A Novel Vehicle Lateral Velocity Observer With Adaptive Feedback Gain via Deep Learning

The lateral velocity of a vehicle is crucial for effective lateral control. However, unlike yaw rate signal, there is a lack of reliable sensors that are affordable enough to be applied in mass production for directly measuring lateral velocity. As a result, numerous studies have focused on estimating lateral velocity through various methods. In this study, we propose a novel approach that combines a vehicle lateral dynamics model with lateral acceleration kinematic models, utilizing an AI-powered state observer. Whereas many previous studies have relied on empirical tuning by domain experts to determine the fusion weights of these models, our research takes a data-driven approach through artificial intelligence methods. Specifically, we designed a new method to calculate the true values of observer feedback gains, which act as appropriate model weights depending on the driving situation. Based on these ground truth values, we developed an AI model capable of predicting optimal observer feedback gains through supervised learning techniques. Finally, verification using both training and test scenario data demonstrated that the lateral velocity estimates from the state observer, powered by AI-generated feedback gains, were significantly more accurate than those obtained using gains designed by domain experts.