Pattern Recognition for Scenario Detection in Real-World Traffic Data

Abstract

The continuous development and integration of Automated Driving Systems (ADS) leads to complex systems. Such systems need to be tested and validated thoroughly for all situations these systems may encounter on the road, to assure their safety and reliability. Test drives with ADS require millions of operational hours which is infeasible. TNO proposes to acquire micro-traffic data (data collected on the level of the individual vehicles) to generate real-world scenarios (situations) for testing and validating ADS. Such scenarios are resembled by typical patterns in the data. To achieve extraction and classification of scenarios from micro-traffic traffic data, TNO already developed knowledge-driven, rule-based techniques. However, a drawback of these techniques is the loss of generalization. This work proposes an unsupervised data mining approach to mine events from real-world traffic data to overcome this limitation. After decimation and discretization of the data, we combine Frequent Itemset Mining and Frequent Sequence Mining over multiple sensor outputs in the form of data streams for recognizing patterns that represent real-world traffic scenarios. The method shows that different configurations can result in different generalizations to satisfy the need of the expert, as different levels of abstraction can be desired by the user. Finally by conducting experiments we conclude which configuration provides the most desirable result in finding events for longitudinal movement.