摘要
由于传统车道线检测方法存在计算量大、无视觉线索和车道线遮挡等主要问题,制约着车道线检测的发展。目前,UFSA(UltraFast Structure-aware)算法的提出可以有效解决上述问题,并在车道线检测领域广泛的应用。而UFSA算法存在网络卷积和池化提取特征会丢失重要信息、边界信息不够敏感等问题,故加入ASPP(Atrous Spatial Pyramid Pooling)空洞金字塔池化与FCANet(Frequency Channel Attention)频率域通道注意力的融合机制定义为FCASPP(Frequency Channel Attention Spatial Pyramid Pooling),上述机制能够有效地在大感受野时,获取更丰富上下文信息并提取更有用和紧致的特征而抑制噪声信息,L-Dice(Lane Dice Loss)函数比Softmax函数更加关注车道边界的信息。通过消融实验验证了上述改进的有效性,且无需添加任何计算量。在TuSimple和CULane两个基准数据集中,检测精度与原文相比,分别提高了0.21个百分点和1.7个百分点,速度与原文相当,所提算法较具竞争力。
Due to the main problems of traditional lane line detection methods,such as large amount of calcula-tion,lack of visual clues and lane line occlusion,the development of lane line detection is restricted.At present,Ul-tra Fast Structure Aware(UFSA)algorithm can effectively solve the above problems,and has been widely used in the field of lane line detection.However,the UFSA algorithm has some problems such as network convolution and feature extraction though pooling can lose important information and boundary information is not sensitive enough.Therefore,the fusion mechanism of Atrous Spatial Pyramid pooling(ASPP)and Frequency Channel Attention(FCANet)is defined as Frequency Channel Attention Spatial Pyramid Pooling(FCASPP),which can effectively obtain richer context information and extract more useful and compact features to suppress noise information in large receptive fields,so that the Lane Dice Loss(L-Dice)function pays more attention to lane boundary information than the softmax function.The effectiveness of the above impronement is verified by ablation experiments without adding any a-mount of caculation.In the two benchmark datasets of TuSimple and CULane,the detection accuracy is improved by 0.21 percentage points and 1.7 percentage points respectively compared with the original text,the speed is equivalent to the original,and the algorithm in this paper is more competitive.
作者
王祥
柯福阳
朱节中
夏德铸
WANG Xiang;KE Fu-yang;ZHU Jie-zhong;XIA De-zhu(School of Automation,Nanjing University of Information Engineering,Nanjing Jiangsu 210000,China;School of Remote Sensing and Surveying and Mapping,Nanjing University of Information Engineering,Nanjing Jiangsu 210000,China;Nanjing University of Information Engineering,Wuxi Research Institute,Wuxi Jiangsu 214000,China;Wuxi University,Wuxi Jiangsu 214000,China)
出处
《计算机仿真》
北大核心
2023年第5期213-219,共7页
Computer Simulation
基金
第十六批次江苏省“六大人才商峰”高层次人才项目(XYDXX-045)
2020年无锡市科技发展资金(N20201011)
西宁市科技计划(2019-Y-12)。
