Self-Supervised Convolutional Neural Networks for Fast Gesture Recognition in Human-Robot Interaction

Abstract

Current autonomous systems (e.g., self-driving cars, autonomous drones, consumer robots, etc.) can already perform a wide variety of tasks and are predicted to be able to collaboratively assist humans in the near future. Thus, the need for efficient human-robot interaction (HRI) methods is greatly increasing. Gesture recognition is an effective HRI approach, since many robots are equipped with cameras and computer vision algorithms have progressed significantly in recent years, with advanced Deep Neural Networks (DNNs) being able to be executed on-board an autonomous system. However, computational/memory limitations are still significant for embedded AI methods, rendering the increase of DNN accuracy without imposing a penalty on runtime requirements a very important research priority. This paper investigates self-supervised DNN pretraining for a novel pretext task, relying on spatiotemporal video frame compression via tensor decomposition and low-rank approximation, as a means to augment gesture recognition performance, without inducing any runtime overhead during the inference stage. Thus, the method permits the use of less complex and much faster neural architectures that are well-suited to robotics applications and HRI. Quantitative evaluation on a gesture recognition dataset for autonomous Unmanned Aerial Vehicle (UAV) handling demonstrates the effectiveness and real-time performance of the proposed method on embedded AI compute hardware.