Deep Learning and SVM-Based Method for Human Activity Recognition with Skeleton Data

Abstract

In recent years, research related to the analysis of human activity has been the subject of increased attention by engineers dealing with computer vision, and particularly that which utilizes deep learning. In this paper, we propose a method for classification of human activities, composed of 3D skeleton data. This data is normalized beforehand and represented in two forms, which are fed to a neural network with parallel convolutional and dense layers. After the network is trained, the training data is propagated again to infer the output from the second last layer. This output is used for training a Support Vector Machine. All hyperparameters were found using the Bayesian Optimization strategy on the PKU-MMD dataset. Our method was tested on the UTD-MHAD dataset, achieving an accuracy of 92.4%