Visual Tracking Via Incremental Lorentzian Discriminant Projection

Visual tracking is a key problem in the field of computer vision. The main challenge of visual tracking is that the target undergoes large appearance, illumination, expression, pose, scale and occlusion variation indoor and outdoor. In recent years, many efficient visual tracking algorithms have been developed such as on-line boosting based tracking algorithm, semi-supervised on-line boosting based tracking algorithm, on-line multiple instance learning based tracking algorithm and fragments based tracking algorithm using integral histogram. Through the research, researchers found that the subspace based tracking methods performed distinguishingly. Because they are good at handling a tracking problem in which the target undergoes large appearance, illumination, expression, pose and scale variation. Recently, researchers have developed many efficient subspace based tracking algorithms like incremental PC A based tracking algorithm, incremental FLD based tracking algorithm and incremental MMC based tracking algorithm by applying the dimensionality reduction methods to the visual tracking problems. But they all contained neither the local structure or the global structure of data when they trained a subspace.Lorentzian Discriminant Projection (LDP) is a manifold learning based linear dimensionality reduction algorithm. Besides the promising discriminant characteristic of LDP, it reflects the inner local structure and the geometry of global data structure that makes LDP a promising algorithm for visual tracking. In this paper, a few classic visual tracking algorithms were studied and summaried at the beginning. Then a variant of LDP called Incremental Lorentzian Discriminant Projection (ILDP) was proposed for efficiently applying LDP to visual tracking problems. In the process of tracking, our tracker incrementally updated the subspace and the sample mean. Our proposed method is able to track the target with large variation, such as large appearance, illumination, expression, pose and scale variation indoor and outdoor. Various experiments are conducted to demonstrate the high performance of our method.