Detección de vehículos en Imágenes Aéreas con Transformadores de Visión (ViT)

Abstract

Abstract: This research implemented and evaluated models based on Vision Transformers for the detection of vehicles in aerial images captured by unmanned aerial vehicles in complex urban environments. Two architectures with a Transformer approach, ViTDet and Deformable DETR, were implemented and compared with the reference convolutional model YOLOv8m using standard performance metrics, Precision, Recall, F1 score, and mAP with a threshold of 0.5, in addition to inference time. The XP methodology was used as an agile strategy for development, and a dataset labeled with the classes Car and Bus was used for the data. In addition, the images were resized to 1280×720 to standardize the input and preserve sufficient detail of the scenes. The results obtained show that Transformer models achieve more accurate detection than the convolutional model, even in low-light scenes. Overall, this work provides experimental evidence on the feasibility and challenges of applying Transformers in urban aerial scenarios.