**1. Introduction**

The accelerometer sensor commonly available in off-the-shelf mobile devices [1,2] measures the acceleration of the movement of the mobile device, enabling the recognition of activities of daily living (ADLs) [3]. After the development and conception of a system architecture for the identification of ADLs, it could be, for example, integrated into the creation of a personal digital life coach [4], essential for the monitoring of elderly persons and persons with impairments, or for the training of certain lifestyles. The accelerometer enables the recognition of several motion activities, including running, walking on stairs, walking, and standing. Following the previous research studies [5–8], several steps are incorporated in the recognition of ADLs, including data acquisition, data processing, data cleaning, feature extraction, data fusion, and data classification.

Several authors studied the automatic recognition of ADLs [9–14]; artificial neural networks (ANN) were widely used [15,16]. The accelerometer was used for the identification of ADLs while comparing some implementations of ANN with different frameworks, such as the multilayer perception (MLP) with Neuroph [17] and Encog [18] frameworks, and the deep neural network (DNN) method with the DeepLearning4j [19] framework. The authors aimed to find the model that achieves the best accuracy in recognition of running, walking, walking downstairs, walking upstairs, and standing. These five ADLs were selected based on the literature review, wherein different studies reported reliable results for these activities, to allow the comparison with the method implemented in this research. The use of data acquired from the accelerometer sensor fused with the data retrieved from the magnetometer and gyroscope sensors is available in the literature [20]. This paper attempts to use different datasets of features with only the accelerometer data that should be analyzed to define the best combination of features. The main objective of this paper is to explore the use of different sets of features obtained using the accelerometer with the same datasets acquired for the previous study. After the comparison performed in [20] about the use of data fusion from the data acquired from the accelerometer, magnetometer, and gyroscope sensors, we verified that one of the major problems is related to the overfitting obtained during the training phase of the ANN.

The frameworks presented in this study were used in the study [20] to verify which the best methods are for the recognition of ADLs using the sensors available in the mobile device. Despite the disadvantages of achieving poor accuracy, MLP implemented with Neuroph and Encog frameworks still have the benefit of the adaption of the low resources of the mobile devices, because these methods need less power processing and memory capabilities than the DNN method implemented with DeepLearning4j. Therefore, the primary motivation of this paper is to verify whether the overfitting problem can be solved using only the accelerometer data. Additionally, the authors aim to verify the accuracy of the proposed method using only one sensor and a smaller number of features for the training of the ANN, in order to use fewer computational resources and reduce the energy expenditure of the mobile device when compared with the use of multiple sensors.

Thus, the main contribution of this paper is to perform a comparison of three different architectures of ANN methods using only the accelerometer data to verify whether the overfitting problems are avoided. This paper presents the use of ANN for ADLs recognition with the data acquired from mobile sensors. In addition, it also presents a comparative study of different implementations to find the most accurate method.

This paper is structured as follows: Section 2 presents work related to the identification of ADLs using the accelerometer sensor. Section 3 describes the steps used for the recognition of ADLs using the accelerometer sensor. Section 4 presents the discussion and results obtained during the research. Finally, Section 5 consists of the presentation of the conclusions regarding the results obtained.

## **2. Related Work**

Several methods can be used for the automatic classification of ADLs with the data acquired from the accelerometer sensor available in the off-the-shelf mobile devices [3,21]. Numerous studies in this field are presented in the literature. Therefore, it is not possible to include them all in this document. Table 1 presents an analysis of 43 studies conducted on ADLs recognition using accelerometer data. The studies were selected according to the following criteria: (1) use of smartphones for data collection; (2) the features being clearly defined; (3) the methods being clearly defined; (4) the accuracy levels being presented. These studies are available in multiple databases such as MDPI, Springer, and ACM

collected using the Google Scholar portal. Still, the vast majority have been found in the IEEE Xplore library. Following the different works analyzed, the methods that reported the best accuracies for the recognition between 1 and 8 ADLs are the different types of ANN, including MLP and DNN methods, using statistical features.

The studies presented in Table 1 reported that the most recognized ADLs with reported average accuracies high than 85% are walking, standing, walking upstairs, walking downstairs, and running. Therefore, these activities are considered in the proposed method. In total, 31 studies use smartphones located in the user's pocket. However, some studies also located the smartphone around the waist, forearm, and wrist. Moreover, some studies combine the use of smartphones with other wearable sensors.

The ADLs recognition indicates an average accuracy between 87.93% and 88.80% using different methods. In addition, the ADLs reporting better accuracies in the analyzed studies are walking, standing, walking upstairs, walking downstairs, and running. In total, 91% (N = 39) of the analyzed papers support walking recognition reporting an average accuracy of 88.80%. The standing activity is included in 29 studies which represent 67% of our literature review and provide an average accuracy of 88.65%. Walking upstairs and downstairs activities are supported by 25 (58%) and 23 (53%) studies, respectively. The first reports an average accuracy of 85.88% and the second reports an average accuracy of 85.5%. Finally, the running activity is assessed by 42% (N = 18) of the evaluated studies and reports 87.93% average accuracy.

Regarding the ADLs recognized in the analyzed studies, the mean, standard deviation, maximum, minimum, correlation, variance, and median are the most used features in the literature. In total, 86% (N = 37) of the analyzed papers use the mean feature, reporting an average accuracy of 85.74%. The standard deviation feature is included in 30, representing 70% of the evaluated papers, and provides an average accuracy of 86.70%. The maximum and minimum values are included in 19 (44%) and 17 (40%) studies, respectively. The maximum feature reports an average accuracy of 87.47%, and the minimum feature reports 88.50%. The median and correlation features are used in 10 studies (23%) each and report average accuracies of 87.44 % and 91.52%, respectively. Eight studies include the variance as a feature for ADLs recognition reporting and average accuracy of 90.15%.

The implementations that reported an accuracy higher than 88% are ANN, multi-column bidirectional long short-term memory (MBLSTM), Bayesian network, and random forest methods, reporting an average accuracy between 88.65% and 91.29%. In total, 40% (N = 17) of the analyzed papers use ANN methods reporting the average accuracy of 91.29%. Eight studies propose the random forest for ADLs recognition, reporting 90.53% average accuracy. The MBLSTM method provides 89,4% average accuracy, and the Bayesian Network is used by three studies reporting an average accuracy of 88.65%.

In summary, the number of ADLs recognized with the different methods used, as well as the particular dataset, influenced the accuracies reported. The identification of a lesser amount of ADLs reported the best results in the literature. Following the ADLs and methods that reported the best results, our research is focused on the implementation of ANN for the recognition of five ADLs, including standing, walking, running, and walking upstairs and downstairs. These ADLs were selected for our implementation because they are the most recognized in the literature, reporting reliable accuracies.


**Table 1.** Summary of the studies available in the literature.


**Table 1.** *Cont*.




**Table 1.** *Cont*.

11

#### *Electronics* **2020**, *9*, 509



#### *Electronics* **2020**, *9*, 509


