Uploaded June 2025 | Updated September 2026, 2 hours ago
This is a video about Ammar Hammad's bachelor's thesis at the Bern University of Applied Sciences in Switzerland. It explores and demonstrates the usability of a neural network for traffic lane detection using color camera input in autonomous vehicles within a simulated environment created in the Unity game engine. Several deep learning-based approaches for lane detection are examined and evaluated for their applicability in Unity. To improve prediction performance and ensure full Unity compatibility, a neural network was implemented from scratch, using the architecture of BezierLaneNet (Z. Feng et al., 2022) as a reference. A Bézier-based road generation system was developed to facilitate both the generation of training data and the deployment of simulations. After iterative testing and adjustments, the custom network achieved performance comparable to BezierLaneNet, as demonstrated by the conducted evaluations. To enable downstream use of the network output, additional processing steps were developed, including the projection of predicted Bézier control points from image to world coordinates and collision detection based on tracked lane geometry. Finally, to showcase and assess the practical applicability of the network, a reinforcement learning environment was built using the Bézier-based road system, in which the trained network is deployed to provide perceptual input to an autonomous driving agent. The resulting system demonstrates an end-to-end pipeline comprising the Bézier-based road generation system, a customized BezierLaneNet, post-processing modules, and a reinforcement learning component for control.
The underlying music was also composed by Ammar Hammad. See also
soundcloud.com/user-365397860/spring-outside
This is a video about Ammar Hammad's bachelor's thesis at the Bern University of Applied Sciences in Switzerland. It explores and demonstrates the usability of a neural network for traffic lane detection using color camera input in autonomous vehicles within a simulated environment created in the Unity game engine. Several deep learning-based approaches for lane detection are examined and evaluated for their applicability in Unity. To improve prediction performance and ensure full Unity compatibility, a neural network was implemented from scratch, using the architecture of BezierLaneNet (Z. Feng et al., 2022) as a reference. A Bézier-based road generation system was developed to facilitate both the generation of training data and the deployment of simulations. After iterative testing and adjustments, the custom network achieved performance comparable to BezierLaneNet, as demonstrated by the conducted evaluations. To enable downstream use of the network output, additional processing steps were developed, including the projection of predicted Bézier control points from image to world coordinates and collision detection based on tracked lane geometry. Finally, to showcase and assess the practical applicability of the network, a reinforcement learning environment was built using the Bézier-based road system, in which the trained network is deployed to provide perceptual input to an autonomous driving agent. The resulting system demonstrates an end-to-end pipeline comprising the Bézier-based road generation system, a customized BezierLaneNet, post-processing modules, and a reinforcement learning component for control.
The underlying music was also composed by Ammar Hammad. See also
soundcloud.com/user-365397860/spring-outside
