改进Deeplab V3+网络在视觉SLAM三维地图构建应用  

Improved Deeplab V3+Network Application for Visual SLAM 3D Map Construction

作　　者：屈航 嵇启春[1] 段中兴[1] QU Hang;JI Qi-chun;DUAN Zhong-xing(Collegel of Information and Control Engineering,Xi′an University of Architecture and Technology,Xi′an 710055,China)

机构地区：[1]西安建筑科技大学信息与控制工程学院,西安710055

出　　处：《小型微型计算机系统》2022年第10期2174-2178,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(51678470)资助.

摘　　要：针对传统同时定位与地图构建(SLAM)算法在构建三维地图过程中缺少语义信息问题,本文在视觉SLAM算法基础上,结合基于深度学习的Deeplab V3+语义分割模型,构建包含几何信息与语义信息的三维稠密语义地图.对Deeplab V3+模型结合视觉SLAM造成语义地图构建难以满足实时性问题,精简Deeplab V3+模型参数,主干网络选用轻量级卷积网络MobileNetV3进行特征提取,同时对空洞空间金字塔池化模块中卷积层采用非对称卷积运算.最后利用贝叶斯更新方法将对RGB图像分割后获得的语义信息增量融合进三维地图,实现在三维空间对不同物体进行语义标注,最终完成三维稠密语义地图构建.实验采用NYU v2数据集进行语义地图构建,结果表明,改进后的Deeplab V3+可以精确快速进行语义分割,应用于三维稠密语义地图构建,满足系统实时性要求.In order to solve the problem that traditional simultaneous localization and map construction(SLAM)algorithm lacks semantic information in the process of constructing 3D maps,this paper,based on the visual SLAM algorithm,combined with the Deeplab V3+semantic segmentation model based on deep learning,constructs 3D dense semantic maps containing geometric information and semantic information.The combination of Deeplab V3+model and visual SLAM makes it difficult to construct semantic map in real time.The parameters of Deeplab V3+model are simplified.The lightweight convolutional network MobileNetv3 is selected for feature extraction in the backbone network,and asymmetric convolution operation is adopted for the convolutional layer in the pooling module of void space pyramid.Finally,the semantic information obtained from the RGB image segmentation is incremental fused into the 3D map by using Bayesian updating method to achieve semantic annotation of different objects in 3D space,and finally the construction of 3D dense semantic map is completed.NYU V2 dataset is used in the experiment to construct semantic map.The results show that the improved Deeplab V3+can segment semantic map accurately and quickly,and can be applied to construct three-dimensional dense semantic map to meet the real-time requirements of the system.

关 键 词：视觉SLAM Deeplab V3+ 语义地图 MobileNetV3 贝叶斯更新 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP242[自动化与计算机技术—计算机科学与技术]