RepViT-Based UAV Small Object Detection on VisDrone and UAVDT

The Unmanned Aerial Vehicles (UAVs), popularly referred to as drones, have gained increasing significance with respect to their use in applications like traffic monitoring, surveillance, disaster management, environmental observation, and precision agriculture. Despite substantial progress in aerial image sensing technologies, accurate object detection in UAV imagery has remained elusive owing to the issues related to small object sizes, object scale variance, motion blur, occlusions, and complexity of the background. The proposed study describes a RepViT architecture to enhance the detection of small objects. The proposed framework makes use of RepViT architecture, Feature Pyramid Network (FPN), and attention mechanism for enhancing discriminative features and suppressing background noise. The effectiveness of the proposed model was evaluated using benchmark UAV object detection datasets such as VisDrone and UAVDT using widely adopted object detection metrics. The proposed framework has sustained the inference rate of 48 frames per second (FPS).