Although numerous online learning strategies have beenproposed to handle the appearance variation in visualtracking, the existing methods just perform well in certaincases since they lack effective appearance learning mechanism. In this paper, a joint model tracker (JMT) is presented, which consists of a generative model based on MultipleSubspaces and a discriminative model based on improvedMultiple Instance Boosting (MIBoosting). The generativemodel utilizes a series of local constructed subspacesto update the Multiple Subspaces model and considersthe energy dissipation of dimension reduction in updatingstep. The discriminative model adopts the GaussianMixture Model (GMM) to estimate the posterior probabilityof the likelihood function. These two parts supervise eachother to update in multiple instance way which helps ourtracker recover from drift. Extensive experiments on variousdatabases validate the effectiveness of our proposedmethod over other state-of-the-art trackers.
展开▼