This project focuses on convex methods for human movement understanding, the task of estimating and quantifying human motion and movement in videos. Compared with previous approaches, the convex methods explicitly model complex inter-component correlations and global constraints and are able to achieve more reliable results. The convex formulations are automatically generated using learning methods. By taking advantage of their special structures, efficient algorithms are devised. This project studies human movement understanding from different perspectives and provides convex solutions to global movement estimation, body part tracking, and local patch motion trajectory estimation. The key research components include convex articulated graph matching with complexity decoupled from the sizes of target candidate sets, convex structure learning and efficient decomposition methods for solving large-scale problems. With robust movement estimation, this project further addresses performance recognition, a new application that quantifies the style and performance of actions.
This project benefits surveillance, robotics, human computer interaction, entertainment, sports and medical applications. The convex framework developed in this project can also be extended to many other areas to construct optimal models for information retrieval, control systems, scheduling and network analysis, and to solve large-scale problems in operational research and simulation.
This project develops a new technology for understanding human movements; a benchmark dataset is also developed. The results are disseminated by peer reviewed publications, web page downloads, and by university courses.
