Infrared pedestrian detection utilizing entropy-edge weighted local gradient orientation descriptor

Detecting infrared pedestrian in outdoor smart video surveillance is always a challenging and difficult problem. Although there have been many methods based on histograms of oriented gradients (HOG) to solve this problem, they would probably fail because of shelter and poor quality of image. To overcome this problem, we propose a robust feature to describe pedestrian which is called entropy-edge weighted local gradient orientation (EEWLGO) descriptor. This feature firstly extracts the “orientation image” to depict pedestrian. Then “pixels” of “orientation image” is reshaped to a vector and it is combined with edge histogram to generate the final proposed EEWLGO descriptor. The descriptor outperforms other methods in not only some kinds of shelters but also robustness to noisy clutters. What's more, the processing time is also approximately identical to others, which fulfils the general real time property of surveillance. Cross validation and test on other datasets demonstrate the high accuracy and good robustness of our algorithm.