Obstacle detection on a track from a quiescent image of a tram’s front view

Abstract

Tram systems are being re-evaluated worldwide, in order to provide a barrierfree, low-building cost transport system. New systems are opening and the existing lines expanding. However, there are many anxieties about the occurrence of traffic accidents with tram vehicles. The safety of the tram vehicle is still solely dependent upon the observation of the driver. However, ITS technology for road transport has advanced enough to assist with this issue. Therefore, we construct a safe support system for tram drivers, to detect obstacles using front view images of trams by image processing. In this paper, we propose a method to detect obstacles on or around tracks from the quiescent images. There are private or common tracks for tram systems; the more dangerous objects are on the common tracks, because the tram vehicle runs on its own track only and cannot avoid the obstacles. A tram vehicle can only run on tracks. This fact means that an exact path in front of the tram is known. So once the track is detected, the safety of trams is maintained by observation around the tracks. On the images, head on tracks are vivid and the rut is dark, the tracks are detected by edge extraction in image processing. If an object's shadow exists on the tracks, the object is detected as an obstacle. If an obstacle has height, the object projects a wide area on a bird's view map, but in the case where tracks have no height, they have a small area on the map; the obstacles are judged by these characteristics. This method can detect both moving and quiescent obstacles, because this process uses a quiescent image. We show some good results by the presented process for a 616 real tram's quiescent front view images.

Railroad Crossing Surveillance and Foreground Extraction Network: Weakly Supervised Artificial-Intelligence Approach

“…In recent decades, classical image-processing algorithms have been widely used to detect various railway-related objects. Šilar and Dobrovolny and Ukai et al used optical flow for railway monitoring, but the optical flow was unable to detect static objects ( 13 , 14 ). Gajbhiye et al proposed a five-frame variance foreground-segmentation model; they only used a single image as the background reference and did not have any background update scheme ( 15 ).…”