Development of a neural network based markerless motion capture system

Markerless motion capture technology has the potential to revolutionize the way biomechanics researchers collect data. By moving data collection out of the lab, new research opportunities previously unavailable will be opened to the greater biomechanics community. Our work utilizing a combination of deep convolutional neural networks (dCNN) and recurrent neural networks (RNN) and unique methodology for training data collection has shown the accuracy potential of a markerless motion capture system utilizing four off the shelf video cameras. Accuracy was compared to gold standard Vicon measurements with root mean squared errors (RMSE) of between 2-3 degrees. This work has shown the accuracy potential of marker less systems and lays the groundwork for developing a more refined system that can be deployed in multiple environments.