To achieve the above result , in the video clips, the skeletal joints of the player in each frame in the two views were marked by hand. From this data it was possible to automatically reconstruct the 2D data in 3D and produce a qualitatively nice and pleasing visualization.
To achieve the above result, in the video the location of the skeletal joints of the player were marked by hand and their corresponding 3D reconstruction in all a small proportion of the frames. From this data it was possible to automatically reconstruct the rest of the sequence.
As can be seen the reconstruction, though qualitatively accurate, is not as precise as the one achieved from the fully manual one using two views in the previous example. However, because of the reduced demands on the user we are able to reconstruct longer sequences.
 |
Here, at CVAP we have created a multi-camera system that allows us to video wide-angle footage of events and in particular football fields. Thus we have complete views of a football field at decent resolutions. At every frame in the video the location of each player is found. Then the problem is to associate the identities from one from to the next.
The above video shows the result of applying our identity association algorithm to football player. Displayed is one player being tracked over several minutes. The video is fast forwarded while he is isolated and he is bounded by a black rectangle when he is indistinguishable (via or person detection algorithm) from a neighbouring player.