Object Detection in UAV Images Based on Improved YOLOv5

Abstract

Aiming at the problems of small targets, many instances, and complex backgrounds in UAV aerial images, in this paper, an improved YOLOv5-based algorithm for detecting objects in UAV images is presented. This paper enhances the robustness of an algorithm for recognizing aerial images by incorporating a spatial pyramid pooling network with a probability pooling method, introducing an upsampling network structure based on deconvolution and convolution attention mechanism. As a result, the issue of invalid features negatively impacting recognition accuracy is resolved, and a higher accuracy in recognizing aerial images is achieved. By conducting experiments on the VisDrone public dataset, it was found that the enhanced algorithm achieved an average accuracy of 34.9%, which is 3.21% higher than the average accuracy achieved by the original, unimproved algorithm.