Self-Collected LiDAR Dataset for Semantic Probabilistic Voxel Mapping
Description
This dataset contains self-collected LiDAR point clouds and synchronized RGB images, captured in diverse indoor office environments using portable Livox AVIA and MID360 LiDAR platforms. The dataset includes semantically annotated 3D point clouds, covering planar surfaces, thin structures, and non-planar cluttered regions. The data supports research in semantic-aware probabilistic voxel mapping, residual routing, and LiDAR-based localization under geometrically degraded or narrow-field-of-view conditions. All sequences include start and end ground-truth alignment using AprilTags for drift evaluation, while continuous full-trajectory ground truth is not available. The dataset is intended for evaluating and benchmarking semantic-enhanced voxel-based localization and mapping algorithms.
Files
Steps to reproduce
The dataset is provided in two formats. First, a ROS bag version includes LiDAR scans (`livox_ros_driver2/CustomMsg`), synchronized RGB images (`sensor_msgs/CompressedImage`), and IMU measurements (`sensor_msgs/Imu`) for end-to-end replication of the pipeline. Second, an extracted offline version contains preprocessed files: `points/` (raw and segmented LiDAR point clouds with timestamps), `color/` (RGB images), and `intrinsic/` (camera intrinsics, extrinsics, distortion coefficients, and sensor shapes). Both versions include time-stamped and calibrated data to allow reproducible experiments for semantic-aware probabilistic voxel mapping, residual routing, and LiDAR localization under degraded indoor conditions.