关于ICE-BA的结构图.pdf

发布时间：2022-05-31 发布人：admin 分类：说明书资料大小：0.66M 资料格式：pdf 举报版权申诉

weixin_31823677-10969980-4744302542914837150.pdf-第1页.png

第1页 / 共12页

weixin_31823677-10969980-4744302542914837150.pdf-第2页.png

第2页 / 共12页

weixin_31823677-10969980-4744302542914837150.pdf-第3页.png

第3页 / 共12页

weixin_31823677-10969980-4744302542914837150.pdf-第4页.png

第4页 / 共12页

weixin_31823677-10969980-4744302542914837150.pdf-第5页.png

第5页 / 共12页

weixin_31823677-10969980-4744302542914837150.pdf-第6页.png

第6页 / 共12页

weixin_31823677-10969980-4744302542914837150.pdf-第7页.png

第7页 / 共12页

weixin_31823677-10969980-4744302542914837150.pdf-第8页.png

第8页 / 共12页

文本预览

ICE-BA架构总结 Backend App main() Config param.txt BundleAdjustment GlobalBundleAdjustor LocalBundleAdjustor Geometry Camera: info, intrinsic, depth IMU Coordinate Trans Map Feature Frame LocalMap GlobalMap Utility Aligned Vector/Matrix IBA (Incremental Bundle Adjustment) SIMD IBA: info, keyframe, point, … etc Parameter Set LinearAlgebra Matrix MXN.h VectorN.h MultiThread Timer …… Visualization Frontend cameras Camera Base Distortion Feature track & operation ORB feature pyramid quaternion timer parameter load draw and view

new image & imu 均值滤波3X3 main() build_img_pyramids build_pyramids fast_pyra_down 快速建⽴4层⾦字塔, ratio = 2 IntegrateQuaternion optical_flow_and_detect 仅预积分四元数部分光流跟踪old feature并新增new feature, 在key_pnts中更新 slave_img_feat_propagator.PropagateFeatures 寻找Right Image对应feature, 在key_pnts_slave中更新 create_iba_frame PushCurrentFrame PushCurrentFrame 建⽴ new CF, 更新 mp measurement zs; 若超过20个new feature被发现或者跟踪到的feature总数⼩于20，创建KF; 建⽴新KF, 更新zs，push_back new MP to Xs; 更新m_ILF，预处理光流跟踪的MP并转化为m_ILF的m_zs, m_Zs imu预处理并更新m_us， PushKeyFrame 更新m_IKF，预处理new MP并转化为m_IKF的GlobalMap:Point m_Xs 更新所有index关联信息，m_xs、m_ds信息 m_LM.IBA_PushCurrentFrame 更新IBA中的m_CsLF, m_CsKF, m_iKF2d, m_ds，为新帧处理做准备 display m_LBA.WakeUp 唤醒LocalBA线程

optical_flow_and_detect() 标记所有feature点为未跟踪 mark_all_feature_tracks_dead () 光流跟踪 propagate_with_optical_flow () 在⾦字塔level 1进⾏跟踪，计算harris response并反畸变, 选择response⾼的strong features 判断是否需要再检测新feature点判断条件：特征点过于集中；特征点过少 XPTrackerInvoker() 光流法跟踪feature XPcalcOpticalFlowPyrLK() redo 若为true，检测新feature点 detect->detect_orb_features() 在level1层⾦字塔,利⽤fast在25*25⽅格内寻找feature，⾼斯平滑，之后⽤harris 重新计算response，根据光流移动dist排除outlier点 filter_outlier_flows() 清除未被跟踪到的feature点 feature_tracks_map_.erase(ft_it) 排序并排除过于靠近的feature点 XpEnforceKeyPointUniformity() 若需要orb描述⼦，则计算feature⽅向 computeOrientation() 计算前后帧亮度变化特征响应排序并排除分布集中或重合在⼀起的feature点; 在level0根据⼩范围response做加权平均修正feature; refine_kp_in_larger_img() 滤除前后帧位置变化过⼤feature; 将⻓时间跟踪的feature更新为new feature; 更新feature track状态在level1 计算response并排除weak feature; 计算orb描述⼦计算orb描述⼦

Incremental, Consistent and Efficient Bundle Adjustment (ICE-BA) Global BA: 并⾏优化所有被选为keyframe的帧 Local BA: 并⾏优化在sliding window中的50个local frame

Cost Function: where we solve δφ by normal equation: Thus, we incrementally update: where By Schur complement, we divide δφ into δφc and δφp We efficiently compute S & s corresponding to (i1, i2) camera pair: We only update Sij & si which has changed after last iter: Then by PCG, we solve δφc, back-substituting δφc, δφp solved:

An improvement for Local BA: To make S sparser and keep the number of features to be re-liearization smaller, we split the origin long feature track into several short overlapping sub-tracks. we solving Schur complement with this sub-tracks point and finally update 3D feature points with origin point. Result: 2~10 times faster than original IBA. Incremental PCG for IBA We use PCG to iterately solve function above. For one frame, if δφc is not large enough, the result of the next iteration δφc+ will be very close to the presions one δφc-. To accelerate convergence of PCG, we initialize these kinds of frames δφc+ = δφc- instead of δφc+ =0. Result: 20% due to better convergence. Relative Marginalization

LocalBundleAdjustor::Run() SynchronizeData() synchronize preintegration, iLF/iKF info, Feature measurement info, push keyframe to GBA init or update new MP/iLF/iKF m_d, LF/KF matching info, MP tracking info, ST info, many u while (! max_iteration) UpdateFactors() update vision/imu/prior factors UpdateSchurComplement() 更新与frame pos相关的Schur部分 SolveSchurComplement() 求解Schur Complement SolveBackSubstitution() SolveGradientDescent() SolveDogLeg() UpdateStatesPropose() ComputeReduction() UpdateStatesDecide() EmbeddedPointIteration() UpdateData() m_callback()

preintegration, m_C updated 判断new MP的m_d是否可⽤，通常不可⽤ while (!new frame/keyframe .empty()) resize and update m_ic2LF find measurement matches in last ST and update ST in these pre_frames and corresponding first find KF if frame if keyframe create new ST 更新ILF的m_d为其所有 feature点深度的加权平均根据相同的 frame，更新IKF的参数 SearchMatchingKeyFrames() SearchMatchingKeyFrames() update preintegration delta D MakeZero PushKeyFrame() resize and init IKF, update us 寻找关联最紧密的 keyframe, update m_iKFNearest PushLocalFrame() m_GBA->WakeUp() 将SlidingWindow中所有frame的坐标转换到Nearest KeyFrame坐标系更新相关的us 若为同时的LF, update LF new MP, add new ST; 若为其他LF, 更新两者之间的match关系 m_GBA->PushKeyFrame()

分享到：

赞收藏

资料库

关于ICE-BA的结构图.pdf

相关推荐

人工智能

热门标签

最新资料