Lifelong SLAM 论文解读合集（3）：针对长时间重复运行SLAM地图更新问题

(RAL 2020)Learning Matchable Image Transformations for Long-Term Metric Visual Localization

论文传送门: https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8963763.

(ICRA 2019)Long-Term Occupancy Grid Prediction Using Recurrent Neural Networks

论文传送门: https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8793582.

概述

这篇文章采用了RNN来做占据栅格地图的长期预测。RNN的输入包括一个对整个场景的bird eye view，including occupancy and velocity。深度CNN网络运用了convLSTM来把动态和静态的场景分开，且预测未来的动态场景。这篇文章把DOGMs（动态占据栅格地图）作为卷积和循环神经网络的输入，以预测将来的占据栅格地图信息。convLSTM能够运用上下文并且捕捉连续帧时间上的相关性。他的RNN中Encoder-LSTM用于进行输入序列的滤波处理，Decoder-LSTM被用来为未来几帧做预测。

动态占据栅格地图

IJRR里提出了 D. Nuss, et al., “A random finite set approach for dynamic occupancy grid maps with real-time application,” The International Journal of Robotics Research, vol. 37, no. 8, pp. 841–866, 2018. 传统占据栅格地图采用贝叶斯二值滤波的方法，每个栅格的速度在静态栅格里被假设为0，动态栅格就删除了这个假设。栅格的状态被用粒子群代替，越多的粒子进入了同一个栅格，这个栅格被占据的概率就越高。

基本架构

由于动态栅格地图削弱了栅格之间的独立假设，我们可以采用上下文信息进行综合。总体网络采用了VGG和反卷积网络，在两者中间加入了两层编码解码LSTM。The task of the Encoder-LSTM is to observe the input sequence and save information of several time steps in its internal states, which can be seen as sequential filtering of the scene. 长短期记忆被用来存储一些关键帧。In addition, the output of the Encoder-LSTM is used as static prediction ystat. 编码LSTM的输出被作为静态预测。Each time step, the Decoder-LSTM will be initialized with the current internal states of the Encoder-LSTM and produces a sequence of future predictions. In this work, the Decoder-LSTM is applied four times to produce the predictions for t0 + 0:5 s, t0 + 1 s, t0 + 1:5 s and t0 + 2 s, where t0 indicates the current time. 在down和upscaling之间，还添加了一些convLSTM的skip connection，这一操作能够让一些高层次特征加入到upscaling过程中，去减少遮挡带来的影响。

(RAL 2019)1-Day Learning, 1-Year Localization: Long-TermLiDAR Localization Using Scan Context Image

论文传送门: https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8633942.

三维点云描述

这个人发明了一种叫做SCI的描述子集合(???)，这个描述子可以作为图片形式进行保存。类似于直接将极坐标系下的点映射到直角坐标(不是变换，而是强行映射过去)。点的极坐标位置决定其在二位图像中的位置，其高度决定图像对应像素的颜色。

(ICRA 2020)Long-Term Robot Navigation in Indoor Environments Estimating Patterns in Traversability Changes

论文传送门: https://arxiv.org/pdf/1909.12733.pdf.