A pedestrian detection apparatus of the present invention extracts a whole-body region and a prospective head region based on a pattern matching between an input image picked up by an infrared camera and whole-body/head model images stored in a model-image storage unit. If the whole-body region has not been recognized, the prospective head region which is at the closest position to a prospective head region recognized from a preceding input image is recognized from the current input image, and an image matching with the head model image of the recognized prospective head region is determined to be a head.