Human Action Recognition from Motion Capture Data Using Machine Learning Techniques

Human action analysis is an actively developing area of research. A successful solution to the problem of action analysis can contribute to a number of tasks in various areas of human activity, such as diagnostics in medicine, maintaining public safety, developing human-machine interfaces, training, etc. In this paper, we solve the problem of recognizing human actions based on motion capture data obtained using special means of recording human movements. This allows us to analyze motion based only on information about the position of the controlled parts of a human body on which the sensors are attached. To recognize actions, we use a classical pattern recognition scheme consisting of feature extraction and classification stages. At first, we preprocess initial motion data and generate feature descriptions in a reduced space. Then we use classical machine learning methods to classify extracted features with the subsequent aggregation using simple voting. In our experiments, we use the Berkeley Multimodal Human Action Database as the initial data and achieve a recognition accuracy of 91.6%, which is more than 10% better than the baseline technique.