Trajectory Based Human Pose Estimation Using Convolutional Neural Network
This first animation shows the linear force validation trajectory (solid model) and the network’s best estimate of pose after 1 second increments.
This demo shows the model performing fairly well when applied to a totally new type of trajectory data- that of a time varying noisy endpoint force exerted at the hand. Despite the fact that the network was trained entirely on constant linear hand force motion, the model was able to still estimate the pose of human (with some outliers) to a reasonable degree of accuracy. This is encouraging because is suggests that a more advanced rigid body model similar to this one trained entirely in simulation could be applied to real world data.
