期刊文献+

基于可变形卷积和数据增强的三维多目标检测

3D Object Detection Based on Deformable Convolution and Data Augmentation
下载PDF
导出
摘要 近年来,自动驾驶受到越来越多的关注,以点云为输入数据的三维目标检测在该领域中发挥着至关重要的作用。然而,点云目标的尺度差异性以及变换性等问题,导致了目标检测精度的下降。以CenterPoint网络为框架,提出了一种基于可变形卷积和数据增强的三维多目标检测优化算法,该方法提取点云特征后生成地图视角的特征图谱,在检测头网络加入可变形卷积层,并引入图像翻转方法进行数据增强,提高网络对于目标的检测能力。在公开数据集nuScenes上的实验结果表明,该网络与其他方法相比,在汽车、公交车以及行人等类别的检测精度上有一定程度的提升。 Automated driving has received increasing attention in recent years,and 3D target detection using point clouds as input data plays a crucial role in automated driving.However,problems such as scale variability of point cloud targets and transformability have led to degradation of target detection accuracy.In this paper,a 3D target detection algorithm based on deformable convolution and data enhancement is proposed using the CenterPoint network as a framework.The method extracts point cloud features and then generates a feature map of the map view,adds a deformable convolution layer to the detection head network and introduces an image flipping method for data enhancement to improve the network's detection capability for targets.Experimental results on the publicly available nuScenes dataset show that the network achieves a certain degree of improvement in detection accuracy compared to other methods in the categories of car,bus and pedestrian.
出处 《工业控制计算机》 2023年第3期22-24,共3页 Industrial Control Computer
关键词 三维目标检测 可变形卷积 数据增强 中心点检测 3D object detection deformable convolution data augmentation CenterPoint

相关主题

;
使用帮助 返回顶部