The Computer Vision Department is concerned with the study and the development of innovative algorithms and advanced systems which, making use of vision sensors, perform automatic non-contact dimensional measurements with high accuracy. Within this context several aspects of computer vision have been carried out in recent years. They span from sensor and optics calibration to automatic object description and recognition, stereo measurements, sensor pose estimation and 3D-scene reconstruction. This work is stimulated by applications in the fields of robotics, autonomous navigation, surveillance and cultural heritage preservation.
The problem of object recognition has been further investigated in order to identify autonomously goals to be reached and to detect obstacles which can hinder the path of the vehicle. Namely, the development of geometrical models for the interpretation of 3D scenes containing polyhedral objects has been carried on facing the problem of labelling planar projections of polyhedra. Geometrical feature grouping, object detection, description and recognition algorithms have been developed to contribute to an accurate localisation of goals and obstacles.
Top: the environment where the autonomous vehicle moves. Bottom: the scene as seen by the camera on the robot (left); the segmented scene with the recognized object in red (right).
In the framework of a contract with Alcatel Alenia Spazio Italia, a specific module of the vision system has been developed for the ESA Eurobot Flight Model (EFM), a robotic system designed to support Extravehicular Activity on the International Space Station (ISS) and future planetary missions. In particular, algorithms have been developed and tested for the visual registration of the robot pose with respect to the ISS. The registration methodology relies on the comparison of the actually observed scene with the rendering of the CAD model of the environment from the planned robot pose, and works both with a single camera and with a multiple-camera head.
Hardware setup and man-machine interface for pose registration with a camera on the robot wrist, simulating a stereo head.
Research has been carried on in the area of sequential SLAM algorithms based on stereo vision and occupancy grid maps. In particular, an original approach has been developed strongly reducing the incremental pose estimation error, by combining visual odometry from stereo with grid map correlation. In this approach, the angular motion of a rover is estimated by tracking and registering clouds of 3D points of the environment, obtained from stereo while the robot moves. Positional accuracy is provided by correlating panoramic local maps, acquired by panning the stereo head at successive key positions, with an incrementally updated global map.
[
full size image
-
watch the video
]
Top: The Pioneer 3DX robot used for testing SLAM algorithms (left); panoramic view of the robot environment (right).
Bottom: estimated robot trajectory on a CAD model of the environment (left); accumulated 3D point clouds and estimated robot poses (right).
A standard progressive video camera, equipped with an appropriate focal length objective, was used for territory monitoring applications. Experimental results show that it is possible to reach an accuracy of a few millimeters in the measurement of displacements at a distance of 200 m. They make use of natural light invariant reference points as features for the motion evaluations while succeeding images of the same location are evaluated. Thanks to the use of natural targets, this technology can be particularly useful for the monitoring of geological instabilities in unreachable areas. Results achieved, namely the detection of a small movement of about 75 mm registered between 12 and 23 December 2004 on the Gardiola landslide, have suggested an improvement in the equipment used with the addition of high precision rotating tables, which allow automatic bearing of the camera, and so the evaluation of large landslide areas.
The new automatic equipment designed for scanning large areas, during the performance tests carried out in INRIM laboratories.