基于单目深度估计的低功耗视觉里程计  

Low Power Visual Odometry Technology Based on Monocular Depth Estimation

作　　者：马榕 陈秋瑞 张晗 梅铮 王锐[2] 魏伟 Ma Rong;Chen Qiurui;Zhang Han;Mei Zheng;Wang Rui;Wei Wei(Science and Technology on Special System Simulation Laboratory,Beijing 100854,China;BeiHang University,Beijing 100191,China;The Fourth Military Representative Office of the Air Force Armament Department in Beijing,Beijing 100041,China)

机构地区：[1]北京仿真中心航天系统仿真重点实验室,北京100854 [2]北京航空航天大学,北京100191 [3]空军装备部驻北京地区第四军事代表室,北京100041

出　　处：《系统仿真学报》2021年第12期3001-3011,共11页Journal of System Simulation

摘　　要：随着人工智能、精密机械和计算技术的发展,微小型无人系统在未来战场上将会扮演重要的角色。为解决单目视觉里程计尺度缺失以及微型机器人自身体积和载荷限制带来的功耗问题,引入单目深度估计技术,构建了一个低视角数据集,通过搭建卷积神经网络从单张图像中预测深度信息,对神经网络模型进行结构优化,将深度估计与单目视觉里程计融合并部署到JetsonNano平台。实验表明,融合后的单目视觉里程计能够在特定环境下恢复尺度信息,在JetsonNano上的功耗能够保持在较低水平,可为微型无人系统在未来战场上的隐蔽化、轻量化部署提供一定的研究基础。With the development of artificial intelligence,precision machinery and computing technology,micro-unmanned system will play an important role in the future battlefield.To solve the lack of monocular visual odometry scale,micro robot power consumption and load limits,the monocular depth estimation technology is introduced and a low view dataset is collected.A convolutional neural network to predict depth information from a single image is built,and the structure of neural network model is optimized.The depth estimation with monocular visual odometry are combined and deployed on JetsonNano.Experiments show that the combined monocular visual odometry can recover scale information in a specific environment,and the power consumption on Jetson Nano can be kept a low level,which can provide some research basis for the concealable and lightweight deployment of micro-unmanned system in the future battlefield.

关 键 词：单目视觉里程计 单目深度估计 卷积神经网络 神经网络结构优化 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]