Amelia Maddox

Date of Award

2021

Document Type

Thesis

Degree Name

Bachelors

Department

Natural Sciences

First Advisor

Gillman, David

Area of Concentration

Computer Science

Abstract

Dance is the art of translating meaning and emotion into symbolic sequential movement. This thesis aims to model such creativity in the form of an Artificial Intelligence (AI) choreographer; given a word, the AI generates sequences of movement that capture the “feel” or emotion of the word. The AI has two main components: sentiment analysis of the word and translating the inferred sentiment into body movement, the latter of which is further broken down into a classification model and a generation model. The backbone of the word analysis is handled by a pre-trained model from a third party. The classification and generation of dance is handled by a random forest (RF) classifier and custom LSTM-RNN, respectively, trained on motion capture data from various contemporary dances. The location features of the motion capture data were relativized before normalizing both the location and the rotation features. This preprocessing technique and the decision to use a RF classifier was determined through a performance comparison of 5 classification models handling 15 differently preprocessed motion capture data. A user study to measure the realism of the artificially choreographed dances was constructed, though unfortunately not conducted due to time constraints. Prior to development on the AI, a survey of professional contemporary dancers was conducted to understand the dance genre and its relationship with emotion. The goal of this thesis was to develop an AI which not only choreographed realistic dances, but replicable ones too. Therefore, the dances were visualized using 3D animation software and pilot testing was conducted to gather feedback on modeling techniques which enhanced perceived replicability.

Download

Share

COinS