The drone detection method utilizes infrared camera data and a YOLOv13 deep learning model.

Abstract

The invention relates to a method for detecting unmanned aerial vehicles using image data from infrared cameras and a YOLOv13 deep learning model, aiming to enhance the detection of small targets in environments with complex thermal interference.The method described in the invention includes the following steps: acquiring infrared images from a sensor; pre-processing the images to normalize the data and reduce thermal background noise; feeding the image data into an improved YOLOv13-based object detection model; and performing inference to determine the location, label, and reliability of the UAV target. The difference lies in the fact that the YOLOv13 model, optimized for infrared data, includes: (i) integrating a coordinate-based attention module in the Backbone via the AMC-C3K2 architecture to enhance the ability to represent the features of small targets; (ii) adding an SPPF module to expand the sensing field and exploit multi-scale features; (iii) improving the Neck architecture to reduce and optimize the feature fusion flow to minimize spatial information loss of the target; and (iv) proposing a specific IR-NCIoU loss function for infrared data, which improves the accuracy of small object localization and increases stability during training.Furthermore, the invention proposes a process for handling and enhancing infrared data, including the addition of thermal noise simulating the real-world environment to improve the model's generalizability. Thanks to synchronized improvements in network architecture, feature merging mechanism, and loss function, the method described in the invention allows for the accurate detection of small, low-contrast UAV targets in complex thermal environments, while ensuring real-time processing and suitability for deployment on edge computing systems.

Images