Research On Target Recognition Method For UAV Application

In recent years,UAV has been widely used in the field of military livelihood,especially in the military,there is an urgent need for target identification from the perspective of UAV,and it is of strategic significance to conduct autonomous target identification in war.The object recognition technology combining deep learning and UAV vision has great development prospect.This thesis mainly studies the target recognition method for UAV application.In view of the large number of small targets in the UAV perspective,it is necessary to improve the identification accuracy of small targets.Considering the computing power limitation and power loss of the onboard processor of UAV,the target recognition algorithm embedded in UAV needs to meet the requirements of fewer algorithm parameters and high real-time performance.Aiming at the above two problems,this thesis has carried out a series of work,including:An improved method is proposed to solve the problem of low accuracy of UAV recognition caused by the large number of small targets.YOLOv5 s was selected as the basic model for improvement.The multi-scale fusion mechanism was used to add P2 layer of small target recognition.The shallow feature layer with rich detail information and deep feature layer with rich semantic information were integrated to improve the small target recognition accuracy.On the boundary frame loss,the SIo U loss function was used to replace the CIo U loss function,and the total degree of freedom was reduced from the perspective of vector.The unnecessary trial and error process of the prediction frame was reduced,and the accuracy of the model was further improved without increasing the complexity of the model.Kmeans++algorithm is used to cluster prior boxes,and 12 anchor boxes suitable for Vis Drone data set are obtained.The improved model improved 6.8% on the Vis Drone test set m AP@.5 and 9.4%on the homemade small-target dataset.Lightweight research on the improved model is carried out.C3 Ghost module is used to replace the C3 model of the original model.After studying the influence of width factor and depth factor on the model,it is concluded that it is the best choice to change the width of the model from 0.5 to 0.3125 and keep the depth unchanged.The final model in the Vis Drone test set mAP@.5 increased by 0.8%,and in the small target data set m AP@.5 increased by 1.1%,both of which showed improved accuracy.The number of model parameters is 30% of the initial model,and the weight of the model is only 8.11 MB.It can be deployed to the embedded equipment of UAV,and the detection speed can reach 44.32 FPS,which meets the real-time requirement of target recognition.The model is applied to Hiss embedded device and its performance is observed.The preparation work on PC includes model training,transformation,etc.In terms of hardware,the system equipment with Hi3519V101 processor as the core was built,and the model was transplanted to the embedded device.The recall rate reached 89.1%,the FPS reached 20,and the target recognition from the perspective of UAV was completed.