YOLOv7-UAV: Improved YOLOv7 Algorithm for Small Object Detection in UAV Image Scenarios

Object detection in unmanned aerial vehicle (UAV) images has become a popular research topic in recent years. However, UAV images are captured from high altitudes with a large proportion of small objects and have dense object regions, posing a significant challenge to small object detection. To slove this issue, we propose an efficient YOLOv7-UAV algorithm in which a low-level prediction head P2 which is added to detect the small targets from the shallow feature map, and a deep-level prediction head P5 which is removed to reduce the effect of excessive down-sampling. To mitigate the mismatch between the desired ground truth box and the prediction box, the SCYLLA-IoU function is used in the regression loss to speed up the training convergence process. Moreover, the proposed YOLOv7-UAV algorithm is quantified and compiled in the Vitis-AI development environment and validated in terms of power consumption and hardware resources on the FPGA platform. The experimental results show that the mean average precision of YOLOv7-UAV is improved by 33% compared to traditional YOLOv7, and the FPGA implementation improves the energy efficiency by 12 times compared to the GPU.

I. Introduction

With the advantages of flexibility and maneuverability, unmanned aerial vehicles (UAVs) are regarded as a potential technology for both commercial and military applications, including disaster surveillance, traffic patrol, aerial base station and navigating a military battlefield [1] [2]. Investigating the precious object detection in UAV image scenarios is an on-going research topic due to the high altitude of UAVs. Although the existing object detection algorithms provide effective detection of common targets in natural scenes, it is difficult to detect the UAV images since small targets have large proportions and occupy fewer pixels than regular objects [3] [4] [5]. Meanwhile, UAVs has limited energy and hardware resources to carry out the algorithm [6]. The YOLO algorithms have the lightweight network architecture that balances speed and accuracy performance, making them widely used in real-time object detection on edge devices [7] [8] [9].