Markerless Motion Analysis from Synchronized 2D Camera Views: A Convolutional Neural Network Approach

In this work, we propose a CNN-based methodology for the clinically accurate tracking of a human subject. Making use of a state-of-the-art convolutional pose estimator, joint center locations were independently evaluated within the 2D views produced by synchronized calibrated cameras, and lifted to 3D space by means of triangulation. The precision of the resulting pose estimates was then increased through an original refinement routine, leveraging subject-specific kinematic information.