Accurate, Real-time 3D Reconstruction of Human Postures from Orthogonal Images Based on Neural Networks(Internationa Session 5)

概要

論文の詳細を見る
This paper deals with a real-time method for reconstructing 3D posture parameters of the human from orthogonal images by using artificial neural network techniques with higher accuracies than conventional computer vision based methods. To sample the data used for training the neural network, a human wearing the sensing devices of a motion capture system is synchronously observed by orthogonal cameras so that at each frame the motion capture data as well as the positions of the located feature points of the human areas in the camera images are obtained at the same time. During the training phase, the positions of the located feature points and the motion capture data are respectively used for the input and output vectors of the neural network. During the reconstruction phase, the positions of the located feature points at each frame of the video sequence to be reconstructed are inputted into the trained neural network so that the inputted data at each frame are converted to the motion capture data. The proposed method is implemented on a personal computer and runs in real-time (30 frames/sec). Experimental results confirm both the feasibility and the effectiveness of the proposed method for estimating 3D human postures.
社団法人電子情報通信学会の論文
2007-10-18