Research On Detection And Tracking Algorithm Of Motion Targets For Passive Millimeter-wave And Terahertz Imaging

Passive millimeter wave and terahertz imaging technology forms its images according to the difference of radiation from the target and background.When combined with security scan tasks,this technology brings not only the ability for the scanner to detect various forbidden objects including gasoline,alcohol,metal and so on,but also it brings efficiency and safety to traditional security check.Therefore,passive imaging technology especially using the widely existing terahertz electromagnetic wave to get photos of certain target is broadly concerned by scientific research institutions all over the world.And how to carry out effective detection and tracking of passive imaging results has also been closely watched by researchers.Based on actual research project,this thesis analyzes the characteristics of the passive imaging results of millimeter wave and terahertz,and probes into the problem of how to effectively detect and track the targets in passive imaging.The specific research contents and results are as follows:(1)This thesis studies the passive millimeter wave and terahertz imaging theory and analyzes the relationship between the performance indexes of all modules in passive terahertz imaging system and the final brightness temperature resolution of the system.The advantages of passive terahertz imaging technology are also deduced theoretically in this thesis.(2)As traditional target detection algorithms based on segmentation is easy to introduce the problem of false alarm,This thesis also studies how to use the target detection algorithms based on features to reduce the misjudgments of algorithm.Especially we study how to use convolution neural network to learn the target characteristics itself to provide high quality features for the subsequent detection module to improve the accuracy of the detection.(3)According to the prior knowledge of the imaging scene that the target scale is under the premise of not dramatically change,We studies a simplified neural network architecture.The new architecture can help to reduce the complexity of the whole network without affecting its accuracy.We also study using a rectified cross entropy loss function to help train our convolutional neural network and analyze how different factors will affect its performance to improve the generation performance of the whole network.(4)In object tracking process of passive imaging scene,we are often confronted with problems of shade and disappearance.This thesis studies tracking methods based on semi-supervised learning.,We analyze the misdetection of traditional TLD algorithm in high noise,low resolution scenario,then a segmentation algorithm is used to improve the detect module.This thesis names the new tracking learning and detection algorithm based on segmentation TLD-BS for short.By reconstructing the training sample of classifier,the new algorithm can obtain more stable feature representation vectors of the noise image and accomplish better target tracking result.