Human Activity Recognition using deep LSTM ar...

Human Activity Recognition using deep LSTM architecture

Author Name : Chada Bhanu Prakash Reddy, Harsha Vardhan Thokala, Asrith Raghavendra Madanala

ABSTRACT

Since the past decade, inertial sensors have been dramatically decreased in size, making them easy to include in smartphones. Human actions may be recognised by utilising various methods, such as machine learning. This article describes a system that utilises accelerometer and gyroscope data acquired from a smartphone as inputs to a long short-term memory (LSTM) network, which is then used as the primary source of human activity recognition (HAR). Six actions that humans do were discovered: sitting, standing, lying down, walking, climbing stairs, and going downstairs. Various network topologies were put through a grid search in an attempt to optimise. Laying down was determined to be the simplest action to identify, while standing was shown to be the most difficult. Going downstairs and walking upstairs (two linked actions) may be detected properly thanks to the ability of stacked LSTM networks to process past and future information of a signal. The method suggested by HAR achieved an overall accuracy of 92.3 percent.

Keywords: human activity recognition, long short-term memory, sensors, classification