双目立体视觉——华为云-云享MindTalks第十二期

灰太狼 2022-10-27 01:38 385阅读 0赞

## 前言 ##

前段时间在华为云-云享MindTalks第十二期中分享了双目视觉，个人的水平比较菜，分享时间也比较短，只能简单地分享一下基础和思路；制作PPT时主要参考了一些论文，同时也参考了一些网上的博客；

**目录**

单目视觉

双目视觉

1）双目摄像头

2）打开双目摄像头

3）双目测距

4）立体视觉系统

双目立体匹配（额外讲解）

参考

--------------------

## 开始分享 ##

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70][]

## 单目视觉 ##

下面几款是比较常见的单目摄像头；有的用于工业、有的用于学习开发、有的是用于监控。

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 1][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 2][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 3][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 4][]

## 双目视觉 ##

### 1）双目摄像头 ###

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 5][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 6][]

### 2）打开双目摄像头 ###

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 7][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 8][]![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 9][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 10][]

[链接][Link 1]

### 3）双目测距 ###

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 11][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 12][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 13][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 14][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 15][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 16][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 17][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 18][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 19][]

### 4）立体视觉系统 ###

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 20][]

双目标定

*  摄像头由于光学透镜的特性使得成像存在着径向畸变，可由三个参数k1,k2,k3确定。
 *  由于装配方面误差，传感器与光学镜头之间并非完全平行，因此成像存在切向畸变，可由两个参数p1,p2确定。
 *  单个摄像头的定标主要是计算出摄像头的内参（焦距f、成像原点，五个畸变参数即：一般只需要计算出k1,k2,p1,p2，对于鱼眼镜头等径向畸变特别大的才需要计算k3。
 *  双目摄像头定标不仅要得出每个摄像头的内部参数，还需要通过标定来测量两个摄像头之间的相对位置（即右摄像头相对于左摄像头的旋转矩阵R、平移向量t）

[参考： 基于MATLAB的双目视觉标定 ][_MATLAB_]

双目标定参数

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 21][]

这里说一些旋转矩阵，它是3x3的矩阵（9个参数）；有的朋友标准出来是一个旋转向量（3个参数），需要把旋转向量转换为旋转矩阵；相关代码对于开头注释的地方。

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 22][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 23][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 24][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 25][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 26][]

## 双目立体匹配（额外讲解） ##

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 27][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 28][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 29][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 30][]

**代价函数 **

*  AD/BT
 *  AD+Gradient
 *  Census transform
 *  SAD/SSD
 *  NCC
 *  AD+Census
 *  CNN

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 31][]

**端到端视差计算网络**

 Disp-Net (2016)

 GC-Net (2017)

 iRestNet (2018)

 PSM-Net (2018)

 Stereo-Net (2018)

 GA-Net (2019)

 EdgeStereo (2020)

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 32][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 33][]

双目测距总结

优势：

（1）成本比单目系统要高，但尚处于可接受范围内，并且与激光雷达等方案相比成本较低；

（2）没有识别率的限制，因为从原理上无需先进行识别再进行测算，而是对所有障碍物直接进行测量；

（3）直接利用视差计算距离，精度比单目高；

（4）无需维护样本数据库，因为对于双目没有样本的概念。

难点：

（1）计算量大，对计算单元的性能要求高，这使得双目系统的产品化、小型化的难度较；（芯片或FPGA）

（2）双目的配准效果，直接影响到测距的准确性；

（3）对环境光照非常敏感；（光照角度、光照强度）

（4）不适用于单调缺乏纹理的场景；（天空、白墙、沙漠）

（5）相机基线限制了测量范围。（基线越大，测量范围越远；基线越小，测量范围越近）

## 参考： ##

1）\[Wang 2015\] Wang W, Yan J, Xu N, et al. Real-time high-quality stereo vision system in FPGA. IEEE Transactions on Circuitsand Systems for Video Technology, 2015, 25(10): 1696-1708.2）

2）\[Kim 2016\] K.-R. Kim and C.-S. Kim. Adaptive smoothness constraints for efficient stereo matching using texture and edgeinformation. ICIP 2016.

3）\[Zbontar 2016\] Zbontar J, LeCun Y. Stereo matching by training a convolutional neural network to compare image patches.Journal of Machine Learning Research, 2016.

4）\[Park 2017\] Park H, Lee K M. Look wider to match image patches with convolutional neural networks. IEEE Signal ProcessingLetters, 2017.

5）Leonid Keselman, et al. Intel R RealSenseTM Stereoscopic Depth Cameras. CVPRW. 2017.

6）立体匹配算法原理与应用.奥比研究院.徐玉华

7）基于双目视觉的空间非合作目标姿态测量技术研究.颜坤

8）https://www.bilibili.com/video/BV1ka4y1L7xT?from=search&seid=5727123941116684431

9）https://blog.csdn.net/u011808673/article/details/90641589 10）https://www.cnblogs.com/polly333/p/5130375.html

[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70]: /images/20221024/4c3b15696e6c4f9ca9799f63f4d138b9.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 1]: /images/20221024/e104d6d186f9408caa33503ea4e9c7b4.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 2]: /images/20221024/32b65a5620354ef2aa5dc71cabc06258.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 3]: /images/20221024/98c9b44bbf2b4c98a43d6d937cd5494f.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 4]: /images/20221024/27601027d1b84a04843e81e16c12d38f.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 5]: /images/20221024/68d7fc65715946b68cad732731d0abac.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 6]: /images/20221024/088e0c9454b14e35bb3ad668c1cb3e95.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 7]: /images/20221024/e73ffcb9086c48bb966b1dd306fd2e3f.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 8]: https://img-blog.csdnimg.cn/20210202212758972.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 9]: /images/20221024/e4810d9cc73c479395a5884ac2335a4a.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 10]: /images/20221024/6756f373ae264ec9b9609a591307ffda.png
[Link 1]: https://blog.csdn.net/qq_41204464/article/details/113090604?ops_request_misc=%25257B%252522request%25255Fid%252522%25253A%252522161219462316780299013005%252522%25252C%252522scm%252522%25253A%25252220140713.130102334.pc%25255Fblog.%252522%25257D&request_id=161219462316780299013005&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2~blog~first_rank_v2~rank_v29-2-113090604.pc_v2_rank_blog_default&utm_term=%25E5%258F%258C%25E7%259B%25AE
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 11]: /images/20221024/92e6f5b2aed049198d976bcb047faa30.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 12]: /images/20221024/e16e90beefae4f4d8d40f9b2ad9abb52.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 13]: /images/20221024/8a57dba07d3840d0a7ba0f64f066f204.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 14]: /images/20221024/b594a0f76f3c42b8bec647b046849428.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 15]: /images/20221024/ca05d4c6d05b41ee9cb6155c88a953ea.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 16]: /images/20221024/70bb0527882d47979d4025569457f4d7.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 17]: /images/20221024/0859e5d1d1484b078ac9966bc610026a.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 18]: /images/20221024/ab795184dc10400cbf578703e23aabc5.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 19]: /images/20221024/0f5ffc1600d34f6fa8fa0f2ae82a067b.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 20]: /images/20221024/701441ff2f5b480e89bcfa33a520d43f.png
[_MATLAB_]: https://blog.csdn.net/qq_41204464/article/details/86602452?ops_request_misc=%25257B%252522request%25255Fid%252522%25253A%252522161220938216780274183321%252522%25252C%252522scm%252522%25253A%25252220140713.130102334.pc%25255Fblog.%252522%25257D&request_id=161220938216780274183321&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2~blog~first_rank_v2~rank_v29-3-86602452.pc_v2_rank_blog_default&utm_term=%25E5%258F%258C%25E7%259B%25AE
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 21]: /images/20221024/3adc6087e7834685a9cc155fe06b20a8.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 22]: /images/20221024/02c2af92b83a4683bb4c2362cdd4b8da.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 23]: /images/20221024/4bb72ba864ff4d1a88d9cdbd51663657.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 24]: /images/20221024/5aecd9485c1040a88d34666bd715f53d.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 25]: /images/20221024/c71918b77d44408a94bfecce1f3a670a.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 26]: /images/20221024/fb650679035f4be9a968552150bb8ba3.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 27]: /images/20221024/1f1913fce5524b30a6862b2ab551a9c4.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 28]: /images/20221024/81f380a894da43d3a440fdc77b34f718.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 29]: /images/20221024/7d2a3b06ebd44283890234dc4677ffde.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 30]: /images/20221024/e31eb3baa484447887593c5dc8687b37.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 31]: /images/20221024/0f92608f3a8241bbb33edded5aed65eb.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 32]: /images/20221024/ef99eb666344477bb6646b175641d254.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzQxMjA0NDY0_size_16_color_FFFFFF_t_70 33]: /images/20221024/d34169adee664b1cba18ecf374eddd69.png