Multisensory Wearable Glove for Object Recognition Based Embedded Machine Learning

This paper proposes a low-cost multisensory glove (≈ 140 USD) equipped with commercial piezoresistive force sensors (FSRs) and inertial measurement units (IMUs) for object recognition. A set of 28 daily life objects is used to evaluate the glove by applying grasping actions. The raw signals acquired through these actions are employed directly to train a shallow one-dimensional convolutional neural network. The network is deployed on a low-cost edge system to perform online object recognition, thereby presenting an end-to-end system. The proposed system achieves a classification accuracy of 99.29% while consuming only 59.85 mW and 0.4875 mJ of power and energy per inference respectively. Overall, the proposed system seems a suitable candidate for various biomedical applications such as post-stroke rehabilitation.