Object recognition in Electrical Impedance Tomography

  • Nadya A. Madjid

Student thesis: Master's Thesis


Equipping robots with sensing capabilities similar to the human sense of touch will enhance their adaptability to unexpected changes in the environment and their capability of perceiving objects from their surroundings. Electrical Impedance Tomography (EIT) principles can be used to develop a stretchable and flexible low-cost tactile sensor with a high sensitivity range, which based on the voltage differences measured on the sensor boundary can reconstruct images of the internal impedance distribution. This type of sensors can be used in the context of object recognition, specifically, when the object comes into contact with the surface of the sensor, with its shape and location being inferred from the associated impedance images. Inferring these characteristics of the interacting object from EIT impedance images is considered challenging due to the low spatial resolution of the images, fuzzy object boundaries and the presence of noise in the form of the image artefacts. The solution proposed in this thesis to cope with the low quality of EIT images is a novel object recognition framework. Given a database of the impedance profiles corresponding to the set of objects robots can interact with, the developed system takes as input an impedance image and provides as output the category of the interacting object. The proposed framework consists of two sub-modules, i.e., image segmentation and object classification. For image segmentation, five methods were explored for segmenting EIT images, specifically, three state-of-the-art non-data driven algorithms and two machine learning approaches i.e., CNN and Random forest with textural features. For the object classification subsystem, a set of ML classifiers were trained on selected shape descriptors computed from the segmented images and their performance was evaluated using 10-fold cross validation. The integrated system achieved accuracy of approx. 98% in classifying the set of generated objects using a CNN approach, i.e., MobileNetV2, as the pre-trained encoder for object segmentation, and the Ensemble Subspace KNN method for object classification.
Date of AwardJul 2020
Original languageAmerican English


  • tactile sensing
  • EIT
  • image segmentation
  • boundary descriptors
  • textural features
  • transfer learning

Cite this