The detection of human motion has numerous applications in areas such as film production, animation, medical analysis, sports science and natural human-computer interaction. Traditionally, systems are used to track the 3D position of markers glued to the body and reconstruct the body movement from this information. Such systems are expensive, difficult to install and the movements can only be recorded in a small recording area.
The Institute for Information Processing (TNT) is researching novel methods to reconstruct the motions in arbitrary environments and without special markers. Apart from standard cameras, inertial sensors worn on the body are used, as well. In addition to recording the movements, methods are being researched to estimate the underlying forces and moments in the musculoskeletal system of the human body.
Monocular camera:
In a project for the acquisition of human motion based on a single moving camera, optimization methods are being developed which reconstruct not only posture but also camera movement and additionally estimate anthropometric variables of the actor. This enables the automatic recording of human motion with a standard camera, as found e.g. in smartphones.
Inertial sensors and sensor fusion:
In addition to camera-based approaches, reconstruction methods of human motion using inertial sensors are being researched. The sensors are worn on the body underneath the clothing and thus allow easy recording of movement in everyday situations and are suitable for long-term recordings [2]. In addition, fusion approaches are being developed to compensate for sensor uncertainties by combining them with image information from video data. This increases the accuracy and robustness of motion reconstruction.
Physical modeling:
A further research focus is the physical modeling and analysis of the recorded motion. The used methods include both, traditional approaches, such as forward and inverse dynamics, and machine learning approaches. The reconstruction of the acting forces and moments can be used, for example, to assess movements with respect to their efficiency or load level.
Body:
pose estimation, body shape models, subspace projections, auto encoder, physical models, deep learning, generative adversarial networks, sensor fusion, IMUs
Faces:
Face detection, emotion recognition, extraction of distinctive features, 3D face reconstruction, processing of 3D meshes, motion capture, pose estimation, visual speech synthesis, virtual avatars
