Research on real-time dense small target detection algorithm of UAV based on YOLOv3-SPP

Unmanned aerial vehicle (UAV) image target measurement technology is widely used in military, traffic, environmental monitoring, security and other fields, and has important application value. The characteristics of UAV aerial images, such as many kinds of targets and large-scale changes, bring great challenges to target recognition. Common UAV image target measurement algorithms are complex and cannot be well deployed to UAV platform. Aiming at the above two problems, a lightweight UAV image target detection algorithm was proposed in this study. Based on YOLOv3-SPP, a measuring head suitable for UAV image was designed. Combined with ASFF adaptive feature fusion method, an adaptive feature fusion method ASFF Small was proposed. Meanwhile, a new loss function SIoU was introduced to accelerate the convergence speed and improve the accuracy of the model. Bottleneck module was improved by depth separable convolution to reduce the computational complexity and complexity of the model. A large number of experiments on VisDrone open data set show that the method reduces the parameters of baseline model by 53% and the number of floating-point operations by 46.6%, respectively, and achieves a mean average accuracy (mAP) of 42.5%. Compared with the baseline algorithm, it improves 11.3%, which is superior to the common similar algorithms, and can well complete the target detection task of UAV aerial images.