一种基于图卷积的车载视频对象语义分割方法  

A semantic segmentation method for vehicle video objects based on graph convolution

作　　者：赵江洪 尹利莎[2,4] 陈先昊 杨甲 郭明 ZHAO Jianghong;YIN Lisha;CHEN Xianhao;YANG Jia;GUO Ming(State Key Laboratory of Geo-information Engineering,Xi’an 710054,China;School of Geomatics and Urban Spatial Informatics,Beijing University of Civil Engineering and Architecture,Beijing 102616,China;Key Laboratory for Urban Spatial Information of the Ministry of Natural Resources,Beijing 102616,China;Beijing Key Laboratory for Architectural Heritage Fine Reconstruction&Health Monitoring,Beijing 102616,China;State Key Laboratory of Information Engineering in Surveying,Mapping and Remote Sensing,Wuhan University,Wuhan 430072,China;Beijing Haoyu World Surveying and Mapping Development Co.,Ltd.,Beijing 100039,China)

机构地区：[1]地理信息工程国家重点实验室,西安710054 [2]北京建筑大学测绘与城市空间信息学院,北京102616 [3]自然资源部城市空间信息重点实验室,北京102616 [4]建筑遗产精细重构与健康监测北京市重点实验室,北京102616 [5]武汉大学测绘遥感信息工程国家重点实验室,武汉430072 [6]北京浩宇天地测绘科技发展有限公司,北京100039

出　　处：《测绘科学》2023年第2期157-167,共11页Science of Surveying and Mapping

基　　金：国家自然科学基金项目(41601409,41971350);国家重点研发计划项目(2018YFC0807806);地理信息工程国家重点实验室开放基金课题项目(SKLGIE2019-Z-3-1);北京建筑大学市属高校基本科研业务费专项资金项目(X18063);北京市自然科学基金项目(8172016);武汉大学测绘遥感信息工程国家重点实验室开放基金资助项目(19E01);自然资源部数字制图与国土信息应用重点实验室开放研究基金项目(ZRZYBWD202102);住房和城乡建设部软件学研究项目(R20200287);北京市社会科学基金决策咨询重大项目(21JCA004)。

摘　　要：针对图像语义分割网络(SegNet)在对车载视频分割过程中,因局部特征的丢失造成语义分割精度不高的问题,该文提出一种具有权重系数和图卷积网络的视频分割深度卷积网络(WG-ViSeg)。该网络对SegNet进行改进,在高级特征提取过程中加入图卷积结构,通过扩大节点的感受野减少局部特征的丢失。该网络又利用SE注意力机制改变特征图谱的权重系数进一步提高网络编码能力。对Camvid数据增强验证后结果表明,在满足车载视频对象的快速响应范围内,WG-ViSeg能够很地改善分割过程中出现的碎片化状况,较好地分割出相邻目标对象,对车载视频的整体分割精度达到89.7%,较现有的最优网络提升了5%,尤其对自动驾驶较为重要的车辆、行人等类别的语义分割精度提升了17%。In order to solve the problem that the semantic segmentation accuracy of SegNet was not high due to the loss of local features in the process of vehicle video segmentation,a deep convolutional network with weight coefficient and graph convolutional network for video segmentation(WG-ViSeg)was proposed.That was improved SegNet by adding Graph Convolutional Networks in the process of advanced feature extraction to reduce the loss of local features by expanding the receptive field of nodes.Furthermore,the WG-ViSeg network used Squeeze-and-Excitation to change the weight coefficient of feature map to further improve the network coding ability.After validated by enhanced Camvid data,the results showed that within the fast response range of vehicle video application objects,WG-ViSeg could effectively solve the fragmentation problem in the segmentation process and segment the adjacent objects better.The overall segmentation accuracy of WG-ViSeg network for vehicle video reached 89.7%,and this was a 5%improvement over the existing optimal network.In particular,the semantic segmentation accuracy of vehicle,pedestrian and other categories which were important for autonomous driving,was improved by 17%.

关 键 词：车载视频 视频语义分割 SegNet 图卷积 

分 类 号：P237[天文地球—摄影测量与遥感]