摘要
为了改善现有斑马线检测精度低、检测不全面及实时性差等多重问题,提出了一种改进的SegNet语义分割模型进行斑马线检测,使其更适应于辅助驾驶系统中的预警场景,为无人驾驶车辆提供帮助。以语义分割中的原始SegNet模型为基础,首先将SegNet模型的特征提取网络部分进行改进,融入金字塔池化模块,进行多方面特征提取,获取上下文全局信息,达到减少斑马线细节信息丢失的目的;然后将模型对称的结构改为编码不变、解码减少的不对称结构,减少网络参数,增设细节处理,精准检测斑马线像素点位置,有效地提高了检测准确率和实时性。对获取到的斑马线图像进行亮度增强、添加噪声、图像翻转等随机转换方式来扩充采集的斑马线图像样本,使斑马线图像样本满足各种语义分割模型检测需求。在扩充后的斑马线数据集上进行试验,以精确率,召回率,F1值为检测精度评价标准,以运行时间为实时性评价标准,将改进SegNet模型分别与原始SegNet模型,U-Net模型,PSPNet模型进行对比。改进SegNet模型检测斑马线精度达到了97.6%,与其他模型相比检测精度得到有效提高,且运行速度加快,满足当前实时性检测的需求。因此改进SegNet模型检测斑马线更全面,可应用于辅助驾驶系统中。
In order to improve the multiple problems of existing zebra crossing detection,such as low accuracy,incomplete detection and poor real-time performance,an improved SegNet semantic segmentation model is proposed for zebra crossing detection,which makes it more suitable for the early warning scene in the auxiliary driving system and provides help for driverless vehicles.Based on the original SegNet model in semantic segmentation,first,the feature extraction network part of SegNet model is improved and integrated into the pyramid pooling module to extract various features and obtain the global context information,so as to reduce the loss of zebra crossing detail information.Then,the symmetrical structure of the model is changed to the asymmetric structure with unchanged coding and reduced decoding to reduce network parameters,add detail processing and accurately detect the pixel position of zebra crossing,which effectively improves the detection accuracy and real-time performance.The acquired zebra crossing images are randomly transformed by brightness enhancement,noise addition,image reversal and other methods to expand the collected zebra crossing image samples,so that the zebra crossing image samples can meet the detection requirements of various semantic segmentation models.The experiment is carried out on the expanded zebra crossing data set.Taking the accuracy rate,recall rate and F1 value as the detection accuracy evaluation criterion and the running time as the real-time evaluation criterion,the improved SegNet model is compared with the original SegNet model,U-Net model and PSPNet model respectively.The result shows that the detection accuracy of the improved SegNet model reaches 97.6%,the detection accuracy is effectively improved,and the running speed is accelerated to meet the current needs of real-time detection compared with other models.Therefore,the improved SegNet model is more comprehensive in detecting zebra crossings and can be applied to assisted driving system.
作者
付阳阳
陶建军
王夏黎
李妮妮
袁绍欣
FU Yang-yang;TAO Jian-jun;WANG Xia-li;LI Ni-ni;YUAN Shao-xin(School Information Engineering,Chang’an University,Xi’an Shanxi 710064,China;Shaoxing Transportation Construction Co.,Ltd.,Shaoxing Zhejiang 312000,China)
出处
《公路交通科技》
CAS
CSCD
北大核心
2022年第4期117-122,149,共7页
Journal of Highway and Transportation Research and Development
基金
国家重点研发计划项目(2020YFB1600400)
浙江省交通运输厅科技计划项目(2020026)。
