**4. Discussion**

This review confirms that AdaBoost, and in general boosting ensemble methods, are reliable for the identification of daily activities. Several studies are not well described, and the source code of the algorithms are not publically available. The verification and reproducibility of the obtained results is not easily possible, because of the following reasons: Only some authors shared the datasets; in many cases, the methods are not explained well explained, in particular, the preprocessing of the datasets; and the hyper-parameter tuning is poorly described, or the exact algorithm parameters are not described.

The number of studies using the AdaBoost method for the recognition of daily activities is minimal, and the daily activities mainly recognized are the simple activities, including walking, running, walking upstairs and downstairs, and other quotidian activities.

Following our literature review, most of the analysed studies (85%) report the best results using AdaBoost methods. Only two studies (15%) presented in [49,58] have said that the AdaBoost based methods do not show the best results when compared with the other approaches for daily activities and environments recognition. Nevertheless, the authors of these studies still recognised the reliable applicability of the AdaBoost method for activity and environment recognition activities.

In summary, all reviewed works first perform a feature extraction step, which somewhat varies depending on the used sensor types. In cases of multiple sensors, or multi-channel sensors, the feature extraction is performed independently for each time series (i.e., channel or sensor). Generally, various statistical metrics, as listed in Table 3, are computed on the raw signal in the time domain, and rarely features are deriving from the frequency domain. Then, after the features are extracted from each sensor as a separate time series, the extracted features are fed into the classifiers. Very often, a systematic approach to feature extraction improves the accuracy [23].

The authors used different features, and the average accuracies obtained with them can be comparable. Table 3 presents the average accuracy of the various features extracted, verifying that the features that allow the recognition of daily activities with an accuracy higher than 90% are the mean, standard deviation, pitch, roll, azimuth and median absolute deviation of signal of motion sensors, and the mean of the signal of magnetic sensors.


**Table 3.** Average of the accuracy reported in the studies analysed, grouped by features.

Moreover, Table 4 presents the advantages and disadvantages of the Adaboost method, proving that it can be used for the recognition of daily activities and environments with the recent advancements in the hardware and software of the devices commonly used.

**Table 4.** Advantages and disadvantages of the use of Adaboost method in the different studies analyzed.


In comparison with other algorithms, the Adaboost method uses different algorithms as the weak learner, in which these algorithms will take into account the features extracted from the signals, such as mean, standard deviation, variance, and others. In general, Adaboost made use of complex data, but it can be used with 1D data in comparison with other algorithms. The authors of the research studies analysed used the Adaboost with uni-dimensional data, i.e., they used the features extracted from the data to provide the results, where the results obtained proved its reliability for physical and physiological data.

In conclusion, the use of mobile devices for daily activities recognition using AdaBoost is limited, because of the low power processing and battery capabilities of these devices [59,60]. According to

the reported studies in this review, it is possible to conclude that the use of the AdaBoost method is reliable with mobile devices as verified by the accuracies reported in the different studies, where only two studies reported accuracies lower than 50%.
