Research On Object Tracking Based On Multiple Instance Learning Model

A popular research topic among computer vision is Object Tracking, with variousapplications in robot navigation, human-machine interaction and virtual reality. The objecttracking research is also challenging, especially in scenarios where the lighting, occluding andnon-rigid deformation happen. In recent literature, Multiple Instance Learning (MIL) is aunique method that shows excellent performance. Based on the original MIL, the distributionfield and covariance matrix are employed for the improving of appearance model in objecttracking. The main contents of this research are listed as follows：(1) Research of object tracking algorithm based on distribution field and MIL (DF_MIL).It employs a combination of Distribution Field and Multiple Instance Learning. In this method,images are expanded into Distribution Fields and then Gaussian smoothed through featuredomain and spatial domain, so that the domain of attraction is expanded while the targetinformation is maintained. The discriminative classifier is implemented using online multipleinstance learning, which can deal with inaccurate positive samples in the updating process andallow some flexibility while finding a decision boundary. As a result, drifting fromaccumulation of tracking error can be avoided. Combining with the boosting framework,weak classifiers are selected in the distribution field to form a strong classifier, for updating ofnew appearance model in an on-line manner.(2) Research of object tracking algorithm based on Covariance Matrix and MIL(COV_MIL). Covariance Matrix is introduced into MIL tracking to improve the performanceof target representation. When classifying samples, not only the confidence value ofappearance model but also the COV response value of the samples on the several mostobvious features are employed through a weighting process. The location of object is thenindicated by the max weighted value. Besides, tracking results of the recent frames are addedto the positive bag to increase the reliability of instances.Comparison results of the method proposed in this thesis and the traditional MILmethods show that DF_MIL has stronger performance in selection of weak classifiers whereproblems such as occluding deformation and lighting change etc are overcome. Results alsoshow that with the combination of Harr and Covariance Matrix in COV_MIL so that the ability of target representation and robustness in complex scenes are improved significantly.