Force Myography Hand Gesture Recognition Using Transfer Learning
Authors
Andersen, Thomas Buhl ; Eliasen, Rógvi ; Jarlund, Mikkel
Term
4. term
Education
Publication year
2020
Submitted on
2020-06-11
Pages
11
Abstract
Force myografi (FMG) får stigende opmærksomhed til at genkende håndbevægelser ved at måle ændringer i muskeltryk via sensorer på huden. I dette arbejde vil vi forbedre nøjagtigheden af gestusgenkendelse ved at udnytte data fra flere personer gennem transfer learning, dvs. at bruge viden fra én gruppe til at hjælpe en anden. Vi afprøver to neurale netværkstilgange: et progressivt neuralt netværk og en kombineret progressiv variant, som forsøger at lære mere generaliserede kendetegn fra kildedomæner. Vores resultater viser, at transfer learning kan give bedre resultater end et traditionelt fuldt forbundet neuralt netværk, men at det skal anvendes med omtanke for at undgå negativ transfer, hvor overført viden faktisk svækker ydeevnen. Under arbejdet bemærker vi også, at der mangler offentligt tilgængelige benchmarkdata: Mange studier indsamler egne datasæt med specialbygget hardware og forskellige sæt af gestus, hvilket både begrænser mulighederne for videnoverførsel og gør det svært at sammenligne metoder. For at hjælpe feltet fremad stiller vi et benchmarkdatasæt til rådighed, indsamlet med et kommercielt sensorsetup fra 20 personer, der udfører 18 unikke gestus. Vi håber, at andre kan bruge disse data til lignende transfer learning-tilgange og til mere retfærdige sammenligninger af algoritmer, samtidig med at det bliver lettere at komme i gang med FMG-baseret gestusgenkendelse.
Force myography (FMG) is gaining attention for recognizing hand gestures by measuring changes in muscle pressure with sensors on the skin. In this work, we aim to improve gesture recognition accuracy by using data from multiple people through transfer learning—that is, leveraging knowledge from one group to help another. We test two neural network approaches: a progressive neural network and a combined progressive variant designed to learn more generalized features from source domains. Our results show that transfer learning can outperform a standard fully connected neural network, but it must be applied carefully to avoid negative transfer, where imported knowledge actually harms performance. We also observe a shortage of publicly available benchmark data: many studies collect their own datasets with custom hardware and different gesture sets, which limits both knowledge transfer and fair comparison of methods. To support progress in the field, we release a benchmark dataset collected with a commercially available sensor setup from 20 participants performing 18 unique gestures. We hope others will use these data for similar transfer learning approaches, enable more comparable results, and lower the barrier to entry for FMG-based gesture recognition.
[This abstract was generated with the help of AI]
Keywords
Documents