Study On Infrared Target Tracking Technology Based On Machine Learning Theory

The topic of this thesis comes from the project of national natural science foundation of China,the key scientific research projects of the 12-th five years plan,and the 863 projects.The research mianly solves the key technical problem of infrared image target tracking in complex background.Infrared imaging technology has been widely used in military and civil fields,due to its advantages of good secrecy,anti-interference and climate adaptation.It is a hotspot and difficulty in the field of infrared image processing to early detect the target which is tracked.As the target is far from the image plane,it only occupies a few pixels on the image and there is little text information.In addition,small targets are easy to be submerged in the cluttered background,and the tracking task is disturbed by similar objects.Therefore,how to effectively realize infrared target detection and tracking over complex background conditions is of great significance.Based on the tracking theories in the visible or infrared videos,here,we deal with the tracking problem by utilizing the feature extraction and enlarging the difference between the target and background information.In this paper,the guided image filter and convolutional network without training are first applied into tracking infrared targets,and then proved to be effective.Compared with the traditional template matching method,mean-shift method and particle filter,the proposed algorithms can realize online update of model parameters.On the one hand,the presented algorithms can effectively handle the appearance change and partial occlusion.On the other hand,the updating process only requires the image information of the adjacent frames referring to target’s neighborhood,helping to increase the performance of the real-time.In this dissertation,based on generative and discriminative models in visual tracking,we propose a variety of infrared target tracking algorithms via the image filtering and machine learning methods.Finally,the main research results are as follows,(1)Guided image filtering can remove noise while preserving edges,and the spatiotemporal context learning tracking algorithm has achieved good results.Therefore,we combined them to track dim and small targets.The experimental results show that the presented algorithm does well in suppress the background edge and algorithm efficiency.(2)Taking both the accuracy and speed of the algorithm,we presented a background suppression algorithm based on the SVD method,in order to enhance the small targets.Besides,according to the basic theory of kerlized correlation tracking with high performance,we presented the ‘preprocessing before tracking’ framework that uses the curvature filter method.Experimental results indicate that the presented algorithm achieve high algorithm speed and accurate tracking results.(3)First,we briefly described the particle filter and analyze the two-layer convolutional network in detail.Therefore,we introduced the strong feature into the infrared tracking,and designed two algorithms.Experimental results show that the proposed infrared tracking algorithms with strong classification performance can effectively track infrared targets over kinds of complex backgrounds.(4)In order to enhance the difference of target and background,the guided image filter was utilized to deal with infrared videos.Since the convolutional feature has strong representation,we presented a generative algorithm based on guided filter and this discriminative feature.Moreover,with the help of a sparse representation-based model,a joint model for infrared target tracking is proposed to deal with background clutter and low resolution.Experimental results show that the proposed methods perform well in high background clutter and low resolution,and the joint model is more stable and better than the individual model.