Extraction of cutaneous primitives from tactile sensor images

Abstract

Tactile array sensors give images of contact. Different approaches are possible to analyze these images. In this paper the authors present a new approach to tactile data analysis based on the extraction of tactile primitives. The technique consisting of the extraction of primitives was initially developed for visual images processing, but tactile data are quite different from a camera image. Tactile sensors have low resolution and are not affected by perspective; on the other hand tactile data present high distortion. (due to the sensor protective layer) and a tensorial nature. Moreover, tactile sensors can usually only one primitive, while cameras see more. For this reason they need a different, and still largely unexplored, type of analysis. The system described here uses a set of concurrent neural nets, in order to determine the tactile primitive (and its features) that better describe the contact area between an object and a tactile sensor. Input data are a scalar combination of the normal stress components inside the sensor. Simulations show that the neural system learns and has a high noise immunity. >