Distance Estimation Using a Bio-Inspired Optical Flow Strategy Applied to Neuro-Robotics

Abstract

Movement detection and characterization of a 3D scene are relevant tasks in vision systems and particularly in robotic applications controlled by visual features. One of the challenges to characterize a 3D scene in navigation systems is the depth estimation. In contrast to classical approaches using visual based stereo systems, we propose a monocular distance estimation system using convolutional neural networks (CNN) and a bio-inspired optical flow approach as part of a neuro-robotic system. We train the CNN using optical flow visual features guided by ultrasonic sensor-based measures in a 3D scenario. The datasets used are available in: http://sites.google.com/up.edu.mx/robotflow/. Experimental results confirm that a monocular camera can be applie for controlling the robot navigation and obstacle avoidance.