A simple and cost effective new tactile sensor is presented, based on a camera capturing images of the shading of a deformable rubber membrane. In Computer Vision, the issue of information encoding and classification is well studied. In this paper we explore different ways of encoding tactile images, including: Hu moments, Zernike Moments, Principal Component Analysis (PCA), Zernike PCA, and vectorized scaling. These encodings are tested by performing tactile shape recognition using a number of supervised approaches (Nearest Neighbor, Artificial Neural Networks, Support Vector Machines, Naive Bayes). In conclusion: the most effective way of representing tactile information is achieved by combining Zernike Moments and PCA, and the most accurate classifier is Nearest Neighbor, with which the system achieves a high degree (96.4%) of accuracy at recognising seven basic shapes.
|Name||Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics|
|Conference||Proceedings of the 15th Annual Conference, TAROS, 2014|
|Country||UK United Kingdom|
|Period||1/09/14 → 3/09/14|