A gesture recognition system for mobile robots that learns online

This paper presents a gesture recognition framework designed for controlling mobile robots using dynamic hand gestures. The system is unique in that it can be trained to recognize a new gesture by simply demonstrating the gesture once. From then on, the system improves its recognition accuracy over time by learning from experience. The classification scheme utilizes a nearest neighbor approach with a dynamic time warping distance metric. The approach robustly handles both temporal and spatial variations in executions of highly dynamic and complex gestures. One product of the research is an open source software package that will work with little to no modifications on robots utilizing a 3D depth sensor and the Robot Operating System (ROS). Experimental results show the ability of novice users to train and utilize the system effectively, as well as the system´s ability to increase its recognition accuracy over time by learning from experience.