*2.2. Related Works on Context Impacts*

Since WD is part of human activity recognition (HAR), we surveyed the HAR works that considered the contexts such as placement and personalization, instead of purely WD. We will also introduce some typical SC works that considered contexts.

(1) Placement is the most common context and is the factor that has attracted researchers' attention. Olguin et al. placed one or two accelerometers on three different parts of the body and studied the classification accuracy of activity recognition [17]. Lester et al. studied whether a single accelerometer could generalize well on different locations and the reliability of activity recognition on a novel individual [18]. The works in [2,19] explored the influences of placements on different body parts. The work in [1,3] showed that the negative influence of various placements of the sensor could be mitigated. Cleland et al. studied the optimal placement to detect daily activities [20]. Sun et al. investigated the effects of varying positions and orientations on the accuracy of activity recognition [21].


Kunze [1,2] evaluated the context impact of placements and orientations. The work considered the placement including head, trousers, torso and wrist. It showed that the displacement of sensors could harm the accuracy, but this could be mitigated by extracting placement-independent features and placement recognition. Besides, it showed that the closely related placements usually generated misclassifications.

For step counting, the Pan-Tompkins method (PTM) [27] only mounted the sensor at the foot and reached a high accuracy in SC. In [28], each subject wore the sensor on his/her waist, and then, the activities were classified and steps counted.

It can be seen that previous activity recognition studies mainly conducted the experiments under a few contexts such as classifiers, placements and orientations of the sensor; therefore, in-depth studies considering complete contexts are needed. Besides, The majority of most past research only studied HAR, which made the evaluations on both WD and SC under various contexts necessary. In our paper, we considered more comprehensive application contexts that include classifiers, placements, orientations, window size, sampling rates and personalization to make a complete comparison.
