Tactile image based contact shape recognition using neural network

This paper proposes a novel algorithm for recognizing the shape of object which in contact with a robotic finger through the tactile pressure sensing. The developed algorithm is capable of distinguishing the contact shapes between a set of low-resolution pressure map. Within this algorithm, a novel feature extraction technique is developed which transforms a pressure map into a 512-feature vector. The extracted feature of the pressure map is invariant to scale, positioning and partial occlusion, and is independent of the sensor's resolution or image size. To recognize different contact shape from a pressure map, a neural network classifier is developed and uses the feature vector as inputs. It has proven from tests of using four different contact shapes that, the trained neural network can achieve a high success rate of over 90%. Contact sensory information plays a crucial role in robotic hand gestures. The algorithm introduced in this paper has the potential to provide valuable feedback information to automate and improve robotic hand grasping and manipulation.