Redes Neurais de Grafos aplicadas ao Reconhecimento de Gestos

Abstract

In the recent years the Graph Neural Networks have been spreadly used in applications envolving machine learning, including graph classification, molecular property predictions and pattern recognition. This paper presents a new methodology for action classification using skeleton graphs, that is composed by three blocks: a graph constructor, a feature extractor and an action classifier. The first block constructs a graph for each frame. The feature extraction is executed by a Graph Neural Network which modifies the feature vectors of the nodes through the information of the node itself and neighborhood nodes. These feature vectors are concatenated and generate an embedding to represent the frame. Finally. the video representation is an pseudoimage composed by the frame embeddings. Such pseudoimage is the data imput for the classifier based on convolucional and dense layers.We carried out experiments with the Chalearn dataset and we achieved 91.38% of accuracy. Even with one single Graph Neural Network layer and few parameters to be trained, the results were competitives in relation to papers in the literature with more complex models.