Multi-camera system for traffic light detection: About camera setup and mapping of detections

Traffic light recognition is an object detection task mainly investigated on monocular images from a single camera [1], [2], [3], [4], [5] usually one with a small field of view. They guarantee high object resolution and less distortion. Both properties support common state of the art detectors. However, a single-camera setup only covers a certain distance range. Traffic lights in close range disappear from the field of view. When stopped at the stop line of an intersection, the vehicle can-not recognize the toggle process of the traffic light. In this paper a two-camera setup for traffic light recognition is presented. A stereo camera is used for far and moderate range, while for close range a wide-angle camera is used. The fields of view of both cameras overlap. In particular, the transition of a detection in the stereo camera to the wide-angle camera is presented. Objects detected by the stereo camera are handed over to the wide-angle camera. Since the hand-over is erroneous, a template matching correction presented. The accuracy of the hand-over as well as a subsequent correction method is evaluated.