基于深度学习的室内视觉位置识别技术  

Indoor visual place recognition technology based on deep learning

作　　者：朱会平 钱久超 朱肖光 应忍冬[1] ZHU Hui-ping;QIAN Jiu-chao;ZHU Xiao-guang;YING Ren-dong(Shanghai Jiaotong University, Shanghai 200240, China)

机构地区：[1]上海交通大学

出　　处：《信息技术》2019年第6期77-81,共5页Information Technology

摘　　要：视觉位置识别技术通过将地点图像与数据库中的图像集进行匹配,根据配对图像标签中的位置信息得到定位结果。现有的视觉位置识别网络都是为了应对室外场景而构建和训练的,在室内场景中的识别性能较差。文中提出了一种基于深度学习的室内视觉位置识别卷积神经网络架构,并在室内场景识别数据集上对网络进行了训练,然后在本地室内数据集上对网络参数进行进一步的微调,较好地解决了室内环境中的位置识别问题。和现有的其它视觉位置识别网络相比,文中训练的网络在实际室内环境测试中最大有30%的识别精度提升。Visual place recognition technology matches a place images with images in the dataset, and then gets the positioning result from the location information in the tag of matched images. The existing visual place recognition networks are all constructed and trained to cope with outdoor scenes, which makes them perform poorly in indoor scenes. An indoor visual place recognition convolutional neural network based on deep learning is proposed and trained on a special indoor scene recognition dataset, further fine-tuned on the local indoor dataset, which solves the problem of place recognition in indoor environments. Compared with other existing visual place recognition networks, the network trained in this paper can gain an accuracy improvement of 30% in the actual indoor environment test.

关 键 词：深度学习 卷积神经网络 室内场景 视觉位置识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]