Ballroom dance step recognition by means of video processing and Artificial Intelligence techniques.

Abstract

[Abstract]: Ballroom dancing, or DanceSport is a sport of great complexity due to the number of dance steps that make up every type of dance. This project tries to serve of help to dancers of all levels, both to amateur and professional ones as well as to competition judges in order to evaluate in a more objective manner. To do so, Artificial Intelligence techniques are applied to classify the different dance steps. There were filmed some Rumba choreographies by professional dancers and other videos were collected from external sources. Features from the videos of the datasets are extracted and selected, in terms of body keypoints and movement features for said keypoints to train the models. In order to do so, it is necessary to label all the dance steps in each video, indicating its start and ending times. Thereafter, models are trained with different Artificial Intelligence algorithms, including Multi-Layer Perceptron and Long Short-Term Memory Recurrent Neural Networks, and different feature inputs and parameters. Each model is tested against data from different datasets, such as the custom or the reference dataset. Finally, performance metrics are calculated for each model, and compared against the others.[Resumo]: Baile deportivo ou baile de salón é un deporte de gran complexidade debido ao número de pasos que compoñen cada tipo de baile. Este proxecto trata de servir de axuda a bailaríns de todos os niveis, tanto a principiantes como a profesionais e tamén a xuíces de competición para avaliar de maneira máis obxectiva. Para iso, aplicáronse técnicas de Intelixencia Artificial para clasificar os distintos pasos de baile. Graváronse algunhas coreografías de Rumba bailadas por profesionais e outros vídeos foron recopilados de recursos externos. Se extraen e seleccionan características dos vídeos dos datasets en termos de keypoints do corpo e características de movemento para ditos keypoints para entrenar os modelos. Para poder facer iso, é preciso etiquetar todos os pasos de baile en cada vídeo, indicando o comezo e final. Posteriormente, entrénanse os modelos cos diferentes algoritmos de Intelixencia Artificial, sendo estes Multi-Layer Perceptron e Long Short-Term Memory Recurrent Neural Networks, e diferentes entradas de características e parámetros. Cada modelo é comprobado contra datos de diferentes datasets, como o propio deste proxecto ou o de referencia. Finalmente, se calculan métricas de rendemento para cada modelo, e son comparadas entre elas.