An Investigation into YOLO-v8 Model Optimization for Small Object Detection in UAVs using Attention Mechanism
N S Rupak, N. Rayvanth, Pulipati Kushank, Rimjhim Padam Singh · 2024
The growth of UAV technology has led to advancements in aerial surveillance, resulting in the development of specialized algorithms for object detection in complex environments like university campuses, beaches etc. This study proposes to leverage the recent YOLOv8 model with various sophisticated attention mechanism like Spatial attention, channel attention, squeeze and excitation block attention, Efficient channel attention network, etc. to efficiently detect the tiny person objects in images captured using Unmanned aerial vehicles (UAVs) like drones. For the purpose of this analysis, the work utilizes a recent dataset, Manipal-UAV dataset, that focuses on detecting tiny objects in the form of person. The work also provides a comprehensive comparison of recent YOLO models for model selection and provides a detailed analysis of state-of-art attention mechanism when incorporated into the best baseline YOLOv8x model. The findings from this study will suggest that tailored YOLO models are effective for UAV-based person detection tasks, potentially driving advancements in aerial surveillance technology, particularly in scenarios requiring accurate identification of small objects like individuals. The research over-all aims to highlight the importance of dataset-specific optimizations to enhance the performance of object detection models in UAV imagery with mAP50 of 50.2%, thereby, promoting more reliable surveillance applications in various real-world settings.