DEM-Based Terrain-Aided INS (TAN)

Simulated terrain-aided navigation for a high-speed UAV without GNSS: IMU dead reckoning, radar altimeter, synthetic DEM, terrain profile matching, EKF/UKF fusion, and integrity gating.

Full research spec: docs/PROJECT_SPEC.md · Engineering plan: docs/roadmap.md

---

What it does

flowchart LR
    IMU[IMU] --> INS[INS dead reckoning]
    ALT[Altimeter] --> TM[Terrain matcher]
    DEM[DEM map] --> TM
    INS --> TM
    TM --> GATE[Integrity gate]
    GATE --> FILT[EKF / UKF]
    FILT --> OUT[Corrected position]
    FILT -.->|optional reset| INS

1. Load real SRTM .hgt maps natively, or build a synthetic DEM fallback.
2. Fly a truth trajectory (straight legs + banked turns).
3. Simulate IMU and radar altimeter with noise and bias.
4. Run INS (drifts over time).
5. Match measured height profiles to the DEM (MAD matcher).
6. Gate bad matches (χ² test), then fuse with EKF or UKF.
7. Plot trajectories and errors; run benchmarks and Monte Carlo.

---

Quick start

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

python run.py sim      # plots 01–06 → sim_outputs/[tile_name]/
python run.py bench    # scenario table → results/
python run.py mc       # Monte Carlo → results/ + monte_carlo_*.png
python run.py all      # run all three

Or directly:

python tests/main_sim.py

---

Settings (no code edits needed)

Edit config/settings.yaml (~20 fields). Example:

Group	Keys	Purpose
Flight	start_speed_mps, flight_segments_sec, turn_rate_deg_s	Mission path
DEM	dem_amplitude_m, dem_seed	Terrain roughness
Sensors	imu_*, altimeter_*	Noise / bias
Matching	match_search_radius_m, match_window_size, match_interval_sec	TAN tuning
Filter	filter_type (ekf / ukf), filter_q_pos_m, filter_r_pos_m	Fusion
Integrity	integrity_chi2_threshold, closed_loop_ins_reset	False-match rejection
UI	plot_theme (dark / minimal), output_dir	Plots

Override file path:

export DEM_TAN_SETTINGS=/path/to/my_settings.yaml

---

Project layout

config/settings.yaml     ← all tunables
data/
  dem_tiles/             Place SRTM `.hgt` dataset tiles here (e.g., n34e051.hgt)
src/
  dem/                   DEM loader + synthetic terrain + HGT parser
  simulation/            Truth trajectory
  sensors/               IMU + altimeter models
  navigation/            INS, matcher, EKF, UKF, integrity
  analysis/              NavigationPipeline, scenarios
  plotting/              Dark / minimal themes
tests/
  main_sim.py            End-to-end demo
  run_benchmarks.py      Multi-scenario comparison
  run_monte_carlo.py     Statistical runs
docs/                    Walkthrough logs and project documents
trash/                   Transient discarded files
sim_outputs/             Figures (01–06) separated conditionally per loaded DEM title
results/                 CSV + markdown tables

---

Output figures (sim_outputs/)

Once started, the pipeline runs dynamically over every .hgt dataset placed at data/dem_tiles/. Outcomes are organized neatly per tile layout:

File	Phase	Content
01_trajectory_3d.png	1–2	Truth path over DEM (3D)
02_imu_accel.png	3	Body-frame accelerometer
03_altimeter.png	3	True vs noisy AGL + terrain height
04_navigation_map.png	4–6	Plan view on DEM (truth / INS / fused)
05_full_navigation.png	7	Map + accepted/rejected terrain matches
06_error_over_time.png	Eval	Horizontal position error vs time
monte_carlo_*.png	9	CEP histogram & dispersion (run.py mc)

---

Sample results

After python run.py bench, open results/benchmark_summary.md. Typical patterns:

• Mountainous / hilly terrain → fusion can beat INS-only when matches are accepted.
• Flat terrain → matching fails more often (high gate rejection).
• UKF vs EKF → similar on this 4-state model; compare last benchmark row.

Monte Carlo prints CEP50 / CEP95 and saves histograms in sim_outputs/.

---

Themes

plot_theme	Look
dark	Dark background (default)
minimal	Light, clean lines

---

Limitations (MVP)

• Local flat-Earth INS (not full strapdown 6-DOF).
• Synthetic DEM safely fallbacks natively without external data required.
• Profile MAD matching (not full NCC grid search).
• Particle filter not implemented (see roadmap).

---

License

Academic / research use — see course or lab guidelines.