Real-Time Gesture Recognition Using Deep Learning Towards Alzheimer’s Disease Applications

Abstract

There have been significant efforts in the direction of improving accuracy in detecting human action using skeleton joints. Determining actions in a noisy environment is still challenging since the Cartesian coordinate of the skeleton joints provided by depth sense camera depends on camera position and skeleton position. In a few of the human-computer interaction applications, skeleton position and camera position keep changing. The proposed method recommends using relative positional values instead of actual Cartesian coordinate values. Recent advancements in the Convolution Neural Network (CNN) help us achieve higher prediction accuracy using image format input. To represent skeleton joints in image format, we need to represent skeleton information in matrix form with equal height and width. With some depth sense cameras, the number of skeleton joints provided is limited, and we need to depend on relative positional values to have a matrix representation of skeleton joints. We can show near the state-of-the-art performance on MSR 3-Dimensional (3D) data and the new representation of skeleton joints. We have used image shifting instead of interpolation between frames, which helps us have state-of-the-art performance.