基于长期度量学习和图像匹配变换的机器人视觉定位研究  

Research on Robot Vision Positioning Based on Long-Term Metric Learning and Image Matching Transformation

作　　者：周自更 黄修乾 胡昌斌 黄双得 许保瑜 曹家军 ZHOU Zigeng;HUANG Xiuqian;HU Changbin;HUANG Shuangde;XU Baoyu;CAO Jiajun(Kunming Power Supply Bureau of Yunnan Power Grid Co.,Ltd.,Kunming Yunnan 650000,China;Yunnan Power Grid Co.,Ltd.,Yunnan Kunming 650011,China)

机构地区：[1]云南电网有限责任公司昆明供电局,云南昆明650000 [2]云南电网有限责任公司,云南昆明650011

出　　处：《电子器件》2020年第6期1396-1402,共7页Chinese Journal of Electron Devices

基　　金：云南电网有限责任公司科技项目(YNKJXM20170213)。

摘　　要：长期度量和自定位是自主移动机器人的基本功能,但由于光照、天气或季节变化,导致机器人采集的图像外观存在较大变化,为机器人行进、定位系统的正常运行带来了挑战。虽然基于运行经验的图像映射可以弥补图像外观的差距,但在几天或几个月的较长时间跨度内,就很难满足机器人可靠度量所需的数据量。为此从色彩恒常性理论出发,基于深度神经网络构建了RGB至灰度的非线性映射关系,实现了不同光照和天气条件下,图像内部特征匹配数的最大化;最后利用人工合成数据集与真实图像数据集进行了对比实验,实验结果显示在昼夜周期中,机器人视觉定位性能显著改进,同时显著降低了数据需求;实验发现结合低维图像的纹理特征,可以进一步改善图像外观特征匹配的效果。Long term measurement and self-positioning are the basic functions of autonomous mobile robots.However,due to the changes of light,weather or season,the appearance of the images collected by the robot changes greatly,which brings challenges to the robot moving and positioning system based on machine vision.Although the image mapping based on experience can make up for the gap of image appearance,it is difficult to meet the data needed for reliable measurement of robot in a few days or months.Based on the theory of color constancy,a nonlinear mapping from RGB to gray level is constructed based on the depth neural network,which maximizes the number of image internal feature matching under different illumination and weather conditions.Finally,experiments are carried out through the synthetic data set and the real image data set.The experimental results show that the performance of robot visual positioning is significantly improved in the day and night cycle Now combining the texture features of lowdimensional image can further improve the image appearance feature matching effect,and significantly reduce the data demand.

关 键 词：巡线机器人 色彩空间变换 非线性映射 视觉度量 视觉定位 

分 类 号：TM243.62[一般工业技术—材料科学与工程]