DIR-YOLOv5: a real-time drone-perspective infrared object detection method based on YOLOv5

Abstract

With the advancement of drone technology, object detection from the perspective of drones has found extensive applications in various fields, including surveillance, search operations, and reconnaissance tasks. Currently, most drones in the market are equipped with visible light imagers, while some high-end drones are equipped with infrared imaging detectors capable of performing infrared object detection tasks. Infrared imaging utilizes a passive imaging mode, enabling it to detect thermal radiation emitted by objects. As a result, it offers the distinct advantage of continuous operation without being restricted by daylight conditions. In comparison to visible imaging, infrared imaging uses longer wavelengths and possesses a certain level of penetration capability through clouds and smoke. Consequently, infrared object detection represents a significant research area within the field of object detection. However, detecting infrared objects, especially small ones, remains challenging due to the complexity of background information, lower resolution compared to visible images, and the lack of shape and texture information in infrared images. In response to these challenges, this study proposes a real-time drone-perspective infrared (IR) object detection method based on the YOLOv5 framework, known as DIR-YOLOv5. To effectively address the challenge of infrared vehicles occupying fewer pixels in the drone’s perspective image and making objects difficult to detect, the coordinate attention (CA) for feature enhancement is introduced. we also introduce a Spatial-Channel dynamic and query-aware sparse attention mechanism (SCBiFormer), which is optimized based on BiFormer. Additionally, we redefine the loss function as the Repulsion Loss function to tackle the problem of infrared vehicle objects gathering and overlapping occlusion in scenarios like parking lots. Furthermore, we expand the ISVD infrared image object detection dataset to include multiple scenarios and conduct experiments using this dataset. The experimental results demonstrate the excellent performance of the proposed method in infrared image object detection tasks, showing improved object detection accuracy and reduced false detection rate compared to current mainstream methods.