We propose a novel procedure for estimating blur in a single image corrupted by blur due to a rigid camera motion during the exposure. Often this blur is approximated as space invariant, even if this assumption holds, for example, only on small image region in perspective images captured during camera movement. Our algorithm analyzes separately selected image regions containing a corner and in each region the blur is described by its direction and extent. The algorithm works directly in space domain, exploiting gradient vectors at pixels belonging to the blurred corner edges. The algorithm has been successfully tested both on synthetic and real images showing good performance even on small image regions and in presence of noise.