Motion capture data processing, retrieval and recognition.

Abstract

Character animation plays an essential role in the area of featured film and computer games. Manually creating character animation by animators is both tedious and inefficient, where motion capture techniques (MoCap) have been developed and become the most popular method for creating realistic character animation products. Commercial MoCap systems are expensive and the capturing process itself usually requires an indoor studio environment. Procedural animation creation is often lacking extensive user control during the generation progress. Therefore, efficiently and effectively reusing MoCap data can brings significant benefits, which has motivated wider research in terms of machine learning based MoCap data processing. A typical work flow of MoCap data reusing can be divided into 3 stages: data capture, data management and data reusing. There are still many challenges at each stage. For instance, the data capture and management often suffer from data quality problems. The efficient and effective retrieval method is also demanding due to the large amount of data being used. In addition, classification and understanding of actions are the fundamental basis of data reusing. This thesis proposes to use machine learning on MoCap data for reusing purposes, where a frame work of motion capture data processing is designed. The modular design of this framework enables motion data refinement, retrieval and recognition. The first part of this thesis introduces various methods used in existing motion capture processing approaches in literature and a brief introduction of relevant machine learning methods used in this framework. In general, the frameworks related to refinement, retrieval, recognition are discussed. A motion refinement algorithm based on dictionary learning will then be presented, where kinematical structural and temporal information are exploited. The designed optimization method and data preprocessing technique can ensure a smooth property for the recovered result. After that, a motion refinement algorithm based on matrix completion is presented, where the low-rank property and spatio-temporal information is exploited. Such model does not require preparing data for training. The designed optimization method outperforms existing approaches in regard to both effectiveness and efficiency. A motion retrieval method based on multi-view feature selection is also proposed, where the intrinsic relations between visual words in each motion feature subspace are discovered as a means of improving the retrieval performance. A provisional trace-ratio objective function and an iterative optimization method are also included. A non-negative matrix factorization based motion data clustering method is proposed for recognition purposes, which aims to deal with large scale unsupervised/semi-supervised problems. In addition, deep learning models are used for motion data recognition, e.g. 2D gait recognition and 3D MoCap recognition. To sum up, the research on motion data refinement, retrieval and recognition are presented in this thesis with an aim to tackle the major challenges in motion reusing. The proposed motion refinement methods aim to provide high quality clean motion data for downstream applications. The designed multi-view feature selection algorithm aims to improve the motion retrieval performance. The proposed motion recognition methods are equally essential for motion understanding. A collection of publications by the author of this thesis are noted in publications section.