Fig. 2
From: Scene as Occupancy and Reconstruction: A Comprehensive Dataset for Unstructured Scene Understanding
Data Overview. (a) Hardware system that includes a monocular camera, a Light Detection and Ranging (LiDAR), an Inertial Measurement Unit (IMU) and a Real-Time Kinematic (GNSS) system. (b) Scene Reconstruction, which is characterized by numerous irregularly shaped obstacles and an uneven road surface. Semantic classes are predicted from camera images with the assistance of SAM for annotation. (c) A dense point cloud map is generated from single-frame point clouds using LiDAR odometry. (d) Semantic images are then projected onto the dense point cloud map, creating a dense semantic point cloud. Voxelization is performed on this map to obtain 3D semantic occupancy prediction labels, which distinguish drivable and non-drivable areas and provide guidance for trajectory planning. (e) A rule-based aggregation method converts the dense point cloud map into a Road Elevation Map, where road undulations affect speed planning.
