Bending of cnn model with genetic algorithm to recognize human activities from sensor data

Abstract

Human activity recognition (HAR) is a critical application on wearable devices for fitness tracking, healthcare, and elder care assistance. Inaccurate recognition results, on the other hand, may have a negative impact on users or even result in an unexpected accident. As a result, it is critical to improve the accuracy of human activity recognition. Recent advancements in deep learning (DL) techniques and their automatic feature extraction ability attract many researcher for HAR. However, the large size feature maps generated by these DL models affect the overall classification accuracy, and it also increases the computational cost, as there are many redundant features generated by the DL models. This thesis aims to provide effective and efficient HAR methods to address the major HAR challenges, which can be divided into three contributions. The first contribution is a novel wearable sensor signal to a corresponding activity image encoding algorithm that addresses the problem of capturing information in both the time and frequency domains. The second contribution is to propose a novel feature extractor based on a deep learning approach that addresses the extraction of quality features for activities. The third contribution is a novel feature selection framework that selects the most relevant features and addresses the negative effect of large feature size on the performance of a HAR model. Extensive experiments on publicly available datasets have been conducted for the proposed approaches. Experiments have shown that the proposed methods outperform the state-of-the-art methods.