Investigating effects of window length on 1D-CNN-LSTM and effectiveness of Heuristic features in solving sensor orientation and placement problems in human activity recognition using a single smartphone accelerometer

Abstract

Human Activity Recognition (HAR) using smartphone sensors can offer multiple applications in different spheres. Using deep learning classifiers such as Convolutional Neural Networks (CNN), Short-Term Long Memory (LSTM), or their hybrid showed promising improvement in HAR. However, using these deep learning networks requires segmenting the input data into multiple data windows of similar length. The length of the data windows can significantly affect HAR's performance. Therefore, the influence of the window lengths needs to be investigated to choose an optimal window length. Additionally, the orientation and placement of the smartphone sensor also present significant challenges to HAR. Many approaches have been proposed to solve the orientation and placement problems. In my study, I first evaluated the effects of window length on 1D-CNN-LSTM in HAR for six activities: Lying, Sitting, Walking, and Running at 3-METs (Metabolic Equivalent of Tasks), 5-METs and 7-METs. Subsequently, I evaluated the effectiveness of the heuristic features in HAR in solving sensor orientation and sensor placement problems for three smartphone locations: Pocket, Backpack and Hand. I performed this evaluation using 1D-CNN-LSTM by using the optimal window length found in the first part.