Table 3.
Baseline classifiers included in the datasets’ comparative evaluation.
| OPPORTUNITY Dataset | ||
|---|---|---|
| Challenge Submissions [7] | ||
| Method | Description | |
| LDA | Linear discriminant analysis. Gaussian classifier that classifies on the assumption that the features are normally distributed and all classes have the same covariance matrix. | |
| QDA | Quadratic discriminant analysis. Similar to the LDA, this technique also assumes a normal distribution for the features, but the class covariances may differ. | |
| NCC | Nearest centroid classifier. The Euclidean distance between the test sample and the centroid for each class of samples is used for the classification. | |
| 1NN | k nearest neighbour algorithm. Lazy algorithm where the Euclidean distances between a test sample and the training samples are computed and the most frequently-occurring label of the k-closest samples is the output. | |
| 3NN | See 1NN. Using 3 neighbours. | |
| UP | Submission to the OPPORTUNITY challenge from U. of Parma. Pattern comparison using mean, variance, maximum and minimum values. | |
| NStar | Submission to the OPPORTUNITY challenge from U. of Singapore. kNN algorithm using a single neighbour and normalized data. | |
| SStar | Submission to the OPPORTUNITY challenge from U. of Singapore. Support vector machine algorithm using scaled data. | |
| CStar | Submission to the OPPORTUNITY challenge from U. of Singapore. Fusion of a kNN algorithm using the closest neighbour and a support vector machine. | |
| NU | Submission to the OPPORTUNITY challenge from U. of Nagoya. C4.5 decision tree algorithm using mean, variance and energy. | |
| MU | Submission to the OPPORTUNITY challenge from U. of Monash. Decision tree grafting algorithm. | |
| Deep approaches | ||
| Method | Description | |
| CNN [17] | Results reported by Yang et. al., in [17]. The value is computed using the average performance for Subjects 1, 2 and 3. | |
| Skoda dataset | ||
| Deep approaches | ||
| Method | Description | |
| CNN [23] | Results reported by Ming Zeng et. al., in [23]. Performance computed using one accelerometer on the right arm to identify all activities. | |
| CNN [43] | Results reported by Alsheikh et. al., in [43]. Performance computed using one accelerometer node (id #16) to identify all activities. | |