Research On Target Recognition And Tracking Technology Of Small Four Axis UAV

Due to the rapid development of UAV technology,UAVs have been widely used in people’s daily life and military fields.Life and personal privacy pose a serious threat,and the demand for anti-drone is growing.Since UAV is a typical low-slow and small object,its flight altitude is low,its speed is slow,its effective detection area is small,the Doppler effect is not obvious,and its flight environment is complex.Research on anti-UAV technology based on microwave radar and machine vision face huge challenges.Therefore,the UAV object recognition and tracking technology based on deep learning is studied in this paper,itis focusing on solving the problems of poor UAV visual detection effect and weak UAV object tracking performance in complex backgrounds.The research content includes the following aspects:1.Improve the network structure of the YOLOv3 object detection algorithm.Aiming at the low detection accuracy of the YOLOv3 object detection algorithm for small object UAVs,a method was proposed to improve the traditional feature pyramid structure(FPN)into a bi-directional feature pyramid(Bi-Fusion FPN),so that the relevant network layers can be fused from shallow layers.The feature information of the network and the deep network can improve the detection accuracy of the model small object UAV.Experiments show that the small object UAV detection accuracy of the improved YOLOv3 model is improved by 3.8%.2.Apply data augmentation operations to the UAV dataset.Aiming at the imbalance between the number of small object UAV images and the number of medium and large object UAV images in the UAV dataset,this paper proposes a data enhancement method for the UAV dataset,and the training dataset contains small objects.The over-sampling operation of the UAV image and the copy-paste operation of the small object UAV in the image help the object detection algorithm to obtain more small object features.Experiments show that the detection accuracy of the YOLOv3 algorithm enhanced by the training data set is improved by 1.2%.3.Improve the network structure of Siam BAN object tracking algorithm.Aiming at the problem that the Siam BAN object tracking algorithm has poor performance in distinguishing objects from backgrounds in complex environments,this paper proposes a method to replace the depthwise convolution operation with an asymmetric convolution(AC)operation,and use two independent convolutions instead The direct convolution on the stitched feature map is adopted,so that more parameters are added to the feature fusion operation of the template image and the search image,thereby improving the network tracking performance.Aiming at the deviation between the regression frame output by the Siam BAN algorithm and the real position frame of the UAV object,a Refinement Module is proposed to refine and integrate the classification branch output and the regression branch output of the network to enhance the model tracking effect.Experiments show that the UAV tracking accuracy of the improved Siam BAN model is improved by 2.5%.