Identification of limb movements in bandages

Abstract

This thesis focuses on the recognition of upper-limb movements using capacitive textile sensors embedded in an elastic bandage. The aim was to design and evaluate a system capable of capturing and classifying arm movements in real time. A conductive textile sensor was integrated into an elbow bandage and connected to an ESP32 based data acquisition system. Measurements from 20 participants performing daily activities were used to create a dataset for training classification models. A neural network based approach was applied for activity recognition. The model was chosen in terms of computational cost parameters. The results show that the system achieves an overall inference accuracy of 0.9449 and an F1 score of 0.9586 for high intensity tasks. However, limitations were identified in distinguishing similar activities, and a significant reduction in performance was noted due to model quantization for embedded deployment. Specifically, INT8 quantization noise in earlier non-hierarchical experiments reduced the recall of stationary states like sleep to 0.0000. The finalized hierarchical structure successfully compensated for these inherent accuracy losses, recovering the sleep recall to 0.7550 and maintaining system stability above 76% for all activities. The results confirm that textile capacitive sensors are suitable for wearable motion monitoring, offering a balance between accuracy, comfort, and efficiency. Additionally, robustness tests revealed that the pre-trained model is highly sensitive to physical sensor displacement, highlighting that precise wearable positioning is a strict requirement for cross-user generalization. However, more research needs to be done in order to design deployable embedded models.