationdistribution convergenceattention perception
国家科技期刊平台
NSTL
万方数据
