Human activity classification is assuming great relevance in many fields, including the well-being of the elderly. Many methodologies to improve the prediction of human activities, such as falls or unexpected behaviors, have been proposed over the years, exploiting different technologies, but the complexity of the algorithms requires the use of processors with high computational capabilities. In this paper different deep learning techniques are compared in order to evaluate the best compromise between recognition performance and computational effort with the aim to define a solution that can be executed by an IoT device, with a limited computational load. The comparison has been developed considering a dataset containing different types of activities related to human walking obtained from an automotive Radar. The procedure requires a pre-processing of the raw data and then the feature extraction from range-Doppler maps. To obtain reliable results different deep learning architectures and different optimizers are compared, showing that an accuracy of more than 97% is achieved with an appropriate selection of the network parameters.
Assessment of Human Activity Classification Algorithms for IoT Devices / Ciattaglia, G.; Senigagliesi, L.; Gambi, E.. - 456:(2023), pp. 170-181. (Intervento presentato al convegno 9th EAI International Conference on IoT Technologies for HealthCare, HealthyIoT 2022 tenutosi a prt nel 2022) [10.1007/978-3-031-28663-6_13].
Assessment of Human Activity Classification Algorithms for IoT Devices
Ciattaglia G.
;Senigagliesi L.;Gambi E.
2023-01-01
Abstract
Human activity classification is assuming great relevance in many fields, including the well-being of the elderly. Many methodologies to improve the prediction of human activities, such as falls or unexpected behaviors, have been proposed over the years, exploiting different technologies, but the complexity of the algorithms requires the use of processors with high computational capabilities. In this paper different deep learning techniques are compared in order to evaluate the best compromise between recognition performance and computational effort with the aim to define a solution that can be executed by an IoT device, with a limited computational load. The comparison has been developed considering a dataset containing different types of activities related to human walking obtained from an automotive Radar. The procedure requires a pre-processing of the raw data and then the feature extraction from range-Doppler maps. To obtain reliable results different deep learning architectures and different optimizers are compared, showing that an accuracy of more than 97% is achieved with an appropriate selection of the network parameters.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.