TinyML System for Touch Modality Classification Based on Multisensory Glove

This paper presents a TinyML system based on a multisensory glove for touch modality classification. It employs a glove equipped with five tactile sensors and five IMUs to collect a dataset for five different touch modalities used to train a neural network classifier based on a one-dimensional convolution neural network (1-D CNN). Classification accuracy demonstrated the effectiveness of the proposed model achieving an accuracy of 100% and 98% when float32 and int8 data representation were used respectively. Moreover, the proposed 1-D CNN model is deployed for the low-cost STM32 F401-RE for real-time classification. Hardware evaluation results demonstrate that the system achieves a low energy consumption of 0.636 mJ for int8 with a latency of 1.125 ms. The system paves the way toward intelligent and energy-autonomous wearable systems.