Future work

In practice, a lot of parameters can reduce the accuracy of the results such as: the numbers of pixels in the images, the size of the stripes used for the histogram, the architecture and the size of the neural network, the orientation of the camera and also the accuracy of the odometry. The study of the exact effect of all those parameters can be an useful future work. Especially the change in the environment could be useful to study. These studies could lead to a more general distance estimator that could perform its task everywhere.

The use of a more accurate measure of the distance during the learning stage could also be investigate. We can use other devices or simply use techniques such as the one described in [9] or such as the one described in [6] designed to improve the accuracy of the internal odometry.

In order to build a robot that can follow corridors using a map provided by the user, the robot must not only follow corridors and estimate distances, but it must also be able to estimate its steering angle. A possible future work is to estimate angles when the robot turns. A first step could be to estimate angles when the robot turns without moving. The further step needed to achieve a robot that is able to navigate is to combine the angle estimation and the distance estimation. This will allow a user to build paths for the robot such as: move $10\;m$ forward, then turn $90^\circ$ on the left and move $2\;m$ forward.

As we can see, visual odometry is an open field and there are a lot of work to do to improve the existing results and to provide an usable system.