4.1.3. Experiment Scheme

In order to collect data for our context impact evaluation tasks, it is necessary to have many different kinds of people participating in our experiments. A total of 15 subjects including 10 males and 5 females, with ages ranging from 18 to 28, heights ranging from 1.6 m to 1.85 m and weights ranging from 45 kg to 90 kg, participated in our experiments. Under the assumptions highlighted in Section 4.1.2, we conducted our experiments as follows. Each subject was required to carry multiple smartphones mounted at different positions on the body and walk continuously alongside an indoor track as in Figure 3. This included walking through corridors, going upstairs, going downstairs and going back to the starting point. As we did not use six devices to collect data simultaneously, multiple rounds of walking with the devices mounted at different locations were required to cover all six placements in the data collection. We also employed a camera in this indoor track to obtain the ground truth of step counting. Besides, activities such as running, riding, brushing teeth and driving, as shown in Table 4, do not need to be collected at all six placements, because the signal is similar under different positions. Hand is the subject carries the smartphone in their hand naturally, which is mainly used to simulate the position of the wrist band. Handheld using (HandU) is the subject carries the smartphone, as well as watches the screen, in order to reflect typical walking and using states. The smartphones are not limited to the left or right side; the individual just behaves naturally since we observe that the signal is similar.

Using these assumptions and experiments, we simplified the data collection, and the data could reflect and represent the real contexts well. In the comparison of each context, we only selected data that were generated under the specific context to train the model. For example, we could simply only use the data collected at FrontPocket to train and test the model when we want to get the accuracies of the placements of FrontPocket.


**Figure 3.** Designed indoor trajectory in data collection.
