A Survey of 3D Point Cloud and Deep Learning-Based Approaches for Scene Understanding in Autonomous Driving

Supported by the advancement of deep learning (DL) techniques and a massive procession of sensor technology, feature learning from 3D lidar data has led to rapid development in the field of autonomous driving. Progress in sensor technologies has led to the increased availability of 3D scanners, such as lidar, which are wildly used for a more accurate representation of a vehicle's surroundings. This article aims to provide a comprehensive survey of 3D point cloud and DL-based methods for scene understanding in autonomous driving, which is mainly divided into two subtasks: object detection and semantic segmentation. For each of these, we review existing research works according to point cloud representation methods, including pure point cloud, projective 2D views, voxel grids, and multimodal data fusion. Finally, we summarize the review work and provide a discussion of future challenges of the research domain. © 2009-2012 IEEE.