基于新型匹配神经网络的单目视觉定位与建图系统
首发时间:2022-12-14
摘要:本文提出了一种新的基于特征点法和光流融合的单目视觉定位与建图系统(SOF-VSLAM)。该系统可以比传统的单目视觉定位系统更稳健地估计载具6自由度姿态以及建立可重复使用的点云地图。该方法采用一种基于深度学习的特征点光流跟踪网络-SOFNet为视觉定位系统提供帧间匹配关系同时可以输出特征点的特征向量应用于回环过程。该系统在TUM和Euroc数据集上的实验表明,相比与传统的单目视觉定位系统,该系统表现的更加鲁棒和精确。
For information in English, please click here
Monocular Vision Positioning and Mapping System Based on a Novel Matching Neural Network
Abstract:This paper proposes a new monocular vision localization and mapping system (SOF-VSLAM) based on feature point method and optical flow fusion. The system can estimate the 6-DOF attitude of the vehicle more robustly than the traditional monocular vision positioning system and establish a reusable point cloud map. This method uses a feature point optical flow tracking network based on deep learning-SOF to provide the matching relationship between frames for the visual positioning system, and at the same time, it can output the feature vector of the feature point and apply it to the loopback process. Experiments of the system on TUM and Euroc datasets show that the system is more robust and accurate than the traditional monocular vision positioning system.
Keywords: Visual Localization ,Deep Learning ,Feature Detection ,Optical Flow
基金:
引用
No.****
动态公开评议
共计0人参与
勘误表
基于新型匹配神经网络的单目视觉定位与建图系统
评论
全部评论