Zagreb dataset


The dataset is produced within the Safelog project and is used for the evaluation of the developed algorithm which estimates the location of humans in the warehouse environments.

The localization is based on the fusion of information from stereo visual odometry and distances to the known absolute poses of detected ground-markers. The ground-markers are part of the warehouse environment since they are used for the localization of robots:

<img src="./marker.png" alt="markers" width="200"/>

The dataset contains:


./calibration/   -> recordings used for obtaining calibration parameters
./recordings/    -> sequences used for the evaluation


The Zagreb dataset contains 10 recordings made in the library of the Faculty for Electrical Engineering and Computing, University of Zagreb. In those recordings, we simulated the warehouse environment and recorded different movements that are expected from a human working in a warehouse. The recordings are in .bag format and contain the following topics:


/imu                                 -> IMU measurements
/stereo/image_raw                    -> Stereo pair concatenated in one image, left on top of the right image
/stereo/left/cam_info                -> Parameters of the left stereo pair
/stereo/right/cam_info               -> Parameters of the right stereo pair
/stickers_camera/image_raw           -> Monocular camera image
/stickers_camera/camera_info         -> Monocular camera parameters
/vrpn_client_node/RigidBody4/pose    -> Ground truth pose

The ground truth is obtained with the Optitrack motion capture system.

The images of the testing area and the camera setup are provided in the following images:

<img src="./library1.jpeg" alt="Library1" width="600"/> <img src="./library2.jpeg" alt="Library2" width="600"/> <img src="./library3.jpeg" alt="Library3" width="600"/> <img src="./camera_setup.png" alt="cam_setup" width="400"/>

Sensor setup

The camera setup contains PercepIn's stereo visual-inertial module Ironsides and the monocular camera FLIR Chameleon3 CM3-U3-50S5M-CS. Both the stereo visual-inertial module and the monocular camera are places on the thick aluminum plate which ensures fixed displacement between them.

Camera parameters computed with the Kalibr calibration package

Intrinsic parameters

Left Ironsides camera

 distortion: [-0.364649802468 0.110798178523 -0.002653213578 0.000506244967]
 projection: [436.789112231987 436.536353093514 308.692657193760 229.732604248368]

Right Ironsides camera

 distortion: [-0.354534668262 0.095241521353 -0.003119334324 0.001659493109]
 projection: [423.543206098086 423.214497125122 312.964843718074 236.220888505556]

Mono camera

 distortion: [-0.029214787660 0.247708294212 -0.003649948222 0.006151333995]
 projection: [3498.024168340175 3495.688197364558 1310.502165153987 1015.363709961205]

Extrinsic parameters

Left Ironsides to Right Ironsides

[ 0.99997561610 -0.00251801738 -0.00651358440 -0.06496088026
  0.00247302164  0.99997308949 -0.00690684022 -0.00028908967
  0.00653080066  0.00689056357  0.99995493337 -0.00020397522
  0.0            0.0            0.0            1.0]

Left Ironsides to Mono camera:

[ 0.99971779360 -0.01133198394 -0.02087868008 -0.03061771524
 -0.00435672144  0.77652494541 -0.63007144684 -0.05015109771
  0.02335277543  0.62998459925  0.77625643481 -0.09631863368
  0.0            0.0            0.0             1.0          ]


The calibration files contain an additional camera (Realsense) which is used as a bypass between stereo camera and monocular camera, since they do not share a large area of common field of view.

Calibration parameters provided by manufacturer

The calibration parameters for PIRVS OldIronsides stereo-inertial sensor are provided in calib_PerceptIn_V1_2162.json.