基于语义建图的室内机器人实时场景分类  

Real-time scene category of indoor robot based on semantic mapping

作　　者：张文[1,2] 刘勇[1] 张超凡[1,2] 张龙[1] 夏营威[1] 

机构地区：[1]中国科学院合肥物质科学研究院应用技术研究所,安徽合肥230031 [2]中国科学技术大学科学岛分院,安徽合肥230026

出　　处：《传感器与微系统》2017年第8期18-21,28,共5页Transducer and Microsystem Technologies

基　　金：国家科技支撑计划资助项目(2015BAI01B00);安徽省科技重大专项计划资助项目(15CZZ02019);中国科学院STS项目(KFJ-SW-STS-161)

摘　　要：针对室内环境下的机器人场景识别问题,重点研究了场景分类策略的自主性、实时性和准确性,提出了一种语义建图方法。映射深度信息构建二维栅格地图,自主规划场景识别路径;基于卷积网络建立场景分类模型,实时识别脱离特定训练;利用贝叶斯框架融合先验知识,修正了错误分类并完成语义建图。实验结果表明:机器人能够进行全局自主探索,实时判断场景类别,并创建满足要求的语义地图。同时,实际路径规划中,机器人可以根据语义信息改善导航行为,验证了方法的可行性。Aiming at problems of robot scene recognition in indoor environment,a senmantic mapping algorithm is proposed,autonomy,realtime and accuracy of scene classification strategy are focused on. Two-dimensional grid map is constructed by mapping depth information and autonomously plan recognition path of scene. Convolutional network is applied to set up scene categorization model,recognize semantic classes without specific training in realtime. By Bayesian framework fusing prior knowledge,modify error classification and accomplish semantic mapping.Experimental results show that robot can carry out global autonomous exploration and realtime judge scene category,and set up semantic mapping which meets need. At the same time,in real path planning,robot can improve navigation behavior according to semantic information,feasibility of the method is verified.

关 键 词：自主建图 卷积网络 贝叶斯框架 语义地图 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]