How the
model works,
page by page.
This page explains the actual application flow: the upload path, the detector, the rule engine, the response shape, and the parts of the workspace that make those results useful.
mAP50
92.7%
pipeline
4 stages
modules
detector + rules
output
image + JSON
Upload
JPG · PNG · WebP
YOLOv8
23 classes
Rule engine
8+ checks
Response
JSON + image
92.7%
mAP50
81.0%
mAP50-95
4
Stages
23
Classes
Detector
YOLOv8
object classes and confidence output
Rules
Severity
scene relationships and violation logic
History
Stored
saved analyses feed downstream pages
Analytics
Charts
aggregate trends from the archive
Four operating layers.
The app is intentionally broken into clear stages so the experience feels understandable and inspectable instead of mysterious.
Input handling
Users upload a road frame in JPG, PNG, or WebP. The frontend prepares the payload and sends it to the FastAPI detector.
YOLOv8 detection
The backend runs inference across trained road classes such as cars, trucks, traffic lights, pedestrians, and signs.
Violation logic
A rule layer evaluates object relationships, severity, and scene context to identify traffic issues worth surfacing.
Result delivery
The response includes detections, violation summaries, and an annotated image that can be reviewed or exported.
What each
page is doing.
The app follows one sequence: learn what it does, see how it works, run a detection, review the result, and adjust settings.
What the model sees
Cars, trucks, pedestrians, road signs, signals, and scene objects are returned as structured detections with confidence and bounding boxes.
What the rules produce
Violations are grouped by severity with descriptions and the objects involved, so the review layer reads clearly instead of feeling raw.
What gets stored
History saves file details, generated thumbnails, the structured response, and whether the run came from demo mode.
What users can tune
Settings control autosave behavior and in-app notifications, giving the app real preferences instead of cosmetic toggles.
Instructions for using the app.
Open the detection page and drag in a supported image format.
Run analysis to call the backend or use the sample data path for a local showcase.
Review the annotated frame, detections table, and rule output together.
Open history to revisit prior runs and analytics to inspect aggregate trends.