Skip to main content

Advanced vision based concurrent mapping and localization

Final Activity Report Summary - AVICMAL (Advanced Vision Based Concurrent Mapping and Localization)

The objective of this project was to improve the understanding of how an autonomous mobile agent can localise itself in the environment using computer vision. Initially, we proposed to complement known vision-based concurrent localisation and mapping techniques with recent advances in pure relative localisation (visual odometry). However, in the implementation phase, we mainly addressed separate mapping and localisation in order to take advantage of the results we achieved after the application to this concourse.

The performed work involves five distinct research directions:
1. autonomous navigation based on separate mapping and localisation procedures;
2. performance evaluation of different closed-form algorithms for recovering the relative orientation in two-view geometry;
3. concurrent localisation and mapping with a calibrated stereo rig;
4. assembling and calibrating a 4DOF setup for controlled motion;
5. collaboration with Gerald Schweighofer on his novel algorithm for stereo structure and motion.

The obtained results are as follows:

1. The work in autonomous navigation, which was the subject of the previous position of the postdoctoral fellow, has been completed and presented at CVPR07 and ACIVS07. A more comprehensive account of the performed work has been submitted to CVIU. Autonomous motion of a car-like vehicle over paths several hundreds meters long has been achieved, by relying exclusively on a single perspective camera.

2. To explain somewhat worse navigation results in mainly planar environments, we performed a comprehensive performance evaluation of closed-form algorithms for recovering the relative orientation in two-view geometry. The results indicate that there is performance degradation in the planar case even with the five-point algorithm. We also proposed a novel explanation of the so-called forward bias which arises when eight-point algorithm is applied to a noisy dataset. The results have been presented at OAGM07 and BenCOS07 (held with CVPR07), and submitted to PAMI.

3. A known software application for monocular concurrent localisation and mapping (LibScene by Andrew Davison) has been extended to accommodate the operation