Research Of Object Tracking Based On Video Sequences

With the development of computer technology and machine learning theory,the target tracking which is the key technology in the field of computer vision have made successful applications in military and civil fields such as robot technology,intelligent monitoring,human-computer interaction and intelligent guidance.And with the development of research,a large number of excellent target tracking algorithms have emerged.The correlation filter-based trackers have attracted great attention since its appearance due to its high efficiency and high precision,and the researchers also develop a large number of research improvements based on these algorithms.But challenging factors in the complex video environment such as illumination variation,partial occlusion and motion blur would change object appearance and affect the ability of appearance model to describe the target,thus,the performance of target tracking algorithm is restricted.To solve the problem contained in the correlation filter-based tracker,two improved algorithms are proposed from the establishment of robust target model: the first one is long-term tracking based on multi-feature adaptive fusion for video target and another one is scale adaptive tracking with convolutional features.In this thesis,we first summarize the research background,research significance and research status of object tracking.We introduce the basic framework of correlation filter-based tracker and summarize the complex tracking video environment.Then we introduce the long-term tracking based on multi-feature adaptive fusion.The approach establishes a robust appearance model by fusing multiple features adaptively including HOG,LBP and CN at response map level to solve the problem that the appearance model established by single feature is not robust in complex video environment.Meanwhile,we train a random ferns classifier as re-detector to re-detect the target in case of tracking failure caused by prolonged occlusion or out of view.The experimental results in OTB demonstrate that the algorithm has good performance.And then we introduce the scale adaptive tracking with convolutional features.Because the traditional features adopted in correlation filter-based trackers cannot effectively obtain the semantic information of the target and have limited expression ability,we establish the appearance model by using the convolution feature with high discriminanted ability.And we use principal component analysis to reduce the high-dimensional convolution features to improve the tracking speed.And meanwhile,we construct a scale pyramid at the position of target to train an independent correlation filter to estimate the scale of object.We validate and analyze the algorithm in OTB,The experimental results demonstrate that the algorithm performs well because the appearance model established by convolution feature can effectively improve the accuracy,and the scale estimation mechanism can effectively estimate the scale of the target to improve the success rate.Finally,we summarize the main content in this thesis,point out the shortcoming of the algorithm and look forward to the future research direction.