A tactile glove for object recognition based on palmar pressure and joint bending strain sensing

With the rapid development of flexible electronics, the tactile systems for object recognition are becoming increasingly delicate. This paper presents the design of a tactile glove for object recognition, integrating 243 palm pressure units and 126 finger joint strain units that are implemented by piezoresistive Velostat film. The palm pressure and joint bending strain data from the glove were collected using a two-dimensional resistance array scanning circuit and further converted into tactile images with a resolution of 32×32. To verify the effect of tactile data types on recognition precision, three datasets of tactile images were respectively built by palm pressure data, joint bending strain data, and a tactile data combing of both palm pressure and joint bending strain. An improved residual convolutional neural network (CNN) model, SP-ResNet, was developed by light-weighting ResNet-18 to classify these tactile images. Experimental results show that the data collection method combining palm pressure and joint bending strain demonstrates a 4.33% improvement in recognition precision compared to the best results obtained by using only palm pressure or joint bending strain. The recognition precision of 95.50% for 16 objects can be achieved by the presented tactile glove with SP-ResNet of less computation cost. The presented tactile system can serve as a sensing platform for intelligent prosthetics and robot grippers.