移动机器人激光SLAM导航（五）：Cartographer SLAM 篇

参考

• Cartographer 官方文档
• Cartographer 从入门到精通

1. Cartographer 安装

1.1 前置条件

• 推荐在刚装好的 Ubuntu 16.04 或 Ubuntu 18.04 上进行编译
• ROS 安装：ROS学习1：ROS概述与环境搭建

1.2 依赖库安装

• 资源下载完解压并执行以下指令
  • https://pan.baidu.com/s/1LWqZ4SOKn2sZecQUDDXXEw?pwd=j6cf

  $ sudo chmod 777 auto-carto-build.sh
  $ ./auto-carto-build.sh
  

1.3 编译

本文只编译 cartographer_ros，以下为同时开三个终端操作

$ mkdir -p cartographer_ws/src

$ cd ~
$ git clone https://github.com/xiangli0608/cartographer_detailed_comments_ws
$ mv ~/cartographer_detailed_comments_ws/src/cartographer_ros ~/cartographer_ws/src

$ cd ~/cartographer_ws
$ catkin_make

2. Cartographer 运行

2.1 数据集下载

• 百度网盘链接（rslidar-outdoor-gps.bag、landmarks_demo_uncalibrated.bag）
  • https://pan.baidu.com/s/1leRr4JDGg61jBNCwNlSCJw?pwd=5nkq

2.2 配置文件

• lx_rs16_2d_outdoor.launch

  <launch><!-- bag 的地址与名称（根据自己情况修改，建议使用绝对路径） --><arg name="bag_filename" default="/home/yue/bag/rslidar-outdoor-gps.bag"/><!-- 使用 bag 的时间戳 --><param name="/use_sim_time" value="true" /><!-- 启动 cartographer --><node name="cartographer_node" pkg="cartographer_ros"type="cartographer_node" args="-configuration_directory $(find cartographer_ros)/configuration_files-configuration_basename lx_rs16_2d_outdoor.lua"output="screen"><remap from="points2" to="rslidar_points" /><remap from="scan" to="front_scan" /><remap from="odom" to="odom_scout" /><remap from="imu" to="imu" /></node><!-- 生成 ros 格式的地图 --><node name="cartographer_occupancy_grid_node" pkg="cartographer_ros"type="cartographer_occupancy_grid_node" args="-resolution 0.05" /><!-- 启动 rviz --><node name="rviz" pkg="rviz" type="rviz" required="true"args="-d $(find cartographer_ros)/configuration_files/lx_2d.rviz" /><!-- 启动 rosbag --><node name="playbag" pkg="rosbag" type="play"args="--clock $(arg bag_filename)" /></launch>
  

• lx_rs16_2d_outdoor.lua

  include "map_builder.lua"
  include "trajectory_builder.lua"options = {map_builder = MAP_BUILDER，               -- map_builder.lua的配置信息trajectory_builder = TRAJECTORY_BUILDER， -- trajectory_builder.lua的配置信息map_frame = "map"，                       -- 地图坐标系的名字tracking_frame = "imu_link"，             -- 将所有传感器数据转换到这个坐标系下published_frame = "footprint"，           -- tf: map -> footprintodom_frame = "odom"，                     -- 里程计的坐标系名字provide_odom_frame = false，              -- 是否提供odom的tf，如果为true,则tf树为map->odom->footprint-- 如果为false tf树为map->footprintpublish_frame_projected_to_2d = false，   -- 是否将坐标系投影到平面上--use_pose_extrapolator = false，           -- 发布tf时是使用pose_extrapolator的位姿还是前端计算出来的位姿use_odometry = false，                    -- 是否使用里程计,如果使用要求一定要有odom的tfuse_nav_sat = false，                     -- 是否使用gpsuse_landmarks = false，                   -- 是否使用landmarknum_laser_scans = 0，                     -- 是否使用单线激光数据num_multi_echo_laser_scans = 0，          -- 是否使用multi_echo_laser_scans数据num_subdivisions_per_laser_scan = 1，     -- 1帧数据被分成几次处理,一般为1num_point_clouds = 1，                    -- 是否使用点云数据lookup_transform_timeout_sec = 0.2，      -- 查找tf时的超时时间submap_publish_period_sec = 0.3，         -- 发布数据的时间间隔pose_publish_period_sec = 5e-3,trajectory_publish_period_sec = 30e-3,rangefinder_sampling_ratio = 1.，         -- 传感器数据的采样频率odometry_sampling_ratio = 1.,fixed_frame_pose_sampling_ratio = 1.,imu_sampling_ratio = 1.,landmarks_sampling_ratio = 1.,
  }MAP_BUILDER.use_trajectory_builder_2d = trueTRAJECTORY_BUILDER_2D.use_imu_data = true
  TRAJECTORY_BUILDER_2D.min_range = 0.3
  TRAJECTORY_BUILDER_2D.max_range = 100.
  TRAJECTORY_BUILDER_2D.min_z = 0.2
  --TRAJECTORY_BUILDER_2D.max_z = 1.4
  --TRAJECTORY_BUILDER_2D.voxel_filter_size = 0.02--TRAJECTORY_BUILDER_2D.adaptive_voxel_filter.max_length = 0.5
  --TRAJECTORY_BUILDER_2D.adaptive_voxel_filter.min_num_points = 200.
  --TRAJECTORY_BUILDER_2D.adaptive_voxel_filter.max_range = 50.--TRAJECTORY_BUILDER_2D.loop_closure_adaptive_voxel_filter.max_length = 0.9
  --TRAJECTORY_BUILDER_2D.loop_closure_adaptive_voxel_filter.min_num_points = 100
  --TRAJECTORY_BUILDER_2D.loop_closure_adaptive_voxel_filter.max_range = 50.TRAJECTORY_BUILDER_2D.use_online_correlative_scan_matching = false
  TRAJECTORY_BUILDER_2D.ceres_scan_matcher.occupied_space_weight = 1.
  TRAJECTORY_BUILDER_2D.ceres_scan_matcher.translation_weight = 1.
  TRAJECTORY_BUILDER_2D.ceres_scan_matcher.rotation_weight = 1.
  --TRAJECTORY_BUILDER_2D.ceres_scan_matcher.ceres_solver_options.max_num_iterations = 12--TRAJECTORY_BUILDER_2D.motion_filter.max_distance_meters = 0.1
  --TRAJECTORY_BUILDER_2D.motion_filter.max_angle_radians = 0.004
  --TRAJECTORY_BUILDER_2D.imu_gravity_time_constant = 1.TRAJECTORY_BUILDER_2D.submaps.num_range_data = 80.
  TRAJECTORY_BUILDER_2D.submaps.grid_options_2d.resolution = 0.1POSE_GRAPH.optimize_every_n_nodes = 160.
  POSE_GRAPH.constraint_builder.sampling_ratio = 0.3
  POSE_GRAPH.constraint_builder.max_constraint_distance = 15.
  POSE_GRAPH.constraint_builder.min_score = 0.48
  POSE_GRAPH.constraint_builder.global_localization_min_score = 0.60return options
  

• trajectory_builder_2d.lua

  TRAJECTORY_BUILDER_2D = {use_imu_data = true,            -- 是否使用imu数据min_range = 0.,                 -- 雷达数据的最远最近滤波, 保存中间值max_range = 30.,min_z = -0.8,                   -- 雷达数据的最高与最低的过滤, 保存中间值max_z = 2.,missing_data_ray_length = 5.,   -- 超过最大距离范围的数据点用这个距离代替num_accumulated_range_data = 1, -- 几帧有效的点云数据进行一次扫描匹配voxel_filter_size = 0.025,      -- 体素滤波的立方体的边长-- 使用固定的voxel滤波之后, 再使用自适应体素滤波器-- 体素滤波器 用于生成稀疏点云 以进行 扫描匹配adaptive_voxel_filter = {max_length = 0.5,             -- 尝试确定最佳的立方体边长, 边长最大为0.5min_num_points = 200,         -- 如果存在 较多点 并且大于'min_num_points', 则减小体素长度以尝试获得该最小点数max_range = 50.,              -- 距远离原点超过max_range 的点被移除},-- 闭环检测的自适应体素滤波器, 用于生成稀疏点云 以进行 闭环检测loop_closure_adaptive_voxel_filter = {max_length = 0.9,min_num_points = 100,max_range = 50.,},-- 是否使用 real_time_correlative_scan_matcher 为ceres提供先验信息-- 计算复杂度高 , 但是很鲁棒 , 在odom或者imu不准时依然能达到很好的效果use_online_correlative_scan_matching = false,real_time_correlative_scan_matcher = {linear_search_window = 0.1,             -- 线性搜索窗口的大小angular_search_window = math.rad(20.),  -- 角度搜索窗口的大小translation_delta_cost_weight = 1e-1,   -- 用于计算各部分score的权重rotation_delta_cost_weight = 1e-1,},-- ceres匹配的一些配置参数ceres_scan_matcher = {occupied_space_weight = 1.,translation_weight = 10.,rotation_weight = 40.,ceres_solver_options = {use_nonmonotonic_steps = false,max_num_iterations = 20,num_threads = 1,},},-- 为了防止子图里插入太多数据, 在插入子图之前之前对数据进行过滤motion_filter = {max_time_seconds = 5.,max_distance_meters = 0.2,max_angle_radians = math.rad(1.),},-- TODO(schwoere,wohe): Remove this constant. This is only kept for ROS.imu_gravity_time_constant = 10.,-- 位姿预测器pose_extrapolator = {use_imu_based = false,constant_velocity = {imu_gravity_time_constant = 10.,pose_queue_duration = 0.001,},imu_based = {pose_queue_duration = 5.,gravity_constant = 9.806,pose_translation_weight = 1.,pose_rotation_weight = 1.,imu_acceleration_weight = 1.,imu_rotation_weight = 1.,odometry_translation_weight = 1.,odometry_rotation_weight = 1.,solver_options = {use_nonmonotonic_steps = false;max_num_iterations = 10;num_threads = 1;},},},-- 子图相关的一些配置submaps = {num_range_data = 90,          -- 一个子图里插入雷达数据的个数的一半grid_options_2d = {grid_type = "PROBABILITY_GRID", -- 地图的种类, 还可以是tsdf格式resolution = 0.05,},range_data_inserter = {range_data_inserter_type = "PROBABILITY_GRID_INSERTER_2D",-- 概率占用栅格地图的一些配置probability_grid_range_data_inserter = {insert_free_space = true,hit_probability = 0.55,miss_probability = 0.49,},-- tsdf地图的一些配置tsdf_range_data_inserter = {truncation_distance = 0.3,maximum_weight = 10.,update_free_space = false,normal_estimation_options = {num_normal_samples = 4,sample_radius = 0.5,},project_sdf_distance_to_scan_normal = true,update_weight_range_exponent = 0,update_weight_angle_scan_normal_to_ray_kernel_bandwidth = 0.5,update_weight_distance_cell_to_hit_kernel_bandwidth = 0.5,},},},
  }
  

• pose_graph.lua

  POSE_GRAPH = {-- 每隔多少个节点执行一次后端优化optimize_every_n_nodes = 90,-- 约束构建的相关参数constraint_builder = {sampling_ratio = 0.3,                 -- 对局部子图进行回环检测时的计算频率, 数值越大, 计算次数越多max_constraint_distance = 15.,        -- 对局部子图进行回环检测时能成为约束的最大距离min_score = 0.55,                     -- 对局部子图进行回环检测时的最低分数阈值global_localization_min_score = 0.6,  -- 对整体子图进行回环检测时的最低分数阈值loop_closure_translation_weight = 1.1e4,loop_closure_rotation_weight = 1e5,log_matches = true,                   -- 打印约束计算的log-- 基于分支定界算法的2d粗匹配器fast_correlative_scan_matcher = {linear_search_window = 7.,angular_search_window = math.rad(30.),branch_and_bound_depth = 7,},-- 基于ceres的2d精匹配器ceres_scan_matcher = {occupied_space_weight = 20.,translation_weight = 10.,rotation_weight = 1.,ceres_solver_options = {use_nonmonotonic_steps = true,max_num_iterations = 10,num_threads = 1,},},-- 基于分支定界算法的3d粗匹配器fast_correlative_scan_matcher_3d = {branch_and_bound_depth = 8,full_resolution_depth = 3,min_rotational_score = 0.77,min_low_resolution_score = 0.55,linear_xy_search_window = 5.,linear_z_search_window = 1.,angular_search_window = math.rad(15.),},-- 基于ceres的3d精匹配器ceres_scan_matcher_3d = {occupied_space_weight_0 = 5.,occupied_space_weight_1 = 30.,translation_weight = 10.,rotation_weight = 1.,only_optimize_yaw = false,ceres_solver_options = {use_nonmonotonic_steps = false,max_num_iterations = 10,num_threads = 1,},},},matcher_translation_weight = 5e2,matcher_rotation_weight = 1.6e3,-- 优化残差方程的相关参数optimization_problem = {huber_scale = 1e1,                -- 值越大,（潜在）异常值的影响就越大acceleration_weight = 1.1e2,      -- 3d里imu的线加速度的权重rotation_weight = 1.6e4,          -- 3d里imu的旋转的权重-- 前端结果残差的权重local_slam_pose_translation_weight = 1e5,local_slam_pose_rotation_weight = 1e5,-- 里程计残差的权重odometry_translation_weight = 1e5,odometry_rotation_weight = 1e5,-- gps残差的权重fixed_frame_pose_translation_weight = 1e1,fixed_frame_pose_rotation_weight = 1e2,fixed_frame_pose_use_tolerant_loss = false,fixed_frame_pose_tolerant_loss_param_a = 1,fixed_frame_pose_tolerant_loss_param_b = 1,log_solver_summary = false,use_online_imu_extrinsics_in_3d = true,fix_z_in_3d = false,ceres_solver_options = {use_nonmonotonic_steps = false,max_num_iterations = 50,num_threads = 7,},},max_num_final_iterations = 200,   -- 在建图结束之后执行一次全局优化, 不要求实时性, 迭代次数多global_sampling_ratio = 0.003,    -- 纯定位时候查找回环的频率log_residual_histograms = true,global_constraint_search_after_n_seconds = 10., -- 纯定位时多少秒执行一次全子图的约束计算--  overlapping_submaps_trimmer_2d = {--    fresh_submaps_count = 1,--    min_covered_area = 2,--    min_added_submaps_count = 5,--  },
  }
  

2.3 运行演示

$ source devel/setup.bash
$ roslaunch cartographer_ros lx_rs16_2d_outdoor.launch

3. Cartographer 调参总结

3.1 降低延迟与减小计算量

• 前端

  • 减小 max_range 即减小了需要处理的点数，在激光雷达远距离的数据点不准时一定要减小这个值
  • 增大 voxel_filter_size，相当于减小了需要处理的点数
  • 增大 submaps.resolution，相当于减小了匹配时的搜索量
  • 对于自适应体素滤波 减小 min_num_points 与 max_range，增大 max_length，相当于减小了需要处理的点数

• 后端

  • 减小 optimize_every_n_nodes, 降低优化频率, 减小了计算量
  • 增大 MAP_BUILDER.num_background_threads, 增加计算速度
  • 减小 global_sampling_ratio, 减小计算全局约束的频率
  • 减小 constraint_builder.sampling_ratio, 减少了约束的数量
  • 增大 constraint_builder.min_score, 减少了约束的数量
  • 减小分枝定界搜索窗的大小, 包括linear_xy_search_window,inear_z_search_window, angular_search_window
  • 增大 global_constraint_search_after_n_seconds, 减小计算全局约束的频率
  • 减小 max_num_iterations, 减小迭代次数

3.2 降低内存

• 增大子图的分辨率 submaps.resolution

3.3 常调的参数

TRAJECTORY_BUILDER_2D.min_range = 0.3
TRAJECTORY_BUILDER_2D.max_range = 100.
TRAJECTORY_BUILDER_2D.min_z = 0.2 -- / -0.8
TRAJECTORY_BUILDER_2D.voxel_filter_size = 0.02
TRAJECTORY_BUILDER_2D.ceres_scan_matcher.occupied_space_weight = 10.
TRAJECTORY_BUILDER_2D.ceres_scan_matcher.translation_weight = 1.
TRAJECTORY_BUILDER_2D.ceres_scan_matcher.rotation_weight = 1.
TRAJECTORY_BUILDER_2D.submaps.num_range_data = 80.
TRAJECTORY_BUILDER_2D.submaps.grid_options_2d.resolution = 0.1 -- / 0.02
POSE_GRAPH.optimize_every_n_nodes = 160. -- 2倍的num_range_data以上
POSE_GRAPH.constraint_builder.sampling_ratio = 0.3
POSE_GRAPH.constraint_builder.max_constraint_distance = 15.
POSE_GRAPH.constraint_builder.min_score = 0.48
POSE_GRAPH.constraint_builder.global_localization_min_score = 0.60

4. Cartographer 工程化建议

4.1 如何提升建图质量

• 选择频率高（25 Hz 以上）、精度高的激光雷达
• 如果只能用频率低的激光雷达
  • 使用高频、高精度 IMU，且让机器人运动慢一点
  • 调 ceres 的匹配权重，将地图权重调大，平移旋转权重调小
  • 将代码中平移和旋转的残差注释掉
• 里程计
  • Cartographer 中对里程计的使用不太好
  • 可以将 karto 与 GMapping 中使用里程计进行预测的部分拿过来进行使用，改完后就可达到比较好的位姿预测效果
• 粗匹配
  • 将 karto 的扫描匹配的粗匹配放过来，karto 的扫描匹配的计算量很小，当做粗匹配很不错
• 地图
  • 在最终生成地图的时候使用后端优化后的节点重新生成一次地图，这样生成的地图效果会比前端地图的叠加要好很多

4.2 降低计算量与内存

• 体素滤波与自适应体素滤波的计算量（不是很大）
• 后端进行子图间约束时的计算量很大
• 分支定界算法的计算量很大
• 降低内存，内存的占用基本就是多分辨率地图，每个子图的多分辨率地图都进行保存是否有必要

4.3 纯定位的改进建议

• 将纯定位模式与建图拆分开，改成读取之前轨迹的地图进行匹配.
• 新的轨迹只进行位姿预测，拿到预测后的位姿与之前轨迹的地图进行匹配，新的轨迹不再进行地图的生成与保存，同时将整个后端的功能去掉.
• 去掉后端优化会导致没有重定位功能，这时可将 cartographer 的回环检测（子图间约束的计算）部分单独拿出来，做成一个重定位功能，通过服务来调用这个重定位功能，根据当前点云确定机器人在之前地图的位姿

4.4 去 ros 的参考思路

• 仿照 cartographer_ros 里的操作：获取到传感器数据，将数据转到 tracking_frame 坐标系下并进行格式转换，再传入到 cartographer 里就行

5. 源码注释

• cartographer_detailed_comments_ws