Data Imputation Method for Human Motion Capture from Monocular Images

H. Cao, T. Tanaka, and N. Sugie (Japan)


Motion Capture, Pose Estimation, Data Imputation


Recently the research of human motion capture from monocular images receives increasing attention inspired by the prospect of multimedia applications. However no proposed technique can deal with practical circumstances well up to now. In this paper we propose a subspace learning based method can handle the complicated motions and have capability as a pre-estimation for further refinement. It begins with a Principal Component Analysis (PCA) subspace learning on a set of training hybrid data encoded by 2D silhouette and corresponded human pose. The human pose is regarded as the missing part of the complete hybrid data and then can be somehow inferred from the learned subspace. Motivated by some ideas of statistical data survey, an incremental PCA data imputation method is incorporated to strengthen the power of inference. Subsequently a multiresolutional extension of this method is suggested in order to handle the case that the appearance of input 2D measurement is quite different from that of training data. Performance of the method is evaluated with synthetic video sequences of various motions. For dealing with human motion capture from monocular images (figure 1), there are two fundamental approaches being usually used: motion tracking, pose estimation. Motion Tracking is a strategy to find corresponding objects in successive frames. The difficulties of this task are related to the complexity of the articulated body. The correspondence analysis is often supported by prediction. Based on previously detected objects and possible high level knowledge, the state of the object is predicted and metrically compared with the state found in the actual image. Prediction introduces a region of-interest in both image space and state space. The commonly used methods for prediction involve the Kalman Filter and CONDENSATION.

Important Links:

Go Back