PostDoc (2) 

The second part of my PostDoc deals with further aspects of human motion estimation. The multi-view human motion estimation scenario is visualized below: I assume the knowledge of a 3D object model including joint locations and observe it in images of different calibrated cameras. The aim is to estimate the rigid motion and joint angles wich lead to a best fit between the object model and the images.

I am interested in: Using a multi-view scenario and dealing with partial- and self-occlusions Implementing morphing techniques (local, global, linear and non-linear ones) Comparing the silhouette based pose estimation algorithms with a marker based tracking system

This page gives the state of the art of my research:

Multi-view Based Pose Estimation

The left figure shows an example of a stereo sequence. The figure is splitted in the left camera (top) and right camera (bottom). Each part shows the original image in the upper left, and the used corner features in the lower left. The images in the middle show pose results overlaid with the original images and the right images show pose results in a virtual environment. See (here) for a recent publication.

Morphing during Human Motion Estimation

The left figure shows an example for morphing techniques: Its left image shows the morphed joint transformed arms, and the right one the non-morphed arms. The angles of the upper arms steer the amount of morphing during lowering or raising the shoulders. Here a global morphing is applied, but I have also implemented local morphing techniques by using radial basis functions (RBFs). The left motion (see also the video) appears much more natural. See (here) for a recent publication.

Multi-View Tracking

The left figure shows example images taken from a 4-camera sequence in the GAIT-Lab. As can be seen, we are able to deal with occlusions and self- occlusions during pose estimation (the person rotates around 180 degrees during the sequence!). Also the morphing techniques from above are applied.

Sports Movement Analysis

The left figure shows further example images taken from 4-camera sequences in the GAIT-Lab. As can be seen, we are able to track complex motion patterns to analyse e.g. push ups or sit ups.

Future/Actual work

• Tracking in complex environments (Outdoor scenes, changing lighting conditions, no-homogeneous background).
• Modeling more complex object models (adding legs, etc.).
• Tracking of multiple humans (e.g. dancing couples).