Reducing Uncertainty in User-independent Activity Recognition - A Sensor Fusion-based Approach

In this study, a novel user-independent method to recognize activities accurately in situations where traditional accelerometer based classification contains a lot of uncertainty is presented. The method uses two recognition models: one using only accelerometer data and other based on sensor fusion. However, as a sensor fusionbased method is known to consume more battery than an accelerometer-based, sensor fusion is only used when the classification result obtained using acceleration contains uncertainty and, therefore, is unreliable. This reliability is measured based on the posterior probabilities of the classification result and it is studied in the article how high the probability needs to be to consider it reliable. The method is tested using two data sets: daily activity data set collected using accelerometer and magnetometer, and tool recognition data set consisting of data from accelerometer and gyroscope measurements. The results show that by applying the presented method, the recognition rates can be improved compared to using only accelerometers. It was noted that all the classification results should not be trusted as posterior probabilities under 95% cannot be considered reliable, and by replacing these results with the results of sensor fusion -based model, the recognition accuracy improves from three to six percentage units.