Capabilities of stereo vision systems for future space missions

Future space missions are likely to exploit advanced computer vision systems, providing high accuracy and a rich and comprehensive perception of the scenario/environment. Till now, main tasks of computer vision has been the detection and recognition of objects of an already known shape: this approach was adopted in the automatic rendezvous, carried on by identifying specific markers in acquired images. Improvements in optical sensors' quality, software skills and computational power allow today to attempt a full reconstruction of the observed scene, and this new approach could greatly expand the applications of computer vision in space missions. Stereo vision is a suitable option for these advanced techniques, offering significant advantages in terms of cost and constraints to the host spacecraft with respect to active vision systems, and probably the more convenient solution for small platforms. Binocular configurations, or even multiple points of view, enable a full 3D representation of the scenario, providing the information about distance missing from a pinhole representation. Specific algorithms are obviously needed to combine the different images, matching up extracted features. The paper discusses stereoscopic vision with a focus on the applications to in-orbit robotic activities and rover navigation. The basic relations as well as the technique to match the features captured by the two cameras, leading to the evaluation of the optical depth in the scene, are briefly recalled. With specific reference to the case of a rover, the detail of the implementation of a stereo vision system on board a small vehicle called RAGNO designed and built a the Guidance and Navigation Lab of Sapienza Universita’ di Roma are presented. The system is able to autonomously drive the rover towards a defined target, also selected from gathered images, while identifying and avoiding intermediate obstacles. Indeed, a complete guidance, navigation and control function with not-toosophisticated hardware has been designed and tested. Characteristics and quantitative figures of this successful implementation are presented.