摘要
可行驶区域检测是自动驾驶系统的关键功能,对保障车辆安全运行具有重要意义。然而,现有基于视觉与激光雷达融合的检测方法往往未能充分考虑图像质量的动态变化,导致系统在复杂场景下性能不稳定。针对这一问题,文章提出了一种基于图像质量感知的多模态可行驶区域检测方法。首先,采用TinySAM网络结合“三角形”提示点布局策略进行图像语义分割,同时利用基于高程分析的边界提取方法处理激光雷达点云数据。其次,设计道路类型判断和图像质量评估模块,通过多个特征指标实现对道路类型的后验判断,并基于多尺度分析评估图像质量。最后,构建一种考虑历史检测性能、环境状况和道路特征的动态融合策略,实现了视觉与激光雷达检测结果的自适应融合。为了充分评估算法性能,对KITTI Road数据集进行了数据增强。实验结果表明,该方法在F1分数、准确率和召回率等指标上均优于现有方法,且在不同图像质量条件下表现出良好的鲁棒性。
Drivable area detection is a crucial function in autonomous driving systems,playing a vital role in ensuring vehicle operational safety.However,existing methods that fuse vision and LiDAR da-ta often suffer from performance instability in complex scenarios due to their failure to account for dy-namic variations in image quality.To address this limitation,this paper presents a multi-modal driva-ble area detection framework with an integrated image quality awareness mechanism.The framework processes visual and LiDAR data in parallel.For the visual stream,the TinySAM network is em-ployed with a"triangle"prompt point layout for semantic segmentation.Concurrently,LiDAR point clouds are processed using an elevation-based boundary extraction method.Furthermore,the frame-work incorporates specialized modules for posterior road type judgment based on multiple feature indi-cators and for multi-scale image quality assessment.The core of this approach is a dynamic fusion strategy that adaptively weights the results from each sensor by considering historical detection per-formance,environmental conditions,and road features.To validate the proposed method,the KITTI Road dataset was augmented for comprehensive evaluation.Experimental results on this enhanced dataset demonstrate that the proposed approach outperforms existing methods in terms of F1 score,precision,and recall,while maintaining strong robustness under varying image quality conditions.
作者
赵有金
姜武华
ZHAO Youjin;JIANG Wuhua(Anhui Automotive Vocational and Technical College,Hefei 230601,China;Hefei University of Technology,Hefei 230009,China)
出处
《安徽水利水电职业技术学院学报》
2025年第4期39-45,共7页
Journal of Anhui Technical College of Water Resources and Hydroelectric Power
基金
安徽省高校自然科学研究重点项目(2023AH053343)。
