SqueezeSeg：用于3D激光雷达点云中道路物体实时分割、具有递归CRF的卷积神经网络

每天一分钟，带你读遍机器人顶级会议文章

标题：SqueezeSeg: Convolutional Neural Nets with Recurrent CRF forReal-Time Road-Object Segmentation from 3D LiDAR Point Cloud

作者：Bichen Wu, Alvin Wan, Xiangyu Yue and Kurt Keutzer

来源：IEEE International Conference on Robotics and Automation (ICRA)，2018

编译：曹利浩

审核：颜青松，陈世浪

欢迎个人转发朋友圈；其他机构或自媒体如需转载，后台留言申请授权

摘要

本文讨论了3D激光雷达点云中道路上物体的语义分割。尤其是希望检测和分类出感兴趣物体实例，例如汽车、行人和骑自行车的人。我们把这个问题表述为一个逐点分类问题，并且基于卷积神经网络（Convolutional Neural Networks, CNN）提出一个端到端的算法——SqueezeSeg：该CNN网络以变换后的点云作为输入，直接输出一个逐点标记的地图，然后使用作为RNN层的CRF（Conditional random field）改善该地图。最后使用传统的聚类算法获得实例级的标记。我们的CNN模型使用KITTI数据集中的激光雷达点云数据训练，分割需要的每个点的标签根据KITTI数据集中的3D边界框制作。为了获得额外的训练数据，我们在流行的视频游戏“Grand Theft Auto V（GTA-V）”中构建了一个激光雷达模拟器，以合成大量逼真的训练数据。实验结果表明，SqueezeSeg精度高，并且具有快速的和稳定的运行时间（每帧8.7±0.5ms），非常适合自动驾驶应用。另外，使用合成数据的额外训练提升了SqueezeSeg对真实世界数据的验证准确性。

源码地址：https://github.com/BichenWuUCB/SqueezeSeg

介绍和演示视频地址：https://youtu.be/Xyn5Zd3lm6s

网络结构：

分割结果示例：

精度：

实时性：

Abstract

We address semantic segmentation of road-objectsfrom 3D LiDAR point clouds. In particular, we wish to detectand categorize instances of interest, such as cars, pedestriansand cyclists. We formulate this problem as a point-wise clas-sification problem, and propose an end-to-end pipeline calledSqueezeSeg based on convolutional neural networks (CNN):the CNN takes a transformed LiDAR point cloud as input anddirectly outputs a point-wise label map, which is then refined bya conditional random field (CRF) implemented as a recurrentlayer. Instance-level labels are then obtained by conventionalclustering algorithms. Our CNN model is trained on LiDARpoint clouds from the KITTI [1] dataset, and our point-wisesegmentation labels are derived from 3D bounding boxes fromKITTI. To obtain extra training data, we built a LiDARsimulator intoGrand Theft Auto V (GTA-V), a popular videogame, to synthesize large amounts of realistic training data.Our experiments show that SqueezeSeg achieves high accuracywith astonishingly fast and stable runtime (8.7±0.5ms perframe), highly desirable for autonomous driving. Furthermore,additionally training on synthesized data boosts validationaccuracy on real-world data. Our source code is open-sourcereleased. The paper is accompanied by a video containing ahigh level introduction and demonstrations of this work.