A generalized thresholding algorithm of pedestrian segmentation for far-infrared images

Abstract

Designing a robust and efficient thresholding algorithm for far-infrared (FIR) images under various imaging conditions is one of critical technologies. The existing algorithms are difficult to deal with the images corrupted by noise, if a predefined filter is not used. However, it is difficult to define an appropriate filter beforehand because some prior knowledge about image noise is required. To solve this problem, an improved fast generalized fuzzy c-means (IFGFCM) is proposed to reconstruct a filtered image first regardless of the type of image noise. A novel adaptive thresholding algorithm combining IFGFCM with clustering centers analysis is then used to segment pedestrians from FIR images automatically. Experiments performed on a set of FIR images show that, compared with three other algorithms, the segmentation effectiveness of the thresholding algorithm is more consistent with the ground truth, and the resulting misclassification rate is less than 2%.