TY  - THES
AU  - Ignasi Rius
ED  - Jordi Gonzalez
ED  - Xavier Roca
PY  - 2010//
TI  - Motion Priors for Efficient Bayesian Tracking in Human Sequence Evaluation
PB  - Ediciones Graficas Rey
N2  - Recovering human motion by visual analysis is a challenging computer vision researcharea with a lot of potential applications. Model-based tracking approaches, and inparticular particle lters, formulate the problem as a Bayesian inference task whoseaim is to sequentially estimate the distribution of the parameters of a human bodymodel over time. These approaches strongly rely on good dynamical and observationmodels to predict and update congurations of the human body according to measurements from the image data. However, it is very dicult to design observationmodels which extract useful and reliable information from image sequences robustly.This results specially challenging in monocular tracking given that only one viewpointfrom the scene is available. Therefore, to overcome these limitations strong motionpriors are needed to guide the exploration of the state space.The work presented in this Thesis is aimed to retrieve the 3D motion parametersof a human body model from incomplete and noisy measurements of a monocularimage sequence. These measurements consist of the 2D positions of a reduced set ofjoints in the image plane. Towards this end, we present a novel action-specic modelof human motion which is trained from several databases of real motion-capturedperformances of an action, and is used as a priori knowledge within a particle lteringscheme.Body postures are represented by means of a simple and compact stick guremodel which uses direction cosines to represent the direction of body limbs in the 3DCartesian space. Then, for a given action, Principal Component Analysis is applied tothe training data to perform dimensionality reduction over the highly correlated inputdata. Before the learning stage of the action model, the input motion performancesare synchronized by means of a novel dense matching algorithm based on DynamicProgramming. The algorithm synchronizes all the motion sequences of the sameaction class, nding an optimal solution in real-time.Then, a probabilistic action model is learnt, based on the synchronized motionexamples, which captures the variability and temporal evolution of full-body motionwithin a specic action. In particular, for each action, the parameters learnt are: arepresentative manifold for the action consisting of its mean performance, the standard deviation from the mean performance, the mean observed direction vectors fromeach motion subsequence of a given length and the expected error at a given timeinstant.Subsequently, the action-specic model is used as a priori knowledge on humanmotion which improves the eciency and robustness of the overall particle filtering tracking framework. First, the dynamic model guides the particles according to similarsituations previously learnt. Then, the state space is constrained so only feasiblehuman postures are accepted as valid solutions at each time step. As a result, thestate space is explored more eciently as the particle set covers the most probablebody postures.Finally, experiments are carried out using test sequences from several motiondatabases. Results point out that our tracker scheme is able to estimate the rough3D conguration of a full-body model providing only the 2D positions of a reducedset of joints. Separate tests on the sequence synchronization method and the subsequence probabilistic matching technique are also provided.
SN  - 978-84-937261-9-5
N1  - exported from refbase (http://refbase.cvc.uab.es/show.php?record=1331), last updated on Fri, 17 Dec 2021 13:58:09 +0100
ID  - Ignasi Rius2010
U1  - Ph.D. thesis
ER  -