Virginia Tech
Physics-Informed Monocular Estimation and Tracking of Three-Dimensional Object Position Using Known Geometry and Instrumented Observer Motion
Abstract
dc:description.abstractThis dissertation presents a physics-informed method for estimating and tracking the three-dimensional position of rigid objects with known geometry using a single monocular camera mounted on an instrumented moving observer. The proposed method uses the object's known dimensions, the observer's recorded motion, and classical estimation techniques, rather than learned models or extensive training data. The proposed method relies on three stated assumptions: (1) the target object is planar, (2) the object has known dimensions and geometry, and (3) the object's principal face remains nearly parallel to the image plane during the vehicle's motion. These assumptions align with the characteristics of geometrically regulated roadside signs (by U.S. Department of Transport) encountered along a vehicle's path and define the proposed method's scope. The proposed method comprises four validated processes. First, a calibrated projection chain links the monocular camera's coordinate system to the vehicle's coordinate system. Second, a sub-pixel edge detector identifies the roadside sign's boundaries at the steepest slope of noise-weighted color-channel responses, providing a confidence for each reading and withholding a result when no high-confidence edge is detected. Third, a consensus-gated adaptation aligns a 3-D wireframe of the roadside sign's true physical size with the detected edges using normalized least-mean-squares updates with per-coordinate clamps, subjecting each fit to fit-quality, retry, and physical-displacement tests before acceptance. Finally, a constant-acceleration motion model integrates the vehicle's recorded motion into the update of every coordinate: as a directly measured velocity on the range coordinate, and as position-derived velocity values on the lateral and vertical coordinates. The proposed method is evaluated using four instrumented test-track datasets that include three roadside sign types and velocity profiles from nominally constant velocity to a complete brake to standstill. Ground truth is established through the surveyed roadside sign positions and the vehicle's Global Positioning System (GPS)-aided inertial navigation system. The proposed method achieves centimeter-level accuracy in lateral and vertical tracking throughout each dataset, with range errors constrained by the quadratic measurement sensitivity inherent to monocular size-based ranging. All residual systematic errors are characterized and bounded, in one case through the testing and rejection of a competing speed-dependent model. The ego-vehicle velocity fusion yields the largest single accuracy improvement: lateral estimate errors on the most challenging roadside sign decrease by more than five fold, and up to fourteen fold across the roadside signs. The proposed method is compared against two optical-flow alternatives evaluated on identical inputs and hardware, the proposed method's accuracy advantage is an order of magnitude better on the lateral and range coordinates in every one of the twelve roadside sign approaches, at a fraction of the per-frame computation time, completing all estimation steps in a lightweight CPU implementation suited to embedded systems. These results demonstrate that reliable three-dimensional tracking of known roadside signs can be achieved at significantly reduced computational cost when the observer's instrumented motion and the object's known geometry are treated as measurements in their own right. This capability suits low-cost vehicles and embedded platforms, and serves as a template for estimation problems in any domain where the platform already stores part of the required information.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Mechanical Engineering
- Department dc:contributor.department
- Mechanical Engineering
- Grantor dc:publisher
- Virginia Tech
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Alhendi, Hamad Gh F. A. A.
- Chair dc:contributor.committeechair
-
- Southward, Steve C.
- Committee members dc:contributor.committeemember
-
- Akbari Hamed, Kaveh
- Wicks, Alfred L.
- Abbott, Amos L.
Subjects
dc:subject × 8Rights
dc:rights- Statement dc:rights
-
- In Copyright
- Licence dc:rights.uri
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Other
- vt_gsexam:47585