Research On UAV Target Tracking Method Based On Decision Level Fusion

With the rapid development of uav market,the popularity of UAV has led to the frequent occurrence of "black flying" of UAV,which has also brought some troubles to the society and citizen’s life.In order to reduce the impact of "black flying" on society,this topic will study the uav target monitoring technology,use relevant acquisition equipment to monitor the UAV in different scenes,and track and identify the UAV target in the image processing stage.For uav targets of different scales,this paper proposes a decision level fusion algorithm based on convolution feature.For targets with low visibility such as smoke or night,this paper proposes a decision level fusion algorithm based on infrared and visible images.The specific work is as follows:（1）According to the characteristics of convolutional neural Network,based on Siam RPN（Siamese Region Proposal Network）algorithm,CIRes NET-22 with better performance and wider and deeper Network is used for feature extraction.Meanwhile,multi-layer convolutional features of the Network are extracted for analysis;In order to enhance the capability of feature representation,the convolutional layer at layer 16,19 and 22 is selected for feature extraction and independent tracking of UAV targets.（2）On the basis of obtaining the tracking results of each layer,a decision level fusion algorithm based on multi-layer convolution features is proposed.The algorithm uses the forward and backward tracking algorithm to obtain the backward tracking result of the target,and calculates the fusion weight value of Euclidean distance and overlap rate of each layer based on the historical data in a short time.The fusion weight is weighted according to the fusion weight to obtain the target location information.Through the actual measurement analysis of uav in different scenarios,the proposed algorithm in this chapter improves 3.38% and 10.07% in accuracy and success rate respectively compared with the improved Siam RPN algorithm.（3）For uav targets with low visibility,a fusion algorithm based on infrared and visible images is proposed.Based on the characteristics of infrared and visible light image characteristics of complementary,using multilayer convolution feature decision level fusion algorithm and reverse tracking algorithm for infrared and visible light to follow up on the results,and then using the Euclidean distance and overlap rate two measurement methods to obtain fusion weights,according to the result of fusion weight for tracking and decision level fusion recognition.Through the test and analysis of the anti-UAV Challenge data set of Computer Vision and Pattern Recognition（CVPR）,the accuracy of the algorithm proposed in this chapter reaches 83.13% and the success rate reaches 77.68%.Compared with the improved Siam RPN algorithm before,In this chapter,the accuracy and success rate of the algorithm are improved by 0.71% and10.97% respectively.Finally,the algorithm proposed in this paper is integrated to design and implement a uav target tracking interactive software interface,which can realize one-click tracking and tracking visualization functions.To sum up,in this paper,convolutional features of different layers are extracted by deeper and wider convolutional neural networks,and the decision level fusion algorithm is designed to integrate multi-layer convolutional feature tracking results,which improves the tracking accuracy of targets at different scales and occlusion.At the same time,a decision level fusion algorithm is designed based on the complementary characteristics of infrared and visible images,which improves the tracking accuracy and success rate of targets in complex scenes with low visibility.