Risk Estimation of Critical and Non-Critical Interactions between Right-turning Motorists and Crossing Cyclists by a Decision Tree

Background During the past years, cycling has become an alternative to motorized traffic and this trend is expected to increase. While the number of severe and fatal injured motorists and passengers has decreased over the past decades, the number of severe and fatally injured cyclists has increased. One of the most critical situations in urban areas is when motorists turn right and cyclists cross the street. Many of those crashes result in severe consequences for the cyclists. Consequently, road traffic safety of cyclists, particularly in urban intersections, needs to be improved. Aim Based on a Decision tree (DT), a method for real-time early risk estimation was developed. Decision trees can help to understand (a) features and values leading to critical traffic situations, (b) help to classify new situations, and (c) help to detect risky situations as early as possible. Outcomes of the DT were assessed, results will be presented, and implications of the results will be discussed. Method In summer 2016 and 2018, two months (one month each year) of video material was collected at the AIM Research Intersection in Braunschweig, Germany. Situations of interacting right-turning motorists and crossing cyclists were extracted and experts evaluated the criticality of the interactions based on aspects of the Traffic Conflict Technique. Each situation’s outcome wasThey were classified as non-critical encounter, slight conflict, severe conflict, or crash as well as every time step regarding the current risk for later real-time classification. Those situations and classifications were used to design and train a DT. In operation, the DT provided the risk level of each motorist-cyclist interaction in real-time and sent out I2X (infrastructure-to-X communication) messages to the road users, and an adaptively triggered road instrumented warning element, the so-called Amber Light. The Amber Light was placed close to the common conflict area warning motorists in case of a predicted conflict with a crossing cyclist to inform, warn or even assist the drivers before an upcoming critical situation in time. In addition the evolution of the DT in question is presented in detail and opportunities of improving it are discussed. Results obtained/expected The most essential parameters affecting criticality were distance to conflict point, speed of the interacting partners, and the predicted post encroachment time (pPET a.k.a. Gap time).The developed method in question indicated that motorists can be warned before an upcoming critical situation. Conclusions Although the DT deals with known weaknesses such as instability, it allowed insight in parameters leading to critical situations. Open questions deal with labeling and learning of critical situations, the decrease of instability, and the application of other, possibly more suitable AI methods. In our future work we plan to analyze other use cases (e.g. left turning) and eventually to develop a generalized algorithm without specific concerns about road geometry, etc., yielding a reliable and accurate online risk estimation for road users