Query-Based Lane Detection with Mixed Supervision

Lane detection is essential for autonomous driving, ensuring accurate lane positioning, collision avoidance, and traffic rule adherence. State-of-the-art methods often rely on handcrafted anchors or discrete representations, limiting their applicability. While LSTR introduces a polynomial lane shape model and query-based architecture, it suffers from slow convergence due to unstable Hungarian matching during training. This thesis presents a mixed supervision strategy that combines one-to-one and one-to-many matching losses to stabilize training and accelerate convergence. On the CULane dataset, our method improves overall accuracy by 0.45% over the predecessor method and reduces validation loss convergence time by half. Competitive results are also achieved on the TuSimple dataset without hyperparameter fine-tuning. These findings demonstrate the proposed method's effectiveness in enhancing both accuracy and efficiency, offering a promising advancement in end-to-end lane detection for autonomous driving.