Recursive segmentation based on higher order statistics in thermal imaging pedestrian detection

Automatic pedestrian detection based on thermal imaging is currently performed in two steps. The segmentation step subdivides the image into multiple regions of interest (ROIs) discarding background regions, while the classification step discriminates pedestrians from non pedestrians in each candidate ROI. In this paper a computationally inexpensive new method is proposed for the segmentation step, which recursively subdivides the image into smaller and smaller rectangular ROIs until a candidate pedestrian is identified. ROI boundaries are found on the base of an adaptive threshold updated at each step of the algorithm, while threshold tuning relies on higher order statistics of gray level histograms. Tests performed on OTCBVS database demonstrate significant improvement over a recent literature method in terms of accuracy and efficiency of segmentation.