Tactile sensing systems require embedded processing to extract structured information in many application domains as prosthetics and robotics. In this regard, this paper proposes computationally light strategies to pre-process the sensor signals and extract features, feeding single layer feed-forward neural networks (SLFNNs) that proved good generalization performance keeping low the computational cost. We validate our proposal by integrating a tactile sensing system on a Baxter robot to collect and classify data from three objects of different stiffness. We compare different features extraction techniques and five SLFNNs to show the trade-off between generalization accuracy and computational cost of the whole processing unit. The results show that the processing unit that extracts the mean and standard deviation features from signals and adopts a fully connected neural network (FCNN) with 50 neurons and ReLu activation function achieves a high accuracy (94.4%) in the 3-class classification problem with a low computational cost, leading to the deployment on a resource-constrained device.

Computationally Light Algorithms for Tactile Sensing Signals Elaboration and Classification

Amin Y.;Gianoglio C.;Valle M.
2021-01-01

Abstract

Tactile sensing systems require embedded processing to extract structured information in many application domains as prosthetics and robotics. In this regard, this paper proposes computationally light strategies to pre-process the sensor signals and extract features, feeding single layer feed-forward neural networks (SLFNNs) that proved good generalization performance keeping low the computational cost. We validate our proposal by integrating a tactile sensing system on a Baxter robot to collect and classify data from three objects of different stiffness. We compare different features extraction techniques and five SLFNNs to show the trade-off between generalization accuracy and computational cost of the whole processing unit. The results show that the processing unit that extracts the mean and standard deviation features from signals and adopts a fully connected neural network (FCNN) with 50 neurons and ReLu activation function achieves a high accuracy (94.4%) in the 3-class classification problem with a low computational cost, leading to the deployment on a resource-constrained device.
2021
978-1-7281-8281-0
File in questo prodotto:
File Dimensione Formato  
Computationally_Light_Algorithms_for_Tactile_Sensing_Signals_Elaboration_and_Classification.pdf

accesso aperto

Descrizione: Contributo in atti di convegno
Tipologia: Documento in versione editoriale
Dimensione 3.81 MB
Formato Adobe PDF
3.81 MB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11567/1073659
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 2
  • ???jsp.display-item.citation.isi??? 0
social impact