HILTTI-2022数据集求指教

[keleeven]

您好，最近在利用fast livo调试HILTTI-2022的数据集，激光里程计已经可以成功运行，但是结合相机这块一直调试有问题，无法正常跟踪特征点，以下是我配置的参数表，麻烦郑博可以点拨一二，感谢！！！
feature_extract_enable : 0
point_filter_num : 1
max_iteration : 10
debug: 1
dense_map_enable : 1
filter_size_surf : 0.05
filter_size_map : 0.1
cube_side_length : 20
grid_size : 40
patch_size : 8
img_enable : 1
lidar_enable : 1
outlier_threshold : 300 # 78 100 156
ncc_en: true
ncc_thre: 100
img_point_cov : 100 # 1000
laser_point_cov : 0.001 # 0.001
delta_time: 0.0

common:
lid_topic: "/hesai/pandar"
imu_topic: "/alphasense/imu"

preprocess:
lidar_type: 4 # 1:Livox Avia LiDAR 2:VELO16 3:OUST64 4:XT32
scan_line: 32 # 16 64 32
blind: 1 # blind x m disable

mapping:
acc_cov_scale: 100
gyr_cov_scale: 10000
extrinsic_T: [-0.001, -0.00855, 0.055] # horizon 0.05512, 0.02226, -0.0297
extrinsic_R: [ 0, -1, 0,
-1, 0, 0,
0, 0, -1]

pcd_save:
pcd_save_en: true

camera:
img_topic: /alphasense/cam0/image_raw
Rcl: [0,0,-1,
0,-1,0,
-1,0,0]
Pcl: [-0.052, -0.053, 0.068]

[xuankuzcr]

你用的外参Rcl和Pcl看起来不对，试试这个
Rcl: [ -0.999926, -0.00670802, 0.0101073,
-0.0100912, -0.00242564, -0.999946,
0.00673218, -0.999975, 0.00235777]
Pcl: [-0.0549762, 0.0675401, -0.0520599]

[keleeven]

你用的外参Rcl和Pcl看起来不对，试试这个 Rcl: [ -0.999926, -0.00670802, 0.0101073, -0.0100912, -0.00242564, -0.999946, 0.00673218, -0.999975, 0.00235777] Pcl: [-0.0549762, 0.0675401, -0.0520599]

谢谢郑博的回复，试了您的参数配置后，就可以正常跟踪视觉点了，我跑的是exp10_cupola_2.bag这个数据集，但是每次在上楼梯这段狭窄空间的时候，位姿就会发生严重飘逸，并且单独跑LIO也会存在一样的问题，但是看作者的FAST-LIVO2论文实验部分，FAST-LIO2和FAST-LIVO都是很小的姿态误差，所以不太清楚是哪里存在问题，希望郑博可以指点一二，感激！！以下是我FAST-LIVO和FAT-LIO2的参数配置

FAST-LIVO：
feature_extract_enable : 0
point_filter_num : 1
max_iteration : 10
debug: 1
dense_map_enable : 0
filter_size_surf : 0.05
filter_size_map : 0.1
cube_side_length : 20
grid_size : 40
patch_size : 8
img_enable : 1
lidar_enable : 1
outlier_threshold : 300 # 78 100 156
ncc_en: false
ncc_thre: 0
img_point_cov : 100 # 1000
laser_point_cov : 0.001 # 0.001
delta_time: -0.05

common:
lid_topic: "/hesai/pandar"
imu_topic: "/alphasense/imu"

preprocess:
lidar_type: 4 # 1:Livox Avia LiDAR 2:VELO16 3:OUST64 4:XT32
scan_line: 32 # 16 64 32
blind: 0.5 # blind x m disable

mapping:
acc_cov_scale: 100
gyr_cov_scale: 10000
extrinsic_T: [-0.001, -0.00855, 0.055]
extrinsic_R: [1.11022302e-16, -1.00000000e+00, 0.00000000e+00,
-1.00000000e+00, 1.11022302e-16, -0.00000000e+00,
0.00000000e+00, 0.00000000e+00, -1.00000000e+00]

pcd_save:
pcd_save_en: true

camera:
img_topic: /alphasense/cam0/image_raw
Rcl: [ -0.999926, -0.00670802, 0.0101073,
-0.0100912, -0.00242564, -0.999946,
0.00673218, -0.999975, 0.00235777]
Pcl: [-0.0549762, 0.0675401, -0.0520599]

FAST-LIO2：
common:
lid_topic: "/hesai/pandar"
imu_topic: "/alphasense/imu"
time_sync_en: false # ONLY turn on when external time synchronization is really not possible
time_offset_lidar_to_imu: 0.0 # Time offset between lidar and IMU calibrated by other algorithms, e.g. LI-Init (can be found in README).
# This param will take effect no matter what time_sync_en is. So if the time offset is not known exactly, please set as 0.0

preprocess:
lidar_type: 3 # 1 for Livox serials LiDAR, 2 for Velodyne LiDAR, 3 for ouster LiDAR,
scan_line: 32
timestamp_unit: 3 # 0-second, 1-milisecond, 2-microsecond, 3-nanosecond.
blind: 1

mapping:
acc_cov: 0.1
gyr_cov: 0.1
b_acc_cov: 0.0001
b_gyr_cov: 0.0001
fov_degree: 360
det_range: 150.0
extrinsic_est_en: false # true: enable the online estimation of IMU-LiDAR extrinsic
extrinsic_T: [-0.001, -0.00855, 0.055]
extrinsic_R: [1.11022302e-16, -1.00000000e+00, 0.00000000e+00,
-1.00000000e+00, 1.11022302e-16, -0.00000000e+00,
0.00000000e+00, 0.00000000e+00, -1.00000000e+00]

publish:
path_en: true
scan_publish_en: true # false: close all the point cloud output
dense_publish_en: false # false: low down the points number in a global-frame point clouds scan.
scan_bodyframe_pub_en: true # true: output the point cloud scans in IMU-body-frame

pcd_save:
pcd_save_en: true
interval: -1 # how many LiDAR frames saved in each pcd file;
# -1 : all frames will be saved in ONE pcd file, may lead to memory crash when having too much frames.

[keleeven]

尝试了较多的方法，给激光点云建立噪声模型，加上BALM，修改IMU噪声，但无一例外，运行exp10_cupola_2.bag都在上楼梯这段发生了里程计的严重漂移，无论是FAST-LIO2，还是FAST-LIVO，希望郑博抽空可以指点迷津，感激不尽！！！

[xuankuzcr]

Exp10 Cupola 2这个数据不在Hilti2022的榜单上，livo2论文里只测了打榜需要的sequence。我下载下来看看。

[xuankuzcr]

我刚测了下exp10_cupola_2.bag，lio2我也跑不下来，但livo2可以跑下来。只是顶楼的地面稍微有点厚，但可以接受。livo2里面的LiDAR部分修改自voxel map，参数可以给你参考下。ps: 上下楼梯时很狭窄，voxel_size要设小一点，要不然平面会误判。

  max_iterations: 10
  voxel_map_en: true
  pub_plane_en: false
  dept_err: 0.001 
  beam_err: 0.01
  min_eigen_value: 0.005
  sigma_num: 3
  voxel_size: 0.1 
  max_layer: 2
  max_points_num: 150
  layer_init_num: [5, 5, 5, 5, 5]

[keleeven]

收到！！！谢谢郑博抽出时间调试，并感谢细节上的指点，我去尝试一下。