gps-denied-desktop/_docs/01_solution/solution_draft01.md

Solution Draft

Product Solution Description

A Python-based GPS-denied visual navigation service that determines GPS coordinates of consecutive UAV aerial photo centers using visual odometry, satellite image geo-referencing, and sliding window position optimization. The system operates as a background REST API service with real-time SSE streaming.

Core approach: Consecutive images are matched using learned features (SuperPoint + LightGlue) to estimate relative motion (visual odometry). Periodically, each image is matched against pre-cached Google Maps satellite tiles to obtain absolute position anchors. A sliding window optimizer fuses VO estimates with satellite anchors, constraining drift. The system handles route disconnections by treating each continuous VO chain as an independent segment, geo-referenced through satellite matching.

┌─────────────────────────────────────────────────────────────────────┐
│                        Client (Desktop App)                         │
│   POST /jobs (start GPS, camera params, image folder)               │
│   GET  /jobs/{id}/stream (SSE)                                      │
│   POST /jobs/{id}/anchor (user manual GPS input)                    │
│   GET  /jobs/{id}/point-to-gps (image_id, pixel_x, pixel_y)        │
└──────────────────────┬──────────────────────────────────────────────┘
                       │ HTTP/SSE
┌──────────────────────▼──────────────────────────────────────────────┐
│                     FastAPI Service Layer                            │
│   Job Manager → Pipeline Orchestrator → SSE Event Emitter           │
└──────────────────────┬──────────────────────────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────────────────────────┐
│                     Processing Pipeline                              │
│                                                                      │
│  ┌─────────────┐  ┌──────────────┐  ┌────────────────────────────┐  │
│  │  Feature     │  │  Visual      │  │  Satellite Geo-Referencing │  │
│  │  Extractor   │→│  Odometry    │→│  (cross-view matching)     │  │
│  │ (SuperPoint) │  │ (homography) │  │  (SuperPoint+LightGlue)   │  │
│  └─────────────┘  └──────────────┘  └────────────────────────────┘  │
│         │                │                       │                    │
│         ▼                ▼                       ▼                    │
│  ┌──────────────────────────────────────────────────────────────┐    │
│  │            Sliding Window Position Optimizer                  │    │
│  │    (VO estimates + satellite anchors + drift constraints)     │    │
│  └──────────────────────────────────────────────────────────────┘    │
│                              │                                       │
│  ┌───────────────────────────▼──────────────────────────────────┐    │
│  │                  Segment Manager                              │    │
│  │   (independent segments, satellite-anchored stitching)        │    │
│  └──────────────────────────────────────────────────────────────┘    │
│                                                                      │
│  ┌──────────────────────────────────────────────────────────────┐    │
│  │               Satellite Tile Cache Manager                    │    │
│  │   (progressive download, Google Maps Tiles API, disk cache)   │    │
│  └──────────────────────────────────────────────────────────────┘    │
└──────────────────────────────────────────────────────────────────────┘

Existing/Competitor Solutions Analysis

Solution	Approach	Accuracy	IMU Required	Open Source	Relevance
YFS90/GNSS-Denied-UAV-Geolocalization	VO + satellite matching + terrain-weighted constraint optimization	<7m MAE	No	Yes (GitHub, 69★)	Highest — same constraints, best results
AerialPositioning (hamitbugrabayram)	Multi-provider tile engine + deep matchers + perspective warping	Not reported	Simulated INS	Yes (GitHub, 52★)	High — tile engine and perspective warping reference
NaviLoc	Trajectory-level VPR + VIO fusion	19.5m MLE	Yes (VIO)	Partial	Medium — uses IMU, different altitude range
ITU Thesis (Öztürk 2025)	ORB-SLAM3 + SuperPoint/SuperGlue/GIM SIM	GPS-level	No	No	High — architecture reference
Mateos-Ramirez et al. (2024)	ORB VO + AKAZE satellite + Kalman filter	143m mean (17km)	Yes	No	Medium — higher altitude, uses IMU
VisionUAV-Navigation	Multi-algorithm feature detection + satellite matching	Not reported	Not stated	Yes (GitHub)	Low — early stage

Key insight from competitor analysis: YFS90 achieves <7m without IMU using terrain-weighted constraint optimization. This validates that our target accuracy (20-50m) is realistic and possibly conservative. The sliding window optimization approach is the critical differentiator from simpler VO+satellite systems.

Architecture

Component: Feature Extraction

Solution	Tools	Advantages	Limitations	Requirements	Security	Cost	Fit
SuperPoint	superpoint (PyTorch)	Learned features, robust to viewpoint/illumination changes. Repeatability in aerial scenes. GPU-accelerated. ~80ms per image.	Requires GPU. Fixed descriptor dimension (256).	NVIDIA GPU, PyTorch, CUDA	Model weights from official source only	Free (MIT license)	Best
SIFT	OpenCV cv2.SIFT	Classical, well-understood. Scale/rotation invariant. Good satellite matching (SIFT+LightGlue top on ISPRS 2025).	Slower than SuperPoint. Less robust to extreme viewpoint changes.	OpenCV	N/A	Free	Good fallback
ORB	OpenCV cv2.ORB	Very fast. Many keypoints.	Not scale-invariant. Poor for cross-view matching.	OpenCV	N/A	Free	Only for fast VO

Selected: SuperPoint as primary (both VO and satellite matching — unified pipeline). SIFT as fallback for satellite matching where SuperPoint struggles.

Component: Feature Matching

Solution	Tools	Advantages	Limitations	Requirements	Security	Cost	Fit
LightGlue	lightglue (PyTorch)	Fastest learned matcher (~20-50ms). Adaptive pruning. Best on satellite benchmarks. ONNX/TensorRT support for 2-4x speedup.	Requires GPU for best performance.	NVIDIA GPU, PyTorch	Model weights from official source only	Free (Apache 2.0)	Best
SuperGlue	superglue (PyTorch)	Graph neural network, strong spatial context. 93% match rate (ITU thesis).	Slower than LightGlue (~2x). Non-commercial license.	NVIDIA GPU, PyTorch	Model weights from official source only	Non-commercial license	Backup
GIM	gim (PyTorch)	Best generalization for challenging cross-domain scenes.	Additional model complexity.	NVIDIA GPU, PyTorch	Model weights from official source only	Free	Supplementary for difficult matches

Selected: LightGlue as primary. GIM as supplementary for difficult satellite matches.

Component: Visual Odometry (Consecutive Frame Matching)

Solution	Tools	Advantages	Limitations	Requirements	Security	Cost	Fit
Homography-based VO	OpenCV findHomography, decomposeHomographyMat	Perfect for downward camera + flat terrain. Cleanly gives rotation + translation. Known altitude resolves scale. Simple, fast.	Assumes planar ground (valid for steppe at 400m). Fails at sharp turns (by design).	OpenCV, NumPy	N/A	Free	Best
Essential matrix VO	OpenCV findEssentialMat, recoverPose	More general than homography. Works for non-planar scenes.	Scale ambiguity harder to resolve. More complex. Unnecessary for our flat terrain case.	OpenCV	N/A	Free	Overengineered
ORB-SLAM3 monocular	ORB-SLAM3	Full SLAM with map management, loop closure.	Heavy dependency. Map building unnecessary. Scale ambiguity.	ROS (optional), C++	N/A	Free (GPL)	Too complex

Selected: Homography-based VO with SuperPoint+LightGlue features.

VO Pipeline per frame:

1. Extract SuperPoint features from current image
2. Match with previous image using LightGlue
3. Estimate homography (cv2.findHomography with RANSAC)
4. Decompose homography → rotation + translation (cv2.decomposeHomographyMat)
5. Select correct decomposition (motion must be consistent with previous direction)
6. Convert pixel displacement to meters: displacement_m = displacement_px × GSD
7. GSD = (altitude × sensor_width) / (focal_length × image_width)
8. Update position: new_pos = prev_pos + rotation × displacement_m
9. Report inlier ratio as match quality metric

Component: Satellite Image Geo-Referencing

Solution	Tools	Advantages	Limitations	Requirements	Security	Cost	Fit
Direct cross-view matching with perspective warping	SuperPoint+LightGlue, OpenCV warpPerspective	Pre-warp UAV image to approximate nadir view. Reduces viewpoint gap. Proven approach.	Needs rough camera pose estimate (from VO) for warping.	PyTorch, OpenCV	API key secured	Google Maps API cost	Best
Template matching (normalized cross-correlation)	OpenCV matchTemplate	Simple, no learning required.	Very sensitive to scale/rotation/illumination differences. Poor for cross-view.	OpenCV	N/A	Free	Poor for cross-view
VPR retrieval + refinement (NetVLAD/CosPlace)	torchvision, faiss	Handles large search areas.	Coarse localization only (tile-level). Needs fine-grained refinement step.	PyTorch, faiss	N/A	Free	Supplementary — coarse search

Selected: Direct cross-view matching with perspective warping using SuperPoint + LightGlue.

Satellite Matching Pipeline per frame:

1. Estimate approximate position from VO
2. Fetch satellite tile(s) from cache at estimated position (zoom 18, ~0.4m/px)
3. Crop satellite region matching UAV image footprint (with margin)
4. Warp UAV image to approximate nadir view using estimated camera pose
5. Extract SuperPoint features from warped UAV image
6. Extract SuperPoint features from satellite crop (can be pre-computed and cached)
7. Match with LightGlue
8. If insufficient matches: try GIM, try wider search area, try zoom 17
9. If sufficient matches (≥15 inliers): a. Estimate homography from matches b. Transform image center through homography → satellite pixel coordinates c. Convert satellite pixel coordinates to WGS84 using tile geo-referencing d. This is the absolute position anchor
10. Report match count and inlier ratio as confidence metrics

Component: Sliding Window Position Optimizer

Solution	Tools	Advantages	Limitations	Requirements	Security	Cost	Fit
Constrained sliding window optimization	scipy.optimize, NumPy	Fuses VO + satellite anchors. Constrains maximum drift. Smooths trajectory. Inspired by YFS90 (<7m).	Window size tuning needed.	SciPy, NumPy	N/A	Free	Best
Extended Kalman Filter	filterpy	Standard, well-understood. Online fusion.	Linearization approximation. Single-pass, no backward smoothing.	filterpy	N/A	Free	Good simpler alternative
Pose Graph Optimization	g2o or GTSAM (Python bindings)	Globally optimal. Handles complex factor graphs.	Heavy C++ dependency. Overkill for sequential processing.	g2o/GTSAM, C++	N/A	Free	Over-engineered

Selected: Constrained sliding window optimization (primary), with EKF as simpler initial implementation.

Optimizer behavior:

• Maintains a sliding window of last N positions (N=20-50)
• VO estimates provide relative motion constraints between consecutive positions
• Satellite matches provide absolute position anchors (hard/soft constraints)
• Maximum drift constraint: cumulative VO displacement between anchors < 100m
• Optimization minimizes: sum of VO residuals + anchor residuals + smoothness penalty
• On each new frame: add to window, re-optimize, emit updated positions
• Enables refinement: earlier positions improve as new anchors arrive

Component: Segment Manager

The segment manager is the core architectural pattern, not an edge case handler.

Segment lifecycle:

1. Start condition: First image of flight, or VO failure (feature match count < threshold)
2. Active tracking: VO provides frame-to-frame motion within segment
3. Anchoring: Satellite matching provides absolute position for segment's images
4. End condition: VO failure (sharp turn, outlier, occlusion)
5. New segment: Starts from satellite anchor or user-provided GPS

Segment states:

• ANCHORED: At least one satellite match provides absolute position → HIGH confidence
• FLOATING: No satellite match yet → positioned relative to start point only → LOW confidence
• USER_ANCHORED: User provided manual GPS → MEDIUM confidence (human error possible)

Segment stitching:

• All segments share the WGS84 coordinate frame via satellite matching
• No direct inter-segment matching needed
• A segment without any satellite anchor remains "floating" and is flagged for user input

Component: Satellite Tile Cache Manager

Solution	Tools	Advantages	Limitations	Requirements	Security	Cost	Fit
Progressive download with disk cache	aiohttp, aiofiles, sqlite3	Async download doesn't block pipeline. Tiles cached to disk. Progressive expansion follows route.	Needs internet during processing. First few images may wait for tiles.	Google Maps Tiles API key	API key in env var, not in code	Google Maps API: $200/month free credit covers ~40K tiles	Best
Pre-download entire area	requests, sqlite3	All tiles available at start. No download latency during processing.	Requires known bounding box. Large download for unknown routes. Wasteful.	Same	Same	Higher cost if area is large	For known routes

Selected: Progressive download with disk cache.

Strategy:

1. On job start: download tiles in radius R=1km around starting GPS at zoom 18
2. As route extends: download tiles ahead of estimated position (radius 500m)
3. Cache tiles on disk in {zoom}/{x}/{y}.jpg directory structure
4. Cache is persistent across jobs — tiles are reused for overlapping areas
5. Pre-compute SuperPoint features for cached tiles (saved alongside tile images)
6. If tile download fails or is unavailable: log warning, mark position as VO-only

Tile download budget:

• Initial 1km radius at zoom 18: ~300 tiles (~12MB)
• Per-frame expansion: 5-20 new tiles (~0.2-0.8MB)
• Full 20km flight: ~2000 tiles (~80MB) over the course of processing
• Well within $200/month Google Maps free credit

Component: API & Real-Time Streaming

Solution	Tools	Advantages	Limitations	Requirements	Security	Cost	Fit
FastAPI + SSE	FastAPI ≥0.135.0, EventSourceResponse, uvicorn	Native SSE support. Async pipeline. Excellent for ML workloads. OpenAPI docs auto-generated.	Python GIL (mitigated with asyncio + GPU-bound ops).	Python 3.11+, uvicorn	CORS configuration, API key auth	Free	Best

Selected: FastAPI + SSE.

API Endpoints:

POST /jobs
  Body: { start_lat, start_lon, altitude, camera_params, image_folder }
  Returns: { job_id }

GET /jobs/{job_id}/stream
  SSE stream of:
    - { event: "position", data: { image_id, lat, lon, confidence, segment_id } }
    - { event: "refined", data: { image_id, lat, lon, confidence } }
    - { event: "segment_start", data: { segment_id, reason } }
    - { event: "user_input_needed", data: { image_id, reason } }
    - { event: "complete", data: { summary } }

POST /jobs/{job_id}/anchor
  Body: { image_id, lat, lon }
  Manual user GPS input for an image

GET /jobs/{job_id}/point-to-gps?image_id=X&px=100&py=200
  Returns: { lat, lon, confidence }
  Interactive point-to-GPS lookup

GET /jobs/{job_id}/results
  Returns: full results as GeoJSON or CSV

Component: Interactive Point-to-GPS Lookup

For each processed image, the system stores the estimated camera-to-ground homography (from either satellite matching or VO+estimated pose). Given a pixel coordinate (px, py) in an image:

1. If image has satellite match: use the computed homography to project (px, py) → satellite tile coordinates → WGS84. High confidence.
2. If image has only VO pose: use camera intrinsics + estimated altitude + estimated heading to ray-cast (px, py) to the ground plane → WGS84. Medium confidence.
3. Both methods return confidence score based on the underlying position estimate quality.

Testing Strategy

Integration / Functional Tests

• End-to-end pipeline test using provided 60-image sample dataset with ground truth GPS
• Verify 80% of positions within 50m of ground truth
• Verify 60% of positions within 20m of ground truth
• Test sharp turn handling: simulate turn by reordering/skipping images
• Test segment creation and reconnection
• Test user manual anchor injection
• Test point-to-GPS lookup accuracy against known coordinates
• Test SSE streaming delivers results within 1s of processing completion
• Test with FullHD resolution images (degraded accuracy expected, but pipeline must not fail)

Non-Functional Tests

• Processing speed: <5s per image on RTX 2060 (target <2s)
• Memory: peak RAM <16GB, VRAM <6GB during 3000-image flight
• Memory leak test: process 3000 images, verify stable memory
• Concurrent jobs: 2 simultaneous flights, verify isolation
• Tile cache: verify tiles are cached and reused across jobs
• API: load test SSE connections (10 simultaneous clients)
• Recovery: kill and restart service mid-job, verify job can resume

Security Tests

• API key authentication enforcement
• Google Maps API key not exposed in responses or logs
• Image folder path traversal prevention
• Input validation (GPS coordinates, camera parameters)
• Rate limiting on API endpoints

References

• YFS90/GNSS-Denied-UAV-Geolocalization — <7m MAE without IMU
• AerialPositioning — tile engine and deep matcher integration reference
• NaviLoc (2025) — trajectory-level visual localization, 19.5m MLE
• ITU Thesis (2025) — ORB-SLAM3 + SIM integration
• Mateos-Ramirez et al. (2024) — VO + satellite correction for fixed-wing UAV
• LightGlue (ICCV 2023) — feature matching
• SuperPoint — feature extraction
• DALGlue (2025) — 11.8% improvement over LightGlue
• SCAR (2026) — satellite-based aerial calibration
• DUSt3R/MASt3R evaluation (2025) — extreme low-overlap matching
• FastAPI SSE docs
• Google Maps Tiles API

Related Artifacts

• AC assessment: _docs/00_research/gps_denied_visual_nav/00_ac_assessment.md
• Comparison framework: _docs/00_research/gps_denied_visual_nav/03_comparison_framework.md
• Reasoning chain: _docs/00_research/gps_denied_visual_nav/04_reasoning_chain.md