Infrared Target Tracking Algorithm Based On Machine Learning

Infrared imaging system is widely used in military and civilian fields such as infrared guidance night navigation and intelligent security because of its advantages of working all day,good concealment,strong anti-interference performance and strong ability to penetrate smoke.As the key technology of infrared imaging system,infrared target tracking technology plays an important role in modern defense and alarm tasks.However,due to the low signal-to-noise ratio,serious background clutter and low resolution,infrared target tracking technology becomes a challenging subject,which has important theoretical significance and practical value for its in-depth research.In this thesis,the infrared target tracking algorithm is deeply discussed and studied.Firstly,the infrared image pre-processing technology was analyzed.Then two infrared target tracking algorithms based on machine learning theory were studied,and the effectiveness of the two tracking algorithms was analyzed and verified.For the problem of low resolution,low contrast and lack of texture detail,five typical infrared image enhancement algorithms were analyzed.Through simulation experiments,these five algorithms are compared and analyzed from four aspects: visual effect,distortion,detail enhancement and processing speed.The relatively good guiding filter is selected as the image preprocessing algorithm of the tracker.For the problem that the grayscale features of the infrared image are shortage and change with temperature,this thesis adopted the Histogram of Oriented Gradient(HOG)feature which is effective against temperature changes and target deformation.Based on the analysis of the accuracy of traditional correlation filtering and the real-time performance of kernel correlation filter tracking algorithm,this thesis studied and achieved the scale estimation tracking algorithm based on context-aware.It combines the accurate scale estimation algorithm with the kernel correlation filtering algorithm in the context-aware tracking framework,which can provide more negative samples for the filter so as to suppress the interference of the similar objects in the background and improve the accuracy of the algorithm.For the problem that the correlation filtering algorithm model is simple and thegeneralization ability is poor,this thesis studied and achieved the end-to-end tracking algorithm based on deep learning.The algorithm performs pyramidal fusion on the convolution layer of the full convolution Siamese network,and inputs the merged scale feature layers into the region proposal subnetwork.Then,by parallel computing the similarity between the template frame and the frame to be detected,the target position of the frame to be detected is determined and the target size proposal is outputted.Finally,the non-maximum suppression method was used to select the scale proposal with the highest score,so that the tracking effect with better accuracy is obtained.