Gesture recognition for social service robots in the elderly care

Due to the scarcity of caregivers in nursing homes, there's the need to ensure that the same number of professionals is able to do more and deliver better care services: the introduction of social service robots can reduce the workload, giving the care workers more time for human-specific tasks. With the purpose of enabling more natural and intuitive interactions between humans and robots and offering a customized service to each patient, this thesis presents a gesture recognition system that is able to recognize full-body commands, performed both statically (such as standing and sleeping) and dynamically (such as waving the hands or shaking the head), solely through RGB frames. The system has two main blocks: the extraction block, which extracts the body pose estimations through OpenPose, and the evaluation block, which predicts the gesture performed using a Nearest-neighbor classifier with Fast Dynamic Time Warping as distance metric. The correctness of the prediction of the 12 commands is evaluated: the tests lead to a hit rate of 74%. Additional experiments are conducted to verify the robustness of the system: in different lighting conditions (even in dark environments) the gesture recognition system is able to keep a comparable accuracy with respect to normal conditions; while increasing the distance dropped the hit rate.