Data Processing, Retrieval And Reconstruction For Human Motion Capture

Human motion capture is a rising technique for data acquisition and it is a multi-disciplinary intersection of mathematics,computer graphics,image processing,data processing and so on.In the past decades,many researchers at home and abroad focused on motion capture and obtained lots of fruitful achievements.The technique has been widely used in modern computer animation,film making,virtual game,medical analysis and physical training.The study on motion capture is an important issue with not only theoretical significance but also application value.Since 1990's,with the development of the motion capture,many fruitful results can be found in solutions to various practical problems such as research in capture pattern,design and development of motion capture system,data processing algorithms and relevant software, management and retrieval methods for motion data reuse,comprehension and analysis of motion data,motion reconstruction and so on.The main works of this thesis includes three aspects as follows:1.Scattered data processing for optical motion capture:a data processing method based on rigid structure matching and a dynamic impulse noise model based on semantic parts for noise and missing data processing are presented.According to the development and practical problems of internal industry,the proposed rigid structure matching-based method resolves the problems of recognition,matching and reconstruction of scattered makers by five matching types and topological checking on the hypothesis of articulated rigid structure of human body.Furthermore,by the definition of semantic joint,semantic part and candidate part,the dynamic impulse noise model constructed based on semantic parts rather than traditional trajectories of markers filters the noise data and reconstructs the missing markers by use of median filter algorithm.Finally,data processing system is developed on the basis of algorithm and experiments show that the approaches can resolve the problems with better results,robustness and efficiency.2.Feature representation of human motion and motion data retrieval:feature representation and extraction in high dimension space for human motion are proposed,and then a motion data retrieval algorithm is implemented based on biomimetic pattern recognition.To resolve the problem of feature representation extraction and retrieval of human motion,semantic joint and semantic model are defined,a semantic similarity approach is examined.Moreover,by feature extraction,a method of hyper-sphere cover in high dimension space is presented with evaluation of the cover domain of same type motions' feature vectors.Furthermore,based on biomimetic pattern recognition,improved hyper sausage neurons and hyper sausage neuron chain in high dimension space are constructed for same type motions to cover the feature domain for motion data retrieval.The experimental system examines the proposed algorithms with the CMU free motion database,and results show that the methods achieve better retrieval recall and precision with satisfied efficiency.3.Data driven approaches for human body and facial motion reconstruction:by the combination of motion capture data and NURBS,two data driven approaches for human body and facial motion reconstruction are implemented respectively based on local coordinate interpolation and NURBS mapping.For human body motion simulation,a piece-wise NURBS based human body model is constructed and the relationship between NURBS control points and semantic joints are determined subsequently.Then,the morph of human model is driven based on local coordinate interpolation algorithm by semantic model which extracted from motion data.Especially for facial motion,a MPEG-4 compatible NURBS surface facial model is constructed and then the morph of facial model is driven by motion data using local coordinate decomposition and NURBS mapping.The simulation feasibility and efficiency are proved by experimental systems.To sum up,efficient and practicable data processing methods are proposed,and experimental systems are developed.The robustness and efficiency of methods are examined based on mass experimental data.The study of these problems provides the arithmetic and feasible basis for design and development of software with independent intellectual property.