Skip to content

CHANGELOG.md

Latest commit

 

History

History
57 lines (46 loc) · 2.35 KB

CHANGELOG.md

File metadata and controls

57 lines (46 loc) · 2.35 KB
Raw

Changelog

All notable changes to Edge Traffic Vision are documented in this file.

The project follows semantic versioning for repository releases. The changelog records engineering evidence and does not imply production readiness.

[Unreleased]

Pending

  • Continuous road-video validation using a naturally changing video source
  • Final decision on whether version 0.1.0 is a normal or prerelease release
  • Convert this section to a dated 0.1.0 entry immediately before tagging

Added

  • Repository, package, configuration, and test foundations
  • Deterministic BDD100K ingestion and dataset validation
  • Annotation-balanced pilot selection across ten road-object classes
  • CPU training readiness, smoke training, and controlled 40-epoch training
  • Confidence-threshold operating-point selection
  • Reproducible ONNX export and artifact validation
  • PyTorch and ONNX Runtime raw-backend comparison
  • Low-light, blur, and Gaussian-noise robustness evaluation
  • Config-driven ONNX Runtime video and camera inference
  • Warm-up-aware component and source-to-output video profiling
  • Representative 360p, 720p, and 1080p multi-scene validation
  • CPU-only, non-root Docker packaging
  • GitHub Actions tests, linting, Docker build, and smoke validation
  • Release metadata, release notes, checklist, and package-build workflow
  • MIT License and SPDX package metadata
  • Tag-only release workflow with tag-to-commit verification
  • Pull-request and main-branch package preflight validation

Verified engineering results

  • Balanced pilot: 2,500 images and 45,443 annotations
  • Selected 40-epoch model: mAP50 0.14127 and mAP50-95 0.07561
  • Selected confidence 0.20: precision 0.51974, recall 0.32730, and F1 0.40166
  • Controlled ONNX Runtime raw-backend speedup: 3.317x
  • Largest measured robustness weakness: Gaussian noise
  • Severe Gaussian-noise mAP50 drop: 39.71 percent
  • Representative video validation completed at 360p, 720p, and 1080p
  • Docker image build and smoke checks passed in GitHub Actions

Known limitations

  • The trained model, ONNX artifact, datasets, videos, and generated reports are not bundled with the source package
  • Continuous real road-video validation remains pending
  • Webcam and camera-device forwarding remain environment-dependent
  • Results are based on a limited balanced pilot rather than full-scale production training
  • The current runtime is CPU-only