Open Access

ARTICLE

A Lane Coordinate Generation Model Utilizing Spatial Axis Attention and Multi-Scale Convolution

Duo Cui^*, Qiusheng Wang

School of Automation Science and Electrical Engineering, Beihang University, Beijing, 100191, China

* Corresponding Author: Duo Cui. Email:

Computers, Materials & Continua 2025, 84(1), 417-431. https://doi.org/10.32604/cmc.2025.063507

Received 16 January 2025; Accepted 31 March 2025; Issue published 09 June 2025

Abstract

In the field of autonomous driving, the task of reliably and accurately detecting lane markings poses a significant and complex challenge. This study presents a lane recognition model that employs an encoder-decoder architecture. In the encoder section, we develop a feature extraction framework that operates concurrently with attention mechanisms and convolutional layers. We propose a spatial axis attention framework that integrates spatial information transfer regulated by gating units. This architecture places a strong emphasis on long-range dependencies and the spatial distribution of images. Furthermore, we incorporate multi-scale convolutional layers to extract intricate features from the images. The two sets of feature maps are concatenated and subsequently transformed into an input sequence for the decoder, with the lane marking coordinates considered as a target sequence for coordinate generation. This decoder can directly segment multiple lane markings, eliminating the need for additional post-processing algorithms, thereby significantly streamlining the lane recognition process. The proposed method demonstrates a high degree of accuracy in recognizing lane markings and exhibits robust capabilities in differentiating between occlusions and objects resembling lanes. It shows exceptional performance on the TuSimple and CULane datasets.

---

Keywords

---