基于鸟瞰图的多任务端到端3D目标检测方法

End-to-End Multi-Task 3D Object Detection Method Based on Bird Eye's View Images

由于在自动驾驶过程中激光雷达扫描过程中受到空间限制,采集到的点云数据损失了过多的信息。在聚合表示方法的基础上,提出了基于鸟瞰图的多任务端到端3D目标检测方法。首先,利用纹理特征提取器与语义特征提取器对特征金字塔的高级特征层与低级特征层进行特征融合,获得扩展的特征金字塔层,使得自动驾驶过程中的中等目标(骑自行车的人)与小目标(行人)不再耦合在同一级别的特征金字塔层,提高了点云拓扑信息的区域细节性。其次,在损失函数中引入用于分类的Focal loss与用于回归的CIOU loss,改善了正负样本比例,对目标检测框进行了约束,使得目标框在回归过程中能够以更高精度收敛。实验结果表明,提出的方法在自动驾驶中小目标检测中具有更强的检测能力与更高的检测精度。

Due to spatial limitations during the scanning process of LiDAR during autonomous driving,the collected point cloud data loses too much information.Based on compressed representation,a novel multitask end to end 3D object detection model in Bird's Eye View is presented.Firstly,texture feature extractor and semantic feature extractor were used to fuse the high-level and low-level feature layers of a feature pyramid network(FPN)to obtain an extended feature pyramid layer,which eliminated the coupling of medium target(cyclist)and small target(pedestrian)at the same level of feature pyramid layer during the auto-driving process,improving the regional detail of the point cloud topology information.Secondly,the Focal loss and CIOU loss were introduced in the loss function to improve the proportion of negative sample and limit the bounding box.As a result,the bounding box can converge more accurately in the regression process.Finally,The experimental results show that the proposed method has better detection ability and higher detection accuracy in small and medium targets detection of auto-driving.