DAVAN: DTM-ANCHORED VISION-BASED ABSOLUTE NAVIGATION USING SENSOR FUSION

Navigation systems in modern autonomous aerial platforms heavily rely on Global Navigation Satellite Systems (GNSS). However, the ease of GNSS signals for jamming, spoofing and environmental obstruction require robust, independent alternative navigation strategies. While vision based absolute navigation presents a viable solution, traditional image to map matching methods are highly sensitive to seasonal, weather, structure and illumination variations.

This research addresses these limitations by replacing photographic maps with Digital Terrain Models (DTMs), which offer stable, constant topographical representations. This thesis introduces Digital Terrain Model Anchoring (DTMA), a novel, drift free absolute localization framework for monocular aerial cameras operating in GNSS-denied environments. DTMA iteratively aligns a sparse 3D point cloud, built with Structure from Motion (SfM), with a preexisting DTM. The algorithm uses a number of filters for improving of 3D points cloud quality. The filters based on residuals, mathematical limitations and semantic segmentation, which physically filter out non ground anomalies, such as vegetation and infrastructure, ensuring strict terrain point representation. An Iteratively Reweighted Least Squares (IRLS) solver is then employed to dynamically estimate the 7-Degree of Freedom (7-DoF) transformation, simultaneously recovering the absolute global position, orientation and the missing metric scale of the visual structure.

To further improve system robustness, two advanced architectural variants are proposed. Deep Learning DTMA (DL-DTMA) bypasses classical SfM dependencies by utilizing monocular neural networks to estimate depth maps from single discrete images, enabling localization in feature poor or zero-baseline scenarios. Simultaneously, Factor Graph DTMA (FG-DTMA) reformulates the pose estimation as a non-linear graph optimization, jointly solving for the camera pose and explicit 3D landmarks to improve geometric consistency and convergence stability.

To deliver a continuous, high frequency navigation solution, the visual anchors generated by these algorithms are tightly coupled with an Inertial Measurement Unit (IMU) via an Error State Extended Kalman Filter (ES-EKF). This sensor fusion architecture successfully bounds inertial drift and dynamically calibrates IMU biases in real time. The complete end to end pipeline is evaluated through extensive parameter sweeps in the AirSim simulation environment and validated against the real world aerial ALTO dataset. Experimental results demonstrate that the proposed DTMA frameworks achieve high precision, drift free localization across complex topographies, showing a highly resilient potential and alternative solution to satellite based navigation.

M.Sc. student under the supervision of Prof. Hector  Rotstein.