Obstacle Detection Using Faster R-CNN Oriented to an Autonomous Feeding Assistance System

Abstract

Obstacle detection has been a relevant issue for the implementation of autonomous robotic systems, within which increasingly robust algorithms have begun to be applied, especially Deep Learning techniques. However, these have not been widely used for the detection of obstacles in static robotic agents, contrary to what happens with mobile agents. For this reason, this work explores the use of one of these techniques, which is a neural network based on the Faster R-CNN, focused on detecting a specific obstacle (hands) in an application environment for a food assistance robot. For this purpose, a database containing 6205 training images and 1350 validation images was prepared, where 31 users perform different movements with their hands. To verify the capacity of the network, 3 architectures of different depths were implemented, which were evaluated and compared, resulting in the network of greater depth obtained the highest accuracy, of 77.4%, taking into account that the hands are not only still but also in movement, generating distortion in them and greater difficulty for their detection. Also, the internal behavior of the network was visualized through activations, to verify what it had learned, showing that it managed to focus on the hands, with some activations located in parts of the user's body such as face and arm.