Pre-crash scenarios at road junctions: A clustering method for car crash data

• Published in: Accident analysis and prevention 2017-10, Vol.107, p.137-151
• Main Authors: Nitsche, Philippe, Thomas, Pete, Stuetz, Rainer, Welsh, Ruth
• Format: Article
• Language: English
• Subjects: Accidents, Traffic - prevention & control; Accidents, Traffic - statistics & numerical data; Algorithms; Automated cars; Automobile Driving - statistics & numerical data; Car crashes; Cluster Analysis; Clustering; Computer Simulation; Databases, Factual; Environment Design - statistics & numerical data; Humans; Intersections; Pre-crash scenarios; Risk Assessment; Road safety
• ISSN: 0001-4575; 1879-2057
• DOI: 10.1016/j.aap.2017.07.011

•Data from 1054 junction accidents was clustered to identify pre-crash scenarios.•This resulted in 13 clusters for T-junctions and 6 clusters for 4-legged junctions.•Association rules revealed common crash characteristics to describe the scenarios.•High-injury scenarios have little overlap with high-frequency scenarios.•Results support existing findings and can be used for safety performance studies. Given the recent advancements in autonomous driving functions, one of the main challenges is safe and efficient operation in complex traffic situations such as road junctions. There is a need for comprehensive testing, either in virtual simulation environments or on real-world test tracks. This paper presents a novel data analysis method including the preparation, analysis and visualization of car crash data, to identify the critical pre-crash scenarios at T- and four-legged junctions as a basis for testing the safety of automated driving systems. The presented method employs k-medoids to cluster historical junction crash data into distinct partitions and then applies the association rules algorithm to each cluster to specify the driving scenarios in more detail. The dataset used consists of 1056 junction crashes in the UK, which were exported from the in-depth “On-the-Spot” database. The study resulted in thirteen crash clusters for T-junctions, and six crash clusters for crossroads. Association rules revealed common crash characteristics, which were the basis for the scenario descriptions. The results support existing findings on road junction accidents and provide benchmark situations for safety performance tests in order to reduce the possible number parameter combinations.