Open-World Learning for Traffic Scenarios Categorisation
Lakshman Balasubramanian, Jonas Wurst, Michael Botsch, Ke Deng
Abstract
Categorisation of traffic scenarios is an important component of scenario-based development and validation of automated vehicles. This problem requires an open-world learning approach but most of the machine learning methods used for traffic scenario categorisation work under the closed-world assumption. A closed-world model will classify all the inputs to one of the classes from the training data. An open-world learning method can identify, collect and cluster unknown traffic scenarios and incrementally add new scenario categories to the already existing ones. In this work, a hierarchical architecture for open-world learning method is proposed. The open-world architecture consists of the following components: an open-set recognition model, storage buffer, outlier detection, class-conditioned generative replay model, and clustering method. The components in the architecture contain novel machine learning approaches to address the challenging open-world learning tasks, e.g., Extreme Value Theory (EVT) for open-set recognition, Random Forest Activation Patterns (RFAPs) for clustering, class-conditioned generative models for replay, and self-supervised pre-training for feature generation. The proposed architecture is tested using real-world and simulation-based datasets. The results show the performance advantages of the proposed method. Also, extensive analysis of each component of the hierarchical open-world architecture underlines their importance in the overall architecture.