用于实例分割的数据集自动生成方法  

Automatic data set generation method for instance segmentation

作　　者：王晓年[1,2] 宋梦譞 WANG Xiaonian;SONG Mengxuan(School of Electronics and Information Engineering,Tongji University,Shanghai 201804,China;National Experimental Teaching Demonstration Center of Computer and Information Technology,Shanghai 201804,China)

机构地区：[1]同济大学电子与信息工程学院,上海201804 [2]计算机与信息技术国家级实验教学示范中心,上海201804

出　　处：《实验技术与管理》2023年第7期28-32,40,共6页Experimental Technology and Management

基　　金：融合数据驱动的无人车拟人驾驶机理化方法研究建设项目(61973239);教育部产学合作协同育人项目上海海思课程开发(0800166035)。

摘　　要：近年来自动驾驶技术迅猛发展,无人车产业已逐渐成为世界关注的一个焦点。随着越来越多的人工智能算法用以求解无人驾驶车辆的环境感知、决策和规划中的问题,对标注数据的需求也日益剧增。为了减少数据标注的时间和人工成本,该文提出了一种基于Unity3D的图像和激光数据标注的自动生成方法。首先基于二维地图信息构建三维场景,再利用虚拟相机实现可见光图像和语义分割数据的采集,同时还模拟了多线束激光雷达的标注结果。文中自动生成的标注数据可以作为检测、分割任务的数据集,也可以作为图像和激光融合、补全的数据集,从而满足人工智能算法重新学习或者迁移学习的需求。该文所提出的方法不仅高效、简洁,还保证了数据的多样性,有效地解决了数据标注的难题。With the rapid development of autonomous driving technology in recent years,the unmanned vehicle industry has gradually become a focus of world attention.As more and more artificial intelligence algorithms are used to solve problems in environment perception,decision making and planning of unmanned vehicles,the demand for annotated data is also increasing dramatically.In order to reduce the time and labor cost of data annotation,this paper proposes an automatic generation method for image and laser data annotation based on Unity3D.Firstly,a 3D scene is constructed based on 2D map information,and then a virtual camera is used to realize the visible image and semantic segmentation data acquisition,and the annotation results of multi-line beam LIDAR are also simulated.The automatically generated annotation data in the paper can be used as a dataset for detection and segmentation tasks,as well as for image and laser fusion and complementation,thus satisfying the needs of artificial intelligence algorithms for relearning or migration learning.The proposed method in this paper is efficient and concise while ensuring data diversity and effectively solving the data annotation challenge.

关 键 词：三维场景重建 实例分割 激光点云 标注数据 

分 类 号：TU984.113[建筑科学—城市规划与设计]