Generating Explanatory Saliency Maps for Mixed Traffic Flow Using a Behaviour Cloning Model

Yasin Maan Yousif, Jörg P. Müller

Abstract: Multi-agent mixed traffic modelling and simulation are needed for safety estimation of traffic situations. Many of the most accurate traffic prediction models use deep learning methods that are considered black box models. This means that the output cannot be directly interpreted based on the input. However, such interpretation can be valuable in providing explanatory information about the model predictions for a simulation or a real-world dataset. On the other hand, formulating the prediction problem as states to actions mapping problem namely in a markov decision process (MDP) framework is a more realistic approach to fully imitate the traffic entity behaviour. Therefore, a behaviour cloning approach with memoryless architecture is presented here. As a result, it is easier to link the output to the image input using saliency maps extraction methods. The saliency maps calculated from the trained model highlight the traversable areas for the agent to reach its destination, avoiding collision with other agents and obstacles. They also show the salient roads edges that influence the direction of the predicted movement. These results are based on an analysis of a representative set of examples from the dataset.