Autonomous Space Robotics Lab

Canadian Planetary Emulation
Terrain 3D Mapping Dataset

a100_dome_vo

[Home] [Overview] [Hardware] [Details] [Downloads] [References]

This dataset consists of 50 laser scans obtained using a Clearpath Husky A100 at the University of Toronto Institute for Aerospace Studies (UTIAS) indoor rover test facility, located in Toronto, Ontario, Canada. In addition to the laser scans, this dataset also contains pose-to-pose transformation estimates between consecutive scan stops based on stereo visual odometry. This dataset was collected November 2010.

The UTIAS indoor rover test facility consists of a large dome structure, which covers a circular workspace area 40m in diameter. In this workspace, gravel was distributed to emulate scaled planetary hills and ridges, providing characteristic natural, unstructured terrain. Four large retroreflective sheets were placed outside the workspace to serve as easily-identifiable landmarks for ground truth localization.

The Clearpath Husky A100 used to gather the dataset carried a number of payloads. The relevant payloads for this dataset consisted of a vertically-scanning SICK LMS291-S05 laser rangefinder mounted on a Schunk PW70 pan-tilt unit to produce the 3D scans, a Honeywell HMR3000 inclinometer to measure rover inclination, and a Point Grey Bumblebee XB3 stereo camera for visual odometry.

The laser scanner was configured for a vertical angular resolution of 0.5°, and the scans were constructed by linearly interpolating the pan positions of the laser slices through the slow 360° scan sweep. Based on the manufacturer's specifications, the SICK LMS291-S05 has a range measurement standard deviation of 0.01m.

As depicted in the figure to the right, the scan reference frame was defined to be the laser rangefinder center, with +x pointing forward, +y to the left, and +z up. The inclinometer sat on the rover top plate, and the stereo visual odometry estimates were transformed into the laser sensor frame.

The 50 scans composing for this dataset were obtained with a larger inter-scan spacing. As a result, this dataset may be challenging for dense matching approaches. However, visual odometry estimates are also provided to assist in the alignment process. The covariances of the visual odometry estimates were determined through calibration experiments conducted in an indoor lab prior to the collection of this dataset. The estimated motions were compared to Vicon motion capture data, providing statistical characterization of the estimate accuracy. A number of loops were traversed around the terrain, providing many loop closure opportunities. Since the terrain lay in the middle of an artificial dome structure, the scans were trimmed to remove the points that fell outside the workspace. The following image depicts an overhead view of the terrain, along with the scan locations and odometry measurements.

The ground truth sensor poses were provided by the identification of known landmarks outside the workspace. These consisted of four large retroreflective sheets mounted on the dome structure, which were utilized in a SLAM formulation to determine the six-degree-of-freedom sensor poses at each scan stop. Previous benchmarking experiments involving the same configuration demonstrated centimetre-level accuracy in translation, and half-degree accuracy in orientation [1]. By utilizing the same laser sensor for estimation as well as ground truth, the need for calibration to determine the ground truth pose offset was avoided. The ground truth reference frame was defined to be the same as the first scan pose.

The a100_dome_vo dataset is available at ftp://asrl3.utias.utoronto.ca/3dmap_dataset/a100_dome_vo/.

[Download entire a100_dome_vo dataset (zip) | Compressed size: 104MB, Extracted size: 433MB]

[1] Tong C and Barfoot T D. "A Self-Calibrating Ground-Truth Localization System Using Retroreflective Landmarks." In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 3601-3606. Shanghai, China, 9-13 May 2011.