Pedestrian detection with a vehicle camera using fast template matching based on background elimination and active search

Abstract

Template matching using the silhouette shapes of pedestrians as features is widely used as a method for detecting pedestrians in images taken from an in-vehicle camera. Because of the need to detect a variety of pedestrian shapes, the template matching involves scanning the entire image for a very large number of templates. A large number of computations are needed to utilize multiple templates, and so it is difficult to implement the processing at the video frame rate. Furthermore, because images of outdoor street scenes that are captured by an in-vehicle camera consist of problematic backgrounds, false detections occur easily in the background area. A method is therefore proposed to perform highly efficient, high-speed matching by removing the background region, as this region has a negative impact on the detection capabilities in template matching. The background removal is achieved by using difference calculations that use information about the motion of the observation measured by sensors. Furthermore, an active search is carried out on the images after removing the background to give an efficient search, and a broad reduction in the number of computations is achieved. © 2007 Wiley Periodicals, Inc.