An improved YOLOv5 for real-time UAV detection

Object detection, a core and demanding challenge within the field of computer vision, has recently garnered significant attention. The expanding Unmanned Aerial Vehicle (UAV) market demonstrates considerable potential in various fields, including surveillance and visual navigation. With the rapid advancements in deep learning, object detection techniques based on these methods have significantly advanced detection capabilities. However, UAV detection presents unique challenges compared to traditional object detection, such as smaller target sizes and more complex backgrounds, particularly in real-time environments. To address these challenges, this paper introduces an enhanced YOLOv5 method for real-time UAV detection. In this approach, YOLOv5 serves as the base detector, with PANet neck and mosaic augmentation techniques incorporated enhance the precision in detecting smaller objects. Additionally, a squeeze-and-excitation (SE) block is integrated into the backbone to enhance performance by focusing on channel information. Experimental results on our custom UAV-view dataset demonstrate that the improved YOLOv5 algorithm demonstrates quicker convergence and superior accuracy in object detection when juxtaposed with the original YOLOv5.