**1. Introduction**

AdaBoost is one of the first boosting algorithms developed by Yoav Freund and Robert Schapire that was adapted for practical application in many solving tasks. AdaBoost is a method that uses ensemble learning techniques to combine multiple weak classifiers into a single strong classifier. It is combined with other artificial intelligence methods to increase the accuracy of the recognition [1]. Thus, weak learners, including decision tree and decision boosting, are commonly used with the AdaBoost method. In comparison with other machine learning methods, the AdaBoost method is less susceptible to overfitting.

One of the strategies adopted by the different implementation of Adaboost consists in combination with other methods to reduce the errors obtained [2,3]. The primary purpose of ensemble learning techniques is to improve the results by combining the results of different methods [2,3]. These techniques consist of the combination of several machine learning techniques with a single purpose and model to improve the prediction results [4–6]. It can be divided into two groups, sequential ensemble methods and parallel ensemble methods, where our focus is the sequential ensemble methods, because the implementation of Adaboost consists in the application of a base learner that is generated sequentially [7].

In the last years, several studies have been developed with a focus on the recognition of daily activities using the sensors available in the commonly used mobile devices. These studies conclude that it is possible to accurately detect the daily activities and environments with motion, magnetic, location and acoustic sensors embedded on mobile devices, reporting reliable results available in the literature with different machine learning methods [8–23].

To date, and due to the increasing power processing capabilities of the different mobile devices, the Adaboost method is one of the most used methods, and it reports reliable results [24–32]. The motivation of this systematic review is to evaluate the reliability of the Adaboost method for daily activities and environment recognition using the sensors available in mobile devices for further implementation of a framework [33–42].

Generally, the raw readings of one-dimensional (e.g., blood pressure sensor, thermometer, etc.) or multi-dimensional signals (e.g., accelerometer or gyroscope) can be directly processed by AdaBoost, and other classification and regression algorithms in general. To do that, all sensory readings in a specific time window represent different inputs. For example, if a thermometer reads data with 1 Hz frequency, and the window is 60 s, there will be 60 inputs to AdaBoost. Similarly, a three-dimensional gyroscope would present 180 inputs. Many deep learning methods accept the input data in this format. Be that as it may. Usually, many algorithms benefit from a feature engineering step [43], which significantly improves the accuracy or simplifies the complexity of the models [23,44].

Due to the complex nature of the sensory data collected using the sensors available in mobile devices, the overfitting problem is impacts many machine learning algorithms, including multilayer perceptron neural networks (MLP), deep neural networks (DNN) and feedforward neural networks (FNN) [33–42]. Methods for parameter tuning such as grid search [45] and systematic feature selection [23] are usually applied to mitigate this problem.

Previous studies [33–42] shown that the proposed framework includes the correct modules for the reliable recognition of daily activities and environments. However, the results can be improved with other methods, including ensemble learning methods.

This paper reviews the different studies available in the literature related to the implementation of the AdaBoost method for daily activities recognition. This review is included in the research and development of a framework associated with the identification of daily activities and environments using the sensors available in mobile devices, where the AdaBoost method can increase the accuracy compared to other implementations. The motivation of this paper is to improve the accuracy reported in previous studies for the recognition. This review intends to explore the use of the Adaboost method to verify if it reports better results than MLP, FNN, and DNN methods for the identification of daily activities.

The main contribution of this review is the presentation of a base of study for the readers who deal with the recognition of daily activities and environments using sensors available in mobile devices providing an in-depth survey of several research projects which implement Adaboost method.

This review shows that the features that reported better results are mean, standard deviation, pitch, roll, azimuth and median absolute deviation of the signal of motion sensors, and the mean of the signal of magnetic sensors. According to the results, the Adaboost method provides huge accuracy for the recognition of daily activities and environments.

The following sections are organized as follows: Section 2 presents the methodology of the review. The results obtained are presented in Section 3. Section 4 presents the discussion on the results. Finally, the conclusions are presented in Section 5.
