Hard Point Cloud Localization Dataset

This is a dataset to evaluate the robustness of point-cloud-based localization algorithms in extremely severe situations including aggressive sensor motion, point cloud degeneration, and data interruptions.

The dataset contains five indoor sequences recorded with a Microsoft Azure Kinect and three outdoor sequences recorded with a Livox MID360. Each outdoor sequence is split into two consecutive rosbags.

indoor_easy_01 & 02    : "Easy" sequences without aggressive sensor motion and data interruptions.
indoor_hard_01             : "Hard" sequence involving quick sensor motion and point cloud degeneration.
indoor_kidnap_01 & 02 : Sequences that involve kidnapping situations (the sensor was moved to another room while its view was completely occluded several times)
outdoor_hard_01 & 02  : "Hard" outdoor sequence involving quick sensor motion and point cloud degeneration.
outdoor_kidnap             : Outdoor sequence that involve kidnapping situations.

Each rosbag corresponds to an environmental map file as follows:
indoor_easy       : map_indoor_easy.ply
indoor_kidnap    : map_indoor_easy.ply
indoor_hard        : map_indoor_hard.ply
outdoor_kidnap  : map_outdoor_kidnap.ply
outdoor_hard     : map_outdoor_hard.ply

Each indoor sequence contains the following ROS2 messages:
- /points2/decompressed : sensor_msgs/msg/PointCloud2 : Point cloud data
- /imu                                : sensor_msgs/msg/Imu              : Imu data

Each outdoor sequence contains the following ROS2 messages:
- /livox/points           : sensor_msgs/msg/PointCloud2       : Point cloud data
- /livox/lidar              : livox_ros_driver2/msg/CustomMsg : Point cloud data in the Livox format
- /livox/imu               : sensor_msgs/msg/Imu                    : Imu data

We used the follwoing LiDAR-IMU transformation parameters (T_lidar_imu : [tx, ty, tz, qx, qy, qz, qw]) for the indoor and outdoor sequences:
- indoor  : [0.003, 0.004, -0.051, -0.476, 0.474, 0.524, 0.525]
- outdoor: [0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 1.000]

Groundtruth trajectories were obtained through batch optimization of point cloud registration residuals and IMU motion residuals. Each line in a GT trajectory file represents the LiDAR pose in the map frame [time, tx, ty, tz, qx, qy, qz, qw] (TUM format). We recommend evo toolkit (https://github.com/MichaelGrupp/evo) for quantitative evaluation.