Usually, when a robot is being built, people tend to use optic flow to control the behaviour of the robot. It is a powerful way to perceive distances from a sequence of images. It can be computed using different algorithms [23,5,2]. With optical flow it is possible to build robots able to navigate in complex environments. Obstacle avoidance and corridor following can be done using optic flow. For instance, in [7] the game of tag was programmed on two robots using obstacle avoidance based on optical flow.
Optical flow is also used for automatic road following in vision-based intelligent vehicles [3]. However, optical flow needs a lot of computation. So it requires a lot of processing power which is usually not available on mobile agents. That is why this thesis tries to investigate other methods of navigation which do not require a computation of the optic flow. We want neural networks to estimate the optic flow for us.
Learning methods have already been used a lot in order to mimic animal behaviour. Some of them use optic flow as a preprocessing stage [25,26], other use simpler versions of optic flow (one dimension optic flow for instance [32]), and other do not use optic flow at all. We will focus our work on the third category, this being the vision of the flies has been modeled with neural networks without using optic flow [13]. Our goal is to build a corridor following robot that is able to learn its task.
Another study about animals concern the visual path integration. Srinivasan has focused his study specifically on bees [28,29]. This study has lead to construct of a robot that uses of the visual system of a bee. Unfortunately, the implementation uses two cameras [12]. It has been shown that bees can still estimate how far they travelled even if they only use one eye. So our second goal is to estimate the egomotion of a robot by using a learning method and only one camera.
Our aim is to create a new odometry system and use it to validate the existing one. Such validation techniques are widely used in robotics. Sonars are usually used to do this, but any sensors can be used to reduce odometry errors [15,30]. For instance, the article [24] describes a general method to reduce odometry errors using another sensor.
The following sections will briefly describe the experiments done and the results extracted from these experiments.