Recently, Deep Learning (DL) techniques have shown their effectiveness for Human Activity Recognition (HAR) tasks. However, due to the storage and computational requirements, most of the existing HAR solutions assume that the training and inference phases are offloaded to the cloud or an external server, with a harmful impact on the network load of the mobile/wearable device as well as on the user’s privacy. A promising solution is represented by the emerging Edge Artificial Intelligence (AI) techniques that aim at moving the data analytics closer to the sensing units or directly on them. In this paper, we present our preliminary results about the offloading of HAR inference tasks on low-power microcontroller units. We consider the problem of detecting critical movements (e.g. falling, running) of workers within an industrial environment for safety purposes. The full pipeline of the HAR system is presented by using an Arduino BLE 33 Sense as a wearable unit: for the detection task, a DL model based on a Convolutional Neural Networks (CNN) is trained on the inertial sensor data. A dynamic range quantization technique is used to reduce the size of the model which is then loaded on the firmware. Preliminary results show that the accuracy of the CNN model is 97% and overcomes baseline, non-DL techniques, while the quantization technique ensures a reduction of 53% of the model size.
Ghibellini A., Bononi L., Di Felice M. (2022). Intelligence at the IoT Edge: Activity Recognition with Low-Power Microcontrollers and Convolutional Neural Networks. NY : IEEE [10.1109/CCNC49033.2022.9700665].
Intelligence at the IoT Edge: Activity Recognition with Low-Power Microcontrollers and Convolutional Neural Networks
Ghibellini A.;Bononi L.;Di Felice M.
2022
Abstract
Recently, Deep Learning (DL) techniques have shown their effectiveness for Human Activity Recognition (HAR) tasks. However, due to the storage and computational requirements, most of the existing HAR solutions assume that the training and inference phases are offloaded to the cloud or an external server, with a harmful impact on the network load of the mobile/wearable device as well as on the user’s privacy. A promising solution is represented by the emerging Edge Artificial Intelligence (AI) techniques that aim at moving the data analytics closer to the sensing units or directly on them. In this paper, we present our preliminary results about the offloading of HAR inference tasks on low-power microcontroller units. We consider the problem of detecting critical movements (e.g. falling, running) of workers within an industrial environment for safety purposes. The full pipeline of the HAR system is presented by using an Arduino BLE 33 Sense as a wearable unit: for the detection task, a DL model based on a Convolutional Neural Networks (CNN) is trained on the inertial sensor data. A dynamic range quantization technique is used to reduce the size of the model which is then loaded on the firmware. Preliminary results show that the accuracy of the CNN model is 97% and overcomes baseline, non-DL techniques, while the quantization technique ensures a reduction of 53% of the model size.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.