Abstract: A model-based filtering approach for real time human motion data [DE]
Felix Paul, A model-based filtering approach for real time human motion data, Hochschule Düsseldorf, Bachelor thesis, 14.6.2018.
Acquiring human motion data from video images plays an important role in the field of computer vision. Ground truth tracking systems require markers to create high quality motion data. But in many applications it is desired to work without markers. In recent years affordable hardware for markerless tracking systems was made available at a consumer level. Efficient depth camera systems based on Time-of-Flight sensors and structured light systems, have made it possible to record motion data in real time. However, the gap between the quality of marker-based and markerless systems is significant. In this thesis, markerless motion tracking systems and its applications for virtual studio live productions are discussed with an emphasis on the necessary quality needed, to create a robust system. Also the error sources of the markerless motion tracking pipeline are considered and a filtering framework proposed. The proposed method is then proven to be more robust and accurate than the unfiltered data stream and can be used to visually enhance the presence of an actor within a virtual environment in live broadcast productions.
Kinect, noise, latency, markerless motion tracking, pose estimation, Holt double exponential smoothing filter, virtual studio, shadows
Prof. Jens Herder, Dr. Eng./Univ. of Tsukuba
Prof. Dr. Christian Geiger
The research took place at the Virtual Sets and Virtual Environments Laboratory.