**7. Conclusions**

This research presents different approaches found in the literature of modular systems in autonomous vehicles. We focused on modular systems since it is an easier form to solve the problem of autonomous driving. One of the main advantages is redundancy. This type of system needs to be redundant and reliable since it can be dangerous consequences in case of error like human fatalities. Alternatively,

end-to-end systems have become more studied in the latest years. In the future it is expected to have more proposals using this approach, however, at this moment there is more limitation with this type of system like the lack of hardcoded safety measures.

Our approach is based on the terrain perception stage of modular systems. We select an existing model that obtains similar results to the one found in the literature. Our model is lightweight to be run on mobile devices but still robust enough. Two checkpoints were compared, obtaining 83.88% of weighted IoU for the best result.

We expect to improve the semantic segmentation results by augmenting the dataset for future experiments so the network has more data to learn. Further research could explore different parameters and hyperparameters of the model and their influence on the results. Since the architecture selected is oriented to run in mobile devices, we look for implementing and testing the video segmentation in smartphones.
