1 / 1

we consider situations in which the object is unknown

4 nd INTERNATIONAL SYMPOSIUM ON IMPRECISE PROBABILITIES AND THEIR APPLICATIONS, ISIPTA’05 Carnegie Mellon University, Pittsburgh, USA, July 20-23 2005. EVIDENTIAL MODELING FOR POSE ESTIMATION Fabio Cuzzolin, Computer Science Department, UCLA; Ruggero Frezza, DEI, Universita’ di Padova.

zaria
Download Presentation

we consider situations in which the object is unknown

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. 4nd INTERNATIONAL SYMPOSIUM ON IMPRECISE PROBABILITIES AND THEIR APPLICATIONS, ISIPTA’05 Carnegie Mellon University, Pittsburgh, USA, July 20-23 2005 EVIDENTIAL MODELING FOR POSE ESTIMATION Fabio Cuzzolin, Computer Science Department, UCLA; Ruggero Frezza, DEI, Universita’ di Padova TRAINING A BOTTOM-UP MODEL • MODEL-FREE POSE ESTIMATION • BOTTOM LINES • we want to estimate the configuration (pose) of an • unknown object from a sequence of images • model-free pose estimation requires to build maps from • image measurements (features) to poses • different features have to be integrated to increase • accuracy and robustness of the estimation • the evidential model provides such a framework • pose estimation: reconstruction of the actual pose of a moving object by processing the sequence of images taken during its motion • model-based pose estimation: a kinematic model of the body is known and used to help the estimation • we consider situations in which the object is unknown • the only way of doing pose estimation is then building a map between image measurements (features) and configurations (poses) directly from the data • this can be done in a training stage, where an approximate parameter space is collected as ground truth configuration values produced by an “oracle” (motion capture) • example: Rehg and Kanade • pose = angles between • links of the fingers • TRAINING : the body moves in front of the camera(s), while • a sequence of poses is provided by a motion capture system. Then • some features are computed from the images • these feature sequences are passed to an HMM with n states yielding: • the approximate feature spaces • the maps between features and poses • EVIDENTIAL MODEL • FEATURE-POSE MAPS AND HMMs 1 • hidden Markov models (HMMs) provide a way to the learn feature-pose maps • automatically through the EM algorithm • for each state of the HMM a Gaussian likelihood is set up on the feature range • this partitions the feature range into n regions (approximate feature space) • and is associated with a refining from Y to Q • the collection of approximate parameter • space and feature spaces, linked by the • refining maps learned in the training stage • form an evidential model of the object • the evidential model can then be used • thereafter to estimate the pose of the • body when new images become available • ESTIMATION : the body performs new movements in front of the camera(s), • and for each available image • the features are computed as before • the likelihoods of the features are transformedinto belief functions • those measurement functions are projected onto Q and combined • a pointwise estimate of the object pose is computed by pignistic or relative plausibility transformation 2 • PERFORMANCES • EXPERIMENTS ON HUMAN BODY TRACKING • FEATURE EXTRACTION • two views acquired • through DV cameras • pose estimate for the leg exp • the estimate of component 9 of the pose vector shows a • neat improvement when using the comprehensive model 1 3 • comparison between visual • estimate and real view 2 • the silhouette of the body of interest is detected by colorimetric analysis • the bounding box containing the object is found • feature vector = collection of the coordinates of the vertices of the box • two experiments: four markers on right arm, eight markers on legs • we built evidential models for the two separate views and • compared them with an overall model

More Related