Research On Small Sample Target Tracking Technology For Mobile Agents

In recent years,with the rapid development of science and technology,artificial intelligence technology has made rapid progress.Computer vision,as a vital part of artificial intelligence,has attracted the attention of more and more scholars and experts.Target tracking technology,as one of the major research topics in the field of computer vision,has also received widespread attention.The application of target tracking technology is very extensive.It can be applied to air target tracking,drone reconnaissance,etc.in the military,and can be applied to security surveillance,traffic,and criminal reconnaissance in civilian use.This article focuses on the research of target tracking related technologies,and aims to explore a small sample target tracking method for mobile agents through artificial intelligence and deep learning technology.This paper proposes a method for target tracking based on small samples.The main work and contributions can be summarized as the following three points:(1)For neural networks,small samples are not enough for network training,so the sample information needs to be transformed first.The use of data augmentation techniques on small samples is a preparation for subsequent neural network training.According to the morphological changes of the observation and analysis of the target,the samples are processed to a certain extent by color adjustment,noise addition,translation,scaling,rotation,etc.to generate more samples.By simulating the morphological change of the target during the movement process,it provides a data basis for the subsequent training.(2)A method of target tracking based on Siamese network is proposed.The network model is trained using Siamese network,and the characteristics of Siamese network are used to learn the points of association between the target and the sample.By improving the Siamese network,the input scale at both ends of the network is changed,with one end as the target size and the other end as the background size.The similarity matrix of the target is obtained through convolution operation,and multiple scales are added to train in the network to predict the possible position of the target in the next frame of image.Neural network training can get the initial weight value well,and has good anti-noise ability and good robustness.Provide a basis for subsequent target positioning.(3)In order to locate the target accurately,a target localization method based on density clustering is proposed.This method builds a weighted operation matrix based on the density and distance based on the network's preliminary prediction of the target position.By clustering the coordinates and width and height in the prediction structure and filtering the edge anomalies predicted by the network,the target position of the next frame is finally locked to complete the target positioning.