Scene Video Text Tracking with Graph Matching
Wei-Yi Pei,Chun Yang,Li-Yu Meng,Jie-Bo Hou,Shu Tian,Xu-Cheng Yin
DOI: https://doi.org/10.1109/access.2018.2797181
IF: 3.9
2018-01-01
IEEE Access
Abstract:Video has become one of the dominant data resources with the development of the Internet. As a result, the structured sorting of videos, which can be used for storage and extraction, represents a growing concern in the community. In particular, the text within videos can carry rich semantic information, leading to many novel studies wherein text tracking and recognition are performed. One essential step in text tracking involves template matching. In general, the adjacent matrices are modeled to represent the extracted tracking object features. Then, often, the Hungarian algorithm is applied to find the correspondence pairs between consecutive frames. In many works, text features are extracted based on morphological features such as color histograms and aspect ratios. However, under those features, similar text objects are not sufficiently distinguishable to make a distinction between them. To address this issue, we regard the template matching task as a graph matching problem. The main novelty involves a graph matching approach that utilizes the relationship between two trajectories or two objects, whereby a graph matching solver can be readily used in our tracking system. By utilizing the content information, the mismatch between the same object among different frames is effectively reduced. The experimental results demonstrate that the tracker with the graph matching method tends to increase the valid correspondence of trajectories and candidate objects.