Ostatnia modyfikacja podstrony: 18.07.2018 12:40

Pedestrian detection in infrared images



This page contains a description of the research project conducted in thDivision of Signal Processing and Electronic Systems regarding the detection of pedestrians in thermal and in the near infrared images. The scope of the project apply to Advanced driver-assistance systems (ADAS)and Intelligent Transportation Systems (ITS) research fields.

Currently, we are carrying out research on optimizing the classifiers resolutions for real-time applications and we propose a "performance index" to assess the performance of the classifier.


Pedestrian Databases:
To experimentally verify certain assumptions, we perform experiments on various publicly available databases. For this purpose, we create appropriate training and testing subsets to optimize the final result. In this part, we provide the direct links to our prepared training/testing subsets and present the original databases used, along with the references to original sources of those databases.
NTPD (Night-time Pedestrian Dataset)
The NTPD consists of images of pedestrians stored by the NIR active system with resolution of 64×128 pixels and is divided originally into two sub-sets: training and testing. 
Positive training subset: 1998
Negative training subet: 8730 
Positive testing subset: 2370
Negative testing subset: 12600
The negative samples was prepared based on original full resolution images, Direct links to negative subsets: part1, part2, part3.
Y. Zhang, Y. Zhao, G. Li, R. Cheng, “Grey Self-similarity Feature for Night-time Pedestrian Detection,” Journal of Computational Information System 10: 7, 2014, pp. 2967 – 2974.
OSU (Ohio State University) Thermal Pedestrian Dataset
The OSU (Ohio State University) Thermal Pedestrian Dataset contains 10 video sequences of size 320´240 pixels taken at the university campus walkway intersection with a street.
Positive training subset: 1004
Negative training subet: 1932
Positive testing subset: 964
Negative testing subset: 1932
To obtain negative samples, frames with no pedestrians were cut with a window size of 32x64 with spacing by 8 pixels. Moreover, with the use of rotation and vertical or horizontal mirroring the 3864 of negative samples was produced. Direct link to negative subset.
J. Davis, M. Keck, “A two-stage approach to person detection in thermal imagery,” Proc. of Workshop on Applications of Computer Vision, Breckenridge, Colorado, USA,  vol. 1, pp. 364 – 369.
CVC-09 (Computer Vision Center, FIR Sequence Pedestrian Dataset)
The CVC-09 dataset is divided into two subsets: one recorded in day and the second at night recorded with thermal camera.
Night-time subset of CVC-09 dataset
Positive training subset: 6998
Negative training subet: 30030
Positive testing subset: 7862
Negative testing subset: 72985
Day-time subset of CVC-09 dataset
Positive training subset: 11839
Negative training subet: 25410
Positive testing subset: 6711
Negative testing subset: 75398
We prepared positive datasets of samples by cutting pedestrians from original images based on annotations. Negative samples were prepared by cutting out areas, which do not contain pedestrians. They were extracted by a window of size of the largest used classifier (i.e. 64´128 pixels). 
Y. Socarras, et al. “Adapting Pedestrian Detection from Synthetic to Far Infrared Images,” ICCV – Workshop on Visual Domain Adaptation and Dataset Bias Australia, 2013.

Please note that the original databases: LSIFIR, OSU, NTPD, CVC-09 are not our private property. We share only training or testing subsets prepared by us, which are based on the original databases.


Karol Piniarski, MSc karol.piniarski@put.poznan.pl

Pawel Pawlowski, Ph. D pawel.pawlowski@put.poznan.pl

Adam Dabrowski, Professor, Ph. D adam.dabrowski@put.poznan.pl