Wearable sensor-based human activity and demographic group recognition with deep learning

With the advancement of technology, the use of wearable mobile devices, especially in the fields of health and daily activity tracking, has gained significant importance. The data collected from these devices play a crucial role in the early diagnosis of diseases. Moreover, wearable sensor technologies have the potential to offer more personalized health information monitoring by classifying user activities, gender and age range. This enables an effective response to the specific health needs of individuals based on their age and gender, making personal health management more efficient. This thesis addresses the problems of human activity recognition and demographic group classification on datasets collected through wearable mobile devices. To solve these problems, Naive Bayes, Random Forest, k-Nearest Neighbors (kNN), Convolutional Neural Networks (CNN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Transformer classification algorithms were designed and their performances were analyzed in detail. As a result of the analysis, due to the success of LSTM, CNN, and Transformer models, a new hybrid model based on an attention-based fusion of these models is proposed. The developed model has significantly improved the classification of gender and age groups in human activities compared to the methods present in the current literature. In the experimental studies conducted within the scope of this thesis, the effects of various hyperparameters on classification performance have been analyzed. The classifiers designed within the scope of the thesis were evaluated on HPAwIO, MHEALTH, and UCI-HAR datasets. The proposed fusion-based attention-based model achieved average accuracy rates of 85.15% in indoor and outdoor activity classification, 84.04% in gender classification, and80.54% in age range classification on the HPAwIO dataset. Additionally, this model attained success rates of 85.35% in activity classification on the MHEALTH dataset and 86.07% onthe UCI-HAR dataset. These results demonstrate the effectiveness and generalizability of the model in classification tasks, as well as its adaptability across different datasets and scenarios.