QuasittudeNet: Self-Supervised Network for Vehicle Attitude Estimation

Abstract

Precise attitude estimation is fundamental for the safety and control of intelligent vehicles. Although Inertial Measurement Units (IMUs) are ubiquitous, differentiating between gravitational tilt and proper acceleration measurements remains a challenge for classical kinematic filters, particularly during dynamic maneuvers. Recent approaches have explored deep learning to build better estimation models. However, such approaches typically rely on large volumes of expensive ground-truth-labeled data. To address these limitations, this paper introduces QuasittudeNet, a lightweight self-supervised framework for estimating vehicle roll and pitch. By combining standard IMU readings with odometer speed data, the network learns to decouple suspension dynamics from inertial forces without requiring supervision. A physics-guided learning framework is used with a U-Net backbone and attention mechanisms that dynamically weigh the relevance of the sensors. Extensive validation across five public datasets (including NuScenes, Kitti, and A2D2) demonstrates that QuasittudeNet reduces pitch estimation error by up to 76% compared to baselines, successfully bridging the gap between data-driven performance and physics-based reliability.