In this paper, we present a novel sequence generation-based framework for lane detection, called Lane2Seq. It unifies various lane detection formats by casting lane detection as a sequence generation task. This is different from previous lane detection methods, which depend on well-designed task-specific head networks and corresponding loss functions. Lane2Seq only adopts a plain transformer-based encoder-decoder architecture with a simple cross-entropy loss. Additionally, we propose a new multi-format model tuning based on reinforcement learning to incorpo-rate the task-specific knowledge into Lane2Seq. Experimen-tal results demonstrate that such a simple sequence generation paradigm not only unifies lane detection but also achieves competitive performance on benchmarks. For example, Lane2Seq gets 97.95% and 97.42% F1 score on Tusimple and LLAMAS datasets, establishing a new state-of-the-art result for two benchmarks.

Lane detection, one of the crucial foundations of the autonomous driving of Rubber-Tired Gantries (RTGs), plays a vital role in automating manual container terminals. Deep-learning-based lane detection methods have robust and generalized global feature extraction capabilities to deal with complex scenarios well. However, the high preparation cost of large-scale labeled data has limited their application in RTG lane detection. Therefore, this paper presents a cost-effective, scalable incremental learning-based detection method. Specifically, some lane images are collected online, with reliable segmentation labels generated by an image-processing-based lane detection method. Next, a semi-supervised clustering approach is employed to construct a dynamically expanding sample pool, ensuring that samples are representative and diverse. Finally, a lane detection network model is self-trained by using all labeled and unlabeled samples. Extensive experimental results show that our proposed method outperforms existing methods and can achieve a lane detection accuracy of 94.87% and a detection success rate of 99.06%, with the potential for further performance improvement as data size increases..