摘要
军事领域离不开导航技术,而定位是导航的基础。尽管目前以卷积神经网络为代表的机器学习技术,已经在有关单张图像的六自由度位姿回归方面取得显著进展,其精度达到甚至超过基于手工特征提取的传统方式。然而,单张图像的位姿回归问题只参考了场景的结构化信息,缺乏序列图像间的时序关系,导致其位姿回归的损失函数不包含时空约束,最终限制了算法定位精度。针对单张图像六自由度位姿回归缺乏时空约束的问题,依靠长短期记忆网络(LSTM)对时序关系的捕捉优势,提出一类基于LSTM与孪生网络的视觉定位技术,其综合孪生网络与LSTM各自优势,将图像间运动视差信息与位姿时序信息同时作用于位姿预测,构建从序列图像到图像各自对应六自由度位姿信息的端对端深度神经网络,并通过开源数据库和仿真数据验证了算法的准确性和精度,为军事领域的视觉定位和协同跟踪奠定了基础。
The military field is inseparable from navigation technology,and positioning is the basis of navigation.Although the current machine learning technology represented by convolutional neural networks has made significant progress in the six-degree-of-freedom pose regression of single images,its accuracy reaches or exceeds the traditional method based on manual feature extraction.However,the pose regression problem of a single image only refers to the structured information of the scene,and lacks the time series relationship between the sequence images.The loss function of the pose regression does not contain the space-time constraint,which ultimately limits the accuracy of the algorithm.Aiming at the lack of space-time constraints in the six-degree-of-freedom pose regression of a single image,relying on the long-short-term memory network (LSTM) to capture the timing relationship,a kind of visual positioning technology based on LSTM and siamese networks is proposed,which integrates the siamese network and The advantages of LSTM are that the inter-image motion disparity information and the pose timing information are simultaneously applied to the pose prediction,and the end-to-end depth neural network corresponding to the six-degree-of-freedom pose information from the sequence image to the image is constructed,and through the open source database and simulation.The data verifies the accuracy and accuracy of the algorithm and lays a foundation for visual positioning and coordinated tracking in the military field.
作者
梁欣凯
宋闯
郝明瑞
赵佳佳
郑多
LIANG Xin-kai;SONG Chuang;HAO Ming-rui;ZHAO Jia-jia;ZHENG Duo(Science and Technology on Complex System Control and Intelligent Agent Cooperation Laboratory,Beijing 100074,China)
出处
《现代防御技术》
2019年第5期65-70,共6页
Modern Defence Technology
关键词
视觉定位
位姿估计
深度学习
长短期记忆网络
孪生网络
visual localization
pose estimation
deep learning
long short-term memory (LSTM)
siamese network
