Multi-View Human Activity Recognition in Ambient Assisted Living Using Lightweight Deep Learning Models

Human Activity Recognition (HAR) is crucial for the development of intelligent assistive technologies in Ambient Assisted Living (AAL) environments. This paper proposes an innovative method for Multi-View Human Activity Recognition (MV-HAR) using lightweight deep learning models, specifically MobileNet and Cyclone-CNN (CCNet), to achieve quick and precise activity detection. Utilizing the Robot House Multi-View Human Activity Recognition (RHM-HAR) dataset, which contains four different views-front, back, ceiling (omni), and mobile robot-our models effectively address challenges related to viewpoint variation and motion dynamics. The dataset includes 14 multi-view daily living action classes, providing a balanced set of synchronized human actions suitable for multi-domain neural network learning. MobileNet and CCNet are employed for their high recognition accuracy, computational efficiency, and real-time application capabilities in AAL scenarios. We propose a Mutual Information (MI)-based method to assess the redundancy and relevance of each viewpoint, ensuring the fusion of multi-view data with minimum redundancy and maximum relevance. Benchmarking results demonstrate that multi-view combinations significantly enhance recognition performance compared to single-view models, particularly in complex activities involving high levels of movement.