Monocular Visual Map Registration and Sensor Fusion for Cost-Effective Land Vehicle Positioning

Abstract

Accurate positioning is a critical component of autonomous driving systems, particularly in GNSS-challenged environments such as urban canyons, underground parking garages, and indoor facilities. Traditional high-accuracy positioning solutions, such as high-resolution Light Detection and Ranging (LiDAR)-based positioning solutions and high-end inertial navigation systems (INS), are effective but prohibitively expensive and impractical for large-scale deployment in consumer and commercial vehicles. This thesis presents a cost-effective and scalable solution that leverages monocular vision and low-cost onboard sensors, such as Inertial Measurement Unit (IMU) and On-Board Diagnostics II (OBDII), to achieve reliable drift-free vehicle positioning using existing 3-D digital maps. The proposed system introduces a pipeline that generates dense metric 3-D point clouds from monocular images using recent advances in deep learning-based depth estimation, specifically the Metric3D model. The system applies a structured set of preprocessing and refinement steps, including semantic and transient object removal (STOR), scale correction, depth-based cropping, and statistical outlier filtering, to ensure robust alignment between the generated point clouds and prior 3-D maps. For accurate pose correction, the processed point cloud is registered with the map using Generalized Iterative Closest Point (GICP) in an environment-dependent Visual Map Registration (VMR) pipeline, and the resulting transformation is integrated into an Error-State Extended Kalman Filter (ES-EKF) for multi-sensor fusion with inertial and odometry measurements to ensure a high-rate, continuous, and smooth integrated positioning solution. Furthermore, a dynamic tuning mechanism is introduced to adaptively adjust measurement noise based on observed corrections, enhancing filter stability and resilience. The performance of the proposed solution is validated on real-world indoor and outdoor trajectories performed by the Navigation and Instrumentation Research Lab (NavINST)’s road test vehicle using low-cost sensors. Results show that the proposed method achieves sub-meter horizontal accuracy, consistent heading estimation, and robustness in both indoor and urban environments. Comparative analysis against baseline methods demonstrates substantial improvements in positioning precision, confirming the system’s potential for scalable deployment in navigation systems for autonomous vehicles, delivery robots, and smart infrastructure.