Dimension reduction analysis for the recognition of human physical activities using multimodal fusion

Abstract

Automatic recognition of human activities is an important task in computer vision applications. Robust approaches using one or more sensors generally link redundant features that consume computing resources and computation time during the sub-activity classification process. This article explores dimension reduction for recognizing human activities and primitive movements by merging data from Kinect visual depth sensors, IMUs inertial sensors, and electrodes of electromyographic record (EMGs). It shows a comparative study where different techniques of dimension reduction from the state of the art are evaluated, and their behavior is analyzed based on activity recognition performance and the computation time of the methods. The results show that state-of-the-art methodologies could have lower temporal costs when implementing them without significantly affecting performance when recognizing the activity.