To provide a technology which quickly detects an obstacle on a crosswalk even when position and direction of an image pattern of the crosswalk in an image taken by a vehicle-mounted camera are not coincident with position and direction of a reference pattern representing the image pattern of the crosswalk on which no obstacle exists.

When the position and direction of the image pattern of the crosswalk in the image taken by a monocular camera 3 are not coincident with the position and direction of the reference pattern, stored in a memory means 4, of the crosswalk on which no pedestrian exists, conversion processing is performed for a power spectrum by a conversion means 5a so that the position and direction of the image pattern of the crosswalk in the image taken by the monocular camera 3 become coincident with them of the reference pattern. Consequently, the pedestrian 11 on the crosswalk taken by the monocular camera 3 is quickly detected by a detection means 5b based on a difference between the image pattern of the crosswalk including the reference pattern of the converted power spectrum and the reference pattern.