ETHZ ASL

Logo

Datasets and open-source code from the Autonomous Systems Lab (ASL) at ETH Zurich, led by Prof. Roland Siegwart.

Urban Dynamic Objects LiDAR Dataset

Year: 2021
DOI: 10.3929/ethz-c-000788171

Urban Dynamic Objects LiDAR Dataset

Hauptgebaeude Niederdorf Shopville Station

Data Access

➡️ ETH Research Collection (landing page & downloads): https://doi.org/10.3929/ethz-c-000788171

Dataset Description

The dataset contains more than 12,000 scans recorded in:

  • The main hall of ETH Zurich (Hauptgebaeude)
  • Two levels of Zurich main train station (Station, Shopville)
  • A touristic pedestrian zone (Niederdorf)

Each location includes 2 sequences. The point clouds include a significant number of pedestrians.

Data was recorded using a handheld Ouster OS1 64 (Gen 1) LiDAR at 10 Hz with 2048 points per revolution.

We provide globally bundle-adjusted high-frequency poses from a VI-SLAM pipeline. Trajectory lengths range from 100–400 m, and sequence durations are 100–200 s.

Annotations

A subset of the dataset was manually annotated:

  • 10 temporally separated point clouds per sequence (8 sequences total)
  • Pedestrians and objects associated with them (suitcases, bicycles, dogs, etc.)
  • Annotation was based on appearance only (not motion state)

Simulated Dataset

We additionally provide a simulated sequence of a small town featuring:

  • Moving cars, planes, pedestrians, animals
  • Cylinders, spheres, cubes of varying sizes
  • Objects moving horizontally and vertically at different velocities
  • A simulated LiDAR (similar to OS1 64 Gen 1) moving along a closed trajectory → Ground truth poses and object annotations available for all point clouds

Structure of the Dataset

Sequences are provided as rosbag files.

Topics

Description Topic Message Type
Point Clouds /os1_cloud_node/points sensor_msgs/PointCloud2
Transformations /tf tf2_msgs/TFMessage
Transformations /T_map_os1_lidar or /T_map_os1_sensor geometry_msgs/TransformStamped

The world coordinate frame is map. Point clouds are recorded in:

  • os1_lidar frame (real datasets)
  • os1_sensor frame (simulated dataset)

Corresponding poses are provided via /tf or /T_map_os1_lidar(sensor).

Distortion & Undistortion

Because the sensor was handheld, point clouds are distorted by ego motion. The provided high-frequency poses can be used for undistortion.

Dynamic Object Indices

For each sequence, annotated dynamic objects are stored in an indices.csv file:

  • Each line begins with the timestamp (ns) of the point cloud
  • Followed by all annotated dynamic point indices

Reference Paper

Patrick Pfreundschuh, Hubertus F.C. Hendrikx, Victor Reijgwart, Renaud Dubé, Roland Siegwart, Andrei Cramariuc, Dynamic Object Aware LiDAR SLAM based on Automatic Generation of Training Data, IEEE International Conference on Robotics and Automation (ICRA), 2021.

@inproceedings{pfreundschuh2021dynamic,
  title={Dynamic object aware lidar slam based on automatic generation of training data},
  author={Pfreundschuh, Patrick and Hendrikx, Hubertus FC and Reijgwart, Victor and Dub{\'e}, Renaud and Siegwart, Roland and Cramariuc, Andrei},
  booktitle={2021 IEEE International Conference on Robotics and Automation (ICRA)},
  pages={11641--11647},
  year={2021},
  organization={IEEE}
}

Tags: mapping, lidar, dynamic, pedestrian, slam, urban, outdoor