This page introduces two new datasets which can be employed on to test pedestrian detection and tracking algorithms on real world scenes.
The position of the camera is fixed, recording the vehicular traffic and pedestrians crossing the street.
The original size of the color images is 1600 x 1200 pixels, and we cropped the captures to a size of 1319 x 339 pixels in order to analyze the zebra cross.
The following figures show several examples of this dataset.
The objective of the dataset is the study of pedestrian behavior on street corners and the interaction between vehicles and persons.
It can be assumed the state of the pedestrian red light, because is visible the signal of the perpendicular zebra crossing.
As an example, the second figure (GS06 frame 735) of the GS06 sequence shows one pedestrian crossing the street with red light, at the same time that vehicles are approaching at high speed.
The dataset is composed of two sequences captured on the same day, with a delay of almost two hours.
Each sequences has 1,000 images, within them, there are 27 pedestrians for the GS06 set, and 51 pedestrians for GS54 set.
The video frame rate is in the order of 7-8 fps, and there are some glitches of one or two frames.
The average displacement of a person walking in this sequence is 15 pixels between to captures; however, it may increase to more than 25 pixels due to those gaps.
Each pedestrian on the sequence is identified with a rectangle and a unique label in order to test detection and tracking systems.
This test sequence has a high complexity.
First, pedestrians are not the only objects in motion in the image, existing vehicles and bicycles along the street.
There also are many occlusions between pedestrians and between pedestrians and vehicles.
Moreover, as can be appreciated in third and fourth figure, separated by 56 frames (7 seconds), lighting conditions are not controlled and can vary significantly.
Finally, there are chromatic aberrations, and a over-sharpening which saturates color transitions.
Datasets
Both datasets were zipped and splited in several files to long downloads from the ftp site.
Zipped files are not encrypted.
GS06 dataset:
GS54 dataset:
- GS54.part1.rar (104,857,600 bytes, main encripted zip file),
- GS54.part2.rar (104,857,600 bytes),
- GS54.part3.rar (104,857,600 bytes),
- GS54.part4.rar (61,154,253 bytes).
GT_GS_06.dat (87Kb), and
GT_GS_54.dat (178K): The hand labelled ground truth data with the following format:
frameNumber, personID, bodyLeft, bodyTop, bodyWidth, bodyHeight
frameNumber - The frame number (counted from 0)
personID - A unique identifier for the individual person
bodyLeft,bodyTop,bodyWidth,bodyHeight - The body bounding box in pixels
Homography_GS_dataset.dat (1K), is the homography matrix to project image points to real world coordinates.
As the matrix were computed for the original image, in order to project the points in the cropped file, you must add 48 to the X position, and 272 to the Y position.
The following
MATLAB code sample draws the street floor of the view in the real world coordinates, as shown this figure:
These datasets can be referred as "GS06-PANKit" and "GS54-PANKit".
Any commercial use of the datasets is strictly forbidden.
The methodology for automatically detect the trajectories are detailed in [DYNA2017].
Please give appropriate acknowledgements when using these databases and quote
the following paper :
[DYNA2017]
Pablo Negri, Damian Garayalde
"Pedestrian Tracking Using Probability Fields and a Movement Feature Space"
, DYNA Colombia, Vol. 84 (200), pages 217-227, January 2017
[
pdf].
The dataset GS06 and GS05 where recorded under license Nro. DI-2012-1554-DGINC of the 24/08/2012 of the City of Buenos Aires.
This work was supported by PICT-2283 of ANPCyT, CONICET (Argentina), and the ACyT A14T13 of UADE.
