@MASTERSTHESIS\{IMM2011-06121,
    author       = "M. S. Christiansen",
    title        = "Markerless motion capture for biomechanical applications",
    year         = "2011",
    school       = "Technical University of Denmark, {DTU} Informatics, {E-}mail: reception@imm.dtu.dk",
    address      = "Asmussens Alle, Building 305, {DK-}2800 Kgs. Lyngby, Denmark",
    type         = "",
    note         = "Supervised by Associate Professor Erik Simonsen at Division of Biomechanics, University of Copenhagen and Associate Professor Henrik Aan{\ae}s, haa@imm.dtu.dk, at {DTU} Informatics",
    url          = "http://www.imm.dtu.dk/documents/ftp/ep2011/MartinSandauChr_MasterThesis.pdf",
    abstract     = "Motion capture is widely used for gait analysis. Today most motion capture systems are marker based, which is both time consuming and imprecise. Many markerless approaches have been presented in the past years and the majority of the approaches are based on articulated models. One of the most promising approaches is presented in (Corazza, et al., 2006), in which a subject specific articulated models is applied. However the subject specific model is obtained by a laser scanner that is expensive to purchase. This master thesis presents an approach to obtain a full body model by using a photogrammetric approach that costs a fraction of a laser scanner. The proposed approach is based on the PhotoModeler® Scanner software and eight cameras. Tests of the precision and the impact of varying spatial resolutions have been performed. The results of the precision tests showed that a 3±1 mm resolution can be obtained with 10 Mega pixel {SLR} cameras, if ideal illumination and texture is obtained. The precision obtained is comparable to the laser scanned models used in (Corazza, et al., 2006).
In addition, a pose estimation approach for markerless motion capture system is proposed to replace the subject specific articulated model approach. The advantage of replacing the subject specific articulated models is that the expenses and time consumption of synthesizing the models are avoided. The key elements of the approach are as follows:
· Acquire detailed {3D} models of the test subject by using a Patch Based Multi-view stereo algorithm proposed in (Furukawa, et al., 2010).
· Track the {3D} points in the model over time and perform full body segmentation into rigid parts.
· Use the rigid segments to calculate the joint centers. 
Since the novelty of the approach is the pose estimation represented in the second item, this thesis has focused on finding a method to solve this problem. The approach is tested on various {3D} models of flexing limbs. The registrations were promising for small angular differences, but failed when the differences became large. Improvement of the registration algorithm would therefore be an obvious objective for a future work. However the final results of the test illustrated satisfying segmentations."
}