Optimization Of Siamese Neural Network Object Tracking Algorithm For Rotorcraft UAV

UAVs are unmanned aircraft operated using radio remote control equipment and self-provided program control devices.In recent years,with the rapid development of civil rotor drones,it has brought convenience and regulatory requirements,while the supervision of drones requires technical support for object tracking.Thanks to the rapid growth of data volume and the continuous improvement of computer performance worldwide,deep learning has achieved better performance than traditional machine learning in many fields.In the field of object tracking,deep learning has also gained more and more attention.As a representative,the Siamese neural network tracking algorithm(Siamese-RPN,referred to as SiamRPN),under the premise of ensuring real-time tracking,turns the tracking problem into a similarity solving problem by using the Siamese network,and realizes the end-to-end object tracking.However,when the SiamRPN algorithm encounters occlusion problems,similar background interference,and tracking small target UAVs on UAV tracking tasks,there is a lack of robustness.In view of these problems,this paper proposes an improved SiamRPN algorithm tracking framework,which has two major improvements.First,the object detection algorithm RetinaNet is used for auxiliary detection,and the tracking results of the SiamRPN algorithm are modified according to a certain mechanism.The second is to improve the structure of the RetinaNet model for small targets to improve its auxiliary tracking effect for small target UAVs.In the tracking test experiment of UAVs,it can be seen that the improved SiamRPN algorithm tracking framework improves the robustness and accuracy of the SiamRPN algorithm,proving the effectiveness of this improved SiamRPN tracking framework.