| Summary: | Digital Air Traffic Control (ATC) Towers represent a significant advancement in airport infrastructure, allowing for remote control capabilities without direct visual surveillance. These towers rely on real-time video feeds from digital cameras, necessitating accurate mapping between camera images and real-world coordinates. In this study, an optimal bidirectional deep learning model is designed, implemented and evaluated for the camera calibration task. The performance of this model is compared against two reference models: a baseline mathematical model and a benchmark non-invertible machine learning (ML) model. A hybrid model was achieved using the stacked ensemble technique, combining an Artificial Neural Network (ANN) base-model with a Random Forest meta-model. Using this, the compounded error from repeated image-world conversions converges onto a stable value, effectively functioning as a locally invertible model. The mean converged error of the final hybrid model were 10.19m and 55.58px for the world and image locations respectively. For world error, this was an approximately 40% improvement from the baseline model, but a 75% increase from the non-invertible benchmark model, while for image error, it was 55% and 30% respectively. It can thus be concluded that the converging hybrid model combines the benefits of invertibility with the complexity and predictive accuracy of ML. While highlighting a trade-off between invertibility and accuracy, its ability to facilitate stable back-and-forth conversions offers valuable opportunities for enhancing digital ATC tower operations. These results pave the way for enhanced situational awareness and decision support, such as proactive runway incursion detection. Future research could further optimize the hybrid model and explore its applications beyond aviation, such as in autonomous driving and augmented reality, driving advancements in computer vision and automation.
|
|---|