Closing the Loop in Automated Driving: Dataset Design, Scenario Synthesis, and Deployed Perception

This thesis investigates data-centric methodologies for automated driving and quantifies their impact from simulation to embedded deployment. First, we study decision transformer in offline reinforcement learning on a lightweight 2D platform, highlighting the limitations of high-level driving simulators. We hence transition to a more complex, reliable 3D pipeline, introducing a custom 60 km highway map engineered for long, reproducible runs, designed to overcome the limitations of the predefined driving environment within CARLA simulator. On this infrastructure, two complementary problems are addressed under matched conditions. For lane change intention, human-in-the-loop sessions with 50 drivers yield >3,400 annotated maneuvers as deep-learning-ready time series. Formulating intention as time-to-lane-change regression enables anticipation up to 4 s; across 1D-CNN, LSTM, GRU, and Transformer baselines the best model attains RMSE = 0.51 s and MAE = 0.30 s, while a compact 1D-CNN offers favorable accuracy–efficiency trade-offs. For scenario understanding, an ASAM OpenSCENARIO 2.0-aligned generator produces labeled MP4 clips spanning nine highway classes; a 3D residual CNN establishes reference performance and highlights systematic confusions among visually similar behaviors. Finally, we address robust traffic light detection on embedded hardware. A modular architecture comprising a frame checker, frame cleaner, takeover Request logic, automated driving revert logic, and a quantized YOLOv11 detector sustains near-nominal accuracy under sensor anomalies. On NVIDIA Jetson Orin Nano, the system achieves 84.6\% mAP50 in clean conditions and 84.4–84.8\% mAP50 under simulated rain and blur, operating in real time. The contributions include: (i) a reproducible CARLA highway environment; (ii) a public lane change framework comprising a dataset and a data generation pipeline; (iii) a configurable scenario generator and benchmark; and (iv) an embedded, anomaly-aware traffic lights detection pipeline with explicit takeover request integration. Results demonstrate that rigorous dataset design and controlled simulation substantially improve evaluation fidelity and support deployable automated driving functions.