A comparison of several approaches to perform a vision-based long range navigation

摘要

In this paper, we deal with the problem of realizing a vision-based long range navigation task in a cluttered environment. To perform such a task, we have already developed two main controllers: a visual servoing one in charge of the navigation in the free space, and an obstacle avoidance one able to guarantee non collision. We have added a topological map made of several characteristic landmarks to realize large displacements. To deal with the occlusions, we have designed an algorithm which can compute the necessary visual data when they are temporarily lost. However, this algorithm requires initial conditions not only on the visual features but also on their depth. If the first ones are given by the last image before the occlusion, the second one is not available on our robot. Thus, in this paper we first propose a supervision algorithm able to select the right controller at the right instant and to switch smoothly between the different control laws. Second, we address the problem of the depth reconstruction and we compare two interesting methods from a theoretical and practical point of view. Simulation results in a noisy context and a table summarizing the advantages and drawbacks of both methods are provided.