"Motion capture is a technique that has recently gained a lot of attention in the domain of understanding human motion through computer vision. This thesis takes a machine learning based approach to analyze 3D posture and movement from motion capture data. It makes use of the motion capture data, from eight camera Vicon motion capture system, to derive mathematical models that allow the recovery of full body configurations directly from it. The approach is completely data-driven. This makes the inference extremely fast. This thesis describes a pipeline to reconstruct raw motion capture marker position data into the animation of the motion, which requires skeleton fitting, gap filling, data smoothing and inverse kinematics. We have formulated a GUI based algorithm that helps in recovering occluded data and filling the gap. The GUI uses Kalman Filter to extrapolate missing marker data and then fills the gaps. Using a host of classifiers to distill a large training database of motion capture that estimates the 3D configurations of human body posture to classify the data into general posture and abnormal posture.
The method relies on using appropriately developed robust motion capture features that help the learning models to clearly distinguish general from abnormal human motion. This is done with the help of 30 selected features from a list of 285 mathematically calculated features that generalizes posture prediction. The classification accuracy achieved using these features in various scenarios pertaining to multiple classification algorithm is more than 97%. Keywords: Motion capture system, Vicon, Nexus, Machine learning, Classification, Anthropometry, Body segment parameters, Gap Filling."