A Vision-Based Deep Learning Approach to UAV Altitude Estimation

In recent years, unmanned aerial vehicles (UAVs) have become increasingly important in various critical missions, necessitating accurate and reliable altitude estimation. Traditional altitude measurement methods such as GPS and barometric sensors often face limitations due to environmental factors. This study proposes a vision-based deep learning approach for UAV altitude estimation using NADIR images captured by onboard cameras. A large and diverse dataset comprising over 300,000 images under different environmental conditions was utilized. A pre-trained ResNet50 convolutional neural network was adapted for regression tasks to estimate altitude from these images. The model achieved a mean absolute error (MAE) of 11.804 meters and an R2 score of 0.9712 on the test dataset, demonstrating high predictive accuracy. The results suggest that the proposed method can serve as a reliable alternative or complement to traditional sensors, enhancing UAV navigation and mission planning capabilities. Future work will focus on improving model performance in challenging conditions and integrating additional sensor data.