Development of a system for measuring flow speeds by analysis of aerial images

Abstract

This thesis presents a complete pipeline for detecting, tracking, and calibrating the trajectories of floating objects in outdoor environments. The proposed approach addresses the challenges posed by uncontrolled acquisition conditions, including strong illumination changes, surface reflections, and partial occlusions. Object detection is performed through non-parametric background subtraction (ViBe), complemented with noise-removal rules to filter out spurious detections caused by reflections and artifacts.

A two-phase tracking strategy is then introduced. The first phase relies on nearest-neighbor association to incrementally build trajectories while discarding implausible detections. The second phase adopts a graph-based formulation that models blob interactions such as fusion and separation events, thereby recovering continuous trajectories even when objects temporarily overlap.

Finally, a calibration procedure is developed to ensure a reliable mapping between 2D image detections and 3D world coordinates. Intrinsic, distortion, and extrinsic parameters are successively estimated, enabling accurate reprojection of real-world positions. The resulting system demonstrates the capacity to reconstruct object trajectories with sufficient accuracy for subsequent hydrodynamic analyses.

Beyond highlighting the strengths and limitations of each component, this work establishes a framework within which lightweight, interpretable methods can operate robustly in outdoor monitoring scenarios. Perspectives for improvement include the integration of Kalman filtering with global assignment strategies and the adoption of acquisition hardware to mitigate reflection artifacts.