HAPT datasets provide a large extracted features extracted by prepossessing the raw
signals generated from sensors.
3.2 Results
These algorithms are tested under MATLAB environment and the WSVM algorithm is
tested with implementation LibSVM [25] using Gaussian kernel is used for all the
datasets. Each training dataset is normalized before classification within a range of
[−1, 1]. We optimized the SVM hyper-parameters (r, C) for all training sets in the
range [0.1, 0.2, 0.5, 1] and {0.1, 1, 5, 10, 100}, respectively, to maximize the error rate
of five fold- cross validation technique. The optimal parameters r opt = 0.9, 0.9, and 0.8
are found to be optimal the training dataset of HAR, HAPT, and WISDM, respectively.
We show in the Table 2 that the fusion of principal component features with WSVMHMM makes the model more robust, achieving better performance. One also notices
for HAR dataset that the multi-class WSVM method improves the classification results
over MC-SVM, MC-HF-SVM and HMM classifiers used alone. On the other hand, the
results also show that WSVM outperforms HMM for recognizing activities for all
datasets except for the HAPT dataset.
In terms of reducing the datasets, the feature reduction identifies the most relevant
features for the learning process. We notice that PCA features can improve the discrimination between different activities than the original features. For WISDM the
performances of activity recognition are low than HAR and HAPT datasets with 561
features. This is explained by the number of features (6) for WISDM is not sufficient
when using PCA algorithm. Another reason to the lowest accuracy in WISDM dataset
is attributed to the use only the accelerometer sensor comparatively to the HAR and
HAPT that use the both accelerometer and gyroscope sensors.
Table 2. The micro-averaged measures: Recall, Precision, F-measure and Accuracy for all
approaches in (%). Bold values are the results for our approach for each dataset.
Datasets Approach
Recall Precision F-measure Accuracy
HAR
MC-SVM [8]
89.6 89.9
89.7
89.3
MC-HF-SVM [8]
89.3 89.2
89.2
89.0
WSVM
92.4 91.6
91.9
93.9
HMM
89.2 90.2
89.7
93.7
Proposed
94.0 96.7
95.3
94.9
HAPT
WSVM
96.0 92.4
94.1
86.1
HMM
98.3 97.1
97.7
96.5
Proposed
97.3 99.0
98.1
96.8
WISDM J48 [26]
81.7 –
–
85.1
LogisticRegression [26] 68.4 –
–
78.1
MultilayerPerceptron [26] 80.4 –
–
91.7
WSVM
83.4 76.5
79.8
81.4
HMM
79.4 80.0
79.7
84.9
Proposed
91.9 79.8
85.4
92.3
Human Activities Recognition in Android Smartphone Using WSVM-HMM Classifier
391
signals generated from sensors.
3.2 Results
These algorithms are tested under MATLAB environment and the WSVM algorithm is
tested with implementation LibSVM [25] using Gaussian kernel is used for all the
datasets. Each training dataset is normalized before classification within a range of
[−1, 1]. We optimized the SVM hyper-parameters (r, C) for all training sets in the
range [0.1, 0.2, 0.5, 1] and {0.1, 1, 5, 10, 100}, respectively, to maximize the error rate
of five fold- cross validation technique. The optimal parameters r opt = 0.9, 0.9, and 0.8
are found to be optimal the training dataset of HAR, HAPT, and WISDM, respectively.
We show in the Table 2 that the fusion of principal component features with WSVMHMM makes the model more robust, achieving better performance. One also notices
for HAR dataset that the multi-class WSVM method improves the classification results
over MC-SVM, MC-HF-SVM and HMM classifiers used alone. On the other hand, the
results also show that WSVM outperforms HMM for recognizing activities for all
datasets except for the HAPT dataset.
In terms of reducing the datasets, the feature reduction identifies the most relevant
features for the learning process. We notice that PCA features can improve the discrimination between different activities than the original features. For WISDM the
performances of activity recognition are low than HAR and HAPT datasets with 561
features. This is explained by the number of features (6) for WISDM is not sufficient
when using PCA algorithm. Another reason to the lowest accuracy in WISDM dataset
is attributed to the use only the accelerometer sensor comparatively to the HAR and
HAPT that use the both accelerometer and gyroscope sensors.
Table 2. The micro-averaged measures: Recall, Precision, F-measure and Accuracy for all
approaches in (%). Bold values are the results for our approach for each dataset.
Datasets Approach
Recall Precision F-measure Accuracy
HAR
MC-SVM [8]
89.6 89.9
89.7
89.3
MC-HF-SVM [8]
89.3 89.2
89.2
89.0
WSVM
92.4 91.6
91.9
93.9
HMM
89.2 90.2
89.7
93.7
Proposed
94.0 96.7
95.3
94.9
HAPT
WSVM
96.0 92.4
94.1
86.1
HMM
98.3 97.1
97.7
96.5
Proposed
97.3 99.0
98.1
96.8
WISDM J48 [26]
81.7 –
–
85.1
LogisticRegression [26] 68.4 –
–
78.1
MultilayerPerceptron [26] 80.4 –
–
91.7
WSVM
83.4 76.5
79.8
81.4
HMM
79.4 80.0
79.7
84.9
Proposed
91.9 79.8
85.4
92.3
Human Activities Recognition in Android Smartphone Using WSVM-HMM Classifier
391