Riassunto analitico
The exponential growth of electronic systems integrated into modern cars has made testing and validation processes increasingly complex. While automatic tests are widely adopted in the industry to speed up these time-consuming procedures, certain tasks still necessitate the involvement of human drivers. To achieve the ultimate objective of fully automating testing and validation, a new simulation-based approach to driving may prove instrumental. This approach implies the creation of a virtual driver capable of emulating human behaviour across a wide spectrum of driving tasks.
Numerous physics-based models exist, primarily focusing on human behaviour within specific traffic scenarios such as car following and lane changing. These traditional methodologies have demonstrated good results in the designed case studies but exhibit limitations in terms of generalization and adaptability. Recent advancements in the field of machine learning have also explored human driving behaviour. Innovative data-driven models exploiting deep learning techniques have been proposed to contend with the complexity and uncertainty inherent in driving. Neural networks designed to mimic human driver behaviour have exhibited strong performance in trajectory prediction. Nonetheless, relying solely on car following and lane changing fails to encapsulate the entirety of human driver decisions and manoeuvring actions. Some attempts have been made to combine these two behaviours, but few studies have delved into this area.
This thesis focuses on creating a general purpose virtual driver that can simulate human behaviour in common driving scenarios. It leverages behavioural models to encompass both free-flow and traffic scenarios within a unified framework. A novel architecture for simulating human driving behaviour is proposed, with real-time predictions generated through deep learning techniques, particularly by exploiting Long Short-Term Memory neural networks. Extensive data provided by Maserati, recorded from a vehicle produced by the Stellantis group, served as the basis for model training. The thesis enumerates driver interfaces and outlines an ad-hoc methodology for data processing. Additionally, it presents the evaluation methods employed during the iterative process of training neural networks.
Ultimately, the thesis conclude with one case study conducted at Maserati facilities, showing promising results coming from this innovative approach.
|
