A terrain segmentation network for navigable areas with global strip reliability evaluation and dynamic fusion

• Published in: Expert systems with applications Vol. 265; p. 125964
• Main Authors: Li, Wei, Liao, Muxin, Zou, Wenbin
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 15.03.2025
• Subjects: Deep neural networks; Dynamic fusion; Global strip reliability evaluation; Navigable areas; Terrain segmentation; Navigable areas; Deep neural networks; Global strip reliability evaluation; Dynamic fusion; Terrain segmentation
• ISSN: 0957-4174
• DOI: 10.1016/j.eswa.2024.125964

Accurate segmentation of safe navigable areas is crucial for scene parsing in autonomous driving systems. However, existing segmentation methods often fail to fully leverage the complementary nature of multiscale features in complex environments, resulting in inadequate information extraction. To overcome this limitation, a Terrain Segmentation Network (TSNet) is proposed for navigable area segmentation, which introduces reliable global semantic information as fusion guide while dynamically exploiting the complementary relationships between multiscale features for progressively guided fusion. Specifically, TSNet is composed of the Global Strip Reliability Evaluation Module (GSREM) and the Dynamic Fusion Module (DFM). The GSREM, which includes the Global Strip Unit (GSU) and the Reliability-Evaluation Unit (REU), is designed to capture the global geometric information of obstacles with long-range contextual characteristics. The information from GSREM serves as prior knowledge to guide the feature fusion. Additionally, we propose the DFM, comprising the Attention Fusion Unit (AFU) and the Contribution Distribution Unit (CDU), to explore the complementary relationships among multiscale feature interactions and to obtain comprehensive scene information. Extensive experiments on diverse wild datasets demonstrate that TSNet outperforms the state-of-the-art methods in accurately identifying navigable areas. The code for TSNet will be available at https://github.com/lv881314/2TSNet. •A terrain segmentation network is presented for navigable areas.•The Global Strip Reliability Evaluation Module serves as prior knowledge to guide feature fusion.•The Dynamic Fusion Module aggregates multiscale features.