Deep Learning-Based Hybrid Approach for Vehicle Roll Angle Estimation

Abstract

Existing methods for vehicle roll angle estimation, which often based on control theory and rule-based design, face challenges in maintaining estimation accuracy across various driving scenarios, such as parking tower driving and rollovers. To address these limitations, this paper proposes an improved hybrid approach that combines a Luenberger-like observer with a deep neural network (DNN). The proposed method enhances roll angle estimation performance by utilizing a DNN-based observer gain to fuse vehicle roll kinematics with the static roll angle-traditionally designed using expert knowledge in the automotive industry. This fusion ensures robust performance across diverse driving situations. Experimental data collected from a real production sport utility vehicle in scenarios including slaloms, banked roads, parking towers, and rollovers are used to train and validate the DNN. Key factors influencing vehicle roll angle are extracted from the data and utilized as inputs for the DNN. To train and validate the DNN, our method uses a loss function based on the target roll angle to improve estimation performance, unlike a conventional DNN-based hybrid approach for vehicle state estimation that employs a loss function based on the target observer gain. Comparative analysis with a rule-based method and the conventional hybrid approach demonstrates significant performance improvements in both typical and challenging driving situations. As a result, the proposed method reduces the roll angle estimation error by more than 0.5 degrees on average in terms of RMSE compared to both the rule-based method and the conventional hybrid approach, confirming an improvement in roll angle estimation performance.