Transformer-based computer vision technology empowers drones

With the rapid development of drones. The traditional “manual feature extraction + classifier-based” object detection algorithm can no longer meet the accuracy requirements. Aiming at the problem that the reasoning speed is slowed down due to the complex background of UAV aerial images, we propose a target detection model based on a multi-head attention mechanism: Swim-Transformer. This model introduces a multi-head attention mechanism based on the YOLOv5, and introduces a decoupling head method in the head to improve detection accuracy and speed up network convergence. Experiments show that the new target detection framework is superior to the traditional target detection algorithm on the UAV aerial photography data set VisDrone, and increasing mAP by about 0.0067%.