This project will focus on reconstructing the 3D motion of football players from real matches as opposed to situations acted out in a studio environment. The figure below displays the expected developments within this project relative to existing systems.
The figure shows the characteristics - quality of visualization, production time and level of automation - of the systems currently in use and those we can develop within in this project while still advancing the frontier of markerless human motion capture.
The approach we propose to adopt within this project is as follows. We intend to build a hierarchy of increasingly quantitatively accurate 3D reconstructions of the motion in a clip. At the lowest level of the hierarchy, corresponding to the most basic 3D reconstruction, the trajectory of the player on the ground plane will be mapped to a prototypical 3D reconstruction of a single action such as walking, running or shooting. Without any prior information, it is reasonable to assume that these trajectories will be the most reliable indication of the performed action and therefore the 3D reconstruction of the motion. This basic qualitative reconstruction will then act as a guide to the extraction of a more detailed representation of the player’s appearance in the image at the next level of hierarchy. This more detailed representation will then be mapped to a more accurate 3D reconstruction. However, due to the generalization limitations of the learned mappings at some level of the hierarchy it will be necessary to apply a model based reconstruction method if a certain accuracy is required. At this stage though the measurement process, for instance skeletal joint localization in the image, may have a fighting chance of success due to the quality of the existing knowledge of the underlying 3D motion.