**Abstract.**

Autonomous driving is a growing research area; however, there are no fully autonomous vehicle (AV) in the world. Existing AVs have different capabilities and can drive by themselves only in specific scenarios with several constraints. This paper discusses several studies from the point of view of a modular system approach. This approach perceives autonomous driving as separate tasks to solve. Studies are classified in object/pedestrian detection, road detection, obstacle avoidance, terrain perception, mapping of the environment, and path planning. Furthermore, various perception sensors are reviewed and compared. Important datasets and metrics found in the literature are presented. Finally, one of our experiments obtained a weighted IoU of 83.88% in the segmentation of five classes. Since this is a work in progress, more research needs to be done, but our proposal shows promising results in terrain perception in off-road environments.

**Keywords:** autonomous driving, terrain perception, semantic segmentation

## **1. Introduction**

Autonomous driving is a growing research area; recently, it has received a lot of attention due to its many advantages. According to a study by Morgan Stanley Research, autonomous vehicles (AVs) can save money from reduced labor costs, improved productivity, lower fuel consumption, and fewer accidents [1]. There are two types of scenes in autonomous driving, on-road and off-road. In the first type, we can find pavement roads, lane markings, defined cues, etc. In the second, there are uneven surfaces, not clear delimiters, vegetation, and different terrains. Several projects have brought significant advances and had a meaningful impact on the state of the art of autonomous driving. The DARPA Challenge held in 2004 was one of the first important competitions initially; it was oriented for military applications, but then the focus change to civilian purposes in urban scenarios. None of the contestants finished; in the second edition, five teams complete the challenge without human intervention.

Autonomous vehicles (AVs) are complex systems. The Society of Automotive Engineers (SAE) [2] defines six autonomy levels in cars, starting from 0 to 5. In level 0, there is no driving automation; the human driver performs all driving tasks. In level 1, some tasks are performed by the car, like adaptive cruise speed control, stability control, and anti-lock braking systems. Partial driving automation is level 2. In this level, there are combined automated functions like acceleration and deceleration in defined situations. Level 3, known as conditional automation, is when the vehicle can control some functions under limited conditions for a certain period. In level 4, the vehicle is capable of fulfilling all driving tasks under certain

conditions. Level 5 is full automation; there is no need for a human driver, the car can drive under all conditions.

There are two main classifications for system architecture in AVs, based on their connectivity and their algorithmic design [3]. In the first, we find ego-only systems and connected systems; for the second, there are modular and end-to-end systems. The majority of the research focuses on modular systems since it is an easier form to implement an AV. In this work, we present several proposals found in the literature from the point of view of modular systems. In addition, there is a short review of the most common sensors used to perceive on AVs. Finally, we implement an existing model for segmenting different terrain types using a lightweight and fast network that can be used in mobile devices.
