Uso de técnicas de aprendizaje profundo en la detección vehicular sobre imágenes aérea

Abstract

Abstract: Vehicle detection is key to certain applications such as road safety, traffic surveillance, and autonomous vehicles. This research evaluates deep learning models applied to vehicle detection using images captured by drones in urban environments. Models such as YOLOv5, YOLOv8, YOLOv10, TensorFlow SSD, and Roboflow COCOs and COCOn were used. The dataset included 1,600 images processed with preprocessing and data augmentation techniques to increase training diversity and robustness. Evaluation metrics included precision, recall, mAP, and F1 score. The results show that Roboflow COCOs is the most robust model, with an mAP of 93.0% and an F1 score of 91.60, while YOLOv5 strikes a balance between accuracy and training time. In contrast, YOLOv10 performed worse despite being faster. The research highlights the impact of combining UAVs and deep learning, although challenges remain, such as detection in extreme conditions and optimizing training times.