Research On Object Tracking Algorithm And Deployment For UAV Aerial Image

With the rapid development of artificial intelligence technology,computer vision technology based on deep learning has been widely used in practical task requirements,such as object tracking,target recognition,anomaly detection and other technologies in the field of security,which can replace human resources for efficient identification and detection.With the rapid development of Unmanned Aerial Vehicle field,UAV has penetrated into every aspect of life.Aerial photography UAV can effectively replace people to carry out security inspection,unmanned monitoring,abnormal warning and other tasks.Therefore,the combination of computer vision technology and UAV will greatly improve work efficiency,reduce human resource consumption and security risks.Therefore,this thesis carries out the research on the object tracking algorithm for UAV aerial photography images,and realizes the deployment and application of neural network algorithm in the airborne edge computing equipment of UAV,which improves the performance and efficiency of aerial photography UAV in practical application.This work has certain research significance and application prospect.At present,there are still some challenges in object tracking algorithm and deployment application of UAV aerial photography: the small size of foreground target in UAV aerial photography perspective leads to low tracking accuracy;The object is easy to be blocked and the ambient illumination changes,resulting in poor robustness of the algorithm.The size and power consumption of UAV airborne edge computing equipment are strictly limited,resulting in the unsatisfactory practical application effect of neural network algorithm.Therefore,in view of the above problems,this thesis conducts research on object tracking algorithm and deployment for aerial images taken by UAVs.The main research contents and work achievements are as follows:(1)Aiming at the problems of small size of object and easy occlusion in aerial photography perspective,an infrared tiny object tracking method based on multiattention is proposed.The feature extraction backbone network was optimized to improve the feature extraction ability of the algorithm for tiny objects.The feature fusion module was added to integrate the multi-scale feature information,and a multiattention module was designed to model spatial context information and cross branch dependencies to improve the foreground and background resolution ability of the model.The comparative experimental results on real infrared tiny object tracking dataset demonstrate the effectiveness of the proposed method.(2)Aiming at the problems of significant illumination changes in aerial view angles,the RGB-T object tracking method based on multi-modal feature augment is proposed to improve the robustness of the algorithm under various illumination conditions.The Transformer structure is used to jointly encode visible and infrared modal data to integrate global information and enhance feature characterization;a Transformer structure-based feature fusion network is designed to integrate potential correlation information of multiple modal data from multiple branches of the siamese network.Experimental results on multiple mainstream RGB-T object tracking public datasets show that the proposed method achieves a competitive tracking accuracy while meeting the real-time requirement.(3)Aiming at the problem of the limited size and running power consumption of the onboard computing equipment of UAV,which leads to the difficulty in the practical application of the algorithm,the deployment research of neural network algorithm based on FPGA edge computing equipment is carried out,and the parallel computing part of the neural network algorithm is accelerated by using FPGA equipment.The algorithm deployment process mainly includes four main parts: neural network algorithm model design and training,neural network algorithm model quantization,compilation and generation of Deep learning Processing Units,and computing inference through hardware platform.The algorithm model needs to be designed,trained,and optimized for calculation,and then combined with hardware equipment to perform actual inference calculation.After deployment,compare the time consuming of FPGA edge computing device deployment algorithm with other mainstream devices,and compare the size and running power consumption of different devices;comparative experiments were conducted on the accuracy of neural network algorithms before and after deployment.The experimental results show that the algorithm deployment method for FPGA edge computing devices can significantly reduce the computing time,and the running power consumption of the devices can also meet the actual use requirements on the premise that the algorithm calculation accuracy is basically unchanged.Comparison and ablation experiments on multiple public datasets prove the effectiveness of the proposed algorithm.The deployment effect of edge computing equipment and comparison experiments verify the feasibility of the deployment scheme in this thesis,and prove that it has certain advantages in practical application.