*5.1. SDA Dataset*

The batch mode recognition is tested by a leave-one-out cross-validation (LOOCV) strategy, i.e., taking one subject for testing and the others for training, then make the test for all the subjects. The sliding window length of feature extraction is set to 5. The length is set based on our experience; normally it depends on the shortest gait or leg phase of the activities, and a window length larger than the shortest duration may conceal some information of the phase, while too small length may cause a badly calculated standard deviation. Both the SemiTMC-GMM and TMC-GMM model are involved in the validation, and the GMM mixture number *κ* is set to 1, 3, 6, and 9, respectively. Particularly when *κ* = 1, the conditioned observation density yields to the conventional Gaussian distribution.

The overall accuracy of batch mode recognition on SDA dataset is shown in Figure 4a. As it can be seen, SemiTMC-GMM achieves an accuracy improvement of about 2–3% compared to TMC-GMM. The proposed model reaches the highest accuracy of 86.00% when *κ* = 6, while the one of TMC-GMM is 83.76%. Meanwhile, TMC-HIST obtains the lowest accuracy of 77.91%. Figure 4b is the plot of Receiver Operating Characteristics (ROC) *w.r.t.* different *κ*. The ROC space takes the true positive rate (TPR) as vertical axis and false positive rate (FPR) as horizontal axis, a better recognition performance is more closer to the upper left corner. So we can see that SemiTMC-GMM (red markers) obtains a better performance than the one of TMC-GMM (cyan markers). Table 1 shows the sensitivity, specificity, F1 score, and Matthews correlation coefficient (MCC) of each individual activity. Particularly for the sensitivity of each individual activity, it equals to the accuracy of corresponding activity. Activities

D1A1 to D1A3 are recognized with relatively poor performance, it is because that these three activities are all walking and are very easily misclassified.

**Figure 4.** The batch mode recognition performance of the Sports and Daily Activities (SDA) dataset, of the SemiTMC-GMM, and TMC-GMM models, according to different GMM mixture number *κ*. ROC = Receiver Operating Characteristics.

**Table 1.** The sensitivity, specificity, F1 score, Matthews correlation coefficient (MCC) value of the batch mode recognition, for each activity of SDA dataset, using the sensor placed on right thigh. **Up**: TMC-HIST; **middle**: TMC-GMM when *κ* = 6; and **down**: SemiTMC-GMM when *κ* = 6.


As reported in [20], the classifiers of kNN, SVM and decision tree are tested on SDA dataset using all the five sensors. The accuracies are 78.97%, 84.03%, and 84.63%, respectively. In [21], the authors used SDA dataset and showed single sensor recognition accuracy of four classifiers: kNN, decision tree, discriminant analysis, and Naive Bayes. Specifically for the right leg sensor that is used in our paper, the four classifiers obtained accuracies of 81.72%, 78.78%, 87.03%, and 76.93%. Therefore, we can state that SemiTMC-GMM outperforms the generic classifiers like kNN, SVM, decision tree, and Naive Bayes, and obtains a similar performance of discriminant analysis. On the other hand, the authors in [31] used CNN to recognize human daily activities in OPPORTUNITY dataset [32], which contains activities, such as open (close) door, open (close) drawer, clean table, and drink cup. They obtained an accuracy of 85.8% by using 23 body-worn sensors, 12 object sensors, and 21 ambient sensors. In addition, for the OPPORTUNITY dataset, [18] used CNN to obtain an accuracy of 77.99% by using the body-worn sensors only, while in [33], CNN obtained an accuracy of 93.75% on six activities: walking, stair ascent, stair descent, sitting, standing, and laying. Because of the prevalent CNNs can generate high dimensional features that suit for the recognition task, then CNNs may probably be suited for sophisticated activities. But it requires a huge quantity of data to train the network, and it is difficult to make CNN work for an adaptive on-line scenario. So, maybe CNN could obtain higher accuracy than SemiTMC-GMM, but we still believe that our proposed model is competent in some scenarios.
