Hum Rahi

User Story 2 (BE/AI): Image-Based Incident Classification¶

Story: As a Backend System, I want to utilize Computer Vision to identify the incident type from the photo, so that the report can be categorized even if the user provides no text.

• In-Scope:
  • Computer Vision model integration.
  • Confidence score generation for each category.
  • Fallback mechanism to "User Selected Category" if AI confidence is low (< 0.70).
• Out-of-Scope: Video analysis; facial recognition of individuals.
• Acceptance Criteria (AC):
  1. AI must classify photos into the predefined list (Pothole, Accident, etc.).
  2. Model must return a primary category and a confidence percentage.
  3. Processing must initiate immediately upon file storage confirmation.
• Validation Rules:
  • Confidence Score: 0.0 to 1.0.
• Error Messages:
  • "AI Engine: Low image quality, unable to classify."
• DoD: Model accuracy >85% on test dataset, unit tests for AI worker.