An Embedded ANN Raspberry PI
383
provided from the Raspberry, are very close, which rounds our prototype reliable and robust model. Moreover, to evaluate the performance of the considered
approaches, we compare our models to other research works. We find that our
trained models are more robust in term of accuracy compared to McCalmont,
G. et al. [3] which achieve an accuracy above 80% for ANN and 70% for the
KNN. Moreover, we compare our trained LSTM-RNN model which reached of
an accuracy performance 99%, to Mi Lee, S. et al. [4] and Abdulmajid Murad
et al. [5] which use a Convolutional Neural Network (CNN) algorithm and
(Sequential ELM, SVM, CNN and RNN) respectively, the best accuracy was
achieved with the RNN-LSTM algorithms (99%) The focus of next paper will be
on a the implementation of LSTM-RNN in the Raspberry PI with a comparative
study between ANN and LSTM-RNN for real-time HAR.
6 Conclusion
In this paper, a deep learning algorithm named Recurrent Neural Network
(RNN) with LSTM memory units and keras was tested using accelerometer
and gyroscope to classify and analyze human activities such as sitting, walking
and running. This produced an overall accuracy of 99%. Furthermore, we have
exported the ANN model using accelerometer and gyroscope to be implemented
in a Raspberry PI. In addition, we have tested the model with data collected
from the MPU6050 and we have successfully shown that the model provides good
results in term of real-time prediction and classification with 86% of accuracy.
For the future work direction, an IoT smart device of human activity recognition
based on embedded deep learning will be developed. In addition, the deep learning algorithm in the medical fields to implement a real-time fall detection system
and anomaly detection system for elderly monitoring, and disease prevention will
be investigated.
References
1. Ramasamy, S.M., Roy, N.: Recent trends in machine learning for human activity recognition-a survey. Wiley Interdisc. Rev.: Data Mining Knowl. Discov. 8(4),
e1254 (2018)
2. Sukor, A.S.A., Zakaria, A., Rahim, N.A.: Activity recognition using accelerometer
sensor and machine learning classifiers. In: IEEE 14th International Colloquium
Signal Processing and its Application, CSPA 2018, no. March, pp. 233–238 (2018)
3. McCalmont, G., et al.: eZiGait: toward an AI gait analysis and sssistant system. In:
2018 IEEE International Conference on Bioinformatics and Biomedicine (BIBM),
Madrid, Spain, pp. 2280–2286. IEEE (2019)
4. Lee, S.M., Yoon, S.M., Cho, H.: Human activity recognition from accelerometer
data using convolutional neural network. In: 2017 IEEE International Conference
on Big Data and Smart Computing (BigComp), Jeju, South Korea, pp. 131–134.
IEEE (2017)
5. Murad, A., Pyun, J.Y.: Deep recurrent neural networks for human activity recognition. Sensors 17(11), 2556 (2017). https://doi.org/10.3390/s17112556
383
provided from the Raspberry, are very close, which rounds our prototype reliable and robust model. Moreover, to evaluate the performance of the considered
approaches, we compare our models to other research works. We find that our
trained models are more robust in term of accuracy compared to McCalmont,
G. et al. [3] which achieve an accuracy above 80% for ANN and 70% for the
KNN. Moreover, we compare our trained LSTM-RNN model which reached of
an accuracy performance 99%, to Mi Lee, S. et al. [4] and Abdulmajid Murad
et al. [5] which use a Convolutional Neural Network (CNN) algorithm and
(Sequential ELM, SVM, CNN and RNN) respectively, the best accuracy was
achieved with the RNN-LSTM algorithms (99%) The focus of next paper will be
on a the implementation of LSTM-RNN in the Raspberry PI with a comparative
study between ANN and LSTM-RNN for real-time HAR.
6 Conclusion
In this paper, a deep learning algorithm named Recurrent Neural Network
(RNN) with LSTM memory units and keras was tested using accelerometer
and gyroscope to classify and analyze human activities such as sitting, walking
and running. This produced an overall accuracy of 99%. Furthermore, we have
exported the ANN model using accelerometer and gyroscope to be implemented
in a Raspberry PI. In addition, we have tested the model with data collected
from the MPU6050 and we have successfully shown that the model provides good
results in term of real-time prediction and classification with 86% of accuracy.
For the future work direction, an IoT smart device of human activity recognition
based on embedded deep learning will be developed. In addition, the deep learning algorithm in the medical fields to implement a real-time fall detection system
and anomaly detection system for elderly monitoring, and disease prevention will
be investigated.
References
1. Ramasamy, S.M., Roy, N.: Recent trends in machine learning for human activity recognition-a survey. Wiley Interdisc. Rev.: Data Mining Knowl. Discov. 8(4),
e1254 (2018)
2. Sukor, A.S.A., Zakaria, A., Rahim, N.A.: Activity recognition using accelerometer
sensor and machine learning classifiers. In: IEEE 14th International Colloquium
Signal Processing and its Application, CSPA 2018, no. March, pp. 233–238 (2018)
3. McCalmont, G., et al.: eZiGait: toward an AI gait analysis and sssistant system. In:
2018 IEEE International Conference on Bioinformatics and Biomedicine (BIBM),
Madrid, Spain, pp. 2280–2286. IEEE (2019)
4. Lee, S.M., Yoon, S.M., Cho, H.: Human activity recognition from accelerometer
data using convolutional neural network. In: 2017 IEEE International Conference
on Big Data and Smart Computing (BigComp), Jeju, South Korea, pp. 131–134.
IEEE (2017)
5. Murad, A., Pyun, J.Y.: Deep recurrent neural networks for human activity recognition. Sensors 17(11), 2556 (2017). https://doi.org/10.3390/s17112556