Incident Detection from Low-angle Images of Heavy Traffics in Tunnels

Accidents or abnormally stalled vehicles in tunnels are liable to induce additional incidents that would be more fatal. They also would induce heavy traffic congestions by disturbing the following traffics. Therefore, it is important to detect such the primary incidents in tunnels as soon as possible, and to inform traffic management officers about them. However, it is difficult to detect incidents correctly distinguishing from pure congestions. In particular, it will become more difficult to detect incidents from low-angled and seriously occluded images as in tunnels. In this paper, a dedicated method for precise segmentation of such the occluded vehicles is described. The proposed algorithm was examined by experiments using two year video images obtained from three tunnels, and it was proved to be effective for quite ill conditions such as heavy traffics in tunnels.