【作者】 许华；

【导师】 康崇； 唐茂峰；

【作者基本信息】 哈尔滨工程大学 ， 计算机技术（专业学位）， 2024， 硕士

【摘要】 车道线检测是感知智能技术领域内一项核心且颇具难度的研究课题。在复杂场景中,如建筑物行人遮蔽、雨雪天气等恶劣气候、路面磨损褪色以及车道线自身不清晰等问题都会影响车道线的检测,同时,实际需求中的计算成本问题也亟待解决。为了能够应对复杂场景下的挑战,同时满足实时性要求。本文对复杂道路条件下的车道线检测展开分析,从车道线先验角度进行研究。为了充分利用车道线的位置先验信息,设计了一种基于混合锚分类的车道线检测算法。缓解行锚下的定位误差问题,将车道定位视为有序分类。针对车道线在不同的距离、角度和光照条件下呈现出不同的尺度特性,采用Res2Net架构适应车道线的各种形态变化,捕获不同层次和尺度下的特征信息;对于边界模糊、遮挡或路面条件复杂的场景,设计了行列空间联合注意力机制,同时利用到行、列以及空间维特征之间关系,关注与车道线相关的位置特征,同时利用到长距离依赖性关系。通过在常用数据集Tusimple以及自制数据集IRLane上进行算法的验证工作,证实了利用车道线位置先验信息的优势以及各个模块的合理性。为了充分利用车道线的形状先验信息,设计了一种基于跨层细化先验的车道线检测算法。通过将权重分配机制融合到特征金字塔中,提出了一种用于车道线检测的权重特征金字塔网络,允许模型根据实际输入图像数据分布的变化自适应地调整各层特征的贡献程度,尤其是在复杂场景下,能够更好地捕捉到车道线的关键细节,并减少无关背景噪声的影响。同时,在权重特征金字塔网络中引入了跨层细化块,跨层细化块以特征图和预设先验为输入,从高级特征图到低级特征图逐渐细化先验,完成先验的更新。在业界公认的Tusimple、CULane数据集以及自制数据集IRLane上进行实验,实验表明,所提出的方法能够使用高级特征来粗略预测车道,而利用局部细节特征来提高定位精度,同时满足实时性要求。从车道线的位置先验信息到形状先验信息,提出的车道线检测算法,在多个数据集上进行验证工作,证实了算法的有效性,完成了复杂场景下车道线检测算法的提升。更多还原

【Abstract】 Lane detection is a core and challenging research topic in the field of perceptual intelligence.In complex environments,such as those obscured by pedestrians or buildings,compounded by adverse weather conditions like rain and snow,as well as issues arising from worn-out or faded road surfaces and indistinct lane markings,the accuracy of lane detection is significantly compromised.Furthermore,the practical demand for minimizing computational costs poses an additional challenge that urgently needs addressing.To address the challenges in complex scenarios and meet real-time requirements,this thesis analyzes lane detection under complex road conditions,focusing on lane priors.To fully leverage the prior information of lane positions,a lane detection algorithm based on hybrid anchor classification is proposed.To mitigate the problem of localization errors under anchor alignment,lane positioning is treated as an ordered classification.Addressing the scale variations of lane under different distances,angles,and lighting conditions,the Res2 Net architecture is adopted to adapt to various morphological changes of lane,capturing feature information at different levels and scales.For scenes with blurred boundaries,occlusions,or complex road Environments,a row column spatial joint attention mechanism is designed.This mechanism simultaneously exploits relationships between row,column,and spatial dimension features,focusing on position features relevant to lane while utilizing long-range dependencies Validation work on commonly used datasets such as Tusimple and a custom dataset called IRLane confirms the advantages of leveraging prior information on lane positions and the rationality of each module.To fully utilize the prior shape information of lane,a lane detection algorithm based on cross layer refined priors is designed.By integrating a weight allocation mechanism into the feature pyramid,a weighted feature pyramid network for lane detection is proposed.This network allows the model to adaptively adjust the contribution of each layer’s features based on the distribution of actual input image data,particularly in complex scenarios,enabling better capture of key lane details and reducing the impact of irrelevant background noise.Additionally,cross layer refinement blocks are introduced in the weighted feature pyramid network.These blocks take feature maps and predefined priors as input,gradually refining the priors from higher-level feature maps to lower-level feature maps to update the priors.Experiments conducted on well-recognized datasets such as Tusimple,CULane,and a custom dataset named IRLane demonstrate that the proposed method can use high-level features to roughly predict lanes while utilizing local detailed features to improve localization accuracy,meeting real-time requirements.The lane detection algorithm proposed in this thesis,which transitions from position prior information to shape prior information,is validated on multiple datasets.The effectiveness of the algorithm is confirmed,achieving advancements in lane detection in complex environments.更多还原