Active learning for active safety improving Autonomous Emergency Braking

Autonomous Emergency Braking (AEB) significantly improves automotive safety by braking the vehicle autonomously before a collision occurs. Yet current commercial systems underperform, particularly in the presence of uncertainty about the vehicle or road surface.The method presented in this work namely, Active Learning for Active Safety AEB (ALASAEB), addresses these issues by developing a complete AEB system that focuses on three key components. First, a Dual Control for Exploration-Exploitation (DCEE) approach is applied to solve the maximum braking problem, achieving effective braking performance across a wide range of road surfaces and velocities. A significant contribution of this work is that a modified Regularized Particle Filter is used to estimate vehicle states and parameters of the Magic Formula tyre model including the peak friction coefficient for the environment.Compared with popular Extremum Seeking methods and the Bosch ABS model, this DCEE approach to maximal braking reduces stopping time by up to 25%. Secondly, an online driver-in-the-loop solution to identify critical parameters such as the current and maximum friction coefficients is required for effective AEB operation in unknown environments. A method, namely Torque Vectoring for Active Learning (TVAL) is introduced that can perform state and parameter estimation whilst following the driver’s input. This is designed to ensure minimal disruption to the driver, allowing them to maintain full control of the vehicle. Then, a scheme for practically implementing TVAL is introduced, which considers powertrain efficiency, safety, and feasibility in an online fashion. Using a high-fidelity vehicle model and drive cycle, the functionality of the TVAL controller is demonstrated across changing road surfaces, where the road surface is succesfully identified whenever possible.Finally, by the virtue of modelling each component as a Discrete Event System, this thesis exploits Supervisory Control Theory to design a monolithic AEB supervisor, supported by an active learning approach, to prevent all possible collisions. This brings together all aspects of the Autonomous Emergency Braking (AEB) functionality into a unified model, namely ALAS-AEB. ALAS-AEB is verified by the ISO22733 test standard and further evaluated using a City Scenario that fully demonstrates the benefits of the proposed method.