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

摘要

In this paper, we address semantic segmentation of road-objects from 3D LiDARpoint clouds. In particular, we wish to detect and categorize instances ofinterest, such as cars, pedestrians and cyclists. We formulate this problem asa point- wise classification problem, and propose an end-to-end pipeline calledSqueezeSeg based on convolutional neural networks (CNN): the CNN takes atransformed LiDAR point cloud as input and directly outputs a point-wise labelmap, which is then refined by a conditional random field (CRF) implemented as arecurrent layer. Instance-level labels are then obtained by conventionalclustering algorithms. Our CNN model is trained on LiDAR point clouds from theKITTI dataset, and our point-wise segmentation labels are derived from 3Dbounding boxes from KITTI. To obtain extra training data, we built a LiDARsimulator into Grand Theft Auto V (GTA-V), a popular video game, to synthesizelarge amounts of realistic training data. Our experiments show that SqueezeSegachieves high accuracy with astonishingly fast and stable runtime (8.7 ms perframe), highly desirable for autonomous driving applications. Furthermore,additionally training on synthesized data boosts validation accuracy onreal-world data. Our source code and synthesized data will be open-sourced.