Omni-stereo vision system for an autonomous robot using neural networks

Autonomous robots' vision is one the most important parts in their navigation system and omni-directional stereo vision is an approach to improve this module. Conventional cameras have restricted field of view while omni-directional cameras provide 360-degree field of view. Depth information obtained from stereo vision is very useful for robot navigation in complex environments. This paper presents an omni-directional stereo vision system for Arvand robots constructed by Sharif CE Middle size RoboCup team. We have used two catadioptric cameras aligned vertically. To compute depth, we have used a neural network. In our approach, we trained the neural network without unwrapping the images and without calibration, by using a small training set and acceptable training time. This is an important factor when such systems need to be used in research labs to make prototype versions. Our experimental results were satisfactory and we could map images to depth map in real time.