Using Machine Learning to Locate People Indoors

The article investigates the problem of automated data processing to record the presence of students in classes. It is proposed to use machine learning methods, since they allow predicting the location of students in the premises even in the presence of anomalies in the data. The solution to this problem will help to increase the efficiency of the educational process and reduce dependence on traditional methods of recording presence, which require time and human resources. Experiments were conducted using various machine learning methods for regression and classification tasks. Prediction accuracy was used as a measure to compare different methods. Among the regression methods, the following were considered: SVR, LinearSVR, NuSVR, PLSRegression, KernelRidge, RidgeCV, BayesianRidge, DecisionTreeRegressor, and ExtraTreeRegressor. The best accuracy was obtained by DecisionTreeRegressor, KernelRidgeRegression and ExtraTreeRegressor methods — 92.5, 93.9 and 95.5 %, respectively. However, regression methods require continuous data, such as user coordinates, which limits their use in environments where technical means do not allow obtaining such data. As an alternative, classification methods were considered, namely: SVC, KNeighborsClassifier, DecisionTreeClassifier and RandomForestClassifier. The initial results showed lower accuracy compared to regression methods, which was due to the lack of representativeness of the training data. To solve this problem, a step-by-step algorithm was applied, which gradually predicts the building, floor and specific room. This algorithm led to a significant improvement in accuracy, with the best result being achieved by the RandomForestClassifier method — 94.3 %. It was concluded that the choice of a machine learning method depends on the technical means used. If they allow you to obtain continuous data, such as coordinates, it is optimal to use the ExtraTreeRegressor, DecisionTreeRegressor, or KernelRidgeRegression regression methods. If continuous data cannot be obtained, it is optimal to use the RandomForestClassifier classification method with the proposed step-by-step algorithm.