Identifying hazardous locations at intersections by automatic traffic surveillance

Kurzfassung

Due to the trend that more and more people are using bicycles, especially in urban areas, the focus is to put on their safety. Except for accident data there is almost no knowledge about non-accidental conflicts possibly indicating hazard potential. There is a lack of data which would provide insights in the safety-related processes occurring in the absence of accidents. In order to gain more detailed and precise data than the available accident data from public authorities, we conducted a measuring campaign to validate our methods and to have possible proof-of-concept. The measuring campaign surveilling the traffic for 12 hours continuously from 6 a.m. till 6 p.m. includes the peaks of traffic volume at rush-hours. The campaign was intentionally conducted at a roundabout being a hot-spot for accidents with involvement of cyclists in Berlin, Germany. Through our video surveillance system we are able to extract automatically trajectories from road users at a very high chronological sampling rate. Thus, it is possible not just to detect accidents, but also conflicts and locate them precisely. Additionally, pre-conflict and pre-accident situations can be analyzed and provide further in-depth understanding. We use state-of-the-art computer vision technology to extract trajectories and to classify road users. The evaluation of the measuring campaign comprises the investigation of traffic parameters, as well as traffic-safety related parameters based on common metrics like TTC, Delta v, locations of strong braking operations etc. Our goal is to identify locations and times with high-conflict potential, as well as identifying the conflict types and types of road users involved. Furthermore, we can test the performance of our system by validating all acquired data with the groundtruth, i.e. quantize object detection performance etc. This enables us to proof the validness of the system referring the automatic surveillance of traffic at intersections. In the future, we plan to analyze the relation of conflicts and accidents by long-term surveillance in order to identify dangerous situations before accidents happen.