Research Of Object Tracking In Complex Video Environment

Object tracking has developed rapidly in recent years,and numerous research results have been produced in many fields such as intelligent medical detection and analysis,intelligent transportation and so on,which have promoted the society with a lot of progress.In the process of continuous development of technology,the two ideas of correlation filter and deep learning have gradually become the mainstream thinking of object tracking.Based on two mainstream ideas,researchers have proposed many tracking algorithms with better performance.However,the interference factors such as occlusion,deformation,scale changes,motion blur,and background clutter in the complex video environments can significantly reduce the accuracy and robustness of the tracking algorithm.In this thesis,two kinds of object tracking algorithms are proposed for the influencing factors in the above complex video environment: the first one is multi-scale robust tracking based on context information and re-detection mechanism,another one is dynamic strategy fusion tracking based on multi-layer convolutional features.First of all,this thesis expounds the research background and significance of the object tracking,and the research status at home and abroad.The related theories and methods involved in the two algorithms proposed in this thesis are introduced.We also summarize the influencing factors of the complex video environment and the solutions.Then we introduce the multi-scale robust tracking based on context information and re-detection mechanism.Firstly,aiming at the problem of missing semantic information between the object and adjacent regions in tracking algorithm,the four regions contained by upper,down,left and right of the object are utilized as the negative samples to train the tracking model,which reduces the effects of similar interferences in the video sequences.Secondly,to solve the problem that the algorithm is prone to tracking failure,the object re-detection mechanism is introduced to train the SVM to reposition the occluded object.Finally,considering the object scale changes greatly in the complex video environment,the scale pyramid is used to construct a onedimensional scale correlation filter,which predicts the object size accurately.The experimental results in the standard database OTB-2013 show that the MRCR algorithm improves the tracking accuracy and robustness through acquiring the context information around the object,using the re-detection mechanism and multi-scale.And then we introduce the dynamic strategy fusion tracking based on multi-layer convolutional features.Focusing on the problem that the single-layer apparent model of convolutional network in the tracking algorithm can not fully represent the object,the features of multi-layer networks are used to represent the object information.In the process of capturing the object,we firstly model the object through the layer of conv4-3,conv4-4,conv5-3 and conv5-4 from the VGGNet-19 network.Then the multi-layer features are performed to calculate the corresponding kernelized correlation filter responses,and the position with the maximum response value indicates the predicted object position of each tracker.Finally,the object position in all responses information are dynamically merged to locate the object through the dynamic strategy fusion algorithm.We experiment in the standard database OTB-2013.The results show that the DSFM algorithm has excellent tracking performance.Finally,we systematically comb the main innovations and analyze the problems of the proposed algorithm in this thesis.The next step is also clarified.