A tensor-based approach to touch modality classification by using Machine Learning

P. Gastaldo, L. Pinna, L. Seminara, M. Valle, R. Zunino

Classification of 3 touch modalities (brushing, sliding, rolling) from a single individual touching of a 4x4 piezoelectric sensor array. The sensor array provides 16 charge vs time output readings as a result of the stimulus-to-charge transduction of the overall sensing system.

This dataset is open access for different algorithms to be applied and compared to each other to distinguish at best the 3 different modalities of touch during sensor operation. All the material is packed into a password protected zip file. Please contact Paolo Gastaldo (paolo.gastaldo@unige.it) to get the password.

ACQUISITION SYSTEM

The acquisition system is designed to record the output signals provided by a tactile sensor array in the presence of a mechanical stimulus.

The tactile sensor array is based on a Polyvinylidene Fluoride piezoelectric polymer film. PVDF has been mainly chosen for its large electro-mechanical transduction frequency bandwidth (few Hz - 1kHz). The film is provided with 16 patterned square electrodes (taxels) on the lower surface and with a continuous layer (ground) on the upper surface. Corresponding contact geometry is found on the underlying PCB substrate, to extract the lower PVDF signals. The piezoelectric film is protected by a PDMS elastic layer. The mechanical-to-electrical transduction by each PVDF taxel is measured as a generated charge and converted to a voltage by a 16 channel charge amplifier. The interface bandwidth ranges from 2.5 Hz to 1.5 kHz. The signals from filter outputs are sampled at the frequency rate of 3kHz and acquired by a DAQ board (NI PCI-6071E). A 4-th order IIR digital filter is implemented by a LabVIEW standard block and a graphic user interface is developed to save and visualize data in the time domain.

EXPERIMENTAL CAMPAIGN

The dataset is generated by predetermining three possible stimuli (see Experimental Protocol )

In the data-collection process, each of the 70 participants is asked to touch the top surface of the sensor array once for each touch modality - first gesture, moving horizontally over a random line, then repeating the same gesture over a vertical random line. Horizontal and vertical gestures are acquired separately. Upon conclusion of the first run, each participant is asked to repeat the above gestures, under the assumption that second-run gestures prove more natural with respect to the first run. According to the experimental protocol, each participant is allowed to complete every single touch within a time window of 7 seconds. No particular indications are given to the participants about the duration of the stimuli; as a result, the position of the stimulus within the time window and its length can vary. Likewise, the experimental protocol does not provide any directive about the pressure level to be applied. The overall experiment results in a total of 840 patterns, as each participant provided 4 patterns for each stimulus (horizontal and vertical gestures, two runs each). Only the data acquired with the 2nd run should actually be used in the pattern-recognition analysis.

DATASET:

DOWNLOAD