To date, pedestrian collision warning methods detect pedestrians without obtaining information about the context of objects in the scene. Lacking contextual information, these warning methods easily confuse objects (e.g., poles, trees, signs) with pedestrians, which results in false detections and inadvertent warnings. Current stereo camera systems can enable rapid and accurate detection of large- and medium-sized objects; however, higher levels of accuracy have not been achieved for pedestrian detection in day/night operating systems. Prior detection methods have focused on developing single-image-based classifiers using two-dimensional (2D) shape/appearance cues for detecting pedestrians. These approaches have high false alarms and add complexity and cost. The researchers developed an in-vehicle stereo vision-only solution that uses stereo cameras with a 50-degree field-of-view that will be able to look 40 meters ahead of the vehicle and detect pedestrians in the vehicle's path with a detection rate of 98 percent; pedestrians out of the vehicle's path, but within 40 meters of the vehicle and moving inward, will be detected with a rate of 90 percent.
To create a real-time vision system that will:
(1) Detect pedestrians or other vulnerable road users at designated crossing locations (e.g., crosswalks and intersections) and midblock or unexpected areas (e.g., between vehicles parked at a road edge, between street corners).
(2) Discriminate between pedestrians and vehicles.
(3) Determine the danger levels of potential collisions.
(4) Provide warning methods to avoid collisions.