In this thesis, we introduce novel approaches for increasing the stability, accuracy, and efficiency of marker-less human motion tracking and 3D human pose reconstruction. As one common underlying concept, the presented approaches contain a retrieval component making use of previously recorded marker-based motion capture (mocap) data. In particular, we contribute to three different areas dealing with various types of sensors including video cameras, optical mocap systems, inertial sensors, and depth cameras. Firstly, we introduce content-based retrieval techniques for automatically segmenting and annotating mocap data that is originally provided in form of unstructured data collections. Secondly, we show how such robust annotation procedures can be used to support and stabilize video-based marker-less motion tracking. Thirdly, we develop algorithms for reconstructing human motions from noisy depth sensor data in real-time. In all these contributions, a particular focus is put on efficiency issues in order to keep down the run time.