Real-time Motion capture of multiple persons in community videos
Tracking multiple persons in 3D with high accuracy and temporal stability in real-time with monocular RGB camera is a challenging task which has a lot of practical applications like 3D human character animation, motion analysis in sports, modeling human body movements and many others. The optical human tracking methods often require usage of multi-view video recordings or depth cameras. Systems which work with monocular RGB cameras are mostly not in real-time, track single person and require additional data like initial human pose to be given. All this implies a lot of practical limitations and is one of the major reasons why optical motion capture systems have not yet seen more widespread use in commercial products. The DFKI research department Augmented Vision presents a novel fully automatic multi-person motion tracking system. The presented system works in real-time with monocular RGB video and tracks multiple people in 3D. It does not require any manual work or a specific human pose to start the tracking process. The system automatically estimates a personalized 3D skeleton and an initial 3D location of each person. The system is tested for tracking multiple persons in outdoor scenes, community videos and low quality videos captured with mobile-phone cameras.