Non-Verbal Human-Robot Interaction Using Neural Network for The Application of Service Robot

Abstract

Service robots are prevailing in many industries to assist humans in conducting repetitive tasks, which require a natural interaction called Human Robot Interaction (HRI). In particular, nonverbal HRI plays an important role in social interactions, which highlights the need to accurately detect the subject's attention by evaluating the programmed cues. In this paper, a conceptual attentiveness model algorithm called Attentive Recognition Model (ARM) is proposed to recognize a person’s attentiveness, which improves the accuracy of detection and subjective experience during nonverbal HRI using three combined detection models: face tracking, iris tracking and eye blinking. The face tracking model was trained using a Long Short-Term Memory (LSTM) neural network, which is based on deep learning. Meanwhile, the iris tracking and eye blinking use a mathematical model. The eye blinking model uses a random face landmark point to calculate the Eye Aspect Ratio (EAR), which is much more reliable compared to the prior method, which could detect a person blinking at a further distance even if the person was not blinking. The conducted experiments for face and iris tracking were able to detect direction up to 2 meters. Meanwhile, the tested eye blinking model gave an accuracy of 83.33% at up to 2 meters. The overall attentive accuracy of ARM was up to 85.7%. The experiments showed that the service robot was able to understand the programmed cues and hence perform certain tasks, such as approaching the interested person.

ABSTRAK: Robot perkhidmatan lazim dalam banyak industri untuk membantu manusia menjalankan tugas berulang, yang memerlukan interaksi semula jadi yang dipanggil Interaksi Robot Manusia (HRI). Khususnya, HRI bukan lisan memainkan peranan penting dalam interaksi sosial, yang menonjolkan keperluan untuk mengesan perhatian subjek dengan tepat dengan menilai isyarat yang diprogramkan. Dalam makalah ini, algoritma model perhatian konseptual yang dipanggil Model Pengecaman Perhatian (ARM) dicadangkan untuk mengenali perhatian seseorang, yang meningkatkan ketepatan pengesanan dan pengalaman subjektif semasa HRI bukan lisan menggunakan tiga model pengesanan gabungan: pengesanan muka, pengesanan iris dan mata berkedip. . Model penjejakan muka telah dilatih menggunakan rangkaian saraf Memori Jangka Pendek Panjang (LSTM), yang berdasarkan pembelajaran mendalam. Manakala, pengesanan iris dan mata berkelip menggunakan model matematik. Model mata berkelip menggunakan titik mercu tanda muka rawak untuk mengira Nisbah Aspek Mata (EAR), yang jauh lebih dipercayai berbanding kaedah sebelumnya, yang boleh mengesan seseorang berkelip pada jarak yang lebih jauh walaupun orang itu tidak berkelip. Eksperimen yang dijalankan untuk pengesanan muka dan iris dapat mengesan arah sehingga 2 meter. Sementara itu, model berkelip mata yang diuji memberikan ketepatan 83.33% sehingga 2 meter. Ketepatan perhatian keseluruhan ARM adalah sehingga 85.7%. Eksperimen menunjukkan bahawa robot perkhidmatan dapat memahami isyarat yang diprogramkan dan seterusnya melaksanakan tugas tertentu, seperti mendekati orang yang berminat.