Laser Range Finder and Image Based Vehicle Classification Dataset

Today it is very important to know the number and type of vehicles on roadways. This information is used to record vehicular traffic data, which is a fundamental for Intelligent Transport Systems applications. Automatic vehicle classification systems (AVC) techniques have been widely considered to identify vehicles as they pass on highways without using any type of electronic pins or chips installed on the vehicle. AVC systems are used in many applications such as automatic tolls, bridge/tunnel clearance verification and road surveillance. This dataset includes range and visual images of different kind of vehicles captured at one of the Universidad del Valle entry points, each range and visual image were labeled with the corresponding class for validation purposes. The sensors used were a SICK LRF LMS200 and a Cannon VC-C50i camera.

The laser SICK LMS200 is connected to a laptop using a RS485/USB interface. For the terms of this project, visual data is also acquired using a Canon VC-C50i analog camera which is connected to a laptop using a RCA/USB converter. In order to test our approach, a dataset was captured on at Universidad del Valle vehicle access entrance at Calle 16. The distribution sensor at this place is shown in the left Figure. The sensor arrangement was set by placing the laser SICK LMS200 vertically in such a way that vehicles were scanned transversely. The LMS200 was placed at 2.75m high in order to get a frontal view of vehicles. The VC-C50i camera was installed at the opposite car lane to obtain a wide field view. As for the VCC50i, it was positioned at 2.82m high. Concerning the LMS200 basic specifications this has a range of 0º to 180º and an angular resolution of 0.25º, 0.5º or 1º while its range precision is ± 15mm. As for the video camera, it used the following settings, NTSC analog images, video output BNC, 3.5mm to 91mm focus and an aperture from f/1.6 to f/4.0. The camera’s video output was acquired by an EasyCap converter getting images of 720 x 450 pixels.The dataset acquisition assumptions included the captured dataset in real world situations comprising different illumination conditions; moreover, scenes displaying pedestrian presence and also more than one vehicle views were inserted. The dataset did not include bad weather conditions due possible damage to the sensors involved. Data acquisitions with occlusions in range images and vehicles in motion at high velocities were discarded. Once the range and visual images were captured, a manual procedure started in order to classify the vehicles between nine different categories.

Camera Calibration - The Canon VC-C50i camera was calibrated using OpenCV libraries. The resulting calibration data is listed as follows:

• • [fx, fy] - [849.26593 ± 2.75870,   879.84977 ±2.84829]

  • [cx, cy] - [299.45146 ±6.59032,   209.08019 ±4.84782]

  • [k1, k2, k3, p1, p2] - [−0.18342 ± 0.03253,   0.50427 ± 0.0407450,   −0.00556 ±0.00165,   0.00080 ±0.0001340,   0]

LRF Calibration - Using "Borenstein, J. (2002). Characterization of a 2D laser scanner for mobile robot obstacle negotiation. Proceedings 2002 IEEE International Conference on Robotics and Automation (Cat. No.02CH37292), 3(May), 2512–2518. doi:10.1109/ROBOT.2002.1013609" the range model of the LRF was found. It is a lineal model with the following parameters:

• • Slope: 0.98453

  • Intersect: 15.233 (mm)

Filenames and data format – The dataset contains 12 different categories of vehicles. Each category holds different range and visual image sequences stored in specific folders. Within each folder three different files are found: the video footage, the range image and the video time-stamps. These directory structure is better explained as follows:

• • Dataset folders - The main dataset folders are:

    • Bus, Bicycles, Camper, Hatchback, Motorcycle,  Pickup, Sedan, Station-wagon, SUV, Trucks, Van, Volkswagen-beatle

  • Sub-folder data - The format to name the sub-folders data was as follows:

    • Data(N) (DD-MM-YYYY_ hh_mm_ss), where N is the acquisition index, DD-MM-YY_hh_mm_ss is the acquisition date and time. NOTE: some sub-folders data has not acquisition date and time.

• Data Files:

  • • data(N).mp4 is the video footage.

    • data(N).dat  is the laser range finder data. This file has the following data format:

      • Sensor_type (0), Scan_angle(180), Measure_mode(1), Sensor_resolution(1), Number_measures(181), time-stamp, Data1, ..., DataN.

    • data(N)Vid.dat is the image data times-tamps.

Files to Download - The data files were separated by categories and they should requested by email to download. These categories correspond with the dataset folders, which are listed below:

• • Bus.

  • Bicycles.

  • Camper.

  • Hatchback.

  • Motorcycle.

  • Pickup.

  • Sedan.

  • Station-wagon.

  • SUV.

  • Trucks.

  • Van.

  • Volkswagen-beatle.

Citing - If you use the LRF and Image based Vehicle Classification dataset in your scientific work, please cite this paper:

@article{Gomez2015,

author = {Gomez, Andres and Hernandez, Pablo and Bladimir, Bacca-Cortes},

journal = {Ingenier\'{\i}a y Competitividad},

number = {1},

pages = {49--61},

title = {{Vehicle Classification Based on a Bag of Visual Words and Range Images Usage}},

url = {http://revistaingenieria.univalle.edu.co/index.php/inycompe/article/view/725/504},

volume = {17},

year = {2015}

}

If you have questions, please send an email to Bladimir Bacca Cortes, Andrés Felipe Gómez o Pablo Hernández.