面向雷达目标检测网络的点云改进多尺度栅格渲染
摘要
先将点云转换为栅格表示再应用卷积神经网络的架构在基于雷达的目标检测中取得了良好性能。然而,由于点的离散化与聚合,从不规则点云数据到稠密栅格结构的转换常伴随信息损失。本文提出一种新颖架构multi-scale KPPillarsBEV,旨在缓解栅格渲染的负面影响。具体而言,我们提出一种新颖的栅格渲染方法KPBEV,其利用核点卷积的描述能力改善栅格渲染过程中局部点云上下文的编码。此外,我们提出一种通用多尺度栅格渲染公式,以将多尺度特征图纳入采用任意栅格渲染方法的检测网络卷积骨干中。我们在nuScenes数据集上进行了广泛实验,并从检测性能与计算复杂度方面评估了这些方法。所提出的multi-scale KPPillarsBEV架构在nuScenes验证集上的Car AP4.0(匹配阈值为4米的平均精度)指标上比基线高出5.37%,比先前最优方法高出2.88%。此外,所提出的单尺度KPBEV栅格渲染在保持相同推理速度的同时,将Car AP4.0较基线提升了2.90%。
引用
@article{arxiv.2305.15836,
title = {Improved Multi-Scale Grid Rendering of Point Clouds for Radar Object Detection Networks},
author = {Daniel Köhler and Maurice Quach and Michael Ulrich and Frank Meinl and Bastian Bischoff and Holger Blume},
journal= {arXiv preprint arXiv:2305.15836},
year = {2023}
}
备注
(c) 2023 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works