*2.2. Activity Recognition with Custom Devices*

There have been numerous works that use a custom made wrist-worn wearable device for the purpose of activity recognition. A custom wearable device removes any potential limitation that may be imposed by a COTS wearable device, such as adding additional sensors or collecting other information that a COTS wearable device API may not expose. Furthermore, HAR approaches that feature custom wearables devices commonly use additional on-body sensors or wearables to enhance activity classification performance.

One interesting use of a custom wrist-worn wearable device is the approach proposed in [23], where a wrist-worn device is coupled with wearable inertial rings to aid in increasing the accuracy of nine high level activities using DT and SVM classification algorithms. The approach is successful, as using only the wrist-worn device provides an accuracy of 68.85% for DT and 65.03% for SVM, while the whole system provides an accuracy of 89.06% for DT and 91.79% for SVM. Similarly, the authors in [24] use custom wearables at the elbow and wrist positions for the training phase for RF and CRF classifiers for smoking and eating sessions. For three high level smoking activities, the authors demonstrate that using the additional sensor at the elbow position results in an accuracy of 93% for RT and 95.74% for CRF. Furthermore, in [14], the authors evaluate and compare wrist- and body-worn sensors for DT, RT, NB, SVM and KNN classification algorithms with accelerometer data, in the context of fall detection. They show that RF achieves the best overall accuracy among the classifiers while the wrist worn device achieves 72% accuracy, which was marginally better than devices worn on other body locations such as the elbow and chest, which achieved 67% accuracy when classifying ten basic activities. However, sensors positioned at the ankle, knee and belt achieved an accuracy of 77%. This is also shown in the work conducted by [11], who compare hip and wrist sensor placements with an LR classifier using accelerometer data to classify seven basic activities. The authors demonstrate

that the hip position provides better accuracy for four activities with an overall accuracy of 91%, while the wrist position provides better accuracy for the remaining three activities with an overall accuracy of 88.4%.

Other approaches using custom wearable devices at the wrist position exclusively include the work conducted by [25], who investigate how the combination of six classification algorithms (NB, SVM, DT, ANN with MLP, KNN and RF) can achieve better accuracy. The authors show that a combination of KNN and RF classifiers for four basic activities only using accelerometer data gives the best accuracy. Furthermore, the authors in [26] compare the ANN with MLP, NB and SVM classifiers using a custom wrist-worn wearable featuring a nine-axial IMU, showing the MLP-based ANN to be the best classifier for their approach. Similarly, the authors in [27] compare the performance of four classifiers (NB, ANN, DT and LR) for identifying eight basic sporting activities, when using a single wrist-worn custom wearable device fitted with a single accelerometer. They show that ANN is the best classifier, achieving an accuracy of 86.7%. The authors in [28] use Emerging Pattern (EP), which is a threshold classifier. EP has low computation requirements, allowing the authors to run the classification algorithm locally on the custom wearable device, which provided an overall accuracy of 86.2% when attempting to classify four basic activities. Lastly, in [29], the authors develop their own classification algorithm that is based on sign-of-slope and threshold evaluation to be used in conjunction with their custom wearable device featuring an accelerometer. They also compare their custom wearable against other COTS devices, specifically the iPhone 6 smartphone, Mi band and SKT smartbands and the Moto360 and Samsung Gear S smart watches. Though the authors' approach is shown to provide better accuracy, it should be noted that data gathering for all of the COTS devices was performed simultaneously, with the participant holding the smartphone and wearing all four COTS devices. This could potentially prevent the participant from performing the activities under a real-life scenario due to the combined weight of the wearables.
