KAS-YOLOv5: A Novel Target Detection Algorithm for Aerial Images

Target detection in aerial image is a challenging task due to the high proportion of small targets, severe occlusion, and large scale variations. In this paper, a novel KAS-YOLOv5 algorithm is proposed based on YOLOv5. First, the K-Means++ clustering algorithm is used to regenerate the initial anchor boxes suitable for the target size of the aerial image. Secondly, Adaptively Spatial Feature Fusion (ASFF) is used to solve the inconsistency problem between features of different scales in aerial images. Finally, a dynamic label assignment strategy Simplified Optimal Transport Assignment (SimOTA) is used to transform the label assignment into an optimal transmission problem, thereby improving the severe occlusion phenomenon. The experimental results demonstrate a high accuracy rate of 42.2% on the VisDrone2019 dataset, outperforming other algorithms.

I. Introduction

With the upgrade of computer hardware, the use of artificial intelligence algorithms combined with unmanned aerial vehicle (UAV) technology for target detection has been increasingly widely applied in urban planning, traffic monitoring, emergency rescue, military reconnaissance, and other fields^[1][2]. UAV has the advantages of strong maneuverability, wide detection range, low safety risk coefficient of operators, and obvious advantages in performing various tasks in complex environments. However, due to the influence of the UAV shooting environment, aerial images have problems such as small target size, severe occlusion, and insignificant features^[2], which makes the target detection algorithm in the general scene still have some defects for aerial image target detection. Solving the problem of low target detection accuracy and high missed detection rate in aerial scenes is a major challenge for current aerial image detection algorithms.