ORB_SLAM3启动流程以stereo_inertial_realsense_D435i为例

概述

ORB-SLAM3 是第一个同时具备纯视觉（visual）数据处理、视觉+惯性（visual-inertial）数据处理、和构建多地图（multi-map）功能，支持单目、双目以及 RGB-D 相机，同时支持针孔相机、鱼眼相机模型的 SLAM 系统。

主要创新点：

1.在 IMU 初始化阶段引入 MAP。该初始化方法可以实时快速进行，鲁棒性上有很大的提升（在大的场景还是小的场景，不管室内还是室外环境均有较好表现），并且比当前的其他方法具有 2-5 倍的精确度的提升。
2.基于一种召回率大大提高的 place recognition（也就是做回环检测，与历史数据建立联系）方法实现了一个多子地图（multi-maps）系统。ORB-SLAM3 在视觉信息缺乏的情况下更具有 long term 鲁棒性，当跟丢的时候，这个时候就会重新建一个子地图，并且在回环的时候与之前的子地图进行合并。ORB-SLAM3 是第一个可以重用历史所有算法模块的所有信息的系统。

主要结论：

在所有 sensor 配置下，ORB-SLAM3 的鲁棒性与现在的发表的各大系统中相近，精度上有了显著的提高。尤其使用Stereo-Inertial SLAM，在 EuRoC 数据集的平均误差接近 3.6 cm，在一个偏向 AR/VR 场景的 TUM-VI 数据集的平均误差接近 9mm。
ORB_SLAM3系统框图

ORB_SLAM3启动入口在Examples文件夹中，包含单目、双目、RGB-D，及其惯性组合。本文以双目+惯性作为例子，介绍stereo_inertial_realsense_D435i启动流程。

启动

./Examples/Stereo-Inertial/stereo_inertial_realsense_D435i ./Vocabulary/ORBvoc.txt ./Examples/Stereo-Inertial/Realsense_D435i.yaml

从启动参数可以看到，参数个数argc = 3，参数./Examples/Stereo-Inertial/stereo_inertial_realsense_D435i 为stereo_inertial_realsense_D435i main()函数入口，./Vocabulary/ORBvoc.txt 训练好的ORBvoc字典，参数./Examples/Stereo-Inertial/Realsense_D435i.yaml为D435i的配置文件，可以根据自己手上的D435i进行相机内外参、imu数据修改，具体可以参考系列博文Realsense d435i驱动安装、配置及校准

进入main函数后

一、利用librealsense2 rs接口对d435i就行设备读取，读取得到的句柄放sensors 中。

   ...
    string file_name;

    if (argc == 4) {
        file_name = string(argv[argc - 1]);
    }

    struct sigaction sigIntHandler;

    sigIntHandler.sa_handler = exit_loop_handler;
    sigemptyset(&sigIntHandler.sa_mask);
    sigIntHandler.sa_flags = 0;

    sigaction(SIGINT, &sigIntHandler, NULL);
    b_continue_session = true;

    double offset = 0; // ms

    rs2::context ctx;
    rs2::device_list devices = ctx.query_devices();
    rs2::device selected_device;
    if (devices.size() == 0)
    {
        std::cerr << "No device connected, please connect a RealSense device" << std::endl;
        return 0;
    }
    else
        selected_device = devices[0];

    std::vector<rs2::sensor> sensors = selected_device.query_sensors();
    int index = 0;

二、根据传感器的类型、名字进行相应的设置，如使能自动曝光补偿、曝光限制、关闭结构光等。

for (rs2::sensor sensor : sensors)
    if (sensor.supports(RS2_CAMERA_INFO_NAME)) {
        ++index;
        if (index == 1) {
            sensor.set_option(RS2_OPTION_ENABLE_AUTO_EXPOSURE, 1);
            sensor.set_option(RS2_OPTION_AUTO_EXPOSURE_LIMIT,5000);
            sensor.set_option(RS2_OPTION_EMITTER_ENABLED, 0); // switch off emitter
        }
        
        get_sensor_option(sensor);
        if (index == 2){
            // RGB camera (not used here...)
            sensor.set_option(RS2_OPTION_EXPOSURE,100.f);
        }

        if (index == 3){
            sensor.set_option(RS2_OPTION_ENABLE_MOTION_CORRECTION,0);
        }
    }

三、此部分可以看出，设置了相机左右目的图片大小640*480，帧率30hz，accel/gyro数据类型为float型。

// Declare RealSense pipeline, encapsulating the actual device and sensors
rs2::pipeline pipe;
// Create a configuration for configuring the pipeline with a non default profile
rs2::config cfg;
cfg.enable_stream(RS2_STREAM_INFRARED, 1, 640, 480, RS2_FORMAT_Y8, 30);
cfg.enable_stream(RS2_STREAM_INFRARED, 2, 640, 480, RS2_FORMAT_Y8, 30);
cfg.enable_stream(RS2_STREAM_ACCEL, RS2_FORMAT_MOTION_XYZ32F);
cfg.enable_stream(RS2_STREAM_GYRO, RS2_FORMAT_MOTION_XYZ32F);

四、读取d435i传感器数据，主要包括左右目相机数据cam_left/cam_right，imu数据imu_stream左目外参Rbc/tbc，右目外参Rbc/tbc，左右目内参intrinsics_left/intrinsics_right

    auto imu_callback = [&](const rs2::frame& frame)
    {
        std::unique_lock<std::mutex> lock(imu_mutex);

        if(rs2::frameset fs = frame.as<rs2::frameset>())
        {
            count_im_buffer++;

            double new_timestamp_image = fs.get_timestamp()*1e-3;
            if(abs(timestamp_image-new_timestamp_image)<0.001){
                // cout << "Two frames with the same timeStamp!!!\n";
                count_im_buffer--;
                return;
            }

            rs2::video_frame ir_frameL = fs.get_infrared_frame(1);
            rs2::video_frame ir_frameR = fs.get_infrared_frame(2);

            imCV = cv::Mat(cv::Size(width_img, height_img), CV_8U, (void*)(ir_frameL.get_data()), cv::Mat::AUTO_STEP);
            imRightCV = cv::Mat(cv::Size(width_img, height_img), CV_8U, (void*)(ir_frameR.get_data()), cv::Mat::AUTO_STEP);

            timestamp_image = fs.get_timestamp()*1e-3;
            image_ready = true;

            while(v_gyro_timestamp.size() > v_accel_timestamp_sync.size())
            {

                int index = v_accel_timestamp_sync.size();
                double target_time = v_gyro_timestamp[index];

                v_accel_data_sync.push_back(current_accel_data);
                v_accel_timestamp_sync.push_back(target_time);
            }

            lock.unlock();
            cond_image_rec.notify_all();
        }
        else if (rs2::motion_frame m_frame = frame.as<rs2::motion_frame>())
        {
            if (m_frame.get_profile().stream_name() == "Gyro")
            {
                // It runs at 200Hz
                v_gyro_data.push_back(m_frame.get_motion_data());
                v_gyro_timestamp.push_back((m_frame.get_timestamp()+offset)*1e-3);
                //rs2_vector gyro_sample = m_frame.get_motion_data();
                //std::cout << "Gyro:" << gyro_sample.x << ", " << gyro_sample.y << ", " << gyro_sample.z << std::endl;
            }
            else if (m_frame.get_profile().stream_name() == "Accel")
            {
                // It runs at 60Hz
                prev_accel_timestamp = current_accel_timestamp;
                prev_accel_data = current_accel_data;

                current_accel_data = m_frame.get_motion_data();
                current_accel_timestamp = (m_frame.get_timestamp()+offset)*1e-3;

                while(v_gyro_timestamp.size() > v_accel_timestamp_sync.size())
                {
                    int index = v_accel_timestamp_sync.size();
                    double target_time = v_gyro_timestamp[index];

                    rs2_vector interp_data = interpolateMeasure(target_time, current_accel_data, current_accel_timestamp,
                                                                prev_accel_data, prev_accel_timestamp);

                    v_accel_data_sync.push_back(interp_data);
                    v_accel_timestamp_sync.push_back(target_time);

                }
                // std::cout << "Accel:" << current_accel_data.x << ", " << current_accel_data.y << ", " << current_accel_data.z << std::endl;
            }
        }
    };

    rs2::pipeline_profile pipe_profile = pipe.start(cfg, imu_callback);

    vector<ORB_SLAM3::IMU::Point> vImuMeas;
    rs2::stream_profile cam_left = pipe_profile.get_stream(RS2_STREAM_INFRARED, 1);
    rs2::stream_profile cam_right = pipe_profile.get_stream(RS2_STREAM_INFRARED, 2);


    rs2::stream_profile imu_stream = pipe_profile.get_stream(RS2_STREAM_GYRO);
    float* Rbc = cam_left.get_extrinsics_to(imu_stream).rotation;
    float* tbc = cam_left.get_extrinsics_to(imu_stream).translation;
    std::cout << "Tbc (left) = " << std::endl;
    for(int i = 0; i<3; i++){
        for(int j = 0; j<3; j++)
            std::cout << Rbc[i*3 + j] << ", ";
        std::cout << tbc[i] << "\n";
    }

    float* Rlr = cam_right.get_extrinsics_to(cam_left).rotation;
    float* tlr = cam_right.get_extrinsics_to(cam_left).translation;
    std::cout << "Tlr  = " << std::endl;
    for(int i = 0; i<3; i++){
        for(int j = 0; j<3; j++)
            std::cout << Rlr[i*3 + j] << ", ";
        std::cout << tlr[i] << "\n";
    }



    rs2_intrinsics intrinsics_left = cam_left.as<rs2::video_stream_profile>().get_intrinsics();
    width_img = intrinsics_left.width;
    height_img = intrinsics_left.height;
    cout << "Left camera: \n";
    std::cout << " fx = " << intrinsics_left.fx << std::endl;
    std::cout << " fy = " << intrinsics_left.fy << std::endl;
    std::cout << " cx = " << intrinsics_left.ppx << std::endl;
    std::cout << " cy = " << intrinsics_left.ppy << std::endl;
    std::cout << " height = " << intrinsics_left.height << std::endl;
    std::cout << " width = " << intrinsics_left.width << std::endl;
    std::cout << " Coeff = " << intrinsics_left.coeffs[0] << ", " << intrinsics_left.coeffs[1] << ", " <<
        intrinsics_left.coeffs[2] << ", " << intrinsics_left.coeffs[3] << ", " << intrinsics_left.coeffs[4] << ", " << std::endl;
    std::cout << " Model = " << intrinsics_left.model << std::endl;

    rs2_intrinsics intrinsics_right = cam_right.as<rs2::video_stream_profile>().get_intrinsics();
    width_img = intrinsics_right.width;
    height_img = intrinsics_right.height;
    cout << "Right camera: \n";
    std::cout << " fx = " << intrinsics_right.fx << std::endl;
    std::cout << " fy = " << intrinsics_right.fy << std::endl;
    std::cout << " cx = " << intrinsics_right.ppx << std::endl;
    std::cout << " cy = " << intrinsics_right.ppy << std::endl;
    std::cout << " height = " << intrinsics_right.height << std::endl;
    std::cout << " width = " << intrinsics_right.width << std::endl;
    std::cout << " Coeff = " << intrinsics_right.coeffs[0] << ", " << intrinsics_right.coeffs[1] << ", " <<
        intrinsics_right.coeffs[2] << ", " << intrinsics_right.coeffs[3] << ", " << intrinsics_right.coeffs[4] << ", " << std::endl;
    std::cout << " Model = " << intrinsics_right.model << std::endl;

从一到四部分可以看出，这部分主要是和d435i的交互，主要包括对其检测、设置、读取数据，这部分可以看做是d435i驱动部分。

五、进入slam系统，主要用于初始化系统线程和准备工作

ORB_SLAM3::System SLAM(argv[1],argv[2],ORB_SLAM3::System::IMU_STEREO, true, 0, file_name);

/**
 * @brief 系统的构造函数，将会启动其他的线程
 * @param strVocFile 词袋文件所在路径
 * @param strSettingsFile 配置文件所在路径
 * @param sensor 传感器类型
 * @param bUseViewer 是否使用可视化界面
 * @param initFr initFr表示初始化帧的id,开始设置为0
 * @param strSequence 序列名,在跟踪线程和局部建图线程用得到
 */
System::System(const string &strVocFile, const string &strSettingsFile, const eSensor sensor,
               const bool bUseViewer, const int initFr, const string &strSequence):

1.输出welcome信息

// Output welcome message
cout << endl <<
"ORB-SLAM3 Copyright (C) 2017-2020 Carlos Campos, Richard Elvira, Juan J. Gómez, José M.M. Montiel and Juan D. Tardós, University of Zaragoza." << endl <<
"ORB-SLAM2 Copyright (C) 2014-2016 Raúl Mur-Artal, José M.M. Montiel and Juan D. Tardós, University of Zaragoza." << endl <<
"This program comes with ABSOLUTELY NO WARRANTY;" << endl  <<
"This is free software, and you are welcome to redistribute it" << endl <<
"under certain conditions. See LICENSE.txt." << endl << endl;

cout << "Input sensor was set to: ";

2.读取配置文件

读取配置文件对应的就是启动./Examples/Stereo-Inertial/Realsense_D435i.yaml文件。针对文件版本不同对于mStrLoadAtlasFromFile、mStrSaveAtlasToFile不同。

//Check settings file
cv::FileStorage fsSettings(strSettingsFile.c_str(), cv::FileStorage::READ);
// 如果打开失败，就输出错误信息
if(!fsSettings.isOpened())
{
   cerr << "Failed to open settings file at: " << strSettingsFile << endl;
   exit(-1);
}

// 查看配置文件版本，不同版本有不同处理方法
cv::FileNode node = fsSettings["File.version"];
if(!node.empty() && node.isString() && node.string() == "1.0")
{
	settings_ = new Settings(strSettingsFile,mSensor);

	// 保存及加载地图的名字
	mStrLoadAtlasFromFile = settings_->atlasLoadFile();
	mStrSaveAtlasToFile = settings_->atlasSaveFile();

	cout << (*settings_) << endl;
}
else
{
	settings_ = nullptr;
	cv::FileNode node = fsSettings["System.LoadAtlasFromFile"];
	if(!node.empty() && node.isString())
	{
		mStrLoadAtlasFromFile = (string)node;
	}

	node = fsSettings["System.SaveAtlasToFile"];
	if(!node.empty() && node.isString())
	{
		mStrSaveAtlasToFile = (string)node;
	}
}

3.是否激活回环

node = fsSettings["loopClosing"];
bool activeLC = true;
if(!node.empty())
{
	activeLC = static_cast<int>(fsSettings["loopClosing"]) != 0;
}

4.词袋文件赋值给mStrVocabularyFilePath，对应启动参数./Vocabulary/ORBvoc.txt

mStrVocabularyFilePath = strVocFile;

5.多地图管理功能

根据mStrLoadAtlasFromFile文件中是否有Atlas，进行相应处理。

（1）若文件中存在：

建立一个新的ORB字典，读取预训练好的ORB字典并返回成功/失败标志，创建关键帧数据库，创建多地图；
######（2）若文件中不存在：
建立一个新的ORB字典，读取预训练好的ORB字典并返回成功/失败标志，创建关键帧数据库，创建关键帧数据库，导入Atlas地图，创建新地图。

bool loadedAtlas = false;

if(mStrLoadAtlasFromFile.empty())
{
	//Load ORB Vocabulary
	cout << endl << "Loading ORB Vocabulary. This could take a while..." << endl;

	// 建立一个新的ORB字典
	mpVocabulary = new ORBVocabulary();
	// 读取预训练好的ORB字典并返回成功/失败标志
	bool bVocLoad = mpVocabulary->loadFromTextFile(strVocFile);
	// 如果加载失败，就输出错误信息
	if(!bVocLoad)
	{
		cerr << "Wrong path to vocabulary. " << endl;
		cerr << "Falied to open at: " << strVocFile << endl;
		exit(-1);
	}
	cout << "Vocabulary loaded!" << endl << endl;

	//Create KeyFrame Database
	// Step 4 创建关键帧数据库
	mpKeyFrameDatabase = new KeyFrameDatabase(*mpVocabulary);

	//Create the Atlas
	// Step 5 创建多地图，参数0表示初始化关键帧id为0
	cout << "Initialization of Atlas from scratch " << endl;
	mpAtlas = new Atlas(0);
}
else
{
	//Load ORB Vocabulary
	cout << endl << "Loading ORB Vocabulary. This could take a while..." << endl;

	mpVocabulary = new ORBVocabulary();
	bool bVocLoad = mpVocabulary->loadFromTextFile(strVocFile);
	if(!bVocLoad)
	{
		cerr << "Wrong path to vocabulary. " << endl;
		cerr << "Falied to open at: " << strVocFile << endl;
		exit(-1);
	}
	cout << "Vocabulary loaded!" << endl << endl;

	//Create KeyFrame Database
	mpKeyFrameDatabase = new KeyFrameDatabase(*mpVocabulary);

	cout << "Load File" << endl;

	// Load the file with an earlier session
	//clock_t start = clock();
	cout << "Initialization of Atlas from file: " << mStrLoadAtlasFromFile << endl;
	bool isRead = LoadAtlas(FileType::BINARY_FILE);

	if(!isRead)
	{
		cout << "Error to load the file, please try with other session file or vocabulary file" << endl;
		exit(-1);
	}

	loadedAtlas = true;

	mpAtlas->CreateNewMap();
}

6.此部分根据是否存在imu数据进行初始化

// 如果是有imu的传感器类型，设置mbIsInertial = true;以后的跟踪和预积分将和这个标志有关
if (mSensor==IMU_STEREO || mSensor==IMU_MONOCULAR || mSensor==IMU_RGBD)
	mpAtlas->SetInertialSensor();

7.依次创建跟踪、局部建图、闭环、显示线程

（1）创建用于显示帧和地图的类，由Viewer调用

//Create Drawers. These are used by the Viewer
mpFrameDrawer = new FrameDrawer(mpAtlas);
mpMapDrawer = new MapDrawer(mpAtlas, strSettingsFile, settings_);

（2）创建跟踪线程（主线程）,不会立刻开启,会在对图像和imu预处理后在main主线程种执行，(main)SLAM.TrackStereo()–>mpTracker->GrabImageStereo–>Track()开启跟踪。

//Initialize the Tracking thread
//(it will live in the main thread of execution, the one that called this constructor)
cout << "Seq. Name: " << strSequence << endl;
mpTracker = new Tracking(this, mpVocabulary, mpFrameDrawer, mpMapDrawer,					 mpAtlas, mpKeyFrameDatabase, strSettingsFile, mSensor, settings_, strSequence);

（3）创建并开启local mapping 线程，线程入口LocalMapping::Run

//Initialize the Local Mapping thread and launch
mpLocalMapper = new LocalMapping(this, mpAtlas, mSensor==MONOCULAR || mSensor==IMU_MONOCULAR,
		mSensor==IMU_MONOCULAR || mSensor==IMU_STEREO || mSensor==IMU_RGBD, strSequence);
mptLocalMapping = new thread(&ORB_SLAM3::LocalMapping::Run,mpLocalMapper);
mpLocalMapper->mInitFr = initFr;

（4）设置最远3D地图点的深度值，如果超过阈值，说明可能三角化不成功，丢弃

if(settings_)
	mpLocalMapper->mThFarPoints = settings_->thFarPoints();
else
	mpLocalMapper->mThFarPoints = fsSettings["thFarPoints"];
	
if(mpLocalMapper->mThFarPoints!=0)
{
	cout << "Discard points further than " << mpLocalMapper->mThFarPoints << " m from current camera" << endl;
	mpLocalMapper->mbFarPoints = true;
}
else
	mpLocalMapper->mbFarPoints = false;

（5）创建并开启闭环线程，程序入口LoopCloing::Run

//Initialize the Loop Closing thread and launch
// mSensor!=MONOCULAR && mSensor!=IMU_MONOCULAR
mpLoopCloser = new LoopClosing(mpAtlas, mpKeyFrameDatabase, mpVocabulary, mSensor!=MONOCULAR, activeLC); // mSensor!=MONOCULAR);
mptLoopClosing = new thread(&ORB_SLAM3::LoopClosing::Run, mpLoopCloser);

（6）设置线程间指针

//Set pointers between threads
mpTracker->SetLocalMapper(mpLocalMapper);
mpTracker->SetLoopClosing(mpLoopCloser);

mpLocalMapper->SetTracker(mpTracker);
mpLocalMapper->SetLoopCloser(mpLoopCloser);

mpLoopCloser->SetTracker(mpTracker);
mpLoopCloser->SetLocalMapper(mpLocalMapper);	

（7）创建并开启显示线程，程序入口Viewer::Run

//Initialize the Viewer thread and launch
if(bUseViewer)
//if(false) // TODO
{
	mpViewer = new Viewer(this, mpFrameDrawer,mpMapDrawer,mpTracker,strSettingsFile,settings_);
	mptViewer = new thread(&Viewer::Run, mpViewer);
	mpTracker->SetViewer(mpViewer);
	mpLoopCloser->mpViewer = mpViewer;
	mpViewer->both = mpFrameDrawer->both;
}
// Fix verbosity
// 打印输出中间的信息，设置为安静模式
Verbose::SetTh(Verbose::VERBOSITY_QUIET);

至此，Systemf完成。

6.此部分清空imu数据向量

// Clear IMU vectors
v_gyro_data.clear();
v_gyro_timestamp.clear();
v_accel_data_sync.clear();
v_accel_timestamp_sync.clear();

7.开启while(!isShutDown())主循环

8.根据时间戳对加速度进行插值

while(v_gyro_timestamp.size() > v_accel_timestamp_sync.size())
{
	int index = v_accel_timestamp_sync.size();
	double target_time = v_gyro_timestamp[index];

	rs2_vector interp_data = interpolateMeasure(target_time, current_accel_data, current_accel_timestamp, prev_accel_data, prev_accel_timestamp);

	v_accel_data_sync.push_back(interp_data);
	// v_accel_data_sync.push_back(current_accel_data); // 0 interpolation
	v_accel_timestamp_sync.push_back(target_time);
}

九、压入数据放入vImuMeas

for(int i=0; i<vGyro.size(); ++i)
{
	ORB_SLAM3::IMU::Point lastPoint(vAccel[i].x, vAccel[i].y, vAccel[i].z,
						  vGyro[i].x, vGyro[i].y, vGyro[i].z,
						  vGyro_times[i]);
	vImuMeas.push_back(lastPoint);
}

十、正式开启双目追踪线程

// Stereo images are already rectified.
SLAM.TrackStereo(im, imRight, timestamp, vImuMeas);

至此，stereo_inertial_realsense_D435i启动流程就全部完成。其中，跟踪Track、局部建图LocalMapping、闭环LoopCloing、显示线程Viewer后续再分子序列详细叙述。

参考：

1.https://blog.csdn.net/u010196944/article/details/128972333?spm=1001.2014.3001.5501
2.https://blog.csdn.net/u010196944/article/details/127240169?spm=1001.2014.3001.5501文章来源地址https://www.toymoban.com/news/detail-533162.html

到了这里，关于ORB_SLAM3启动流程以stereo_inertial_realsense_D435i为例的文章就介绍完了。如果您还想了解更多内容，请在右上角搜索TOY模板网以前的文章或继续浏览下面的相关文章，希望大家以后多多支持TOY模板网！