GNSS-Denied Visual-Inertial Localization

Flagship robotics perception demo for localization when satellite fixes disappear. The project simulates a planar vehicle trajectory, removes GNSS measurements for a configurable outage, and compares a visual-inertial EKF against an IMU plus wheel-odometry dead-reckoning baseline.

What This Demonstrates

• Simulated GNSS dropout from 28s to 62s in a repeatable synthetic route.
• IMU-style prediction with accelerometer and gyro bias drift.
• Wheel-odometry speed correction as a local motion constraint.
• EKF fusion of GNSS fixes, visual-odometry pose corrections, and inertial prediction.
• Trajectory and error plots generated from the actual run, not hand-drawn assets.
• A reproducible Python command and Docker path for reviewers.
• Optional ROS2 C++ wrapper skeleton for moving the estimator toward a live robotics stack.

Quick Start

python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -e .
python -m gnss_denied_vio.simulate --seed 7 --output results/example

On Linux or macOS:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .
python -m gnss_denied_vio.simulate --seed 7 --output results/example

The run writes:

• results/example/trajectory_overview.png
• results/example/position_error.png
• results/example/yaw_error.png
• results/example/sensor_timeline.png
• results/example/metrics.json

Reproducible Result

The committed example was generated with:

python -m gnss_denied_vio.simulate --seed 7 --output results/example

Key metrics from that run:

Metric	Value
GNSS outage	28s to 62s
Fused dropout position RMSE	0.24 m
IMU/odom-only dropout position RMSE	26.65 m
Dropout improvement	99.1%
Visual updates	301
GNSS fixes	56

The sensor timeline shows the estimator still receiving local sensing while GNSS is denied:

Heading drift is also corrected by visual yaw observations:

How It Works

The EKF state is:

[x, y, vx, vy, yaw, accel_bias_x, accel_bias_y, gyro_bias_z]

Prediction uses body-frame accelerometer and yaw-rate measurements, subtracts estimated biases, rotates acceleration into the world frame, and propagates position, velocity, and yaw. The correction steps are:

• GNSS position update when fixes are available.
• Visual-odometry pose update at camera rate.
• Wheel-odometry forward-speed update.

The baseline receives the same IMU prediction and wheel-speed updates but no absolute GNSS or camera pose corrections. That makes the outage behavior easy to inspect.

Docker

docker build -t gnss-denied-vio .
docker run --rm -v "$PWD/results:/app/results" gnss-denied-vio

The container defaults to the same seed and output directory used by the README screenshots.

Project Layout

src/gnss_denied_vio/
  config.py       Scenario and sensor parameters
  ekf.py          EKF state propagation and measurement updates
  sensors.py      Truth trajectory plus synthetic IMU, GNSS, VO, and odometry
  simulation.py   End-to-end run loop, metrics, and result persistence
  plots.py        Pillow-based plot rendering for README screenshots
  simulate.py     CLI entry point
ros2/
  gnss_denied_vio_cpp/  Optional ROS2 C++ wrapper node
tests/
  test_simulation.py    Regression tests for outage behavior

ROS2 Wrapper

The optional ROS2 C++ wrapper is in ros2/gnss_denied_vio_cpp. It is intended as a starting point for live topics:

• /imu/data
• /wheel/odom
• /visual_odometry
• /gnss/pose
• /localization/ekf_odom

Limitations And Next Steps

• The simulation is 2D and does not model roll, pitch, gravity alignment, camera intrinsics, or feature tracking.
• Visual odometry is generated as a noisy drifting pose stream rather than from image frames.
• The EKF uses fixed measurement covariances; a production system should adapt them based on sensor quality.
• The ROS2 wrapper is intentionally lightweight and should be wired to recorded bags or hardware before being treated as deployment code.
• Good next steps: add KITTI/EuRoC bag playback, expose covariance tuning files, add NEES/NIS consistency checks, and extend the state to full 3D inertial navigation.

---

Benchmarks

Reproduced with --seed 7 on the default synthetic scenario (34-second GNSS outage from 28 s to 62 s).

Metric	EKF (fused)	IMU + odometry only
Position RMSE during outage	0.24 m	26.65 m
Outage duration	34 s	34 s
Improvement	99.1%	—
Visual odometry updates	301	0
GNSS fixes (outside outage)	56	56

Reproduce:

python -m venv .venv && source .venv/bin/activate
pip install -e .
python -m gnss_denied_vio.simulate --seed 7 --output results/example
cat results/example/metrics.json