A data mining approach for traffic accidents, pattern extraction and test scenario generation for autonomous vehicles

To effectively fight against traffic accidents, it is of great importance to analyse and understand the conditions that are linked with accidents. Such an analysis can serve as the basis to (i) develop reactive measures by finding the links between the pre-accident conditions (ii) devise proactive strategies that will prevent the occurrence of accidents by making the vehicles safer. This paper contributes to advancement of both approaches. For (i), one needs to identify the patterns in accidents. For (ii), introduction of Connected and Automated Vehicles (CAVs) is a promising solution. However CAVs need to be tested under numerous traffic scenarios to prove their safety before their deployment on public roads. This necessitates a great demand for high quality test scenarios for CAVs.This paper achieves two goals. First, it analyses the past traffic accidents (UK’s STATS19 database) to identify trends in the heterogeneous accident data and unravel the relationships between pre-accident conditions. This is done using a clustering algorithm (ROCK). Seven distinct large clusters emerge as a result. Each of these clusters are then further analysed for their meaning using the frequency analysis and geometric analysis. Secondly the paper underpins the proactive route (ii) by systematically developing, using the information in each cluster, test-case scenarios for CAVs which reflect the risk-prone conditions of the respective clusters. This is done using a data mining method (Market Basket algorithm) and further geometric interpretation of clusters. This way explicit scenarios are developed carrying the characteristics of the clusters that they come from.