A Conceptual Framework for AI-based Explainable Driver Behavior in Human-in-the-loop Simulators

Abstract: This paper shows a conceptual framework for emulating human driving behavior in driving simulators in a generic and adaptable way. It focuses on concepts a) to explicitly include mixed-traffic scenarios on one side and b) to generalize the driver behavior within the simulation in a way to be able to obtain AI-based behavior clusters which are interpretable to human characteristics and human state expressions. The underlying use-cases, its advantages and the underlying setup are intended to feed autonomous driving algorithms with different kinds of pedestrian behaviors like drunk or tired drivers or cyclists. The necessity of this paper is clear: in recent years, the development of autonomous driving has been accompanied by a series of optimistic assumptions. However, despite significant progress, the road to fully autonomous vehicles capable of seamlessly handling all possible driving situations remains an ongoing challenge. As autonomous technology continues to advance, new and complex challenges are emerging. One of the most prominent challenges is navigating in mixed traffic scenarios, in which the road is shared by different entities, including automated and autonomous vehicles, traditional manually driven cars, as well as vulnerable road users, such as cyclists and pedestrians. Understanding, predicting, and replicating human driving behavior in these complex and dynamic environments has emerged as a central but challenging fact of autonomous driving research. The need to address this challenge is not only rooted in safety concerns, but extends to the broader goals of gaining public acceptance and trust in Artificial Intelligence (AI), particularly in the area of self-driving cars. Even assuming that autonomous driving technology is fully mature today, mixed traffic scenarios are expected to persist for several decades. Today, research efforts which aims to model mixed traffic differs in its approaches. Mathematical, mesoscopic and macroscopic approaches exist on one hand for complete traffic flows and usually possess a high level of abstraction of the simulation environment like weather conditions, texture etc. Other conventional approaches use so-called Human-in-the-Loop (HITL) simulations to study driver behavior under different, but “closed” conditions. For example, Kraus developed a behavioral model for lane-changing maneuvers that focused on different psychological aspects of drivers, including fear and happiness, but it did not consider mixed traffic and it focused on the closed scenario of lane changing, so the purpose of the model is not to generalize driving situations. In our approach to generate generic and adaptable mixed traffic scenarios, clustering techniques are first used to categorize drivers with similar behaviors based on variants like k-means, hierarchical clustering, Density Based Spatial Clustering of Applications with Noise (DBSCAN), and Gaussian Mixture Models (GMMs). These techniques use features such as acceleration, braking, lane-changing behavior, and reaction times to form clusters that represent different driving styles, such as aggressive, cautious, or normal driving and different human states like fatigue. Second, the Explainability of AI-based clustering is not always given but necessary in the automotive industry to specifically train and test autonomous cars with dedicated, usually critical driving situations. Consequently, the mapping of AI clusters to driving types will play a further role in this paper and in our overall conceptual framework for emulating human driving behavior.

Keywords: Mixed-Traffic Simulation, Explainable Human Driver Behavior, Human-in-the-loop simulation for vehicle test beds

DOI: 10.54941/ahfe1006899

Cite this paper:

​