| Summary: | Human Activity Recognition (HAR) has been attracting significant research attention because of the increasing availability of environmental and wearable sensors for collecting HAR data. In recent years, deep learning approaches have demonstrated a great success due to their ability to model complex systems. However, these models are often evaluated on the same subjects as those used to train the model; thus, the provided accuracy estimates do not pertain to new subjects. Occasionally, one or a few subjects are selected for the evaluation, but such estimates highly depend on the subjects selected for the evaluation. Consequently, this paper examines how well different machine learning architectures make generalizations based on a new subject(s) by using Leave-One-Subject-Out Cross-Validation (LOSOCV). Changing the subject used for the evaluation in each fold of the cross-validation, LOSOCV provides subject-independent estimate of the performance for new subjects. Six feed forward and convolutional neural network (CNN) architectures as well as four pre-processing scenarios have been considered. Results show that CNN architecture with two convolutions and one-dimensional filter accompanied by a sliding window and vector magnitude, generalizes better than other architectures. For the same CNN, the accuracy improves from 85.1% when evaluated with LOSOCV to 99.85% when evaluated with the traditional 10-fold cross-validation, demonstrating the importance of using LOSOCV for the evaluation.
|
|---|
