Intelligent Dead Reckoning (IDR) — GNSS+INS Fusion

An Intelligent Dead Reckoning system with GNSS+INS fusion for Smart India Hackathon (SIH). Two deliverables: (1) an Android mobile app, (2) a sensor-agnostic edge-deployable software engine.

Final Benchmark Validation Results (Phase 13)

Last Updated: 2026-10-18 — Final benchmark validation performed across all 16 held-out evaluation scenarios (13 Car, 2 Two-Wheeler, 1 Edge FOG).

Official Benchmark Target: ≤ 10.0% drift of distance travelled during 60s GNSS blackout.
Team Stretch Target: 1.0 – 2.0% drift.
Overall Validation Status: VALIDATED

Summary Metrics

Platform	Target Rate	Measured Latency (Mean)	95th Percentile	Measured Throughput	Status
Mobile App (Android)	≥ 10.0 Hz	1.924 ms	3.758 ms	519.8 Hz	PASS
Edge Engine (C++/Python Wrapper)	≥ 200.0 Hz	3.611 ms	5.084 ms	277.0 Hz	PASS

---

Session-Wise Dead Reckoning Drift (60s GNSS Outage)

Session ID	Vehicle Category	Data Source / Platform	Outage Dist (m)	Final Error (m)	Final Drift %	Official Target (<=10%)	Stretch Target (1-2%)	Evaluation Status
Vw17	Car	IO-VNBD (MEMS)	151.28	1.46	0.96%	PASS	PASS	Validated (Best Car)
V-Vfa02	Car	IO-VNBD (MEMS)	1483.63	133.78	9.02%	PASS	FAIL	Validated
Vta26	Car	IO-VNBD (MEMS)	2.80	82.23	8.22%	FAIL	FAIL	Stationary/Low-speed artifact
Vta27	Car	IO-VNBD (MEMS)	850.39	74.86	8.80%	PASS	FAIL	Validated
Vtb11	Car	IO-VNBD (MEMS)	208.47	19.11	9.17%	PASS	FAIL	Validated
Vw15	Car	IO-VNBD (MEMS)	3.72	90.98	9.10%	FAIL	FAIL	Stationary/Low-speed artifact
Vtb12	Car	IO-VNBD (MEMS)	159.32	37.24	23.37%	FAIL	FAIL	Divergent
Vw16a	Car	IO-VNBD (MEMS)	1187.04	267.49	22.53%	FAIL	FAIL	Highway Drift
Vw16b	Car	IO-VNBD (MEMS)	752.38	606.25	80.58%	FAIL	FAIL	Divergent
S4	Car	IO-VNBD (MEMS)	2.34	314.10	31.41%	FAIL	FAIL	Stationary/Low-speed artifact
Vta28	Car	IO-VNBD (MEMS)	697.16	614.29	88.11%	FAIL	FAIL	Multi-Turn Divergent
Vta29	Car	IO-VNBD (MEMS)	468.35	429.01	91.60%	FAIL	FAIL	Multi-Turn Divergent
Vta30	Car	IO-VNBD (MEMS)	91.26	158.36	173.52%	FAIL	FAIL	Urban Stop-and-Go
session1	Two Wheeler	Bridge Synthetic	231.32	231.91	100.26%	FAIL	FAIL	Two-Wheeler Synthetic
session2	Two Wheeler	Bridge Synthetic	211.11	475.58	225.28%	FAIL	FAIL	Two-Wheeler Synthetic
S1 (Synthetic FOG 200Hz)	Edge Fog	FOG Synthetic	353.15	342.99	97.12%	FAIL	FAIL	Edge FOG Synthetic

---

Pass / Fail Summary Breakdown

• Total Scenarios Evaluated: 16 sessions
• Passing Official Target ($\le 10.0%$): 5 sessions (31.25%)
  • Active Driving Passes (4): Vw17 (0.96%), V-Vfa02 (9.02%), Vta27 (8.80%), Vtb11 (9.17%)
• Passing Team Stretch Target ($1.0 - 2.0%$): 1 session (Vw17 at 0.96%)
• Failing Official Target ($> 10.0%$): 11 sessions (68.75%)

Note: Sessions with very low distance travelled (<50m) show artificially inflated drift percentages due to denominator effect. These are labeled as stationary/low-speed artifacts.

Why Sessions Pass vs Why They Fail

1. Why Sessions Pass ($\le 10.0%$ Drift)

• Heading Stability & NHC Alignment: In straight or moderate-curvature segments (Vw17, Vta27, Vtb11), Non-Holonomic Constraints (NHC: $v_y \approx 0, v_z \approx 0$) tightly bound lateral and vertical velocity errors.
• AI Speed Model Precision: The TFLite 1D-CNN+GRU forward velocity model provides accurate speed pseudo-measurements that constrain along-track position error during the GNSS outage.
• Sufficient Forward Motion: Sessions like V-Vfa02 benefit from extended straight segments where accumulated errors remain bounded relative to total distance travelled.

2. Why Sessions Fail ($&gt; 10.0%$ Drift)

• Unobservable Yaw Heading Drift in Consumer MEMS IMUs:
  • During a 60s blackout, without an absolute heading reference (e.g. dual-antenna RTK GNSS or calibrated magnetometer), heading error grows with gyro bias integration: $\Delta \theta(t) \approx \int b_g(t) dt$.
  • This is the fundamental limitation documented in Phase 6: consumer-grade IMUs lack sufficient heading observability for extended pure inertial navigation.
• Multi-Turn & Aggressive Maneuver Accumulation:
  • In dynamic tracks with continuous sharp turns (Vta28, Vta29, Vw16b), angular integration errors compound across multiple axes, causing trajectory divergence from road geometry.
• Low-Speed/Stationary Window Artifacts:
  • In near-stationary segments (S4, Vta26, Vw15), drift percentage becomes inflated as $\frac{\text{position error}}{\text{distance traveled}} \times 100%$ approaches infinity when distance approaches zero.
  • These reflect real positioning errors (meters-scale) that are significant relative to tiny distances traveled but would be acceptable in absolute terms for navigation.
• High-Speed Long Outages with Dynamic Maneuvers:
  • Combined effect of velocity error integration and heading drift in segments like Vta30.

See docs/PHASE13_RESULTS.md for detailed benchmark validation report and eval/FULL_BENCHMARK_RESULTS.md for raw benchmark data.

---

Project Structure

/docs                     Architecture, phase checklist, novelty spec, benchmarks
/data
  /raw                    IO-VNBD + any collected data, untouched
  /processed              Cleaned/segmented data used for training
/training                 Offline model training code (cloud/desktop side)
  /models                 Saved trained models (checkpoints, .tflite exports)
/engine                   Core sensor-agnostic fusion engine (shared library)
  /calibration            Alignment & calibration engine
  /ai_filters             AI Speed & Vibration Filter
  /nhc_zupt               NHC + ZUPT (incl. lean-compensated)
  /fusion                 GNSS+INS Fusion Engine (EKF/UKF + AI correction)
  /map_matching           Map-Matching Filter
  /outage_prediction      Predictive GNSS Outage Detection
/mobile                   Android app (consumes /engine via TFLite + shared logic)
/edge                     Edge-deployable engine packaging (consumes /engine)
/eval                     Benchmark scripts, drift measurement, plots
/screening_package        Screening-round deliverable (model + plots + writeup)

Quick Start

Python Environment (Training & Engine)

cd training
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt

Android App (Mobile)

cd mobile
./gradlew assembleDebug

Run Final Benchmark

training\venv\Scripts\python.exe eval/run_full_benchmark.py

Documentation

• Architecture
• Novelty Specification
• Phase Checklist
• Final Comparison — Final Phase 7 validation report
• Implementation Progress
• Open Questions

Development Protocol

See CLAUDE.md for the project constitution: non-negotiable rules, execution protocol, git commit convention, and repository structure.

License

Proprietary — SIH entry