LSTM神经网络辅助的GNSS/VO组合定位方法  

GNSS and visual odometry integrated localization method based on LSTM neural network

作　　者：林秋良 赵有兵[2] 冯威 黄丁发 LIN Qiuliang;ZHAO Youbing;FENG Wei;HUANG Dingfa(Faculty of Geosciences and Environmental Engineering,Southwest Jiaotong University,Chengdu 610000,China;China Railway Eryuan Engineering Group Co.Ltd.,Chengdu 610000,China)

机构地区：[1]西南交通大学地球科学与环境工程学院,成都610000 [2]中铁二院工程集团有限责任公司,成都610000

出　　处：《导航定位学报》2023年第3期156-164,共9页Journal of Navigation and Positioning

基　　金：国家自然科学基金项目(42171429);中国国家铁路集团有限公司科技研究开发计划重大课题(K2020X017)。

摘　　要：针对当全球卫星导航系统(GNSS)信号失锁时,GNSS与视觉里程计(VO)组合定位方法定位精度下降的问题,提出一种基于长短期记忆(LSTM)神经网络辅助的GNSS/VO组合定位方法:在GNSS工作正常情况下,利用视觉里程计的位移增量和姿态构建LSTM的特征向量,将GNSS解算的位置增量作为输出对LSTM神经网络进行训练;GNSS信号失锁环境中,使用LSTM神经网络输出结果推算得到伪GNSS观测值,并将其与VO的结果进行松组合,实现GNSS/VO组合定位。实验结果表明,在GNSS信号丢失30、60、120s的过程中,所提方法的定位精度可分别提高约62%、64%、69%,证明该方法能够有效地提高GNSS/VO组合定位方法在GNSS拒止环境下的定位精度。Aiming at the problem of positioning accuracy degradation for combination of global navigation satellite system(GNSS)and visual odometry(VO)when GNSS has an outage,the paper proposed a combined GNSS/VO positioning method based on the assistance of long short-term memory(LSTM)neural network:with GNSS’normal work,the feature vectors of LSTM were constructed by using the displacement increment and attitude of VO,and the position increment solved by GNSS was used as the output to train the LSTM neural network;with GNSS’outage,the pseudo GNSS observation was calculated by using the outputs of LSTM neural network,and then was loosely integrated with the results of VO for realizing the combined positioning of GNSS/VO.Experimental result showed that in the absence of the GNSS signal for 30,60,120 s,the proposed method would improve the positioning accuracy by 62%,64%and 69%,respectively,indicating that the method could effectively improve the positioning ability of the integrated navigation system in GNSS rejection environment.

关 键 词：全球卫星导航系统(GNSS) 组合导航 长短期记忆 神经网络 GNSS拒止 视觉里程计(VO) 

分 类 号：P228[天文地球—大地测量学与测量工程]