*2.1. Activity Recognition Using COTS*

A number of researchers are now incorporating COTS wrist-worn wearables into their approach for activity recognition rather than creating custom devices, as shown in Table 1. Additionally, COTS smart watches are usual paired with a smartphone, which also contains its own IMU. Therefore, this creates an optional data source that can be used if required, unlike their experimental wearable device counterparts.

The LG G Watch has been used in a couple of approaches [13,15], producing a reasonable level of accuracy. The authors in [15] use an LG G Watch coupled with a Samsung Galaxy S4 smartphone to identify eating activities for different foods. Their approach features a total of eighteen activities (six low and twelve high level) with seventeen participants. A range of classifiers (ANN with Multilayer Perceptron (MLP), NB and RF) are used, and the authors show that RF produces the best overall accuracy of 93.3% in personal classification. However, we should note that the authors do not provide evaluation for the other classifiers. Similarly, in [16], the authors propose a diet monitoring system that uses a smart watch device to detect fourteen eating and seven non-eating activities with a DT classifier. Using the accelerometer and gyroscope data from the smart watch, the authors show that the proposed system achieves an accuracy of 92% for detecting an eating episode. Furthermore, the authors in [13] use the accelerometer data from an LG G Watch for comparing activity estimation between sensor placement on the wrist and elbow (Myo armband) using KNN, DT, RF and bagging classifiers for eight high level gestures. The presented results show that the smart watch, with an accuracy varying between 86.5% and 96.2% depending on the classifier, does provide an 8% better overall accuracy over the armband, but not for every participant.

Another smart watch that has been shown to give good activity classification [17,18] is the Moto 360. The authors in [17] attempt to derive the activity of a shopper by capturing accelerometer and gyroscope data form a Moto 360 and a smartphone for high and low level activities respectively. Their approach achieves a precision accuracy of 92.26% for the high level activities when using an HMM and DT with Conditional Random Field (CRF) classifiers. The main aim in [18] is to increase the energy efficiency of a smart watch with the classification algorithm running locally. The authors not only use a Moto 360 (Motorola Mobility LLC., Libertyville, IL, USA), but also use a Samsung Galaxy Live (Samsung, Daegu, South Korea), Sony S3 (Sony Corporation, Tokyo, Japan) and LG G Watch R (LG Electronics, Seoul, Korea) for various evaluations throughout their work. The

proposed system also provides a novel approach for semantic abstraction with NB, SVM, DT and LR classifiers, which delivers a good recall accuracy, averaging at 75% over the classifiers. However, it performs poorly in terms of precision accuracy, with the NB and DT classifiers producing 66% and 55%, respectively. Though, the authors do successfully demonstrate that semantic abstraction does improve overall accuracy.

Additionally, the researchers in [19] use a Microsoft Band smart watch for the purpose of identifying the activities of participants in a basketball game. A wide range of different classification algorithms (SVM, KNN, NB, DT and RF) is used. The authors use a personal classifier first used to distinguish the activity of the participant, and then, a collaborative classifier is used to identify the actual participant. With 10-fold cross-validation, the SVM classifier was shown to produce the best precision and recall accuracies of 91.34% and 94.31%, respectively.

The authors in [20] use the GENEActiv [21], which is a more specialised watch featuring a triaxial accelerometer and is geared towards research applications for free-living, sports research and clinical trials. The watch is available commercially and, therefore, still considered COTS technology. The approach in [20] is focused on classification of seven high level daily activities using HMM and CRF algorithms with leave-one-day-out cross-validation for the 21 days of data collected from two participants. It achieves accuracies ranging between 70% and 77% with the use of sub-classing and highlights that the GENEActiv is a feasible device for activity recognition.

A slightly more unorthodox approach is the one used by the authors in [22], who use two Samsung Galaxy S2 smartphones, with one at the pocket position and the other mounted at the wrist position. As a general rule, a smartphone weighs more than a smart watch. As a result, the device will potentially have a higher centre of gravity, which can affect how a participant performs activities. This aside, the authors in [22] do show that good classification accuracy of seven low level and six high level activities can be achieved using KNN, NB and DT classifiers. Additionally, they also show that using accelerometer and gyroscope data from the smartphone at the pocket position does improve the classification of static activities such as standing and sitting.
