**6. Results**

In this work, we study several perspectives on how to attack the problem of autonomous driving. As mentioned before, the two main approaches are modular and end-to-end. Our approach is based on the task of terrain perception. Our approach was to apply transfer learning to retrain an existing model to segment different terrains. Two checkpoints were selected, and five classes were segmented: (0) Object, (1) Vegetation, (2) Sky, (3) Soil, and (4) Grass.

We use Intersection over Union as a metric to evaluate the performance of our approach. Also, the mean IoU (mIoU) is presented, but it is not a good form to evaluate since it does not consider the number of times that a class is presented in the data. mIoU can be skewed by imbalanced datasets giving more importance to classes with more presence. In order to present a more accurate metric, we obtained weighted IoU that gives us an average IoU of each class, weighted by the number of pixels in that class.

**Table 1** presents results for each class, as it can be seen the best results were obtained by the checkpoint pre-trained in MS-COCO with 83.88% of IoU. We believe the reason is the bigger number of images in this set of databases that contained off-road scenes. Nevertheless, class 0, which is objects, was not detected at all in FFD\_MS-COCO. For that specific class, the best results were obtained with FFD\_ADE20K with an IoU of 20.45%. It is important to mention that this class is not found in every image, so the network does not have enough examples to learn and give better predictions.

*Automatic Terrain Perception in Off-Road Environments DOI: http://dx.doi.org/10.5772/intechopen.99973*


### **Table 1.**

*IoU comparison results.*


**Figure 2** shows some qualitative results; the first column shows the original images. The ground truth mask is presented in the second column, and the rest of the columns are the results of the two different experiments. As shown in some of the results with FFD\_ADE20K, the bottom part of the image was detected as the sky and in some results as soil. This problem was persistent in most of the images, in the same way for FFD\_MS-COCO; it was existent but only in a few images.
