Perception algorithms analysis for autonomous drive application in simulated environment

Over the last period, the advancement of autonomous driving technology has made huge steps forward with regard to perception algorithms, i.e. the tools capable of translating sensory data into understandable data. This thesis project aims to delve into the different perception techniques present on the autonomous driving market. In particular, to achieve this goal, CARLA, a cutting-edge open source system for this topic, was used: the main focus is therefore to develop a perception algorithm, obtaining the best possible results in terms of performance within a simulated environment as similar as possible to a complex real-life condition. Understanding and being able to interpret the surrounding environment in real time is one of the most important aspects of autonomous driving, because all the subsequent choices made by the machine itself and consequently the success of the algorithm depend on it. With regard to this in particular, the CARLA platform was chosen because it is able to provide a world within which to make your own simulations that is totally customizable, and therefore can be made as realistic as possible. The fundamental library of this project is OpenCV, because, through various image manipulation tools, it allows you to process with great accuracy and precision the different frames captured during the experiments, and consequently allow a detection of obstacles or surrounding objects. The project was divided into different parts, each functional to the next and in general to the success of the whole, starting from the study of the interface features through which it was possible to simulate everything, i.e. the CARLA application programming interface (API), then moving on to the feasibility and in what terms of the OpenCV library, analyzing the different possibilities such as the possibility of using pre-existing techniques such as object edge detection or visual understanding functions, all tools of fundamental importance in the selected field. In addition to the pure development of a perception algorithm aimed at recognizing road lanes, one of the aspects that have been carried out in parallel is the desire to learn aspects that seem peripheral in software development, but that in reality play a fundamental role in the success of the project and above all in its understanding in the eyes of an outsider: in particular, the use of a correct version control system of the program via Git has been examined, as well as the search for maintaining excellent documentation and therefore in a broader sense the ability to apply the right methodologies for managing a project. The modularity and maintainability of the code were the prerogatives of the entire project, in order to make it reusable for future projects. In this sense, automatic testing methods were used to make everything even more professional, thus integrating a code that is both functional and functional, but also scalable for other projects or situations. In conclusion, the results of the thesis give the possibility to explore new scenarios within the implementation of perception algorithms for Lane Detection, thus promoting the development of different technologies compatible with this. Through the use of simulation tools and modern and advanced libraries, this project aims to set the limit beyond the current "normality", exploring new concepts regarding autonomous driving, thus promoting new improvements in this field.