Unsupervised Clustering of Highway Motion Patterns

Validation and verification of modern autonomous vehicles can be seen as a limiting factor on their way to public roads. The standard road testing is not affordable due to the infinite number of possible real-life situations. Therefore there is wide consensus that it must be complemented by virtual testing. However, also the latter one cannot be performed for all situations, so a finite catalogue of special test cases, so called scenarios, will be used for virtual testing. This catalogue is expected to offer a good coverage of the general intended use of the driving function under test. To this end, it makes sense to derive these scenarios from real data. In this paper we propose a systematic way for building up such a catalogue progressively using sensor data. We use a method based on a variant of an on-line k-means algorithm for time series clustering under their alignment using the dynamic time warping. The advantage of the proposed method is the intuitive representation of the scenarios enabling their easy interpretations. Using experimental data, this paper illustrates how such a catalogue is produced and how it can be used for further scenario detection, for example.