Description

As a base for training models able to predict exact position of doors' hinges and indirectly the position of the door handles we recorded a dataset of different doors in the wild. Aim was to enable robotic manipulation of doors from arbitrary opening state. While we concentrate on solving the localization problem by means of laser range scan data, we additionally recorded RGB-D data as well.

This dataset conists of raw laser range scans of doors taken with a SICK S300 in a height of 20cm over ground. Each door was recorded from both sides (if possible) at three different recording positions of about 1m away from the left door frame, the middle of the door, and the right door frame. There were doors of 7 different buildings, either office and lecture halls of the university or private homes.

In order to overcome the noise of the laser range scanner SICK S300, we averaged 30 scans which is possible because we are in a steady position. Furthermore, the door was recorded in different opening angles ranging form closed state, over unlatched (~ 0-15 deg), over partly open (~ 15-70 deg), to fully open (~ 90 deg). The Labels of the state have been set by eye.

While recording the raw scans we also marked the ground truth hinge, lock, and corner frames in the range scan coordinates. The figure shows all parameters describing the door’s geometry and position in the world.

Additionally, there was a Kinect Azure camera placed at a height of ~ 1.2m coarsly over the SICK. For each sample a depth image and a color image have been recorded. The transformation of the camera to the laser scanner has been manually annotated by aligning the range scan and the points of the central depth image line (these were used as a virtual range scan).

 

Data Format

There are two versions for download. First is the complete dataset including the images, while the second version only comprises the range scans and therefore is much smaller.

The raw data in the respective .zip files consists of separate files for each sample:

  • rangescan.xml containing the serialized mira::robot::RangeScan object this basicaly describes the geometry of the scan by:
    • a StartAngle, which is the direction of the first laser beam in the scan within LaserFrame coordinates.
    • a DeltaAngle, which is the increment of the angle for subsequent scan rays
    • a ConeAngle, which is the width of an individual beam
    • Furthermore, there are the Range values as a list of distances in meters per scan ray,a Valid flag for each ray (0 - ok, 1-BelowMinimum, 2-AboveMaximum, 3-Invalid), and the Reflectance values provided by the sensor.
  • a transforms.xml containing the relative positions of the camera, the SICK range scanner, and the hinge, corner and lock position (labels) all with respect to a reference frame.
  • a .png file containing the color image
  • a color_intrinsic.xml file describing the actual camera intrinsic parameters of the Kinect color camera
  • a .pgm file containing the depth image
  • a depth_intrinsic.xml file describing the actual camera intrinsic parameters of the Kinect depth sensor
     

The door parameters like opening direction and state are provided in the Dataset.csv file, which contains one row for each sample.

The columns are as follows:

ID running number
color intrinsicfilename of the color_intrinsic file for that sample
depth intrinsicfilename of the belonging depth intrinsic file
color imagethe filename of the belonging color image
depth imagethe filename of the belonging depth image
transforms filefilename of the respective transforms.xml file with all the postition annotations
Door State1- closed, 2- unlatched, 3- partly open, 4- fully open
Opening Direction0- left, 1- right
RangeScan filefilename of the belonging rangescan.xml file
DepthInsidedistance of visible door line to lock hinge baseline [m]
DepthOutsidedistance of visible door line to lock hinge baseline [m]
DoorWidthdistance of hinge and lock [m]
BoardWidthInsidewidth of the door panel on inside [m]
BoardWidthOutsidewidth of the door panel on outside [m]
HingeOffset distance of hinge to the border of the panel on inside
DoorIDinstance id for each door
                         

Citations

If you consider using the data sets on this page, please reference the following:

Müller, St., Müller, T., Aamir, A., Gross, H.-M.
Laser-Based Door Localization for Autonomous Mobile Service Robots.
in: Sensors 23 (2023), no. 11, 5247; https://doi.org/10.3390/s23115247

Download

To get access via FTP, please send the completed form  by email to nikr-datasets-request@tu-ilmenau.de (for research purposes only).