Autism Spectrum Disorders (ASD) cover a set of developmental disabilities characterised by some degree of impaired social attitude, communication and language, with the presence of repetitive and stereotyped behaviours or interests. Worldwide, 1 in 160 children has ASD. Intervention during early childhood is important to promote development and long-term positive effects. Children with ASD exhibit significant impairments both in imitation of gestures as well as in their spontaneous use, which some studies have proved being related to the development of social engagement. Newest interventions are based on Human Robot Interaction (HRI) since children with ASD cope well with rule-based and predictable systems rather than human beings. This project's aim is to implement a proper algorithm to automatically recognize gestures within a humanoid-robot therapy for children with ASD. The robot reacts to children's movements when interacting with them. A positive or negative feedback is triggered depending on the algorithm's assessment of the performed gesture. In this way, children's engagement is increased thanks to reinforcement learning, thus empowering the treatment's robustness and decreasing therapist's workload in the Therapist-Robot-Child triad interaction. To achieve the goal, a ResNet was exploited since artificial neural networks are widely used in recognition tasks. Both Offline and Online recognitions were implemented and different settings were experimented to comply with the protocol and its set up. The algorithm was able to recognize 19 different type of gestures and tested with new acquisitions. Healthy subjects' experiments were conducted at NearLab in Politecnico di Milano and reached an accuracy of 97%. 4 ASD children were involved in the therapy at CARElab in Fondazione Don Gnocchi obtaining an accuracy of 83% with 10 type of gestures performed. Children had different ways to approach the therapy, thus the robot, so different engagement levels were detected. Up to date, results have demonstrated successfully the use of a gesture recognition algorithm for the purpose of increasing ASD children engagement and empowering gestures' learning. Outcomes obtained with healthy adults and autistic children were encouraging and pave the way to new developments in the near future.
I disturbi dello spettro autistico (DSA) coprono una serie di deficit neuroevolutivi caratterizzati da difficoltà nello stabilire relazioni sociali, nella comunicazione e nel linguaggio, con la presenza di comportamenti ripetitivi e stereotipati. In tutto il mondo, 1 bambino su 160 ha DSA. L'intervento durante la prima infanzia è importante per promuovere lo sviluppo e gli effetti positivi a lungo termine. I bambini affetti da DSA presentano significativi deficit sia nell'imitazione di gesti sia nel loro utilizzo spontaneo. Alcuni studi hanno dimostrato una correlazione tra questi deficit e lo sviluppo dell'interazione sociale. Le terapie più recenti si basano sulla Human-Robot Interaction (HRI) poiché i bambini con DSA focalizzano più facilmente la loro attenzione su sistemi prevedibili come i robot, piuttosto che sugli esseri umani. Lo scopo di questo lavoro è stato implementare un algoritmo per il riconoscimento di gesti in modo automatico in una robot terapia per bambini con DSA. Il robot analizza i movimenti dei bambini durante l'interazione, generando un feedback positivo o negativo a seconda della valutazione del gesto eseguito, aumentando così il coinvolgimento dei bambini nella terapia. Con l'integrazione dell'algoritmo di riconoscimento, l'efficacia del trattamento aumenta, mentre si alleggerisce il carico di lavoro del terapeuta nell'interazione triadica terapeuta-robot-bambino. Nell'ambito delle tecnologie per il riconoscimento di gesti, le reti neurali artificiali sono ampiamente utilizzate. Per questo motivo, l'obiettivo è stato raggiunto implementando una rete convoluzionale ResNet. Il sistema di interpretazione di gesti è stato sviluppato sia offline che online e sono state sperimentate impostazioni differenti per conformarsi al protocollo e al suo set up. L'algoritmo riconosce 19 tipi di gesti ed è stato testato su nuove acquisizioni. Gli esperimenti su soggetti sani sono stati condotti al Nearlab del Politecnico di Milano e hanno raggiunto una precisione del 97%. 4 bambini con DSA sono stati coinvolti nella terapia presso il CARElab della Fondazione Don Gnocchi ottenendo, fino ad ora, una precisione dell'83% con 10 tipi di gesti eseguiti. Ogni bambino si è approcciato in modo diffrente alla terapia e al robot, pertanto sono stati riscontrati diversi livelli di coinvolgimento. I risultati hanno dimostrato con successo l'uso di un algoritmo di riconoscimento di gesti allo scopo di aumentare la partecipazione di bambini DSA nella terapia e potenziare l'apprendimento dei gesti. I risultati ottenuti con adulti sani e bambini autistici sono stati incoraggianti e aprono la strada a nuovi sviluppi nel prossimo futuro.
Guess my gesture : a gesture recognition algorithm in a robot therapy for ASD children
Giubergia, Alice;IVANI, ALESSIA SILVIA
2019/2020
Abstract
Autism Spectrum Disorders (ASD) cover a set of developmental disabilities characterised by some degree of impaired social attitude, communication and language, with the presence of repetitive and stereotyped behaviours or interests. Worldwide, 1 in 160 children has ASD. Intervention during early childhood is important to promote development and long-term positive effects. Children with ASD exhibit significant impairments both in imitation of gestures as well as in their spontaneous use, which some studies have proved being related to the development of social engagement. Newest interventions are based on Human Robot Interaction (HRI) since children with ASD cope well with rule-based and predictable systems rather than human beings. This project's aim is to implement a proper algorithm to automatically recognize gestures within a humanoid-robot therapy for children with ASD. The robot reacts to children's movements when interacting with them. A positive or negative feedback is triggered depending on the algorithm's assessment of the performed gesture. In this way, children's engagement is increased thanks to reinforcement learning, thus empowering the treatment's robustness and decreasing therapist's workload in the Therapist-Robot-Child triad interaction. To achieve the goal, a ResNet was exploited since artificial neural networks are widely used in recognition tasks. Both Offline and Online recognitions were implemented and different settings were experimented to comply with the protocol and its set up. The algorithm was able to recognize 19 different type of gestures and tested with new acquisitions. Healthy subjects' experiments were conducted at NearLab in Politecnico di Milano and reached an accuracy of 97%. 4 ASD children were involved in the therapy at CARElab in Fondazione Don Gnocchi obtaining an accuracy of 83% with 10 type of gestures performed. Children had different ways to approach the therapy, thus the robot, so different engagement levels were detected. Up to date, results have demonstrated successfully the use of a gesture recognition algorithm for the purpose of increasing ASD children engagement and empowering gestures' learning. Outcomes obtained with healthy adults and autistic children were encouraging and pave the way to new developments in the near future.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_Giubergia_Ivani.pdf
accessibile in internet per tutti
Dimensione
14.3 MB
Formato
Adobe PDF
|
14.3 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/171336