376
A. Jmal et al.
accelerometer, gyroscope etc., while the activity is being performed with the help
of Artificial Intelligence methods. In recent years, several machine learning and
deep learning algorithms for human activity recognition have been proposed.
Sukor et al. [2] used several methods of machine learning such as Support Vector Machine (SVM), Decision Tree (DT), and Multiple Layer Perception-Neural
Network (MLP-NN) to classify activities such as slow sitting, standing, upstairs,
downstairs and lying using the accelerometer sensor embedded in a smartphone.
The obtained results show that the use of Principal Component Analysis (PCA)
to reduce the dimensionality of features obtains higher recognition rate with the
rate of 96.85% for the DT algorithm and 100% for the MLP-NN algorithm which
may have over fitting problems. G. McCalmont et al. [3] also tested the activity
recognition performance. Three classifiers were used, including Artificial Neural
Network (ANN), K-Nearest Neighbor (KNN) and Random Forest (RF) to classify five exercises which are slow walking, normal walking, fast walking, upstairs
and down stairs using accelerometer, gyroscope and magnetometer. They found
that ANN models with many layers achieve an accuracy of 80% while RF and
KNN achieve an accuracy slightly above 70%. Song-Mi Lee et al. presented a
RF algorithm and achieve an accuracy of 89.1% [4]. Furthermore, three human
activity data, walking, running, and staying still, are gathered using smartphone
accelerometer sensor and classified with Convolutional Neural Network (CNN)
and had better performance 92.71% [4]. Furthermore Abdulmajid Murad et al.
[5] use a 3D accelerometer, 3D gyroscope and 3D magnetometer to classify six
activities. Four algorithms Extreme Learning Machine (ELM), SVM, CNN and
RNN are used to classify these activities. The best accuracy was achieved with
the RNN algorithms 96.7%. It could be considered that the ANN is one of the
best machine learning algorithm used for HAR and the RNN is reported to
overperform other deep learning algorithms in term of accuracy and precision
to recognize human activities. In addition, most of research in the field, validate their results with simulations, without comparing theses simulations with
results provided by real-time embedded and hardware based implementations.
So a lack of standalone, sensor based HAR systems, with embedded machine
learning and real-time response is remarked. This research aims is to compare
simulation results with results provided by a real-time implementation and to
judge performance gived by embedded ANN to recognize human activities. The
rest of this paper is arranged as follows: The second section introduces the ANN
architecture and process. In addition, an overview of the LSTM (Long Short
Term Memory) Recurrent Neural Network is presented in the third section. The
next section presents the data acquisition structure for HAR with the database
properties. Furthermore, an evaluation of ANN and LSTM-RNN using Receiver
Operating Characteristic (ROC) are presented. Moreover, to validate our simulations results, the developed ANN model is implemented in a Raspberry PI as
a real-time standalone HAR system.
A. Jmal et al.
accelerometer, gyroscope etc., while the activity is being performed with the help
of Artificial Intelligence methods. In recent years, several machine learning and
deep learning algorithms for human activity recognition have been proposed.
Sukor et al. [2] used several methods of machine learning such as Support Vector Machine (SVM), Decision Tree (DT), and Multiple Layer Perception-Neural
Network (MLP-NN) to classify activities such as slow sitting, standing, upstairs,
downstairs and lying using the accelerometer sensor embedded in a smartphone.
The obtained results show that the use of Principal Component Analysis (PCA)
to reduce the dimensionality of features obtains higher recognition rate with the
rate of 96.85% for the DT algorithm and 100% for the MLP-NN algorithm which
may have over fitting problems. G. McCalmont et al. [3] also tested the activity
recognition performance. Three classifiers were used, including Artificial Neural
Network (ANN), K-Nearest Neighbor (KNN) and Random Forest (RF) to classify five exercises which are slow walking, normal walking, fast walking, upstairs
and down stairs using accelerometer, gyroscope and magnetometer. They found
that ANN models with many layers achieve an accuracy of 80% while RF and
KNN achieve an accuracy slightly above 70%. Song-Mi Lee et al. presented a
RF algorithm and achieve an accuracy of 89.1% [4]. Furthermore, three human
activity data, walking, running, and staying still, are gathered using smartphone
accelerometer sensor and classified with Convolutional Neural Network (CNN)
and had better performance 92.71% [4]. Furthermore Abdulmajid Murad et al.
[5] use a 3D accelerometer, 3D gyroscope and 3D magnetometer to classify six
activities. Four algorithms Extreme Learning Machine (ELM), SVM, CNN and
RNN are used to classify these activities. The best accuracy was achieved with
the RNN algorithms 96.7%. It could be considered that the ANN is one of the
best machine learning algorithm used for HAR and the RNN is reported to
overperform other deep learning algorithms in term of accuracy and precision
to recognize human activities. In addition, most of research in the field, validate their results with simulations, without comparing theses simulations with
results provided by real-time embedded and hardware based implementations.
So a lack of standalone, sensor based HAR systems, with embedded machine
learning and real-time response is remarked. This research aims is to compare
simulation results with results provided by a real-time implementation and to
judge performance gived by embedded ANN to recognize human activities. The
rest of this paper is arranged as follows: The second section introduces the ANN
architecture and process. In addition, an overview of the LSTM (Long Short
Term Memory) Recurrent Neural Network is presented in the third section. The
next section presents the data acquisition structure for HAR with the database
properties. Furthermore, an evaluation of ANN and LSTM-RNN using Receiver
Operating Characteristic (ROC) are presented. Moreover, to validate our simulations results, the developed ANN model is implemented in a Raspberry PI as
a real-time standalone HAR system.