Lane Detection with Versatile AtrousFormer and Local Semantic Guidance. (January 2023)

• Record Type: Journal Article
• Authors: Yang, Jiaxing; Zhang, Lihe; Lu, Huchuan
• Is Part Of: Pattern recognition. Volume 133(2023)
• Publication Date: 2023-01
• Subjects: Lane detection -- Global AtrousFormer -- Local AtrousFormer -- Enhanced feature extractor -- Local semantic guided decoder; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.109053
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose global AtrousFormer to collect information in an atrous yet global way, instilling the essence of the ASPP to the transformer structure and finally enhancing the ability to infer lane distribution. Global AtrousFormer is evolved into a more compact version, local AtrousFormer, by incorporating slice mechanism. Then it is early embedded into feature extractor like ResNet-18 to enhance feature extraction. We propose LSGD to obtain more representative feature vector, which then is used to better the lane existence prediction. Our network achieves favorable results against other state of the arts on the recent challenging datasets, achieving 78.08 F1 score on CULane, 96.71 of accuracy score on TuSimple, and 59.20 of accuracy score on BDD100K Abstract: Lane detection is one of the core functions in autonomous driving and has aroused widespread attention recently. The networks to segment lane instances, especially with bad appearance, must be able to explore lane distribution properties. Most existing methods tend to resort to CNN-based techniques. A few have a try on incorporating the recent adorable, the seq2seq Transformer [1]. However, their innate drawbacks of weak global information collection ability and exorbitant computation overhead prohibit a wide range of the further applications. In this work, we propose Global Atrous Transformer (AtrousFormer) to solve the problem. Its variant local AtrousFormer is interleaved into feature extractor to enhance extraction. Their collecting information first by rows and then by columns in a dedicated manner finally equips our network with stronger information gleaning ability and better computation efficiency. To further improve the performance, we also propose a local semantic guided decoder to delineate the identities and shapes of lanes more accurately. Extensive results on three challenging benchmarks (CULane, TuSimple, and BDD100K) show that our network performs favorably against the state of the arts.