Research On Target Tracking Algorithm For Infrared Image

In recent years,infrared target tracking has been used in military and civil fields.However,the infrared image itself has some disadvantages,such as blurring,low contrast,large background noise,and unclear details,which make it difficult to obtain the features of the infrared target accurately.Some traditional infrared target tracking methods extract the bottom features of the infrared target,which can effectively track the target to a certain extent,but the development in recent years is increasingly limited.With the emergence of deep learning methods for target tracking,to track infrared target more accurately and efficiently,this thesis proposes an infrared multi-target deep learning tracking algorithm combining infrared target saliency detection and siamese FC(Fully-Convolutional Siamese Network).Firstly,aiming at the problems of the low contrast of the infrared image,unclear details of the image and fuzzy edge of the target,this thesis uses the gradient and information entropy of the image to fuse effectively,and adaptively adjusts the fractional differentiation to enhance the edge of the target in the image,then uses the standard deviation and mean value of image pixel gray to fuse to determine the segmentation threshold of the target,so as to distinguish the background and the part of the target in the image,and the target area in the image is linearly enhanced to further highlight the target.Secondly,aiming at the situation that infrared target is interfered by complex background and has many shapes and sizes,this thesis combines multi-scale top-hat transform,processes the enhanced infrared image with corrosion and expansion,extracts the subtraction of light and dark parts to reconstruct the image,so as to reduce the interference of fuzzy background noise,then the reconstructed image is transformed from the time domain to frequency domain,and the three saliency maps of improved spectrum residual,phase spectrum and quaternion Fourier transform are extracted for fusion to generate the final infrared multi-target saliency map.Finally,aiming at the situation that the target is occluded,rotated and deformed in infrared target tracking,this thesis combines saliency target detection and siamese FC network to improve the robustness of target tracking,marks the boundary box of the infrared target by segmentation marks on multiple infrared image saliency detection maps,and then the target's mark information in the current infrared target saliency map and preprocessed infrared image are input into the siamese FC network for training and target tracking,and the continuous tracking of multiple targets in the infrared image is finally realized through some processing including multi-target association matching,target anti-occlusion,and target re-segmentation,etc.Simulation experimental results show that the proposed algorithm in this thesis effectively enhances the infrared image,TBR(Target to Background Ratio)of the local target increases by 0.5 on average,and generates the infrared target saliency map with good effect,AUC(Area Under the Curve)is over 98%.Due to infrared image enhancement and saliency target detection,infrared multi-target tracking is both accurate and stable,and MOTP(Multiple Object Tracking Precision)and MOTA(Multiple Object Tracking Accuracy)is 99.9% and 99.7% respectively.Therefore,the algorithm in this thesis is helpful to the research of the infrared multi-target tracking algorithm with high accuracy and strong robustness.