The activity aims to study and develop techniques addressed at perceiving, understanding and making measurements on the environment where autonomous systems operate using vision sensors. The goal is the interpretation of scenes in order to get an explicit description of their content, detect and recognise objects, measure their dimensions and position and evaluate their space-time behaviour. The research aims to develop sensors, applicable to flexible manufacturing systems, autonomous navigation, robotics, and to develop instrumentation for recording and preserving the cultural heritage and for environment monitoring.

The project on Simultaneous Localisation And Mapping (SLAM) is aimed to study algorithms for an accurate reconstruction of the mobile platform trajectory from visual data only (Visual Odometry). An original approach has been developed where points of interest are extracted and matched in a pair of images taken by the on board stereo head, then tracked along the image sequence grabbed while the robot moves. The motion of the robot is finally estimated by bundle adjustment of the tracked features.

Top: left and right images of a stereo pair.
Bottom: estimated motions of feature points (left); stereo disparities of the same features (right).


The Pioneer 3DX robot used in the tests.


The estimated trajectory.

The cooperation with Alcatel Alenia Spazio Italia has been continued with the development of a vision system for the ESA Eurobot WET Model (EWM), a prototype of a robotic system designed to support Extravehicular Activity on the International Space Station (ISS) and future planetary missions. During 2007 the EWM shall be tested underwater, in order to simulate extraterrestrial microgravity conditions.

An artist's rendition of the Eurobot walking along a series of handrails
(courtesy of Alcatel Alenia Spazio, Italia).
Click here to watch the video.

The video system installed in front of the Gardiola landslide has been improved with the addition of high precision rotating tables, which allow the automatic orientation of the camera and the evaluation of large landslide areas. Starting from the SIFT algorithm developed by Lowe, a new approach has been investigated in order to detect in each image a large number of characteristic points invariant to light changes, that have been used to evaluate possible movements. Preliminary results seem to be encouraging, even if the problem relative to changes in light source direction due to earth rotation must be faced.

The video camera and the rotating tables.

The research activities concern applications of active vision to cultural heritage monitoring and preservation, and to industrial inspection and quality control. The development of prototypes of vision systems for three-dimensional acquisition, reconstruction, and modelling of real objects leads to face problems related to the accurate reconstruction of the geometry and the texture of the objects from their sparse 3D samples, or scans. A fundamental step in this process is the registration, that is the alignment, in a common co-ordinate system, of several scans produced in the acquisition phase. New algorithms based on local curvature of the data set and on geometric invariants have been developed to solve this problem efficiently and robustly. The dimensional accuracy of the final models has been evaluated and compared with other independent acquisition methods.

The Cycladic Thinker, left view curvatures, front view curvatures, and coarse registration of the Thinker scans.