Dynamic Classifier for Nonrigid Human Motion Analysis.pdf

Dynamic Classifier for Non-rigid Human Motion Analysis Huang Fei, Ian Reid Department of Engineering Science, University of Oxford, Oxford,UK. fei@robots.ox.ac.uk Abstract Automatic analysis (parsing) of non-rigid human motion in a cluttered out- door enviroment is a useful but challenging task. In a single view point, the lack of depth order relations causes a major ambiguity of the object identities. Coupled with the non-rigidity of articulation, 3D human motion tracking/pose estimation in one view is a formidable problem. In this pa- per, we present a novel solution that directly address this depth ambiguity, in which we extend a discriminative analysis (Support Vector Machine (SVM)) to non-rigid human motion classification with a temporal generative motion model (Hidden Markov Model (HMM)). This method can discriminate dy- namic depth ordering as well as 3D articulated motion automatically from 2D images. Experiments with this method have demonstrated promising results. 1 Introduction Automatic interpretation (parsing) of non-rigid human motion is essential to motion esti- mation, gait pattern analysis and behaviour understanding. A number of motion parsing algorithms have been put forward in related computer vision research areas: object de- tection [5],[6], visual tracking [7] and behaviour understanding [8]. In these fields, ma- chine learning methods have received increasing attention in recent years. Discriminative- learning based approaches such as SVM or AdaBoost have achieved promising results in face [13] or people [6] detection; Sophisticated motion inference algorithms like Particle filtering [11], Meanshift [12], etc. have demonst