基于注意力和可变形卷积的三维目标检测

3D Target Detection based on Attention and Deformable Convolution

针对自动驾驶场景中物体特征信息有限、模板不适应目标几何形变的问题,提出一种通道注意力机制与可变形卷积结合的网络框架。该方法先将点云划分柱体,再编码为伪图像并送入卷积网络层,结合可变形卷积,引入目标形变偏移量,并引入通道注意力增强网络对目标关键信息的提取能力。实验表明该网络框架能够提高车辆及骑行者的检测精度,在二维模式下,3种类别的检测精度均值提高了1.07%,一定程度上解决了原网络框架存在的漏检问题。

This paper aims at the problem of limited object feature information in the sparse point cloud and the template does not adapt to the geometric deformation of the target in the automatic driving scene,and proposes a network framework combining channel attention mechanism and deformable convolution.First,the network divides spatial cloud points intopillars,then encodes them as pseudo-images and send them to the convolutional network layer.Combined with deformable convolution,this paper introduces the target deformation offset,and introduces the channel attention to enhance the network’s ability to extract key information of the target.Experiments prove that the network framework proposed can effectively improve the detection accuracy of vehicles andriders.In the 2D mode,the average detection accuracy of the three categories is increased by 1.07%,which deals with the problem of missing detection caused by the original network framework to a certain extent.