Autonomous vehicle navigation in outdoor and off-road environments based on computer vision

Abstract

The main goal of this work is to enable an unmanned ground vehicle to maneuver autonomously in an external and non structured environments or loosely structured environments, as for example, a few dense forests or in crop fields. This vehicle must be able to avoid obstacles, autonomously detecting them, and driving itself to a predetermined location, reasoning and defining by its own means the pathway to follow. The navigation system will be based on image processing, in fact, by a pair of cameras (stereo vision), this way building a three dimensional perception of the environment. Our intention is to extract the navigability parameters from the environment, like free paths, obstructions and obstacles, and combine with the positioning, orientation and destiny location (based on GPS coordinates) to build a robust navigation system.We will use in this work the CaRINA I platform (Intelligent Robotic Car for Autonomous Navigation) from LRM-ICMC/USP that is equipped with a stereo camera, a GPS and an electronic compass. At the start of this project, it will be studied and manipulated some algorithms for building the disparity map from the pair of images obtained from the stereo camera. In this process, the algorithm must be compromised on delivering a good response time (due to soft real-time restriction) and quality of the generated map. Using this disparity map, it will be possible to create a navigation map, that represents the navigable (safe) and non navigable (obstacles, hazards to avoid) regions in front of the vehicle. This map will be used in conjunction with the information of actual position amd destination (indicated by GPS and compass) to accomplish the navigation control of the vehicle. We will consider two main approaches for this control, one based on neural networks, like presented in RoBombeiros work (previously developed by our group by Pessin and Osorio), and the other by using the VFH (Vector Field Histogram) algorithm. In both approaches will be considered as inputs the three dimensional information obtained from the navigability map. Beyond that, it will also be necessary develop studies in order to enable the detection of the ground based on the stereo images, and also detect the obstacles in the ground (e.g. holes) and leveling, classifying it as navigable or not.The main applications of this autonomous robot navigation system are:(i) fire combat, like proposed by previous works of Pessin and Osorio;(ii) agricultural applications focusing on the building of autonomous vehicles for land plowing, seeding, pulverizing and harvesting;(iii) military applications or civil ones, on the transportation of duties or supplies on unstructured environments;(iv) the transportation of loads in dangerous environments or where the presence of a driver can be dispensable by the meaning of protecting its security.This kind of technology has several possible applications in our society, being of great relevancy the research and development of new technologies in related areas of this project, as indicates some important international initiatives and research efforts, for example, in the U.S. DARPA Grand/Urban Challenge initiative, the European ELROB, and the world-wide AUVSI/IGVC contest. (AU)