认真的虎ORBSLAM2源码解读(二)：LocalMapping建图

0.前言

注释代码已公开，欢迎交流～～
注释代码已公开，欢迎交流
其他系列文章地址

1.简述

2.头文件

class LocalMapping
{
public:
    LocalMapping(Map* pMap, const float bMonocular);

    void SetLoopCloser(LoopClosing* pLoopCloser);

    void SetTracker(Tracking* pTracker);

    // Main function
    void Run();

    void InsertKeyFrame(KeyFrame* pKF);

    // Thread Synch
    // 当检测到闭环时，请求局部地图停止，防止局部地图线程中InsertKeyFrame函数插入新的关键帧
    void RequestStop();
    void RequestReset();
    bool Stop();
    void Release();
    bool isStopped();
    bool stopRequested();
    //返回mbAcceptKeyFrames，查询局部地图管理器是否繁忙
    bool AcceptKeyFrames();
    // 设置mbAcceptKeyFrames，false表示告诉Tracking，
    //LocalMapping正处于繁忙状态，不能接受插入新的keyframe
    void SetAcceptKeyFrames(bool flag);
    bool SetNotStop(bool flag);

    void InterruptBA();

    void RequestFinish();
    bool isFinished();

    int KeyframesInQueue(){
        unique_lock<std::mutex> lock(mMutexNewKFs);
        return mlNewKeyFrames.size();
    }

protected:

    //返回mlNewKeyFrames是否为空，也就是查询等待处理的关键帧列表是否空
    bool CheckNewKeyFrames();
    /**
    * @brief 处理列表中的关键帧
    * 
    * - 计算Bow，加速三角化新的MapPoints
    * - 关联当前关键帧至MapPoints，并更新MapPoints的平均观测方向和观测距离范围
    * - 插入关键帧，更新Covisibility图和Essential图，以及spanningtree
    */
    void ProcessNewKeyFrame();
    /** 
     * 1.找出和当前关键帧共视程度前10/20（单目/双目RGBD）的关键帧
     * 2.计算这些关键帧和当前关键帧的F矩阵；
     * 3.在满足对级约束条件下，匹配关键帧之间的特征点，通过BOW加速匹配；
     */
    void CreateNewMapPoints();

    // 剔除ProcessNewKeyFrame和CreateNewMapPoints函数中引入在mlpRecentAddedMapPoints的质量不好的MapPoints
    void MapPointCulling();
    // 检查并融合与当前关键帧共视程度高的帧重复的MapPoints
    void SearchInNeighbors();

    // 检测并剔除当前帧相邻的关键帧中冗余的关键帧
    // 剔除的标准是：该关键帧的90%的MapPoints可以被其它至少3个关键帧观测到
    void KeyFrameCulling();

    cv::Mat ComputeF12(KeyFrame* &pKF1, KeyFrame* &pKF2);

    cv::Mat SkewSymmetricMatrix(const cv::Mat &v);

    bool mbMonocular;

    void ResetIfRequested();
    bool mbResetRequested;
    std::mutex mMutexReset;

    bool CheckFinish();
    void SetFinish();
    bool mbFinishRequested;
    bool mbFinished;
    std::mutex mMutexFinish;

    Map* mpMap;

    LoopClosing* mpLoopCloser;
    Tracking* mpTracker;

    std::list<KeyFrame*> mlNewKeyFrames;

    KeyFrame* mpCurrentKeyFrame;

    std::list<MapPoint*> mlpRecentAddedMapPoints;

    std::mutex mMutexNewKFs;

    bool mbAbortBA;

    bool mbStopped;
    bool mbStopRequested;
    bool mbNotStop;
    std::mutex mMutexStop;

    //是否能接受插入新的keyframe标志位
    bool mbAcceptKeyFrames;
    std::mutex mMutexAccept;
};

3.源文件

3.1. Run()

void LocalMapping::Run()
{

    mbFinished = false;

    while(1)
    {
        // Tracking will see that Local Mapping is busy
	// 告诉Tracking，LocalMapping正处于繁忙状态
	// LocalMapping线程处理的关键帧都是Tracking线程发过的
        // 在LocalMapping线程还没有处理完关键帧之前Tracking线程最好不要发送太快
        SetAcceptKeyFrames(false);

        // Check if there are keyframes in the queue
	//返回mlNewKeyFrames是否为空，也就是查询等待处理的关键帧列表是否空
        if(CheckNewKeyFrames())
        {
            // BoW conversion and insertion in Map
	    // 计算关键帧特征点的BoW映射，将关键帧插入地图
            ProcessNewKeyFrame();

            // Check recent MapPoints
	    // 剔除ProcessNewKeyFrame函数中引入的不合格MapPoints
            MapPointCulling();

            // Triangulate new MapPoints
	    /** 
	    * 1.找出和当前关键帧共视程度前10/20（单目/双目RGBD）的关键帧
	    * 2.计算这些关键帧和当前关键帧的F矩阵；
	    * 3.在满足对级约束条件下，匹配关键帧之间的特征点，通过BOW加速匹配；
	    */
            CreateNewMapPoints();

	    // 如果已经处理完队列中的最后的一个关键帧
            if(!CheckNewKeyFrames())
            {
                // Find more matches in neighbor keyframes and fuse point duplications
		// 检查并融合当前关键帧与相邻帧（两级相邻）重复的MapPoints
                SearchInNeighbors();
            }

            mbAbortBA = false;

	    // 已经处理完队列中的最后的一个关键帧，并且闭环检测此时没有请求停止LocalMapping
	    //有空就来次local BA
            if(!CheckNewKeyFrames() && !stopRequested())
            {
                // Local BA
                if(mpMap->KeyFramesInMap()>2)
                    Optimizer::LocalBundleAdjustment(mpCurrentKeyFrame,&mbAbortBA, mpMap);

                // Check redundant local Keyframes
		// 检测并剔除当前帧相邻的关键帧中冗余的关键帧
                // 剔除的标准是：该关键帧的90%的MapPoints可以被其它至少3个关键帧观测到
                // trick! 
                // Tracking中先把关键帧交给LocalMapping线程
                // 并且在Tracking中InsertKeyFrame函数的条件比较松，交给LocalMapping线程的关键帧会比较密
                // 在这里再删除冗余的关键帧
                KeyFrameCulling();
            }
	    // 将当前帧加入到闭环检测关键帧队列中
            mpLoopCloser->InsertKeyFrame(mpCurrentKeyFrame);
        }
        else if(Stop())
        {
            // Safe area to stop
            while(isStopped() && !CheckFinish())
            {
                usleep(3000);
            }
            if(CheckFinish())
                break;
        }

        ResetIfRequested();

        // Tracking will see that Local Mapping is not busy
	//表示现在可以接受插入新的keyframe了
        SetAcceptKeyFrames(true);

        if(CheckFinish())
            break;

        usleep(3000);
    }

    SetFinish();
}

3.2. ProcessNewKeyFrame()

/**
* @brief 处理列表中的关键帧
* 
* - 计算Bow，加速三角化新的MapPoints
* - 关联当前关键帧至MapPoints，并更新MapPoints的平均观测方向和观测距离范围
* - 插入关键帧，更新Covisibility图和Essential图，以及spanningtree
*/

void LocalMapping::ProcessNewKeyFrame()
{
    // 从缓冲队列中取出一帧关键帧
    // Tracking线程向LocalMapping中插入关键帧存在该队列中
    {
        unique_lock<mutex> lock(mMutexNewKFs);
        mpCurrentKeyFrame = mlNewKeyFrames.front();
        mlNewKeyFrames.pop_front();
    }

    // Compute Bags of Words structures
    //计算此关键帧的mBowVec，mFeatVec
    mpCurrentKeyFrame->ComputeBoW();

    // Associate MapPoints to the new keyframe and update normal and descriptor
    //遍历和关键帧匹配的mappoint
    const vector<MapPoint*> vpMapPointMatches = mpCurrentKeyFrame->GetMapPointMatches();

    for(size_t i=0; i<vpMapPointMatches.size(); i++)
    {
        MapPoint* pMP = vpMapPointMatches[i];
        if(pMP)
        {
            if(!pMP->isBad())
            {
		//如果pMP是不在关键帧mpCurrentKeyFrame中，则更新mappoint的属性
                if(!pMP->IsInKeyFrame(mpCurrentKeyFrame))
                {
		    //让mappoint知道自己可以被哪些keyframe看到
                    pMP->AddObservation(mpCurrentKeyFrame, i);
		    //更新此mappoint参考帧光心到mappoint平均观测方向以及观测距离范围
                    pMP->UpdateNormalAndDepth();
		    //在此mappoint能被看到的特征点中找出最能代表此mappoint的描述子
                    pMP->ComputeDistinctiveDescriptors();
                }
                else // this can only happen for new stereo points inserted by the Tracking
                {
		    // 当前帧生成的MapPoints
                    // 将双目或RGBD跟踪过程中新插入的MapPoints放入mlpRecentAddedMapPoints，等待MapPointCulling函数检查
                    // CreateNewMapPoints函数中通过三角化也会生成MapPoints
                    // 这些MapPoints都会经过MapPointCulling函数的检验
                    mlpRecentAddedMapPoints.push_back(pMP);
                }
            }
        }
    }    

    // Update links in the Covisibility Graph
    //更新共视图Covisibility graph,spanningtree
    mpCurrentKeyFrame->UpdateConnections();

    // Insert Keyframe in Map
    //将该关键帧插入到地图map中
    mpMap->AddKeyFrame(mpCurrentKeyFrame);
}

3.3. MapPointCulling()

// 剔除ProcessNewKeyFrame和CreateNewMapPoints函数中引入在mlpRecentAddedMapPoints的质量不好的MapPoints

void LocalMapping::MapPointCulling()
{
    // Check Recent Added MapPoints
    list<MapPoint*>::iterator lit = mlpRecentAddedMapPoints.begin();
    const unsigned long int nCurrentKFid = mpCurrentKeyFrame->mnId;

    int nThObs;
    if(mbMonocular)
        nThObs = 2;
    else
        nThObs = 3;
    const int cnThObs = nThObs;

    //遍历mlpRecentAddedMapPoints
    while(lit!=mlpRecentAddedMapPoints.end())
    {
        MapPoint* pMP = *lit;
	// 已经是坏点的MapPoints直接从检查链表中删除
        if(pMP->isBad())
        {
            lit = mlpRecentAddedMapPoints.erase(lit);
        }
	// 跟踪到该MapPoint的Frame数相比预计可观测到该MapPoint的Frame数的比例需大于25%
	// IncreaseFound / IncreaseVisible < 25%，注意不一定是关键帧。
        else if(pMP->GetFoundRatio()<0.25f )
        {
            pMP->SetBadFlag();
            lit = mlpRecentAddedMapPoints.erase(lit);
        }
	// VI-B 条件2：从该点建立开始，到现在已经过了不小于2个关键帧
	// 但是观测到该点的关键帧数却不超过cnThObs帧，那么该点检验不合格
        else if(((int)nCurrentKFid-(int)pMP->mnFirstKFid)>=2 && pMP->Observations()<=cnThObs)
        {
            pMP->SetBadFlag();
            lit = mlpRecentAddedMapPoints.erase(lit);
        }
        // 步骤4：从建立该点开始，已经过了3个关键帧而没有被剔除，则认为是质量高的点
	// 因此没有SetBadFlag()，仅从队列中删除，放弃继续对该MapPoint的检测
        else if(((int)nCurrentKFid-(int)pMP->mnFirstKFid)>=3)
            lit = mlpRecentAddedMapPoints.erase(lit);
        else
            lit++;
    }
}

3.4. CreateNewMapPoints()

    /** 
    * 1.找出和当前关键帧共视程度前10/20（单目/双目RGBD）的关键帧
    * 2.计算这些关键帧和当前关键帧的F矩阵；
    * 3.在满足对级约束条件下，匹配关键帧之间的特征点，通过BOW加速匹配；
    */

void LocalMapping::CreateNewMapPoints()
{
    // Retrieve neighbor keyframes in covisibility graph
    int nn = 10;
    if(mbMonocular)
        nn=20;
    //在当前关键帧的Covisibility graph中找到共视程度最高的nn帧相邻帧存入vpNeighKFs
    const vector<KeyFrame*> vpNeighKFs = mpCurrentKeyFrame->GetBestCovisibilityKeyFrames(nn);

    ORBmatcher matcher(0.6,false);

    //当前关键帧在世界坐标系下的位姿
    cv::Mat Rcw1 = mpCurrentKeyFrame->GetRotation();
    cv::Mat Rwc1 = Rcw1.t();
    cv::Mat tcw1 = mpCurrentKeyFrame->GetTranslation();
    cv::Mat Tcw1(3,4,CV_32F);
    Rcw1.copyTo(Tcw1.colRange(0,3));
    tcw1.copyTo(Tcw1.col(3));
    // 得到当前关键帧光心在世界坐标系中的坐标
    cv::Mat Ow1 = mpCurrentKeyFrame->GetCameraCenter();

    const float &fx1 = mpCurrentKeyFrame->fx;
    const float &fy1 = mpCurrentKeyFrame->fy;
    const float &cx1 = mpCurrentKeyFrame->cx;
    const float &cy1 = mpCurrentKeyFrame->cy;
    const float &invfx1 = mpCurrentKeyFrame->invfx;
    const float &invfy1 = mpCurrentKeyFrame->invfy;

    const float ratioFactor = 1.5f*mpCurrentKeyFrame->mfScaleFactor;

    int nnew=0;

    // Search matches with epipolar restriction and triangulate
    //遍历相邻关键帧vpNeighKFs
    for(size_t i=0; i<vpNeighKFs.size(); i++)
    {
        if(i>0 && CheckNewKeyFrames())
            return;

        KeyFrame* pKF2 = vpNeighKFs[i];

        // Check first that baseline is not too short
	// 邻接的关键帧在世界坐标系中的坐标
        cv::Mat Ow2 = pKF2->GetCameraCenter();
	// 基线向量，两个关键帧间的相机转移矩阵
        cv::Mat vBaseline = Ow2-Ow1;
	// 基线长度，基线长度
        const float baseline = cv::norm(vBaseline);

	// 判断相机运动的基线是不是足够长
        if(!mbMonocular)
        {
	    // 如果是立体相机，关键帧间距太小时不生成3D点
            if(baseline<pKF2->mb)
            continue;
        }
        else
        {
	    // 邻接关键帧的场景深度中值
            const float medianDepthKF2 = pKF2->ComputeSceneMedianDepth(2);
	    // baseline与景深的比例
            const float ratioBaselineDepth = baseline/medianDepthKF2;
	  
	    // 如果特别远(比例特别小)，那么不考虑当前邻接的关键帧，不生成3D点
            if(ratioBaselineDepth<0.01)
                continue;
        }

        // Compute Fundamental Matrix
        // 根据两个关键帧的位姿计算它们之间的基本矩阵
        cv::Mat F12 = ComputeF12(mpCurrentKeyFrame,pKF2);

        // Search matches that fullfil epipolar constraint
	// 通过极线约束限制匹配时的搜索范围，进行特征点匹配
        vector<pair<size_t,size_t> > vMatchedIndices;
	// 匹配pKF1与pKF2之间未被匹配的特征点并通过bow加速，并校验是否符合对级约束。vMatchedPairs匹配成功的特征点在各自关键帧中的id。
        matcher.SearchForTriangulation(mpCurrentKeyFrame,pKF2,F12,vMatchedIndices,false);

        cv::Mat Rcw2 = pKF2->GetRotation();
        cv::Mat Rwc2 = Rcw2.t();
        cv::Mat tcw2 = pKF2->GetTranslation();
        cv::Mat Tcw2(3,4,CV_32F);
        Rcw2.copyTo(Tcw2.colRange(0,3));
        tcw2.copyTo(Tcw2.col(3));

        const float &fx2 = pKF2->fx;
        const float &fy2 = pKF2->fy;
        const float &cx2 = pKF2->cx;
        const float &cy2 = pKF2->cy;
        const float &invfx2 = pKF2->invfx;
        const float &invfy2 = pKF2->invfy;

        // Triangulate each match
	// 对每对匹配通过三角化生成3D点
        const int nmatches = vMatchedIndices.size();
	//遍历vMatchedIndices
        for(int ikp=0; ikp<nmatches; ikp++)
        {
            const int &idx1 = vMatchedIndices[ikp].first;
            const int &idx2 = vMatchedIndices[ikp].second;

            const cv::KeyPoint &kp1 = mpCurrentKeyFrame->mvKeysUn[idx1];
            const float kp1_ur=mpCurrentKeyFrame->mvuRight[idx1];
            bool bStereo1 = kp1_ur>=0;

            const cv::KeyPoint &kp2 = pKF2->mvKeysUn[idx2];
            const float kp2_ur = pKF2->mvuRight[idx2];
            bool bStereo2 = kp2_ur>=0;

            // Check parallax between rays
	    //将像素坐标转化为归一化平面坐标
            cv::Mat xn1 = (cv::Mat_<float>(3,1) << (kp1.pt.x-cx1)*invfx1, (kp1.pt.y-cy1)*invfy1, 1.0);
            cv::Mat xn2 = (cv::Mat_<float>(3,1) << (kp2.pt.x-cx2)*invfx2, (kp2.pt.y-cy2)*invfy2, 1.0);

            cv::Mat ray1 = Rwc1*xn1;
            cv::Mat ray2 = Rwc2*xn2;
            const float cosParallaxRays = ray1.dot(ray2)/(cv::norm(ray1)*cv::norm(ray2));

            float cosParallaxStereo = cosParallaxRays+1;
            float cosParallaxStereo1 = cosParallaxStereo;
            float cosParallaxStereo2 = cosParallaxStereo;

            if(bStereo1)
                cosParallaxStereo1 = cos(2*atan2(mpCurrentKeyFrame->mb/2,mpCurrentKeyFrame->mvDepth[idx1]));
            else if(bStereo2)
                cosParallaxStereo2 = cos(2*atan2(pKF2->mb/2,pKF2->mvDepth[idx2]));

            cosParallaxStereo = min(cosParallaxStereo1,cosParallaxStereo2);

            cv::Mat x3D;
            if(cosParallaxRays<cosParallaxStereo && cosParallaxRays>0 && (bStereo1 || bStereo2 || cosParallaxRays<0.9998))
            {
                // Linear Triangulation Method
                cv::Mat A(4,4,CV_32F);
                A.row(0) = xn1.at<float>(0)*Tcw1.row(2)-Tcw1.row(0);
                A.row(1) = xn1.at<float>(1)*Tcw1.row(2)-Tcw1.row(1);
                A.row(2) = xn2.at<float>(0)*Tcw2.row(2)-Tcw2.row(0);
                A.row(3) = xn2.at<float>(1)*Tcw2.row(2)-Tcw2.row(1);

                cv::Mat w,u,vt;
		//SVD分解A=u*w*vt
		//实际是要解Ax=0
                cv::SVD::compute(A,w,u,vt,cv::SVD::MODIFY_A| cv::SVD::FULL_UV);

                x3D = vt.row(3).t();

                if(x3D.at<float>(3)==0)
                    continue;

                // Euclidean coordinates
		//归一化
                x3D = x3D.rowRange(0,3)/x3D.at<float>(3);

            }
            else if(bStereo1 && cosParallaxStereo1<cosParallaxStereo2)
            {
                x3D = mpCurrentKeyFrame->UnprojectStereo(idx1);                
            }
            else if(bStereo2 && cosParallaxStereo2<cosParallaxStereo1)
            {
                x3D = pKF2->UnprojectStereo(idx2);
            }
            else
                continue; //No stereo and very low parallax

            cv::Mat x3Dt = x3D.t();

            //Check triangulation in front of cameras
	    //检验3d点投影到第一个关键帧的误差
            float z1 = Rcw1.row(2).dot(x3Dt)+tcw1.at<float>(2);
            if(z1<=0)
                continue;

            float z2 = Rcw2.row(2).dot(x3Dt)+tcw2.at<float>(2);
            if(z2<=0)
                continue;

            //Check reprojection error in first keyframe
            const float &sigmaSquare1 = mpCurrentKeyFrame->mvLevelSigma2[kp1.octave];
            const float x1 = Rcw1.row(0).dot(x3Dt)+tcw1.at<float>(0);
            const float y1 = Rcw1.row(1).dot(x3Dt)+tcw1.at<float>(1);
            const float invz1 = 1.0/z1;

            if(!bStereo1)
            {
                float u1 = fx1*x1*invz1+cx1;
                float v1 = fy1*y1*invz1+cy1;
                float errX1 = u1 - kp1.pt.x;
                float errY1 = v1 - kp1.pt.y;
                if((errX1*errX1+errY1*errY1)>5.991*sigmaSquare1)
                    continue;
            }
            else
            {
                float u1 = fx1*x1*invz1+cx1;
                float u1_r = u1 - mpCurrentKeyFrame->mbf*invz1;
                float v1 = fy1*y1*invz1+cy1;
                float errX1 = u1 - kp1.pt.x;
                float errY1 = v1 - kp1.pt.y;
                float errX1_r = u1_r - kp1_ur;
                if((errX1*errX1+errY1*errY1+errX1_r*errX1_r)>7.8*sigmaSquare1)
                    continue;
            }

            //Check reprojection error in second keyframe
            //检验3d点投影到第二个关键帧的误差
            const float sigmaSquare2 = pKF2->mvLevelSigma2[kp2.octave];
            const float x2 = Rcw2.row(0).dot(x3Dt)+tcw2.at<float>(0);
            const float y2 = Rcw2.row(1).dot(x3Dt)+tcw2.at<float>(1);
            const float invz2 = 1.0/z2;
            if(!bStereo2)
            {
                float u2 = fx2*x2*invz2+cx2;
                float v2 = fy2*y2*invz2+cy2;
                float errX2 = u2 - kp2.pt.x;
                float errY2 = v2 - kp2.pt.y;
                if((errX2*errX2+errY2*errY2)>5.991*sigmaSquare2)
                    continue;
            }
            else
            {
                float u2 = fx2*x2*invz2+cx2;
                float u2_r = u2 - mpCurrentKeyFrame->mbf*invz2;
                float v2 = fy2*y2*invz2+cy2;
                float errX2 = u2 - kp2.pt.x;
                float errY2 = v2 - kp2.pt.y;
                float errX2_r = u2_r - kp2_ur;
                if((errX2*errX2+errY2*errY2+errX2_r*errX2_r)>7.8*sigmaSquare2)
                    continue;
            }

            //Check scale consistency
            //检验尺度的连续性
            cv::Mat normal1 = x3D-Ow1;
            float dist1 = cv::norm(normal1);

            cv::Mat normal2 = x3D-Ow2;
            float dist2 = cv::norm(normal2);

            if(dist1==0 || dist2==0)
                continue;

            const float ratioDist = dist2/dist1;
            const float ratioOctave = mpCurrentKeyFrame->mvScaleFactors[kp1.octave]/pKF2->mvScaleFactors[kp2.octave];

            /*if(fabs(ratioDist-ratioOctave)>ratioFactor)
                continue;*/
            if(ratioDist*ratioFactor<ratioOctave || ratioDist>ratioOctave*ratioFactor)
                continue;

            // Triangulation is succesfull
	    //如果三角化成功
            MapPoint* pMP = new MapPoint(x3D,mpCurrentKeyFrame,mpMap);

            pMP->AddObservation(mpCurrentKeyFrame,idx1);            
            pMP->AddObservation(pKF2,idx2);

            mpCurrentKeyFrame->AddMapPoint(pMP,idx1);
            pKF2->AddMapPoint(pMP,idx2);

            pMP->ComputeDistinctiveDescriptors();

            pMP->UpdateNormalAndDepth();

            mpMap->AddMapPoint(pMP);
	    //新添加的点要加入这里，然后让MapPointCulling()检测
            mlpRecentAddedMapPoints.push_back(pMP);

            nnew++;
        }
    }
}

3.5. SearchInNeighbors()

检查并融合与当前关键帧共视程度高的帧重复的MapPoints

void LocalMapping::SearchInNeighbors()
{
    // Retrieve neighbor keyframes
    int nn = 10;
    if(mbMonocular)
        nn=20;
    //获取当前关键帧与当前关键帧共视程度高的帧放入vpNeighKFs，并标记
    const vector<KeyFrame*> vpNeighKFs = mpCurrentKeyFrame->GetBestCovisibilityKeyFrames(nn);
    vector<KeyFrame*> vpTargetKFs;
    //遍历vpNeighKFs，遍历当前帧essential graph图中权重排名前nn的邻接关键帧
    for(vector<KeyFrame*>::const_iterator vit=vpNeighKFs.begin(), vend=vpNeighKFs.end(); vit!=vend; vit++)
    {
        KeyFrame* pKFi = *vit;
	//如果这个keyframe是坏的，或者已经被当前关键帧标记了
        if(pKFi->isBad() || pKFi->mnFuseTargetForKF == mpCurrentKeyFrame->mnId)
            continue;
        vpTargetKFs.push_back(pKFi);
        pKFi->mnFuseTargetForKF = mpCurrentKeyFrame->mnId;

        // Extend to some second neighbors
	//通过获取与pKFi共视程度高的关键帧来拓展vpNeighKFs
        const vector<KeyFrame*> vpSecondNeighKFs = pKFi->GetBestCovisibilityKeyFrames(5);
        for(vector<KeyFrame*>::const_iterator vit2=vpSecondNeighKFs.begin(), vend2=vpSecondNeighKFs.end(); vit2!=vend2; vit2++)
        {
            KeyFrame* pKFi2 = *vit2;
            if(pKFi2->isBad() || pKFi2->mnFuseTargetForKF==mpCurrentKeyFrame->mnId || pKFi2->mnId==mpCurrentKeyFrame->mnId)
                continue;
            vpTargetKFs.push_back(pKFi2);
        }
    }


    // Search matches by projection from current KF in target KFs
    ORBmatcher matcher;
    //取出当前关键帧的mappoint与pKFi中的mappoint融合
    vector<MapPoint*> vpMapPointMatches = mpCurrentKeyFrame->GetMapPointMatches();
    //遍历vpTargetKFs中的keyframe
    for(vector<KeyFrame*>::iterator vit=vpTargetKFs.begin(), vend=vpTargetKFs.end(); vit!=vend; vit++)
    {
        KeyFrame* pKFi = *vit;

	//将vpMapPoints中的mappoint与pKF的特征点进行匹配，若匹配的特征点已有mappoint与其匹配，
	//则选择其一与此特征点匹配，并抹去没有选择的那个mappoint，此为mappoint的融合
        matcher.Fuse(pKFi,vpMapPointMatches);
    }

    // Search matches by projection from target KFs in current KF
    //获取vpTargetKFs所有良好MapPoints的集合
    vector<MapPoint*> vpFuseCandidates;
    vpFuseCandidates.reserve(vpTargetKFs.size()*vpMapPointMatches.size());

    for(vector<KeyFrame*>::iterator vitKF=vpTargetKFs.begin(), vendKF=vpTargetKFs.end(); vitKF!=vendKF; vitKF++)
    {
        KeyFrame* pKFi = *vitKF;

        vector<MapPoint*> vpMapPointsKFi = pKFi->GetMapPointMatches();

        for(vector<MapPoint*>::iterator vitMP=vpMapPointsKFi.begin(), vendMP=vpMapPointsKFi.end(); vitMP!=vendMP; vitMP++)
        {
            MapPoint* pMP = *vitMP;
            if(!pMP)
                continue;
            if(pMP->isBad() || pMP->mnFuseCandidateForKF == mpCurrentKeyFrame->mnId)
                continue;
            pMP->mnFuseCandidateForKF = mpCurrentKeyFrame->mnId;
            vpFuseCandidates.push_back(pMP);
        }
    }

    //mappoint融合
    matcher.Fuse(mpCurrentKeyFrame,vpFuseCandidates);


    // Update points
    //更新mappoint属性
    vpMapPointMatches = mpCurrentKeyFrame->GetMapPointMatches();
    for(size_t i=0, iend=vpMapPointMatches.size(); i<iend; i++)
    {
        MapPoint* pMP=vpMapPointMatches[i];
        if(pMP)
        {
            if(!pMP->isBad())
            {
		//在此mappoint能被看到的特征点中找出最能代表此mappoint的描述子
                pMP->ComputeDistinctiveDescriptors();
		// 更新平均观测方向和观测距离
                pMP->UpdateNormalAndDepth();
            }
        }
    }

    // Update connections in covisibility graph
    //更新共视图Covisibility graph,essential graph和spanningtree
    mpCurrentKeyFrame->UpdateConnections();
}

3.6. KeyFrameCulling()

// 检测并剔除当前帧相邻的关键帧中冗余的关键帧
// 剔除的标准是：该关键帧的90%的MapPoints可以被其它至少3个关键帧观测到

void LocalMapping::KeyFrameCulling()
{
    // Check redundant keyframes (only local keyframes)
    // A keyframe is considered redundant if the 90% of the MapPoints it sees, are seen
    // in at least other 3 keyframes (in the same or finer scale)
    // We only consider close stereo points
  
    //返回essential graph中与此keyframe连接的节点（即关键帧）
    vector<KeyFrame*> vpLocalKeyFrames = mpCurrentKeyFrame->GetVectorCovisibleKeyFrames();

    //遍历vpLocalKeyFrames
    for(vector<KeyFrame*>::iterator vit=vpLocalKeyFrames.begin(), vend=vpLocalKeyFrames.end(); vit!=vend; vit++)
    {
        KeyFrame* pKF = *vit;
        if(pKF->mnId==0)
            continue;
	// 提取每个共视关键帧的MapPoints
        const vector<MapPoint*> vpMapPoints = pKF->GetMapPointMatches();

        int nObs = 3;
        const int thObs=nObs;
        int nRedundantObservations=0;
        int nMPs=0;
	//遍历该局部关键帧pKF的MapPoints，判断是否90%以上的MapPoints能被其它关键帧（至少3个）观测到
        for(size_t i=0, iend=vpMapPoints.size(); i<iend; i++)
        {
            MapPoint* pMP = vpMapPoints[i];
            if(pMP)
            {
                if(!pMP->isBad())
                {
                    if(!mbMonocular)
                    {
                        if(pKF->mvDepth[i]>pKF->mThDepth || pKF->mvDepth[i]<0)
                            continue;
                    }

                    nMPs++;
		    //如果pMP被其他keyframe观测的数量大于thObs
                    if(pMP->Observations()>thObs)
                    {
                        const int &scaleLevel = pKF->mvKeysUn[i].octave;
                        const map<KeyFrame*, size_t> observations = pMP->GetObservations();
                        int nObs=0;
                        for(map<KeyFrame*, size_t>::const_iterator mit=observations.begin(), mend=observations.end(); mit!=mend; mit++)
                        {
                            KeyFrame* pKFi = mit->first;
                            if(pKFi==pKF)
                                continue;
                            const int &scaleLeveli = pKFi->mvKeysUn[mit->second].octave;

                            if(scaleLeveli<=scaleLevel+1)
                            {
                                nObs++;
                                if(nObs>=thObs)
                                    break;
                            }
                        }
                        if(nObs>=thObs)
                        {
                            nRedundantObservations++;
                        }
                    }
                }
            }
        }  

        if(nRedundantObservations>0.9*nMPs)
            pKF->SetBadFlag();
    }
}