Wing Keypoints Detection Using Computer Vision
Claudio Alexandre da Costa Dias, Luiz Alberto Vieira Dias · 2025
This work presents the development of WingPointsCNN, a convolutional neural network (CNN) designed to automatically detect six keypoints on aircraft wings to support aerodynamic flow analysis through computer vision. Using images captured by fixed cameras during flight, the model accurately identifies the wing's position, enabling the isolation of the region of interest for subsequent tuft-based flow analysis. The CNN architecture is based on a simplified version of NaimishNet and was trained on a dataset of 3,000 manually annotated images. Data augmentation strategies, such as horizontal flipping and random cropping, were essential to improve generalization and reduce overfitting. The model achieved an average intersection-over-ground-truth ratio of 93%, demonstrating high accuracy in wing localization. Additionally, WingPointsCNN was used to generate 120 × 120 pixel image patches containing parts of the wing, which will serve as input for a future CNN dedicated to flow analysis. The results confirm the effectiveness of the approach for real-time applications and its potential for embedded system integration.