Research On Weak Target Tracking Technology Based On Correlation Filtering And Neural Network

Tracking technology of small and weak targets plays an indispensable role in military scale and civil application.Tracking and attacking targets accurately in complex background is one of the focuses of military research in various countries.The main difficulties of tracking small and weak targets are as follows.Firstly,the target itself is too small,occupies less pixels,has no texture information,and has no specific shape.Secondly,the target moves rapidly in the background clutter.The main problems of dim target tracking are low SNR,high false alarm rate and poor real-time performance.Most trackers based on deep learning select a large number of training samples around the target.Then,the feature of convolutional neural network(CNN)is used to identify the target.A large number of training samples are the basis of deep learning tracking algorithm.However,one of the main reasons for immature tracking is the limited number of samples.Secondly,most of these methods mainly use the convolution feature of the last layer,because the convolution feature of the last layer contains a lot of semantic information.In target tracking algorithm technology,semantic information can solve the problem of target tracking loss due to the appearance change of target.But,in the research of tracking algorithm for small and weak targets,small and weak targets occupy few pixels and almost have no shape,so appearance change has little effect on the tracking results of point targets.In this paper,discriminant algorithm is used to track small and weak targets,and three different tracking algorithms are studied.Firstly,the algorithm of dim target tracking based on correlation filtering is proposed.As the main idea of target tracking algorithm,this algorithm mainly extracts the histogram of oriented gradient(HOG)features to judge the correct target,so as to get the maximum confidence value and screen the correct target.The second is boosting algorithm.As an experiment of deep learning into target tracking vision,this algorithm is a new attempt in infrared target tracking research.At last,a tracking algorithm based on neural network and correlation filtering is proposed.The algorithm combines the spatial information and semantic information of CNN,and filters the targets through correlation filtering.When the confidence value of the target is small,the problem that the target may be lost is solved by re-detection mechanism,which increases the accuracy of target tracking.From the experimental simulation and result analysis,it can be seen that the above algorithm can detect and track the target more accurately in their respective algorithm field.However,according to the tracking results of the above three algorithms in different backgrounds,it can be seen that the center location error of the weak small target tracking algorithm based on the fusion of neural network and correlation filtering is the smallest,and the overlap rate is the largest.Compared with the same group of experiments,this algorithm has better stability and higher accuracy.