The rapid advancement in Unmanned Aerial Vehicle (UAV) technology has created new opportunities for real-time monitoring and detection applications, particularly in vehicle detection from aerial imagery. This study presents a robust real-time vehicle detection model, built on YOLOv5, specifically designed for UAV-acquired images and leveraging the VisDrone2019 dataset, which includes annotated categories of cars, vans, trucks, and buses. YOLOv5 serves as the base model, optimized for high-speed processing, and several key enhancements are introduced to improve detection accuracy and performance in complex aerial scenes. An additional prediction head is integrated into YOLOv5 to enhance detection capabilities for smaller-scale objects, addressing challenges in identifying vehicles from high altitudes or dense environments. To retain essential feature information throughout the training process, a Bidirectional Feature Pyramid Network (BiFPN) is employed, enabling efficient feature fusion across multiple scales. Furthermore, Soft Non-Maximum Suppression (Soft-NMS) is utilized as a frame filtering technique, mitigating missed detections by handling cases where vehicles are closely aligned. This combination of YOLOv5 and advanced techniques enables high accuracy in real-time vehicle detection, making it suitable for applications in traffic monitoring, urban planning, and emergency response.

Keywords: Real-time vehicle detection, UAV, YOLOv5, VisDrone2019 dataset, small-scale object detection, Bidirectional Feature Pyramid Network (BiFPN), Soft-NMS, aerial imagery, traffic monitoring.