Abstract: Thermal infrared images are suitable for pedestrian detection in low illumination such as at night. The grayscale distribution of background in an infrared image vehicle-mounted varies obviously and pedestrian is easily confused with background interference and it is difficult to catch a pedestrian in the distance. The system recall rate and false alarm rate requirements can’t be achieved through extracting RoI with double threshold segmentation method. We construct a new RoI extraction method consisting of image preprocessing， RoI generation and RoI post processing etc. An expanding maximum filter is designed to enhance image contrast. Adaptive double threshold segmentation is improved by Haar-like feature. Computing efficiency is raised by designing incremental model. Besides， filters considering gray-scale temporal feature and spatial symmetry feature of a pedestrian are presented to remove false RoI. Comparing with the benchmark method， our method improves recall rate by 49% when the number of false RoIs is less than 40 per frame. And RoI extraction speed isn’t lower than 18 frames per second.

Key words: RoI extraction, expansion maximum filter, Haar-like feature, double threshold segmentation, vehicular infrared pedestrian detection