Deploying a machine learning algorithm (k-medoids) to reduce the multidimensional parameter space of accident data to a set of diverse injury critical scenarios for Automated Driving Systems testing

It has been estimated that millions and billions of miles are required to be driven to showcase the safety of Automated Driving Systems (ADS). Scenario-based test strategies can reduce such distance-based testing efforts. The ISO 21448 suggests amongst other activities, the use of accident data to identify hazardous scenarios. Accident data bases can be very large. Testing each accident under different parameter combinations is not feasible. Hence, a method is required to reduce the accident data search space for the identification of hazardous scenarios. This paper presents a method based on the k-medoids clustering algorithm to reduce the complexity of the multidimensional search space of the British STATS 19 accident data.

This paper has demonstrated that the k-medoids algorithm, is able to reduce the multidimensional parameter space of accident data to a set of diverse injury critical scenarios for ADS testing. Moreover, the method has shown how previously unknown injury critical scenarios occurred under adverse environmental conditions can be identified from accident data. It has been concluded that the proposed method contributes to safety standards which focus to improve the safety performance of ADSs such as the ISO 21448 also known as the SOTIF standard.