Vote3Deep：利用稀疏卷积神经网络实现高效3D点云物体检测

需积分: 16 162 浏览量更新于2024-09-09 收藏 1.52MB PDF 举报

本文主要探讨了在3D点云数据中使用卷积神经网络（CNN）进行快速对象检测的方法。传统的3D点云处理通常依赖于密集的网格结构，这在计算上消耗较大。作者提出了一种名为Vote3Deep的创新方法，该方法引入了特征为中心的投票机制，设计了专门针对输入稀疏性的新型卷积层，实现了对3D点云的高效处理。在Vote3Deep架构中，关键点在于采用了一种特征导向的投票策略，这种策略能够有效地利用输入数据的稀疏特性，减少不必要的计算。通过这种方式，网络能够在保持较高精度的同时，显著提高执行速度。为了进一步优化模型的效率，研究者还提出了在卷积滤波器激活上应用L1正则化，这一策略有助于增强中间表示的稀疏性，从而进一步提升模型的性能。在实验部分，作者将他们的方法应用于KITTI对象检测基准，这是评估3D点云对象检测性能的标准平台。结果表明，即使是只有三层的Vote3Deep模型，也能在激光和激光视觉基础上的检测任务中超越先前的最佳方法，性能提升幅度高达40%。而且，与现有技术相比，Vote3Deep在处理时间上依然保持了竞争力。这篇论文标志着在3D点云数据处理中使用稀疏卷积层和L1正则化的先例，这对于大规模、实时的3D场景理解具有重大意义。它不仅提升了对象检测的准确性，还极大地提高了计算效率，为未来的3D计算机视觉和自动驾驶等领域开辟了新的可能性。这项工作对于那些关注3D点云处理效率和性能优化的科研人员和技术开发者来说，是一篇重要的参考文献。

Vote3Deep: Fast Object Detection in 3D Point Clouds Using Efﬁcient

Convolutional Neural Networks

Martin Engelcke, Dushyant Rao, Dominic Zeng Wang, Chi Hay Tong, Ingmar Posner

Abstract— This paper proposes a computationally efﬁcient

approach to detecting objects natively in 3D point clouds

using convolutional neural networks (CNNs). In particular, this

is achieved by leveraging a feature-centric voting scheme to

implement novel convolutional layers which explicitly exploit

the sparsity encountered in the input. To this end, we ex-

amine the trade-off between accuracy and speed for different

architectures and additionally propose to use an L

penalty

on the ﬁlter activations to further encourage sparsity in the

intermediate representations. To the best of our knowledge, this

is the ﬁrst work to propose sparse convolutional layers and L

regularisation for efﬁcient large-scale processing of 3D data.

We demonstrate the efﬁcacy of our approach on the KITTI

object detection benchmark and show that Vote3Deep models

with as few as three layers outperform the previous state of the

art in both laser and laser-vision based approaches by margins

of up to 40% while remaining highly competitive in terms of

processing time.

I. INTRODUCTION

3D point cloud data is ubiquitous in mobile robotics

applications such as autonomous driving, where efﬁcient and

robust object detection is pivotal for planning and decision

making. Recently, computer vision has been undergoing

a transformation through the use of convolutional neural

networks (CNNs) (e.g. [1], [2], [3], [4]). Methods which

process 3D point clouds, however, have not yet experienced a

comparable breakthrough. We attribute this lack of progress

to the computational burden introduced by the third spatial

dimension. The resulting increase in the size of the input

and intermediate representations renders a naive transfer of

CNNs from 2D vision applications to native 3D perception

in point clouds infeasible for large-scale applications. As a

result, previous approaches tend to convert the data into a 2D

representation ﬁrst, where nearby features are not necessarily

adjacent in the physical 3D space – requiring models to

recover these geometric relationships.

In contrast to image data, however, typical point clouds

encountered in mobile robotics are spatially sparse, as most

regions are unoccupied. This fact was exploited in [5],

where the authors propose Vote3D, a feature-centric voting

algorithm leveraging the sparsity inherent in these point

clouds. The computational cost is proportional only to the

number of occupied cells rather than the total number of

cells in a 3D grid. [5] proves the equivalence of the voting

scheme to a dense convolution operation and demonstrates

its effectiveness by discretising point clouds into 3D grids

and performing exhaustive 3D sliding window detection with

Authors are from the Oxford Robotics Institute, University of Oxford.

{firstname}@robots.ox.ac.uk

Inpu t&Point&Cloud

Object&Detections

CNNs

(a) 3D point cloud detection with CNNs

(b) Reference image

Fig. 1. The result of applying Vote3Deep to an unseen point cloud from

the KITTI dataset, with the corresponding image for reference. The CNNs

apply sparse convolutions natively in 3D via voting. The model detects cars

(red), pedestrians (blue), and cyclists (magenta), even at long range, and

assigns bounding boxes (green) sized by class. Best viewed in colour.

a linear Support Vector Machine (SVM). Consequently, [5]

achieves the previous state of the art in both performance and

processing speed for detecting cars, pedestrians and cyclists

in point clouds on the object detection task from the popular

KITTI Vision Benchmark Suite [6].

Inspired by [5], we propose to exploit feature-centric

voting to build efﬁcient CNNs to detect objects in point

clouds natively in 3D – that is to say without projecting the

input into a lower-dimensional space ﬁrst or constraining the

search space of the detector (Fig. 1). This enables our ap-

arXiv:1609.06666v2 [cs.RO] 5 Mar 2017

下载后可阅读完整内容，剩余6页未读，立即下载

lovefengqiang

粉丝: 2
资源: 16

Vote3Deep：利用稀疏卷积神经网络实现高效3D点云物体检测

sparse-segmentation:RISS 2018-稀疏LIDAR点云的分割

Real-Time Grasp Detection Using Convolutional Neural Networks

3D Fully Convolutional Network for Vehicle Detection in Point Cloud

ImVoteNet - Boosting 3D Object Detection in Point Clouds With Im

yolo开发-3D YOLO End-to-End 3D Object Detection Using Point Clouds

deep hough voting for 3d object detection in point clouds

pointpillars: fast encoders for object detection from point clouds

Multi-View 3D Object Recognition in Point Clouds

FlowNet3D Learning Scene Flow in 3D Point Clouds.pdf

Semantic Segmentation of Point Clouds using Deep Learning

最新资源