Human Detection by Using Centrist Features for Thermal Images

In this paper, we present a new human detection scheme for thermal images by using CENsus TRansform hISTogram (CENTRIST) features and Support Vector Machines (SVMs). Human detection in a thermal image is a difficult task due to low image resolution, thermal noising, lack of color, and poor texture information. For thermal images, contour is one of the most useful and discriminative information, so capturing it efficiently is important. Histogram of Oriented Gradient (HOG) is still the most proven way to capture the human contour. CENTRIST is a computationally efficient technique to capture contour cues as compared to HOG, but so far no one has implemented and tested the accuracy of CENTRIST descriptor for infrared thermal images. We developed CENTRIST based human detection system for thermal images and tested its variants. We also made a new dataset of thermal images, since there was no realistic dataset. Experimental results show that CENTRIST exhibits better detection accuracy than HOG, while reducing the training and the testing time significantly.