| dc.description.abstract | The decision-making and motion planning play a critical role in the autonomous driving by connecting the perception to the vehicle control. It aims at generating available paths in the specific driving environment considering vehicle safety and driving efficiency constraints as well as the ride comfort. The complexity of the decision-making depends on the target driving performances and the driving environment.
The complexity of the future driving environment, due to the coexistence of automated and human-driven vehicles, makes the balance between safety, efficiency, and comfort much more challenging. Therefore, the focus of this research is to provide decision-making algorithms for an autonomous vehicle in the interactive driving environment where the surrounding vehicles are driven by human drivers who are unpredictable due to diverse driving behaviors.
To tackle the above problem, tools from game theory are utilized to analyze the interactions between rational players. To consider the driver intentions of the surrounding vehicles, games with complete information and incomplete information are discussed. The driver behavior is learned during the driving process, based on the Gaussian Mixture Model (GMM) trained by the naturalistic driving data. Then the driver behavior of surrounding vehicles is transmitted to the incomplete information game model, so that the human preferences can be estimated and utilized by the ego vehicle to regulate the predictions of the driving environment. Based on the model of incomplete information game, the uncertainty and the variety of the surrounding human-driven vehicles are both focused. The driving decisions can be made adaptively according to the driving styles of the surrounding vehicles.
The lane change scenario on a highway is selected as the research scene to test the performances of the proposed decision-making model. To make the simulation environment more realistic, the motions of the surrounding vehicles are modelled by the Intelligent Driver Model (IDM), whose driving styles are calibrated and classified in an explainable way based on the real driving data. Multiple scenarios are designed with driving style combinations of various surrounding vehicles. Moreover, the two-player game is extended to the multi-player game with the lateral behavior of the vehicles considered. Finally, the proposed model is validated by comparing the generated driving decisions and trajectories with human drivers’ driving profiles under the same driving conditions extracted from the naturalistic driving data. The results show that the driver aggressiveness estimation could help the ego vehicle change lane more efficiently and guarantee safety under the incomplete information model. The developed game-based decision-making model shows high potential to handle the uncertain interaction between the autonomous vehicle and human-driven vehicles. | en |
