Evaluation of Five Classifiers for Children Activity Recognition with Sound as Information Source and Akaike Criterion for Feature Selection

摘要

The recognition and classification of children activities is a subject of novel interest in which different works have been presented, where the data source to perform this classification is crucial to define the way of working. This work uses environmental sound as data source to perform the activities recognition and classification, evaluating the accuracy of the k-Nearest Neighbor (kNN), Support Vector Machines (SVM), Random Forests (RF), Extra Trees (ET) and Gradient Boosting (GB) algorithms in the generation of a recognition and classification model. In the first stage of experimentation, the complete set of features extracted from the audio samples is used to generate classification models. Then, a feature selection process is performed based on the Akaike criteria to obtain a reduced set of features used as input for a second generation of these models. Finally, a comparison of the results obtained from the data of models of both approaches is carried out in terms of accuracy to determine if the model contained by the features selected improves the performance in the classification of children activities.