Enhanced Long-Range UAV Detection: Leveraging Slicing Aided Hyper Inference with YOLOv8

The increasing use of Unmanned Aerial Vehicles (UAVs) in commercial applications has highlighted the urgent need for advanced detection systems that can reliably identify small drones from long distances. Detecting small drones at extended ranges remains challenging due to their minimal size within the image frame. While higher-resolution cameras can capture more details, traditional object detection methods struggle with resolution constraints, leading to significant degradation in detection accuracy after downscaling. To address these limitations, this research explores the application of the Slicing Aided Hyper Inference (SAHI) method, which enhances object detection by dividing high-resolution images into smaller, overlapping patches that align with the input resolution requirements of detectors like YOLO. By preserving critical pixel information through this approach, SAHI significantly improves the detection accuracy of small, distant drones. To evaluate this, we conducted a series of experiments using various drone datasets, including Long Range Drone Detection (LRDD), Drone vs. Birds, DetFly, and GAN-translated synthetic images. Results show a significant improvement in detecting small drones with SAHI compared to the baseline YOLOv8 model.