Real-Time Vehicle Detection and Classification in UAV imagery Using Improved YOLOv5

Abstract

Recently, Unmanned Aerial Vehicles (UAV) have become useful for various applications. Real-time vehicle detection and classification is an essential task in some real-world applications, such as UAV-based traffic surveillance or disaster management. However, vehicles appear too small in aerial UAV imagery, reducing detection accuracy. While modern UAVs are capable of recording high-resolution videos with higher spatial information, increasing input size reduces inference speed. So balancing the accuracy and inference time is an existing challenge that needs to be addressed. This paper addresses this challenge by proposing an improved YOLOv5 single-stage object detector model, making it more suitable for detecting small objects in high-resolution images. At the same time, we modified the network width and depth to make it suitable for real-time applications that require high inference speed. Experiments conducted on VisDrone and CARPK datasets confirm that compared to baseline YOLOv5 models, the proposed model has 3.7% higher mAP50 and 6.1 FPS higher inference speed on VisDrone dataset. At the same time, its size is 44.6 MB less than YOLOv5L. These results confirm the efficacy of the proposed modifications applied on YOLOv5 to balance accuracy and inference time.