Federated recognition mechanism based on enhanced temporal-spatial learning using mobile edge sensors for firefighters

Abstract

Abstract Background Interest in Human Action Recognition (HAR), which encompasses both household and industrial settings, is growing. HAR describes a computer system’s capacity to accurately recognize and evaluate human activities and behaviors, akin to what humans call perception. Real-time federated activity identification architecture is suggested in this work to monitor smartphone user behavior. The main aim is to decrease accidents happening in an indoor environment and assure the security of older individuals in an indoor setting. The idea lends itself to a multitude of uses, including monitoring the elderly, entertainment, and spying. Results In this paper, we present a new smartphone sensor-based human motion awareness federated recognition scheme using a temporal-spatial weighted BILSTM-CNN framework. We verify new federated recognition based on temporal-spatial data better than existing machine learning schemes in terms of activity recognition accuracy. Several methods and strategies in the literature have been used to attain higher HAR accuracy. In particular, six categories of typical everyday human activities are highlighted, including walking, jumping, standing, moving from one level to another, and picking up items. Conclusion Smartphone-based sensors are utilized to detect the motion activities carried out by elderly people based on the raw inertial measurement unit (IMU) data. Then, weighted bidirectional long short-term memory (BILSTM) networks are for learning about temporal motion features; they are swiftly followed by single-dimensional convolutional neural networks (CNN), which are built for reasoning about spatial structure features. Additionally, the awareness mechanism highlights the data segments to choose discriminative contextual data. Finally, a sizeable dataset of HDL activity datasets is gathered for model validation and training. The results confirm that the proposed ML framework performs 18.7% better in terms of accuracy, 27.9% for the case of precision, and 0.24.1% when evaluating based on the F1-score for client 1. Similarly, for client 2 and client 3, the performance betterment in terms of accuracy is 18.4% and 10.1%, respectively.