A socially intelligent robot is an advanced robotic system that analyzes
user events and recognizes social cues, with which we can estimate the psychophysiological
state, level of attention and cognitive workload of the user.
By incorporating assistive technologies with socially intelligent robots in residential
or healthcare environments we can provide additional care for vulnerable
groups with developmental impairments. We developed an optimized
computer vision pipeline for detection of social cues from a video stream,
which operates faster than in real-time with minimal latency. We compared
the processing time of classic and modern models for face detection and
estimation of the human pose. We trained a machine learning model for binary
classification of attention on the Students Attention Detection dataset
and evaluated it using 10-fold validation. We tested the performance of our
pipeline and attention prediction on our own video dataset featuring elderly
people. Based on the results, we determined the optimal configuration of the
detection model and other parameters with the lowest possible latency.
|
