haopo2005 / kitti-interp

A Interpolation and Noise Tool for KITTI Raw Data

This tool

• Interpolate and re-sync IMU data to give KITTI a 100 Hz IMU and GNSS data and correct the timestamps with vision image.

• Allow user to add noise and bias to GNSS data in given area.

The main entrance is ProcessKITTI.m.

credits: Gong Zheng , Ji Xingwu

Input

This tool uses vision image along with its timestamps from synced+rectified data(like this), then align, interpolate and noise the IMU and GNSS data form the unsynced+unrectified data(like this). Therefore, the tool require two folders contain these two set of data, please make your KITTI folder organized like:

<KITTI_Folder>
├── <Synced_RawData_Folder>
│ └── <Set_Folder>
│ └── <SubSet_Folder>
└── <UnSynced_RawData_Folder>
└── <Set_Folder>
└── <SubSet_Folder>

These paths can be changed in ProcessKITTI.m and ReadData.m.

For example, on my system,

/media/joey/dataset/KITTI #<KITTI_Folder>
├── RawData #<Synced_RawData_Folder>
│ ├── 2011_09_26
│ └── 2011_10_03 #<Set_Folder>
│ ├── 2011_10_03_drive_0027_sync #<SubSet_Folder>
│ ├── 2011_10_03_drive_0034_sync
│ ├── 2011_10_03_drive_0042_sync
│ ├── 2011_10_03_drive_0047_sync
│ ├── 2011_10_03_drive_0058_sync
│ ├── calib_cam_to_cam.txt
│ ├── calib_imu_to_velo.txt
│ └── calib_velo_to_cam.txt
└── RawDataUnsync # <UnSynced_RawData_Folder>
├── 2011_09_26
└── 2011_10_03
├── 2011_10_03_drive_0027_extract
├── calib_cam_to_cam.txt
├── calib_imu_to_velo.txt
├── calib.mat
└── calib_velo_to_cam.txt

Output

This tool will create a path called RawDataFixed under <KITTI_Folder>, with the regular KITTI set name and sub set name. The OXTS data ( which contain IMU and GNSS) is written 3 times in to three folders: oxts-fixed,oxts-interpedand oxts-noised.

• oxts-fixed: oxts data with timestamps and IMU data interpolated. Rest of the data is filled with NaN.
• oxts-interped: oxts data with timestamps, IMU and GNSS data interpolated.
• oxts-noised: oxts data with timestamps, IMU and GNSS data interpolated. Extra artificial noise is add to GNSS.

You should copy the image data (image_00 and image_01) from the synced data set manually .

Processing

Read data

In ReadData.m.

Due to the KITTI's file arrangement, the reading (as well as writing) can be really slow, so be patient.

Interpolate IMU data

In InterpImu.m.

The result will show four pics. Please first do a visual check on the interpolation results on all IMU data: the blue, interpolated dots should be "reasonable".

and then visual check on the time alignment: the square, synced IMU data should overlapped with the blue, aligned unsynchronized IMU data.

Next, do another visual check on the time interpolation: there should be no outliers and negative values on the fixed data (blue point).

Interpolate and noise GNSS data

In InterpAndNoiseGnss.m.

Please tweak the parameters on the top of InterpAndNoiseGnss.m.

The important parameters are:

• regular_std_level,regular_bias: region not in bad_gnss_area or good_gnss_area will use this pair of parameters.

• bad_gnss_area ,good_gnss_area: define the good and bad regions. Column definitions: [ x[m] y[m] z[m] radii[m] white-noise-level[m] multipath-bias-level[m] ]. The x-y-z here is under local ENU, you can determine it on the given figure after the first run.

• grid_size,c1,c2: parameters for Gaussian random field (special bias simulating multipath ). grid_size is the 'resolution' of the field and c1,c2 are the correlation parameters. To simulate a field changes ever 0.2 meters with 10 meters correlation, set to 0.2,10,10 for example.

• time_corr: Gaussian-Markov process correlation parameter.

• MultipathDetectCo: Multipath detection coefficient. Simulating the receiver's over-confidence on multipath detection failure.

Turn on plot_GRF to show the GRF under the trajectory.

Please do visual check on the noise result (blue). The red circle is the bad gnss area.

Write data out to disk

In WriteData.m.

Very, very slow. Please turn off write_fix and write_interp if generating different cases for the same subset to avoid re-writing.