Research On Aerial Target Tracking Based On Convolutional Neural Network

Aerial target tracking is one of the key technologies for aerospace and other types of detection systems.Due to the relatively long observation distance,satellite targets often appear as small targets or even point target states with lack of shape and texture features,which causes tracking difficulties.In particular,noise and other pseudo-objects such as stars and planets block satellite targets make tracking even more difficult in the star-sky background.Therefore,it is challenging to accurately track small targets in the aerial.In this paper,the convolutional neural network is utilized to study the small target tracking in the star-sky background.The main work is as follows:(1)A star-sky image denoising algorithm based on hybrid model is proposed which is aimed at the similarity in size and brightness between target and noise in the sky-sky image.First,the image is denoised by fast and flexible convolutional neural network to preserve the target edge of the image.Secondly,the median filter is utilized to remove the strong noise.Finally,the denoised image is enhanced based on the intuitionistic fuzzy set.The experimental results show that this model method not only has good denoising ability,but also makes the energy of the image at the edge have almost no loss,which lays a foundation for the accurate follow-up of the target.(2)A small-aerial aerial target tracking method based on multi-domain convolutional neural network and autoregressive model is proposed which is aimed at the problem of tracking drift by the obstruction of background and other pseudo-targets such as stars and planets.Firstly,multi-domain convolutional neural network is utilized to adaptively extract the target features and effectively classify the targets.Secondly,the network output target position is utilized as historical parameters to train the autoregressive model,and the target motion trajectory is estimated to achieve the unknown position of the predicted target.Using the bounding box regression model to adjust the target position with the highest confidence.The method fuses the feature information and motion information of the target effectively and successfully solves the problem of drift by occlusion in the tracking process.(3)Design and implementation of the aerial target tracking system.MATLAB is designed and developed of the entire tracking system to achieve accurate tracking of small targets in the aerial.