A Novel Approach for Human Activity Recognition Utilizing Modified Convolutional Neural Networks and Long Short-Term Memory Architectures

Abstract

The use of Human Activity Recognition in smart homes and health monitoring is gaining widespread acceptance as critical components of intelligent systems. State-of-the-art HAR approaches heavily rely on handcrafted features and less advanced learning techniques which might not capture human behavior's complex underlying patterns. This paper presents an approach for the HAR that integrates LSTMs and CNNs. In this model, LSTMs focus on disentangling timing, and ordering information within these features while CNNs learn to extract spatially distributed feature information in a fully automated manner. The recognition of different human activities is better with this fusion approach in terms of accuracy and reliability. Several experiments were conducted to verify the effectiveness of the proposed strategy using OWN and UCF-50 datasets.