• Convolutional Neural Network
• Ultra Fast Deep Lane Detection
• rigid transformation
• ALKS
• monocular camera
• automated lane keeping system
• Advanced Driver Assistance Systems
• CARLA simulator
• autonomous driving
• ADAS
• Inverse Perspective Mapping
• AI safety
• dependable engineering
• pinhole camera model
• ResNet
• Kinematic Bicycle Model
• Symmetrized Segment-Path Distance
• Pure Pursuit Controller
• IPM
• PID controller

The risk of driving accidents is a present issue due to the high intensity of wheeled transportation.
The National Highway Traffic Safety Administration (NHTSA) states that more than 94% of road accidents are due to human error, and considering the European Union, in the six months January-June 2022, there is an increase in the number of road accidents with injuries (+24.7%), wounded (+25.7%), and fatalities within the 30th day (+15.3%) compared to the same period in 2021.
Safety of citizens, in this sense, also goes through the development of robust Advanced Driver Assistance Systems (ADAS) which, even in the current semi-autonomous stage, can help reduce the risk of accidents due to distractions as well as facilitate the driving experience.
The thesis deals with the development of an automated lane keeping system for highway scenarios, exploiting a redundant design for improved safety and robustness of the AI components.
To this end, the system integrates two state-of-the-art neural networks for lane detection, one performing better in terms of speed and the second one characterized by higher latencies but improved dependability. The two are then combined into a single system, where the former network is used in the control loop while the second works behind the scenes and serves to identify failures of the primary system and handle the fallback procedure. If the fallback procedure is activated, the car urges the driver to quickly take control of the vehicle as it begins to decelerate, following the trajectory generated by the most dependable network.