Avaliação de modelos de classificação automática de atividades diárias para dispositivos de baixo custo

Abstract

In this work, machine learning algorithms were used to identify Activities of Daily Living (ADLs) in three public data sets, ARCMA, HMP and UMAFall. The central idea is to define a methodology that is capable of evaluating a set of machine learning methods, in a specific scenario, in order to achieve the best results in the automatic classification of Activities of Daily Living, considering the simplification of the implantation and the necessary structure for its execution.Nowadays the studies with better results use data from sensor networks installed both in the environment and in the body of the volunteers, this type of approach makes the application more complex and makes its implementation difficult.This work will use data generated only by an inertial sensor (accelerometer or magnetometer, depending on the database), this simplification implies less complexity in both development and implemen- tation, in addition to providing cheaper solutions. Thus, the objective of this work is to build a computational framework, which involves parameterization, execution, simulation and validation, thus allowing to find a model capable of performing the classification of Activities of Daily Living using low cost devices. To check the performance of the final model, the Raspberry Pi 3 B was used. To achieve this goal, the proposed methodology performs the data processing, excluding invalid entries, balancing the bases, extracting the relevant characteristics, evaluating the algorithms, search for the best reading window and using a threshold filter to improve accu- racy. To mitigate the results obtained, we use stratified cross-validation to find the F-score and Accuracy of the methods. The following algorithms were evaluated: k-Nearest Neighbors, Naive Bayes, Support Vector Machine, Decision Tree, Random Forest, and Extra-Trees. The Extra-Trees algorithm presented the best results with a final accuracy of 92.06%, 93.97% e 97.79%, and 91.29%, 92.76% e 97.41% of F-score, for the ARCMA, HMP and UMAFall bases, respectively, in a scenario who would be tolerable to exclude two activities from the databases. We use a decision filter that chooses whether the record should be discarded or not, taking into account the probability table returned by each method, this increased the accuracy of the classification by 16.33%, 14.95% and 9.05% on the ARCMA, HMP, and UMAFall bases. All modeling applied in this work is relevant for future studies that aim to implement a real appli- cation to perform automatic detection of ADLs, which is, therefore, the main contribution of this work.