Object recognition based on flexible tactile sensor glove and deep learning

Abstract

In recent years, the development of human-machine interaction systems has increasingly relied on haptic sensing technology, which enables machines to sense and respond to physical stimuli. The flexible tactile sensor glove has attracted attention for its potential applications in robotics, healthcare and assistive technology. These systems offer a new approach to object recognition by capturing tactile data, which can be processed using deep learning models to improve recognition accuracy. This dissertation investigates the design and realization of a flexible tactile glove system for object identification. The experiment involves the fabrication of twenty flexible pressure sensors, which are subsequently embedded into a glove capturing tactile data from eight distinct objects. The data collected from the pressure sensors are pre-processed and utilized to train deep learning models, including LSTM, 1D-CNN, 2D-CNN, 2DCNN-LSTM, and MLP for objects classification. This study focuses on how the configuration of sensors and the selection of network models significantly influence improvements in recognition performance. Experimental results indicate that the 2D-CNN model achieved the highest accuracy, reaching 95.6%. This is attributed to the model's ability to generate superior generalization, making it particularly effective in handling the spatial texture information within tactile data. Similar to previous studies demonstrating the viability of sensor gloves and hybrid neural networks for object identification, this work highlights the issues of model overfitting and emphasizes the need for larger datasets to achieve better generalization. These findings contribute to advancements in human-computer interaction, robotic, and assistive technologies, paving the way for future research into a more refined and adjustable system.