Vida Groznik (2010) Acquisition and processing of sensory data for learning to recognize dance steps. EngD thesis.
The long-term goal of the project, part of which is the current thesis, is to build a dance tutor. This involves learning to recognize basic dance elements, specifically various dance steps. To be able to do this, we must be able to record and preprocess the motion data into machine (and also human) understandable form --- this is the topic of the present thesis. The motion data is obtained using Animazoo motion capture equipment and software. The equipment is based on 19 gyroscopes that record the rotations of the major joints of the human body. For the purposes of the envisioned tutoring system the recorded data needs to be as precise as possible --- less noise generally enables more precise models to be learned. To this end, it was necessary to prepare a precise skeleton of the actor being recorded (the dance expert), find a place with as stable a magnetic field as possible, and also correctly and precisely position the gyroscopes on the actor. The processing of the data mainly involved transforming rotation data into position information in the 3D environment. This is nominally an easy step, however it is subject to various (and plentiful) notations and conventions regarding the axis direction and orientation, making it a time consuming process. The thesis provides ample and precise documentation how to do these transformations on data acquired with the Animazoo equipment. The final processing step is to automatically divide the recorded data into individual dance steps. Several simple (in terms of logic and the number of sensors involved) methods are proposed as well as several measures to determine the quality of said methods. One method is proposed as clearly best for further use, however, if needed, it might be further enhanced, subject to its performance in the next stage of building an intelligent tutor.
Actions (login required)