|More specific subject area||Robotics, Object Detection, Automated Guided Vehicle|
|Type of data||Raw depth data provided by the 2D range sensor|
|Processed 2D bitmap-like image representation of raw data|
|How data was acquired||2D laser rangefinder (SICK S3000 Pro CMS)|
|Data format||Files in text format .txt|
|2D images in .jpg .png ( & pixels)|
|MAT-files in MATLAB format .mat|
|Experimental factors||2D depth data processed offline and converted into 2D images.|
|Images have been manually tagged whether they include a pallet|
|or not, and eventually paired with the respective region of interest.|
|Experimental features||Raw data have been acquired by moving a 2D laser scanner|
|in a realistic reproduction of a factory workshop, featuring|
|pallets, people, robots and other equipment.|
|Data source location||EMARO Lab, Department of Informatics, Bioengineering, Robotics|
|and Systems Engineering, University of Genoa, Genoa, Italy|
|Data accessibility||Dataset and codes are archived in a GitHub repository at:|
|Related research article||"Detection, localisation and tracking of pallets using|
|learning techniques and 2D range data" mohamed2018detection|
Value of the Data
The 2D Laser Rangefinder dataset allows to develop novel techiques for pallet detection, localization and tracking.
The 2D Laser Rangefinder dataset can be used as banchmark to compare the accuracy of different pallet detection, localization and tracking algorithms.
The 2D Laser Rangefinder dataset allows to improve Automated Guidance Vehicles in industrial workshop environments.
The 2D Laser Rangefinder dataset can be used to simulate the 2D range sensor data of a mobile robot moving in an industrial workshop environment.
In this article, we present a dataset of 2D range data obtained from a laser scanner moving inside an industrial workshop environment, where EUR standard pallets (see Fig. 3), people, robots and other equipment are present. Each frame of the sensor trajectory corresponds to: (i) a 2D range scan (see Table 1) obtained from a SICK S3000 Pro CMS laser rangefinder (see Fig. 3); (ii) a 2D image obtained by processing the 2D range scan (see Fig. 4); (iii) a tag attached by a human, indicating whether the scan includes a pallet or not; and (iv), the region of interest of the pallet in the image (if any), also defined by a human.
The laser rangefinder has a resolution of 0.25 deg and a maximum field of view of 190 deg, leading to scans made of 761 ranges. It operates at 16 Hz frequency, and the scans are averaged every frames during the static data acquisition phase in order to reduce noise. There are a total of scans, of which contains a pallet, while the remaining do not. The corresponding 2D images are obtained by converting the range data from polar to cartesian coordinates and resizing them to px. Also, images containing a pallet come with a pallet Region Of Interest (ROI), defined by its upper-left and lower-right vertices. Finally, an additional set of continuous trajectories’ raw range data is also made available, to allow online testing.
2 Experimental design, materials, and methods
2.1 Equipment and Software
In our experiment, the data have been acquired using a commercial 2D laser rangefinder from SICK, in particular the model S3000 Pro CMS111https://www.sick.com/ag/en/s3000-professional-cms-sensor-head-with-io-module/s30a-6011db/p/p31284 pictured in Fig. 3. The sensor has a maximum range of 49 m (20 m at reflectivity), a resolution of 0.25 deg, a 16 Hz refresh frequency, and an empirical error of 30 mm. The maximum field of view of the rangefinder is 190 deg, which is sufficient for the detection of objects in front of an eventual AGV. The sensor generates an array of distances in polar coordinates, i.e., each value in the array correspond to the distance to the closest object for every angle in 0.25 deg increments.
The choice of this sensor was due to its widespread adoption in industrial mobile robotics, where it is mostly employed for safety applications and is appreciated for its robustness and precision. It belongs to the class of sensors based on the Time-of-Flight (TOF) principle, i.e., sensors which measure the delay between the emission of a signal and the moment it hits back a receiver in order to estimate the distance to a surface. This category of systems involves sensing devices such as Laser Measurement Systems (LMS) or LIDARs, radars, TOF cameras, and sonar sensors, and they emit either light or sound waves in the environment. Knowing the speed with which the signal propagates and using precise circuitry to measure the exact TOF, the distance can be estimated with high precision.
The laser rangefinder is then connected to a PC through a RS422-USB converter, which has a transmission rate of 500 kBaud. The PC used to acquire the data is equipped with an Intel® Core i5-4210U 1.70 GHz CPU and 6 GB of RAM, and runs Ubuntu 16.04 64 bit.
On the software side, real-world data is acquired online using an ad hoc software222https://github.com/RobotnikAutomation/s3000_laser running in the Robot Operating System framework333http://www.ros.org/about-ros/ (ROS). Offline processing (i.e., conversion to 2D images and manual definition of the regions of interest) has been perfomed in MATLAB. The scripts employed to that purpose and the resulting .mat files are also provided as part of this dataset.
We performed our experiments for data acquisition in the indoor environment represented in Figs. 1-2, with the sensor moving in the 40 m area highlighted in Fig. 1. Such environmennt has been fitted to reproduce a typycal industrial workshop, featuring industrial pallets, furniture, robots and equipment (e.g., a conveyor belt). People were also included in the scene and allowed to move during data acquisition, which lead to temporary occlusions of the objects in the environment. Between acquisition sessions, the position of several objects was modified to better simulate a dynamic environment. The 2D laser rangefinder was positioned close to the floor, in a way that was both realistic with real world mounting position and able to perceive a pallet laying directly on the ground.
Concerning the type of pallet, we focused on the EUR-pallet standard depicted in Fig.3, which is the European pallet format specified by the European Pallet Association (EPAL)444https://en.wikipedia.org/wiki/EUR-pallet. The size of EUR-pallets is 1200 mm800 mm with a height of 144 mm. Moreover, we defined as operating face of the pallet the one of narrower width. On that face there are two slots, each 227.5 mm wide.
In our experiments, the sensor was moved around the environment. Sensor frames differ from each other by the position and orientation of the pallet with respect of the sensor, but also due to the dynamic nature of the environment, as described in the previous section. In particular, it is possible that the pallet is heavily occluded and only few points belonging to it are visible in the frame.
The acquired raw range data at any time instant represent the array of measured distances from the rangefinder to surrounding objects in the environment in the direction given by the angle . More formally:
where is the maximum number of range points acquired per frame, which is related to the sensor’s field of view and angular resolution. In our case, , as the two values are 190 deg and 0.25 deg respectively. Keep in mind that the sensor employed runs at 16 Hz, which would rapidly lead to a unmanageable amount of data, especially considering the manual labelling steps ahead. For this reason, we decided to effectively reduce the operating frequency to 4 Hz in the static data acquisition phase, thus every is actually the result of the average of raw consecutive frames from the sensor. This also helps reducing noise on the data. An example of such process as well and the structure of the raw range data are reported in Table 1.
In our experiments, we are focusing on the detection of pallets in the environment, hence, the set of all raw range data readings , consisting of 2D range scans, has been manually divided into two classes:
Pallet class represents the case of having a pallet located somewhere in the environment with a free operating face, i.e., it can be eventually be picked up by an AGV as an autonomous forklift. It consists of 340 samples.
NoPallet class represents the case in which no pallet is present in the environment, or there is, but the operating face is too cluttered to allow an AGV such as an autonomous forklift to pick up the pallet. It consists of 225 samples.
This manual labeling step has been performed with the help of an online ROS visualization tool, RViz555http://wiki.ros.org/rviz. An operator checked the screen of the PC while the sensor was being moved, marking frames where a pallet with a free operating face was present in the sensor’s FOV.
This results in a binary 2D image of the operating area’s floor plan, which is then resized to px. An example of the resulting images is given in Fig. 4.
Of course, these images are labeled with the same class as the originating frame. In partcular, images belonging to the Pallet class come with the respective pallet ROI expressed as its upper-left and lower-right vertices (i.e., () and ()), as well as a companion px image containing the pallet only. Such ROIs are the results of the Region Proposal Network we employed in the related research paper mohamed2018detection . The resulting ROIs have been manually labelled to indicate whether they present a pallet or another object. A selection of ROIs not including a pallet is also included in the dataset repository.
We will not further delve here into the details of our specific solution to the problem of pallet localization and tracking, which we present instead in the related research paper mohamed2018detection . We just point out that the data was indeed employed for pallet localization and tracking and that the proposed architecture was tested using
additional continous trajectories, which are also made available on the dataset repository. In particular, localization was performed using the aforementioned Region Proposal Network, cascaded with a Faster Recurrent Convolutional Neural Network classifier that took as input the set of manually labelled ROIsren2015faster
. On the other hand, tracking was performed using a Kalman Filtercuevas2005kalman . The filter was also used to implement a Sequantial Classification procedure, i.e., accepting a ROI as an actual pallet was deferred till it was detected and tracked for a predefined amount of time, eventually reaching a sufficient confidence threshold.
Finally, note that the dataset can be used for multi-pallet detection, but that was not part of our data collection experiment. Indeed, in the related research paper mohamed2018detection we ran a preliminary study on the subject by generating artificial data. We want to stress that given that the EUR-Pallet is an official standard with strict tolerances, differences between any two pallets are not perceivable by the the sensor, due to its characteristics and margin of error. This leads to two major consequences:
It is not possible with this sensor and with the dataset we provide to univocally identify a pallet, yet it is possible to distinguish them from each other if appropriate tracking techniques are put in place, like we did in the related research paper mohamed2018detection .
Artificial 2D images including two or more pallets in every image are easy to generate. This can be achieved by taking an original image and adding the pallet ROI from another image, possibly changing position, orientation, and/or adding noise, and consequently deleting any reading in the original image that would now be occluded by the new pallet. Such artificial images are not provided here, but can easily be generated with the provided materials and tools. Nevertheless, future work on our related research paper will include real world multi-pallet testing, thus an extended dataset will be made available too.
|Frame # ()|
|END of Frame #|
Frame # () 0 3.11 1 3.11 2 3.11 ⋮ ⋮ 100 2.26 101 2.28 ⋮ ⋮ ⋮ ⋮ 757 1.51 758 4.08 759 4.06 760 4.08 END of Frame # Frame # () 0 3.13 1 3.11 2 3.13 ⋮ ⋮ 100 2.26 101 2.28 ⋮ ⋮ ⋮ ⋮ 757 1.51 758 4.08 759 4.08 760 4.05 END of Frame # Frame # () 0 3.13 1 3.11 2 3.00 ⋮ ⋮ 100 2.23 101 2.28 ⋮ ⋮ ⋮ ⋮ 757 1.48 758 4.08 759 4.08 760 4.08 END of Frame # Index Range data 0 3.12 1 3.11 2 3.06 ⋮ ⋮ 100 2.252 101 2.28 ⋮ ⋮ ⋮ ⋮ 757 1.50 758 4.07 759 4.075 760 4.07
2.4 Dataset inspection
The dateset is completely contained in the AllData folder of the provided git repository. The folder is structured as follows:
The Class1 and Class2 folders correspond to Pallet and NoPallet classes, respectively. They include raw laser rangefinder scans in .txt format in total, for the former class and for the latter.
DataSet565.mat is a file containing the whole dataset as a MATLAB matrix.
PalletImages folder containing all the px images in various formats, divided by class and eventually accompained by the relative pallet’s ROI. In particular, the files PalletGrayImages.zip and RGBImages.tar.gz contains the images in .jpg and .png format, respectively.
TrajectoryDataset folder contains additional continous trajectories that we used to test the architecture presented in our related research paper mohamed2018detection . The trajectories are provided in .mat format.
The work by I. S. Mohamed was supported by a scholarship from the ERASMUS+ European Master on Advanced Robotics Plus (EMARO+) programme. The authors would like to thank M.Eng. Yusha Kareem for his helping in data collection process.
Conflict of interest
The authors declare that they have no conflict of interest relevant to this article.
Transparency document. Supplementary material
Transparency data associated with this article can be found in the online version at https://github.com/EmaroLab/PDT.
- (1) I. S. Mohamed, A. Capitanelli, F. Mastrogiovanni, S. Rovetta, R. Zaccaria, Detection, localisation and tracking of pallets using machine learning techniques and 2D range data, arXiv preprint arXiv:1803.11254 (2018).
- (2) A. Geiger, P. Lenz, C. Stiller, R. Urtasun, Vision meets robotics: The KITTI dataset, The International Journal of Robotics Research 32 (11) (2013) 1231–1237.
- (3) W. Maddern, G. Pascoe, C. Linegar, P. Newman, 1 year, 1000 km: The Oxford RobotCar dataset, The International Journal of Robotics Research 36 (1) (2017) 3–15.
- (4) J. Jeong, Y. Cho, Y.-S. Shin, H. Roh, A. Kim, Complex urban lidar data set, in: 2018 IEEE International Conference on Robotics and Automation (ICRA), IEEE, 2018, pp. 6344–6351.
- (5) C. Rennie, R. Shome, K. E. Bekris, A. F. De Souza, A dataset for improved rgbd-based object detection and pose estimation for warehouse pick-and-place, IEEE Robotics and Automation Letters 1 (2) (2016) 1179–1185.
- (6) S. Ren, K. He, R. Girshick, J. Sun, Faster R-CNN: Towards real-time object detection with region proposal networks, in: Advances in neural information processing systems, 2015, pp. 91–99.
- (7) E. V. Cuevas, D. Zaldivar, R. Rojas, Kalman filter for vision tracking, Technical Report, Freie Universität Berlin, Inst. Informatik, Berlin, Germany.