Developing a pipeline for gait analysis with a side-view depth sensor

This thesis presents computational methods for conducting gait analysis with a sideview depth sensor. First, a method to segment human body parts in a depth image is presented. A standard supervised segmentation algorithm is run on a novel graph representation of the depth image. It is demonstrated that the new graph structure improves the accuracy of the segmentation. This contribution is intended to allow fast labelling of depth images for training a human joint predictor. Next, a method is presented to select accurate 3D positions of human joints from multiple proposals. These proposals are generated by a predictor from a side-view depth image. Finally, a gait analysis system is built on the joint selection process. The system calculates standard parameters used in clinical gait analysis. Walking trials have been measured concurrently by a pressure-sensitive walkway and a side-view depth sensor. The estimated gait parameters are validated against the ground truth parameters from the walkway. As future work, the initial segmentation process could be applied to multi-view depth images for training a view-invariant joint predictor. The proposed gait analysis system can then be applied to the predicted joints