Development of robot navigation method based on single camera vision using deep learning

摘要

This paper presents a complete vision guided real-time approach to robot navigation in urban narrow roads based on drivable road area detection using deep learning. In this approach, an illuminant-invariant road database is created from captured images. This database is used to train the Deep Belief Neural Network (DBNN) for road detection. During navigation, a camera takes a snapshot of the road and then the captured image is converted into an illuminant-invariant image. After that, DBNN takes this image as an input. It extracts the road features layer-by-layer for detection. The experimental wheelchair robot follows detected road boundary for navigation. The performance of the developed algorithm is demonstrated by the experiments. In addition, we encompass the large areas of autonomous robot navigation in a single camera.