视觉SLAM实战（二）：ORB-SLAM2 with Kinect2

来源：互联网发布：软件项目团队管理编辑：程序博客网时间：2024/06/07 01:31

前言

　　实战系列很久没有更新了。近期拿到了一台不错的Thinkpad和Kinect v2，前两天orbslam2又放出，于是想要在kinect2下尝试一下orb slam。整个过程没有特别多的技术含量，读者可以把它当成一个实验步骤的总结。

ORB-slam

　　orb-slam是15年出的一个单目SLAM，也可以说是单目中做的非常好的一个实现。另一方面，他的代码也极其清爽，编译十分贴心，十分注重我等程序员的用户体验，受到了广大欢迎。前几天，orb-slam作者推出了orb-slam2，在原来的单目基础上增加了双目和RGBD的接口，尽管地图还是单目常见的稀疏特征点图。于是我们就能通过各种传感器来玩orb-slam啦！这里正巧我手上有一个Kinectv2，咱们就拿它做个实验吧！

　　博主的系统就不多说了，ubuntu14.04, Thinkpad T450。Kinect2 for xbox.

编译Kinect v2

　　要在ubuntu下使用Kinect V2，需要做两件事。一是编译Kinectv2的开源驱动libfreenect2，二是使用kinect2_bridge在ROS下采集它的图像。这两者在hitcm的博客中已经说的很清楚了，咱就不罗嗦了。

　　请参照hitcm的博客安装好libfreenect2和iai_kinect2系列软件：

　　http://www.cnblogs.com/hitcm/p/5118196.html

　　iai_kinect2中含有四个模块。我们主要用它的bridge进行图像间的转换。此外，还要使用kinect2_registration进行标定。没标定过的kinect2，深度图和彩色图之间是不保证一一对应的，在做slam时就会出错。所以请务必做好它的标定。好在作者十分良心地写了标定的详细过程，即使像博主这样的小白也能顺利完成标定啦！

　　kinect2_calibration模块，有详细的标定过程解释：

　　https://github.com/code-iai/iai_kinect2/tree/master/kinect2_calibration

　　标定之后呢，会得到kinect2彩色头、深度头、红外头的内参和外参，它们都以（万恶的）yaml模式存储在你的机器内。kinect2_bridge会自动检测你的标定文件，对深度图进行校正。之后slam过程主要使用彩色头的内参哦！同时，你也可以使用kinect2_viewer模块来看获取的点云和kinect2的图像哦！

编译orb-slam2

　　orb-slam2的github位于：https://github.com/raulmur/ORB_SLAM2 直接clone到本地即可。

　　这版的orb-slam可以脱离ros编译，不需要事先安装ros。（但是由于我们要用kinect2还是装了ros）它的主要依赖项是opencv2.4, eigen3, dbow2和g2o，另外还有一些我没怎么听说过的UI库：pangolin。其中Dbow2和g2o已经包含在它的Thirdparty文件夹中，不需要另外下载啦！opencv的安装方式参见一起做第一篇。（g2o版本问题一直是个坑）所以，只需要去下载pangolin即可。

　　pangolin的github： https://github.com/stevenlovegrove/Pangolin 它是一个cmake工程，没什么特别的依赖项，直接下载编译安装即可。

　　随后，进入orb-slam2的文件夹。作者很贴心的为我们准备了 build.sh 文件，直接运行这个文件即可完成编译。

　　但愿你也顺利编译成功了。orb-slam作者为我们提供了几个example，包括kitti的双目和tum的单目/rgbd。我们可以参照着它去写自己的输入。如果你只想把orb-slam2作为一个整体的模块，可以直接调用include/System.h文件里定义的SLAM System哦。现在我们就把Kinect2丢给它试试。

在Kinect2上运行orb-slam2

　　Kinect2的topic一共有三种，含不同的分辨率。其中hd是1920的，qhd是四分之一的960的，而sd是最小的。博主发现sd的效果不理想，而hd的图像又太大了，建议大家使用qhd的920大小啦！在调用orb-slam时，需要把相机参数通过一个yaml来告诉它，所以需要简单写一下你的kinect参数喽。比如像这样：

%YAML:1.0#--------------------------------------------------------------------------------------------# Camera Parameters. Adjust them!#--------------------------------------------------------------------------------------------# Camera calibration and distortion parameters (OpenCV) Camera.fx: 529.97Camera.fy: 526.97Camera.cx: 477.44Camera.cy: 261.87Camera.k1: 0.05627Camera.k2: -0.0742Camera.p1: 0.00142Camera.p2: -0.00169Camera.k3: 0.0241Camera.width: 960Camera.height: 540# Camera frames per second Camera.fps: 30.0# IR projector baseline times fx (aprox.)Camera.bf: 40.0# Color order of the images (0: BGR, 1: RGB. It is ignored if images are grayscale)Camera.RGB: 1# Close/Far threshold. Baseline times.ThDepth: 50.0# Deptmap values factor DepthMapFactor: 1000.0#--------------------------------------------------------------------------------------------# ORB Parameters#--------------------------------------------------------------------------------------------# ORB Extractor: Number of features per imageORBextractor.nFeatures: 1000# ORB Extractor: Scale factor between levels in the scale pyramid ORBextractor.scaleFactor: 1.2# ORB Extractor: Number of levels in the scale pyramidORBextractor.nLevels: 8# ORB Extractor: Fast threshold# Image is divided in a grid. At each cell FAST are extracted imposing a minimum response.# Firstly we impose iniThFAST. If no corners are detected we impose a lower value minThFAST# You can lower these values if your images have low contrastORBextractor.iniThFAST: 20ORBextractor.minThFAST: 7#--------------------------------------------------------------------------------------------# Viewer Parameters#--------------------------------------------------------------------------------------------Viewer.KeyFrameSize: 0.05Viewer.KeyFrameLineWidth: 1Viewer.GraphLineWidth: 0.9Viewer.PointSize:2Viewer.CameraSize: 0.08Viewer.CameraLineWidth: 3Viewer.ViewpointX: 0Viewer.ViewpointY: -0.7Viewer.ViewpointZ: -1.8Viewer.ViewpointF: 500

　ORB部分的参数我们就不用动啦。然后，对kinect2_viewer进行一定程度的改写，加入ORBSLAM，就可以跑起来喽：

#include <stdlib.h>#include <stdio.h>#include <iostream>#include <sstream>#include <string>#include <vector>#include <cmath>#include <mutex>#include <thread>#include <chrono>#include <ros/ros.h>#include <ros/spinner.h>#include <sensor_msgs/CameraInfo.h>#include <sensor_msgs/Image.h>#include <cv_bridge/cv_bridge.h>#include <image_transport/image_transport.h>#include <image_transport/subscriber_filter.h>#include <message_filters/subscriber.h>#include <message_filters/synchronizer.h>#include <message_filters/sync_policies/exact_time.h>#include <message_filters/sync_policies/approximate_time.h>#include <kinect2_bridge/kinect2_definitions.h>#include "orbslam2/System.h"class Receiver{public:  enum Mode  {    IMAGE = 0,    CLOUD,    BOTH  };private:  std::mutex lock;  const std::string topicColor, topicDepth;  const bool useExact, useCompressed;  bool updateImage, updateCloud;  bool save;  bool running;  size_t frame;  const size_t queueSize;  cv::Mat color, depth;  cv::Mat cameraMatrixColor, cameraMatrixDepth;  cv::Mat lookupX, lookupY;  typedef message_filters::sync_policies::ExactTime<sensor_msgs::Image, sensor_msgs::Image, sensor_msgs::CameraInfo, sensor_msgs::CameraInfo> ExactSyncPolicy;  typedef message_filters::sync_policies::ApproximateTime<sensor_msgs::Image, sensor_msgs::Image, sensor_msgs::CameraInfo, sensor_msgs::CameraInfo> ApproximateSyncPolicy;  ros::NodeHandle nh;  ros::AsyncSpinner spinner;  image_transport::ImageTransport it;  image_transport::SubscriberFilter *subImageColor, *subImageDepth;  message_filters::Subscriber<sensor_msgs::CameraInfo> *subCameraInfoColor, *subCameraInfoDepth;  message_filters::Synchronizer<ExactSyncPolicy> *syncExact;  message_filters::Synchronizer<ApproximateSyncPolicy> *syncApproximate;  std::thread imageViewerThread;  Mode mode;  std::ostringstream oss;  std::vector<int> params;  //RGBDSLAM  slam; //the slam object  ORB_SLAM2::System* orbslam    =nullptr;public:  Receiver(const std::string &topicColor, const std::string &topicDepth, const bool useExact, const bool useCompressed)    : topicColor(topicColor), topicDepth(topicDepth), useExact(useExact), useCompressed(useCompressed),      updateImage(false), updateCloud(false), save(false), running(false), frame(0), queueSize(5),      nh("~"), spinner(0), it(nh), mode(CLOUD)  {    cameraMatrixColor = cv::Mat::zeros(3, 3, CV_64F);    cameraMatrixDepth = cv::Mat::zeros(3, 3, CV_64F);    params.push_back(cv::IMWRITE_JPEG_QUALITY);    params.push_back(100);    params.push_back(cv::IMWRITE_PNG_COMPRESSION);    params.push_back(1);    params.push_back(cv::IMWRITE_PNG_STRATEGY);    params.push_back(cv::IMWRITE_PNG_STRATEGY_RLE);    params.push_back(0);    string orbVocFile = "/home/xiang/catkin_ws/src/walle/config/ORBvoc.txt";    string orbSetiingsFile = "/home/xiang/catkin_ws/src/walle/config/kinect2_sd.yaml";    orbslam = new ORB_SLAM2::System( orbVocFile, orbSetiingsFile ,ORB_SLAM2::System::RGBD, true );  }  ~Receiver()  {      if (orbslam)      {          orbslam->Shutdown();          delete orbslam;      }  }  void run(const Mode mode)  {    start(mode);    stop();  }  void finish()   {  }private:  void start(const Mode mode)  {    this->mode = mode;    running = true;    std::string topicCameraInfoColor = topicColor.substr(0, topicColor.rfind('/')) + "/camera_info";    std::string topicCameraInfoDepth = topicDepth.substr(0, topicDepth.rfind('/')) + "/camera_info";    image_transport::TransportHints hints(useCompressed ? "compressed" : "raw");    subImageColor = new image_transport::SubscriberFilter(it, topicColor, queueSize, hints);    subImageDepth = new image_transport::SubscriberFilter(it, topicDepth, queueSize, hints);    subCameraInfoColor = new message_filters::Subscriber<sensor_msgs::CameraInfo>(nh, topicCameraInfoColor, queueSize);    subCameraInfoDepth = new message_filters::Subscriber<sensor_msgs::CameraInfo>(nh, topicCameraInfoDepth, queueSize);    if(useExact)    {      syncExact = new message_filters::Synchronizer<ExactSyncPolicy>(ExactSyncPolicy(queueSize), *subImageColor, *subImageDepth, *subCameraInfoColor, *subCameraInfoDepth);      syncExact->registerCallback(boost::bind(&Receiver::callback, this, _1, _2, _3, _4));    }    else    {      syncApproximate = new message_filters::Synchronizer<ApproximateSyncPolicy>(ApproximateSyncPolicy(queueSize), *subImageColor, *subImageDepth, *subCameraInfoColor, *subCameraInfoDepth);      syncApproximate->registerCallback(boost::bind(&Receiver::callback, this, _1, _2, _3, _4));    }    spinner.start();    std::chrono::milliseconds duration(1);    while(!updateImage || !updateCloud)    {      if(!ros::ok())      {        return;      }      std::this_thread::sleep_for(duration);    }    createLookup(this->color.cols, this->color.rows);    switch(mode)    {    case IMAGE:      imageViewer();      break;    case BOTH:      imageViewerThread = std::thread(&Receiver::imageViewer, this);      break;    }  }  void stop()  {    spinner.stop();    if(useExact)    {      delete syncExact;    }    else    {      delete syncApproximate;    }    delete subImageColor;    delete subImageDepth;    delete subCameraInfoColor;    delete subCameraInfoDepth;    running = false;    if(mode == BOTH)    {      imageViewerThread.join();    }  }  void callback(const sensor_msgs::Image::ConstPtr imageColor, const sensor_msgs::Image::ConstPtr imageDepth,                const sensor_msgs::CameraInfo::ConstPtr cameraInfoColor, const sensor_msgs::CameraInfo::ConstPtr cameraInfoDepth)  {    cv::Mat color, depth;    readCameraInfo(cameraInfoColor, cameraMatrixColor);    readCameraInfo(cameraInfoDepth, cameraMatrixDepth);    readImage(imageColor, color);    readImage(imageDepth, depth);    // IR image input    if(color.type() == CV_16U)    {      cv::Mat tmp;      color.convertTo(tmp, CV_8U, 0.02);      cv::cvtColor(tmp, color, CV_GRAY2BGR);    }    lock.lock();    this->color = color;    this->depth = depth;    updateImage = true;    updateCloud = true;    lock.unlock();  }  void imageViewer()  {    cv::Mat color, depth;    for(; running && ros::ok();)    {      if(updateImage)      {        lock.lock();        color = this->color;        depth = this->depth;        updateImage = false;        lock.unlock();        if (orbslam)        {            orbslam->TrackRGBD( color, depth, ros::Time::now().toSec() );        }      }    }    cv::destroyAllWindows();    cv::waitKey(100);  }  void readImage(const sensor_msgs::Image::ConstPtr msgImage, cv::Mat &image) const  {    cv_bridge::CvImageConstPtr pCvImage;    pCvImage = cv_bridge::toCvShare(msgImage, msgImage->encoding);    pCvImage->image.copyTo(image);  }  void readCameraInfo(const sensor_msgs::CameraInfo::ConstPtr cameraInfo, cv::Mat &cameraMatrix) const  {    double *itC = cameraMatrix.ptr<double>(0, 0);    for(size_t i = 0; i < 9; ++i, ++itC)    {      *itC = cameraInfo->K[i];    }  }  void dispDepth(const cv::Mat &in, cv::Mat &out, const float maxValue)  {    cv::Mat tmp = cv::Mat(in.rows, in.cols, CV_8U);    const uint32_t maxInt = 255;    #pragma omp parallel for    for(int r = 0; r < in.rows; ++r)    {      const uint16_t *itI = in.ptr<uint16_t>(r);      uint8_t *itO = tmp.ptr<uint8_t>(r);      for(int c = 0; c < in.cols; ++c, ++itI, ++itO)      {        *itO = (uint8_t)std::min((*itI * maxInt / maxValue), 255.0f);      }    }    cv::applyColorMap(tmp, out, cv::COLORMAP_JET);  }  void combine(const cv::Mat &inC, const cv::Mat &inD, cv::Mat &out)  {    out = cv::Mat(inC.rows, inC.cols, CV_8UC3);    #pragma omp parallel for    for(int r = 0; r < inC.rows; ++r)    {      const cv::Vec3b      *itC = inC.ptr<cv::Vec3b>(r),       *itD = inD.ptr<cv::Vec3b>(r);      cv::Vec3b *itO = out.ptr<cv::Vec3b>(r);      for(int c = 0; c < inC.cols; ++c, ++itC, ++itD, ++itO)      {        itO->val[0] = (itC->val[0] + itD->val[0]) >> 1;        itO->val[1] = (itC->val[1] + itD->val[1]) >> 1;        itO->val[2] = (itC->val[2] + itD->val[2]) >> 1;      }    }  }  void createLookup(size_t width, size_t height)  {    const float fx = 1.0f / cameraMatrixColor.at<double>(0, 0);    const float fy = 1.0f / cameraMatrixColor.at<double>(1, 1);    const float cx = cameraMatrixColor.at<double>(0, 2);    const float cy = cameraMatrixColor.at<double>(1, 2);    float *it;    lookupY = cv::Mat(1, height, CV_32F);    it = lookupY.ptr<float>();    for(size_t r = 0; r < height; ++r, ++it)    {      *it = (r - cy) * fy;    }    lookupX = cv::Mat(1, width, CV_32F);    it = lookupX.ptr<float>();    for(size_t c = 0; c < width; ++c, ++it)    {      *it = (c - cx) * fx;    }  }};int main(int argc, char **argv){  ros::init(argc, argv, "kinect2_slam", ros::init_options::AnonymousName);  if(!ros::ok())  {    return 0;  }  std::string topicColor = "/kinect2/sd/image_color_rect";  std::string topicDepth = "/kinect2/sd/image_depth_rect";  bool useExact = true;  bool useCompressed = false;  Receiver::Mode mode = Receiver::IMAGE;  // 初始化receiver  Receiver receiver(topicColor, topicDepth, useExact, useCompressed);  //OUT_INFO("starting receiver...");  receiver.run(mode);  receiver.finish();  ros::shutdown();  return 0;}

编译方面，只要在CMakeLists.txt中加入orb-slam的头文件和库，告诉cmake你想链接它即可。甚至你可以把整个orb-slam放到你的代码目录中一块儿编译，不过我还是简单地把liborb_slam2.so文件和头文件拷了过来而已。

　　实际的手持kinect2运行效果（由于博客园无法传视频，暂时把百度云当播放器使一使）：http://pan.baidu.com/s/1eRcyW1s （感谢也冬同学友情出演……）

　　一起做rgbd slam的数据集上效果：http://pan.baidu.com/s/1bocx5s

　　大体上还是挺理想的。

小结

　　本文主要展现了orbslam2在Kinect2下的表现，大致是令人满意的。读者在使用时，请务必注意kinect2的标定，否则很可能出错。

0 0