Enhanced Vehicle Detection in Aerial Images Using YOLO with Adaptive Scaling Feature Pyramid Network

Abstract

Vehicle detection from aerial images is a crucial task in intelligent transportation systems (ITS) and surveillance due to the distance and scale of vehicles. This paper proposes an enhanced deep learning-based model for vehicle detection by integrating YOLOv8 with an improved Feature Pyramid Network (FPN) using Adaptive Scaling (AS). The proposed model addresses the challenges of detecting vehicles in different environments, including varying scales, illumination, and complex backgrounds. By incorporating Adaptive Scaling into the FPN, the model dynamically adjusts feature extraction across multiple levels, enhancing small and distant object detection. Experimental results on aerial image datasets demonstrate that the improved YOLOv8-AS-FPN model outperforms the YOLOv8 in terms of detection accuracy and robustness, achieving a mAP-50 of 98.6% compared to 97.7% in the original model, with better recall and precision values. The proposed approach provides a significant advancement in aerial vehicle detection, contributing to the development of more reliable ITS applications.