gps-denied-onboard/docs/03_tests/01_sequential_visual_odometry_integration_spec.md

Integration Test: Sequential Visual Odometry (Layer 1)

Summary

Test the SuperPoint + LightGlue sequential tracking pipeline for frame-to-frame relative pose estimation in continuous UAV flight scenarios.

Component Under Test

Component: Sequential Visual Odometry (Layer 1) Technologies: SuperPoint (feature detection), LightGlue (attention-based matching) Location: gps_denied_07_sequential_visual_odometry

Dependencies

• Model Manager (TensorRT models for SuperPoint and LightGlue)
• Image Input Pipeline (preprocessed images)
• Configuration Manager (algorithm parameters)

Test Scenarios

Scenario 1: Normal Sequential Tracking

Input Data:

• Images: AD000001.jpg through AD000010.jpg (10 consecutive images)
• Ground truth: coordinates.csv
• Camera parameters: data_parameters.md (400m altitude, 25mm focal length)

Expected Output:

• Relative pose transformations between consecutive frames
• Feature match count >100 matches per frame pair
• Inlier ratio >70% after geometric verification
• Translation vectors consistent with ~120m spacing

Maximum Execution Time: 100ms per frame pair

Success Criteria:

• All 9 frame pairs successfully matched
• Estimated relative translations within 20% of ground truth distances
• Rotation estimates within 5 degrees of expected values

Scenario 2: Low Overlap (<5%)

Input Data:

• Images: AD000042, AD000044, AD000045 (sharp turn with gap)
• Sharp turn causes minimal overlap between AD000042 and AD000044

Expected Output:

• LightGlue adaptive depth mechanism activates (more layers)
• Lower match count (10-50 matches) but high confidence
• System reports low confidence flag for downstream fusion

Maximum Execution Time: 200ms per difficult frame pair

Success Criteria:

• At least 10 high-quality matches found
• Inlier ratio >50% despite low overlap
• Confidence metric accurately reflects matching difficulty

Scenario 3: Repetitive Agricultural Texture

Input Data:

• Images from AD000015-AD000025 (likely agricultural fields)
• High texture repetition challenge

Expected Output:

• SuperPoint detects semantically meaningful features (field boundaries, roads)
• LightGlue dustbin mechanism rejects ambiguous matches
• Stable tracking despite texture repetition

Maximum Execution Time: 100ms per frame pair

Success Criteria:

• Match count >80 per frame pair
• No catastrophic matching failures (>50% outliers)
• Tracking continuity maintained across sequence

Performance Requirements

• SuperPoint inference: <20ms per image (RTX 2060/3070)
• LightGlue matching: <80ms per frame pair
• Combined pipeline: <100ms per frame (normal overlap)
• TensorRT FP16 optimization mandatory

Quality Metrics

• Match count: Mean >100, Min >50 (normal overlap)
• Inlier ratio: Mean >70%, Min >50%
• Feature distribution: >30% of image area covered
• Geometric consistency: Epipolar error <1.0 pixels