Region-based analysis and retrieval for tracking of semantic objects and provision of augmented information in interactive sport scenes

摘要

This paper introduces techniques for segmentation and tracking based on analysis of region derived descriptors. By partitioning the image into a series of homogeneous regions, the problem of object extraction changes from pixel based to database analysis. A region based approach has distinct advantages over pixel. In particular, it has low dimensionality, is resilient to rotation, shear, and photometric changes, preserves boundaries, gives implied segmentation, provides comprehensive description, has reduced drift, and permits detailed analysis. The method is amenable to the introduction of prior knowledge permitting the simplification of segmentation, which is otherwise an ill-posed problem. Objects are defined as collections of contiguous regions with known statistics, allowing their identification by performing correlations in the database. The interframe difference of region based sequences is high for low-motion scenes due to segmentation noise, but performs consistently better for high motion, making it suitable for tracking. When applied to tracking the method proves robust, occlusion insensitive, and unlike the other techniques, it re-establishes lost lock.