Classification methods based on bayes and neural networks for human activity recognition

The Human Activity Recognition is a context awareness application, which has, for example, sports, security and health monitoring applications. As a way to acquire the human activity data, there are external approaches (e.g. cameras data) and embedded approaches (e.g. accelerometer data). In this area, we can find solutions using multiple sensors simultaneously supporting the real time data acquisition. In the other hand, there are researchers focusing in an optimized and accurate way to classify the activity based on large amount of data. In this paper the Bayes Network and Neural Network classifiers were used. They were built from a public dataset regarding data acquired from accelerometers worn by a set of people. Considering the Data Mining classification steps, the attributes selection, the building and training, the validation tests were done, resulting in the Confusion Matrix, the accuracy rate and the processing time. The goal is to compare different classification methods using different building and training techniques, matching the dataset specifications and quirks, assessing the performance and accuracy rates impacts. A Java program was implemented using the Data Mining public library called Weka. With smaller processing times than the Bayes Network, the implemented Neural Networks had excellent results. The Bayes Network requires a further investigation on configuration and validation techniques, it had the largest processing times, and reached the best results using the k-fold cross-validation technique.