Research On Target Tracking Based On Infrared And Visible Image Fusion

In modern military warfare,it is required to be able to support long-distance highprecision combats,all-weather combats and night combats,and to detect the target in real time from the complex background,so as to ensure accurate target positioning.At present,most of the target tracking algorithms based on machine vision adopt single visible images for detection,and there is no good solution for continuous target tracking under low-light complex dynamic backgrounds like forests.Focused on the low-light complex environment,the main research targets of this thesis are to improve the target tracking accuracy and bring more information of moving direction and trajectory to the observer.The overall target tracking method is carried out from the aspects of infrared and visible image registration,fusion and target detection,the main works are:(1)Image registration algorithm integrating Sift features and slope consistency.The algorithm improves the accuracy of multi-source image registration through multi-level feature point pairs screening strategy to facilitate further image fusion.Firstly,the reference image and the image to be registered are preprocessed,and Sift algorithm is used to extract the feature points of the two images.Secondly,the feature points are matched through algorithm of 2NN to get the rough matching point pairs.Thirdly,according to the prior knowledge that the slope between the correct matching point pairs is consistent,the rough matching point pairs are further screened.Finally,the optimal transformation matrix is selected by RANSAC algorithm to complete the image registration.Compared with other two registration algorithms in terms of mutual information and running time,the accuracy and efficiency of the registration algorithm are both greatly improved.(2)An improved GTF infrared and visible image fusion algorithm based on multicolor space.Firstly,on the basis of infrared and visible image registration,the GTF fusion algorithm based on total variation(TV)is introduced to obtain fused images which contain both texture information of visible images and thermal radiation information of infrared images.Secondly,aiming at the limitation that the GTF algorithm is only applicable to grayscale fusion,the color level image fusion is realized through multi-color space.Finally,compared with other three color level fusion algorithms of Weighted Averaging,RGB-HIS and Wavelet in the aspects of entropy,mutual information and average gradient,the effectiveness of the algorithm is verified.(3)Target tracking algorithm based on color feature and Kalman filter for infrared image sequence.When the target is not occluded,the color feature is used to detect the target position,and when the target is occluded or misdetected,the predicted value of Kalman filter is used as the target position.Secondly,based on the realization of image fusion above,the detected position is mapped to the corresponding fused video frame.Thirdly,the registration algorithm is used to calculate the target motion direction in real time considering the background motion and the target centroids of the upper and lower frames are connected to draw the motion trajectory.Thus,a fused image sequence can be obtained with accurate positioning,showing the movement direction and trajectory.Furthermore,in the face of human body detection in complex forest environment,classical target tracking algorithms are prone to the phenomenon of target box drift,the algorithm in this thesis can deal with occlusion and misdetection fairly well and the positioning accuracy can reach 93%,which is a great improvement compared with tracking algorithms of Camshift?KCF and MOSSE.The positioning accuracy of human eyes is only 87%,so the algorithm proposed can assist the observation of human eyes very well.