一种针对路口监控图像的区域分割方法  

A REGION SEGMENTATION METHOD FOR INTERSECTION MONITORING IMAGES

作　　者：李昕蔚 丁正彦 尚岩峰[4] 祝永新 汪辉[1] 钟雪霞 田犁[1] 黄尊恺 封松林[1] Li Xinwei;Ding Zhengyan;Shang Yanfeng;Zhu Yongxin;Wang Hui;Zhong Xuexia;Tian Li;Huang Zunkai;Feng Songlin(Shanghai Advanced Research Institute,Chinese Academy of Science,Shanghai 201210,China;School of Information Science and Technology,ShanghaiTech University,Shanghai 201210,China;University of Chinese Academy of Sciences,Beijing 101408,China;The Third Research Institute of Ministry of Public Security,Shanghai 201204,China)

机构地区：[1]中国科学院上海高等研究院,上海201210 [2]上海科技大学信息科学与技术学院,上海201210 [3]中国科学院大学,北京101408 [4]公安部第三研究所,上海201204

出　　处：《计算机应用与软件》2020年第3期236-243,共8页Computer Applications and Software

基　　金：国家重点研发计划项目(2016YFC0801304);上海市科学技术委员会科研计划项目(18511111302)。

摘　　要：针对智能交通系统中对交通路口场景理解的需求,提出一种基于线特征先验和凸包损失函数的空间分割网络,目标是对斑马线以及斑马线所围路口区域进行精确检测和分割。利用公安交通管理系统平台采集并标注路口数据集;引入线特征先验,将RGBL图像作为网络输入,为深度学习实例分割提供显著的物体边缘特征以加强深度网络对图像特征学习的针对性;在分割网络中引入SCNN网络结构,构成空间分割网络以增强网络对空间结构的学习;引入凸包二值交叉熵动态损失函数来优化网络的输出精度。实验结果表明,该空间分割网络对斑马线及路口区域的检测正确率和分割完整度和精确度都有了显著的提升。Aiming at the need of understanding the traffic intersection scene in intelligent transportation system,this paper proposes a spatial segmentation network based on line feature prior and convex hull loss function.The task is to accurately detect and segment the zebra crossing and the intersection area surrounded by the zebra crossing.We used the public security traffic management system platform to collect and mark the intersection data set.Then,in order to provide significant object edge features for deep learning instance segmentation,we introduced a line feature prior and used RGBL image as network input,so as to enhance the pertinence of deep network on image features learning.In addition,SCNN network structure was introduced into the segmentation network to form a spatial segmentation network to enhance the learning of the network on the spatial structure.In the end,a convex hull binary cross-entropy dynamic loss function was introduced to optimize the output accuracy of the network.The experimental results show that the detection accuracy,segmentation integrity and accuracy of zebra crossing and intersection area are greatly improved by the spatial segmentation network.

关 键 词：智能交通系统 斑马线检测 实例分割 MASK R-CNN 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]