Research And Realization On Object Tracking Algorithm Using Compressive Hybrid Features

In recent years, human beings are eager to that machines could adapt to the habits of people, which can perceive environmental information and carry out human-computer interaction by the ways of vision, auditory or session. Among these approaches, vision is one important way to realize environmental awareness, which further leads to the rapid development of computer vision technology, and plays an important role in the field of industry, smart home and others relating to robot. Target tracking, as an important part of computer vision technology, has a wide range of requirements and application in kinds of fields like surgery robot, unmanned vehicles, virtual three-dimensional somatosensory games, electronic makeup and so on.In the area of tracking, an algorithm called as the online multi-instance learning tracking(MIL) algorithm is known for its ability of alleviating tracking drift by training classifiers with positive and negative bag recently. However, this algorithm still has two shortcomings:(1)the increased computational complexity results in time consuming due to the lack of consideration of sampling importance when collecting training samples. Additionally,(2)the MIL method, as a 2D feature-based tracking algorithm, performs unsteadily when the object changes poses or rotates seriously in the real 3D tracking scenes.In the present thesis, aiming at the deficiency of the MIL algorithm, a new method, which is called research and realization on object tracking algorithm using compressive hybrid features is proposed.(1)Firstly, a histogram-based feature similarity measurement, as a weighting strategy to select positive samples, is described to reduce the time complexity and solve the problem of the sudden occlusion in the course of selecting samples.(2)Additionally, during the course of tracking, the proposed tracking algorithm is able to improve tracking performance by utilizing features in the depth map. In this process, depth information is used to detect the approximate location of the object at first, and then the very accurate location of the object depending on this approximate location is detected.(3)An obvious cost of adding the depth dimension is the increased time complexity. To this end, during the course of the feature selecting, a sparse measurement matrix is adopted to efficiently extract the features for the appearance model and decrease the feature dimensionThe comparing experimental results, in term Center locating error, success rate and frame rate, demonstrate that the algorithm in this thesis is better in robustness, accuracy, efficiency than six state-of-the-art methods on six challenging video sequences.