A Local Fuzziness-based Active Contour Model for Infrared Human Segmentation
Due to the poor quality of thermal infrared imagery and the great variation of human subjects, segmentation of human subjects is difficult in the thermal infrared images acquired for human detection. In this paper, a local fuzziness-based active contour model is proposed to address this problem. In comparison with such methods as thresholding, it is advantageous to provide more accurate segmentation precision and the enclosed human silhouettes that benefit much incoming human recognition. This model consists of the components named as energy functional and numeric scheme. The energy functional, which decides how a contour evolves over the infrared image domain, is designated from the fuzzy information within local image regions, so that it acquires the ability to differentiate the fuzzy pixels that have similar intensities but actually belong to different regions. The numeric scheme, which evolves the active contour on the image domain by minimizing the energy functional, is carefully formulated via the fusion of the techniques named as direct minimization and narrow band calculation. Solved by this scheme, the minimization of the energy functional is able to reach fast convergence, which ensures the real-time performance of the proposed model. The model is tested on a series of challengeable images that are selected from OSU thermal pedestrian database and the ones acquired by our FLIR A40 thermal infrared camera with the rivals. These images are of different image quality and appearances of human subjects. The results validate its advantages in segmentation accuracy and efficiency, which lead to the adaptation to real-time human detection applications.