Development of a Geometry-Agnostic Deep Learning Model for Detection of Heliostats in Solar Power Tower Plants

Abstract

This thesis investigates the generalization capabilities of deep learning models for the detection of heliostats with varying mirror geometries in aerial images of solar power tower plants. A geometry-agnostic modeling strategy is developed and evaluated in three structured scenarios, each based on different configurations of training and test geometries. The proposed methodology focuses on training models for object and keypoint detection using synthetic datasets and evaluating their performance based on the AP1 and PCK1 metrics. The results indicate that keypoint prediction generalizes more robustly under structural variation than bounding box detection, which is more sensitive to differences in geometric scale and design. Additional comparisons with class-aware models, i.e., models that were trained and tested on the same geometry, show that the geometry-agnostic models can achieve competitive performance in keypoint detection, but consistently underperform in the bounding box accuracy. Selected models trained on a diverse set of geometries, including the test geometry, achieve up to 2 percentage points higher PCK scores than the class-aware baseline while achieving comparable AP performance within a margin of 5 percentage points. Initial experiments on real-world data demonstrate that the sim-to-real transfer remains highly challenging, with performance degrading significantly in the absence of realism-enhancing factors such as soiling or contextual scene elements.