Consistent Labeling Algorithm Research

Consistent labeling is a fundamental problem which frequently arises in computer vision. Consistent labeling can be divided into two parts: sparse labeling and dense labeling. In this paper, a robust sparse labeling based on the local optimal description is proposed. With this sparse labeling algorithm, we present a novel dense labeling algorithm which takes the sparse robust information into the dense frame.This paper gives detail information about the design of algorithms in different application situations.1. Point matching is typical example of sparse labeling, the local robust search is applied to enhancing the performance of the global labeling, and this local search is called the local optimal description. The results quantitatively show that the proposed approach outperforms state-of-the-art methods with the standard datasets which contain the situations of deformation, occlusion, rotation, noise and outliers.2. With the robust sparse labeling algorithm, a new dense labeling method which takes the sparse point's information into the dense frame is tested with the PIV images, the results show that the proposed dense labeling algorithm can handle the situation in which the image information is less robust.3. This paper build the consistent constraints into the three hot fields: background subtraction, multi-object tracking and mesh tracking. The traditional background models use less spatial information, based on this situation, a new background model with consistent constraints is proposed, the detailed results show the proposed background model is more accurate and robust than the VIBe background model, and the propose model can choose the threshold adaptively. Multi-object tracking algorithm gives wrong results when it is designed to track different objects with similar appearance, this paper build the spatial consistent constraints into the multi-object tracking frame, the results show that the proposed multi-object tracking can discriminate different objects with similar appearacne, even can predict the position of occluded objects. Mesh tracking is the key step of object pose measurement, a new mesh tracking based on the spatial constraint is proposed to handle the partial occlusion problem, this paper test the proposed algorithm with simulation experiment about the aircraft pose measurement, the experiment shows that the proposed mesh tracking alogrithm is robust enough to handle the illumination variation, large pose variation and partial occlusion.