Integrated Navigation for INS/GNSS/Vision for Vehicle Positioning
CarViG extends IGNAV with visual information for vehicle positioning.
CarViG adds camera-based information on top of the INS/GNSS navigation stack: monocular/stereo visual odometry, vision-aided INS/GNSS integration, online camera parameter estimation and relative pose estimation. It targets vehicle positioning in GNSS-degraded environments such as urban canyons, tree-lined roads and tunnels. It is an independent CMake project and can be built separately from the main project.
- Vision-aided INS/GNSS coupling (
pos1-posmode = ins-gnsslc-vo-coupled) - Feature pipeline: FAST corners (
ins-libfast.cc), KLT tracking (ins-flow-track.cc), stereo/mono matching (ins-match.cc) with GMS outlier filtering (gms-matcher.cc, optional CUDA acceleration) - Monocular/stereo visual odometry (libviso-derived library in
src/libviso), monocular two-view motion and EPnP pose estimation (ins-vo-mono.cc,ins-pnp.cc) - Online camera parameter estimation: intrinsics (fx, fy, cx, cy), distortion (k1, k2, p1, p2), camera-IMU lever arm and misalignment (
ins-estcam*,ins-unc-cam*) - Relative pose estimation algorithms (
src/relative-pose): Nister five-point, Groebner-basis four-point, seven/eight-point, one-point - MSCKF-based coupling variant and a Python MSCKF implementation (
ins-gnss-vo-msckf.cc,src/msckf-python) - Multi-sensor simulation and post-processing for the VO-coupled chain (
ins-gnss-vo-sim.cc,ins-gnss-vo-postpos.cc) - Extended data formats: euroc, kitti, malaga, m39-mix, oem6-sol/pose/raw, IGV simulation data, etc.
- Inherits the full INS/GNSS capability of IGNAV: loose/tight coupling, motion constraints, alignment initialization, RTS and forward/backward smoothing
CarViG keeps the IGNAV processing core and adds a vision front-end plus a VO-aided coupling filter. Camera models are pinhole with radial-tangential distortion (CAM_PINHOLE / DISTORT_RT in carvig.h); image input can be a directory of images (file-monodir) or dataset streams (euroc, kitti, malaga, m39-mix).
Main modules (files under src/ins-gnss/ unless noted):
| Area | Modules |
|---|---|
| Vision-extended server | carvigsvr.cc |
| Image pipeline | ins-img-filter.cc, ins-libfast.cc, ins-track.cc, ins-flow-track.cc, ins-match.cc, gms-matcher.cc, ins-track-visualizer.cc |
| Motion estimation | ins-vo-mono.cc, ins-pnp.cc, ins-vo-opencv-interface.cc, src/libviso/ |
| Relative pose | src/relative-pose/ (five-point Nister, four-point Groebner, seven/eight-point, one-point) |
| VO fusion | ins-gnss-vo.cc (coupled EKF), ins-gnss-vo-msckf.cc (MSCKF variant) |
| Post-processing / simulation | ins-gnss-vo-postpos.cc, ins-gnss-vo-sim.cc |
| MSCKF (Python) | src/msckf-python/ |
improved/
├── CMakeLists.txt
├── cmake/FindEigen3.cmake
├── include/ carvig.h and common headers
├── src/
│ ├── ins-gnss/ Core integration code (app/, rcv/ and algorithm modules)
│ ├── libviso/ Monocular/stereo visual odometry library
│ ├── relative-pose/ Relative pose estimation algorithms
│ └── msckf-python/ MSCKF implementation in Python
├── data/geomag/ Geomagnetic model data (IGRF12, WMM2015)
├── bin/ Build output: carvignav, tc-postpos
└── lib/ Build output (created by the build)
Prebuilt x86-64 Linux binaries are already shipped in bin/.
Requirements are the same as IGNAV (Linux, CMake, C++11), plus:
- OpenCV 2.4.3 or 3.x (CMake searches for 3.x first, then 2.4.3)
- Eigen3 3.1 or later
- Python development libraries (checked via
find_package(PythonLibs)) - BLAS, LAPACK, zlib, pthread
- libpng, jpeg (CMake references
/usr/local/lib/libpng.so; adjust if your system differs)
cd improved
mkdir -p build && cd build
cmake ..
make -j$(nproc)The build produces improved/bin/carvignav and improved/bin/tc-postpos plus several static libraries.
cd improved
./bin/carvignav -o src/ins-gnss/app/carvig.confcarvignav: main console application; its options are similar to navapp (-s,-p,-m,-o,-w,-r,-t,-sta, ...; run it without arguments for help)tc-postpos: tightly coupled post-processing tool, with-p(console port) and-o(options file)
Four options files are provided in src/ins-gnss/app/:
| File | Purpose |
|---|---|
carvig.conf |
INS/GNSS loosely coupled + vision fusion (default) |
carvig-lc.conf |
INS/GNSS loosely coupled |
carvig-tc.conf |
INS/GNSS tightly coupled |
carvig-sim.conf |
Simulated data (gps/imu/vo simulation files) |
Input paths (inpstr*-path, file-monodir, ...) in these files are absolute paths from the author's machine; update them before running.
- RTKLIB: http://www.rtklib.com
- OpenCV: https://opencv.org/
- Eigen: https://eigen.tuxfamily.org/
- Libviso: http://www.cvlibs.net/software/libviso/
- QQ: 1971129844
- Email: sujinglan@whu.edu.cn