情報処理学会 第76回全国大会講演要旨

Pedestrian detection is a rapidly evolving research area in computer vision with various applications that potentially impact the quality of daily life. A robust feature descriptor that allows the people being discriminated from the background is always the primary task. Generally, pedestrian detection utilizes the features that are　generated from visual images. However, the appearances of people in the images can easily be affected by clothing, illumination,　background, and therefore have greatly impact on the presentation of the descriptor. In this study, a new feature descriptor, namely, statistical distribution of gradients(SDG), is proposed to apply on thermal images. Compared with other feature descriptors, SDG uses the local gradient distribution information with which the object could be　well described along certain directions. Experimental results demonstrate that the proposed feature descriptor has a good　performance in comparison with two well-known feature descriptors, e.g. histogram of oriented gradients (HOG) and Haar wavelets (HWs).