SLAM论文相关

1. SLAM论文

1.1 SOFT-SLAM:Computationally efficient stereo visual simultaneous localization and mapping for autonomous unmanned aerial vehicles（KITTI 双目第一名）

1.2 [CVPR2018]CodeSLAM-Learning a Compact, Optimisable Representation for Dense Visual SLAM

1.2.1 Contributions:

l The derivation of a compact and optimisable representation of dense geometry by conditioning a depth auto encoder on intensity images (深度编码来表达稠密几何结构)

l The implementation of the first real-time targeted

monocular system that achieves such a tight joint optimisation of motion and dense geometry.

1.2.2 depth auto encoder result:

image.png

1.2.3 Illustration of the SfM system

image.png

1.2.4 总结

在后端优化的时候，将光度差和重投影误差一起优化，数据关联使用文中的共视depth auto encoder

1.3 [CVPR2018]ICE-BA: Incremental, Consistent and Efficient Bundle Adjustment for Visual-Inertial SLAM

1.3.1 Contributions:

l a new sliding window based solver that leverages the incremental nature of SLAM measurements to achieve more than 10x efficiency compared to the state-of-the-arts

l a new relative marginalization algorithm that resolves the conflicts between sliding window marginalization bias and global loop closure constraints

1.3.2 Optimization framework

image.png

1.3.3 Relative Marginalization

image.png

1.3.4 总结

本篇文章在local BA的过程中，选定新的reference keyframe,后面的相机位姿将相对于该帧做优化，这和HoloLens里面选定锚点位置有相似之处。

1.4 [CVPR2018] Learning to Find Good Correspondences

1.4.1 Contributions

l being keypoint-based, it generalizes better than image-based dense methods to unseen scenes, which we demonstrate with a single model that outperforms current methods on drastically different indoors and outdoors datasets

l it requires only weak supervision through essential matrices for training

l it can work effectively with very little training data

1.4.2 Result

image.png

1.4.3 Network

image.png

1.4.4 总结

提取局部特征点之后，匹配完扔给神经网络，它能够划分出inlier和outlier，这是一个分类器。

1.5 [CVPR2017]CNN-SLAM: Real-time dense monocular SLAM with learned depth prediction

1.5.1 Contributions

l we illustrate the proposed frameworkfor 3D reconstruction, where CNN-predicted dense depth maps are fused together with depth measurements obtained from direct monocular SLAM

l we show how CNN-predicted semantic segmentation can also be coherently fused with the global reconstruction model

1.5.2 overview

image.png

1.5.3 总结

使用CNN估计单幅视图的深度，并进行语义上的分割，后端优化还是传统的SLAM优化方法。

1.6 [CVPR2017]NID-SLAM: Robust Monocular SLAM using Normalised Information Distance

1.6.1 Contributions

l Robust direct tracking using NID

We present a real-time approach for minimising the NID between a candidate image and a key-frame depth map to recover the sim(3) camera pose. In contrast to previous methods we explicitly incorporate depth uncertainty into the NID score

l Multi-resolution tracking using histogram pyramids

We present a novel histogram-pyramid approach for robust coarse-to-fine tracking using NID which increases robustness and the basin of convergence while reducing computation time at smaller scales

l Direct depth map refinement using NID

We present a per-pixel key-frame depth map refinement approach using NID, which allows for map maintenance and depth updates over successive traversals despite appearance changes over time

1.6.2 Pipeline

image.png

1.6.3 NIC


1.6.4 总结

整个NID-SLAM属于直接法的一种，在跟踪的过程中，设计了上述的NID方法，利用联合熵来表达相似性，来代替传统直接法中跟踪部分。实验结果表明，NID-SLAM比ORB-SLAM和LSD的更加鲁棒，姿态估计效果不相上下。

1.7 [TIP2016] Efficient Non-Consecutive Feature Tracking for Structure-from-Motion

1.7.1 Contributions

l Two-Pass Matching for Consecutive Tracking

l Non-Consecutive Track Matching

1.7.2 Feature matching

image.png

1.7.3 Matching matrix

image.png

1.7.4 总结

l 在匹配阶段，进行两次匹配，第一次通过sift匹配获得少量的较好的匹配结果，然后通过RANSAC 算法生成多个单应矩阵，然后通过单应变换将在同一个平面上的点筛选出来，通过光流做进一步筛选，相当于将提取出来的特征点进行了多个平面归类。

通过连续帧匹配得到的跟踪，对跟踪结果进行评估，生成matching matrix，它表明了非连续帧之间的相关性，对相关性高的帧进行匹配，增加约束。