Tightly-Coupled Multi-Source Visual and Wheel Odometry for Autonomous Parking

Accurate and robust real-time vehicle localization is essential for autonomous parking, yet it is often challenged by GNSS-denial and texture-scarce environments. To address this, we propose a low-cost localization system based on the tight coupling of multi-source visual information and wheel odometry. Built upon a Multi-State Constraint Kalman Filter (MSCKF), the system introduces two key innovations: First, it fuses wheel speed measurements with heterogeneous visual features, which include not only traditional geometric corners but also semantic parking slot corners extracted from a standard parking perception pipeline, allowing us to effectively leverage rich, structured information for localization. Second, it incorporates a global pose correction mechanism that utilizes fully detected parking slots as semantic anchors to effectively suppress long-term cumulative drift. We validated the proposed method on an experimental platform we developed, equipped with low-cost surround-view fisheye cameras and wheel speed sensors. The experimental results demonstrate that, in autonomous parking tasks, the system achieved centimeter-level localization accuracy and outperformed other baselines, indicating its substantial potential for large-scale commercial application.