Using context to improve cascaded pedestrian detection

The design of a fast and accurate pedestrian detector is considered. A system combining a fast cascaded pedestrian detector and a pedestrian validator is proposed. The detector first scans the image of interest and proposes a set of candidate bounding boxes. The pedestrian validator then decides if each proposed bounding box is consistent with a true pedestrian, based on scene context. Experiments show that the resulting system is faster and more accurate than current approaches to pedestrian detection.