Efficiency investigation of artificial neural networks in human activity recognition

Nowadays, human activity recognition (HAR) is an important component of many ambient intelligent solutions where accelerometer and gyroscope signals give the information about the physical activity of an observed person. It has gained increasing attention by the availability of commercial wearable devices such as smartphones, smartwatches, etc. Previous studies have shown that HAR can be seen as a general machine learning problem with a particular data pre-processing stage. In the last decade, several researchers measured high recognition rates on public data sets with numerous "shallow" machine learning techniques. In some cases artificial neural networks (ANNs) produced better performance than other shallow techniques while in other cases it was less effective. After the appearance of deep learning, a significant part of HAR researches turned toward more complex solutions such as convolutional neural networks (CNNs) and they claimed that CNNs can substitute the feature extraction stage in shallow techniques and can outperform them. Therefore in the current state of the art, the efficiency of ANNs against CNNs and other machine learning techniques is unclear. The aim of this study is to investigate the performance of more ANN structures with different hyper-parameters and inputs on two public databases. The result will show that the two key factors in ANN design are the data-preprocessing and the hyper-parameter setup because the accuracy difference between a well and a badly parameterized ANN is huge. A well-tuned ANN with extracted features can outperform other machine learning methods in the HAR problem including CNN classifiers.