Human Pose Estimation is one of the most promising research areas in the field that links together Artificial Intelligence and Computer Vision. Its applications are countless since it mimics one of the principal human capabilities: understanding the physical space that surrounds us, and the interaction of the context with people. Sport is the field that can most benefit from this technology since it involves dynamic movements of individuals.
The project is the research and development of a fully-working application that exploits human pose estimation to perform nonlinear queries to retrieve the exact pose and movements that the user is looking for. I applied this framework to climbing, in which the training phase involves mostly figuring out how to perform a certain pose by looking at how other climbers have done it.
The system uses computer vision techniques to detect and track human keypoints in climbing videos. By indexing these keypoints and their temporal relationships, the application allows users to search for specific movements or positions without having to watch entire videos.
The main components of the system include:
The application successfully demonstrates the concept of nonlinear video querying for climbing videos. Users can find specific movements across a database of climbing videos by sketching or selecting poses, without having to scrub through entire videos manually.
View the slides from the final defense
