YOLOv5 with Dual Attention Network for Object Detection on Drone

Abstract

Detecting the objects in images is a fundamental and crucial research field in computer vision, aiming to locate and classify objects in images. Applying object detection to drones offers numerous advantages. However, to achieve this, the light-weight algorithm is mandatory to operate in real-time on low-cost devices such as Lattepanda and Jetson Nano. In this paper, we propose the YOLOv5 Network with the incorporation of Efficient Residual Bottleneck and DANet. The Efficient Residual Bottle-neck can decrease the parameters, making the model lightweight and reducing computational complexity. DANet emphasizes important image features while suppressing noise and unnecessary information, thereby improving object detection performance. We trained our model on the VisDrone dataset, and compared to YOLOv5, the proposed model achieved approximately a 12% improvement in mAP (mean Average Precision) with an mAP value of 22.8, while reducing the number of parameters by approximately 55% to 3,897,433.