Global road-damage vision
YOLOv8n ONNX trained on 25,619 images across seven RDD2022 source sets. OpenCV 5 performs decoding, preprocessing, CPU DNN inference and class-aware suppression. Four road-damage classes.
OpenCV 5ONNXMulticountry training
R
RoadGuard
Evidence-driven road inspection · Cloud Inspector
AWS-hosted inferenceOpenCV 5 · Global YOLO ONNXMumbai / ap-south-1
FROM VIDEO TO AUDITABLE EVIDENCE
Detect potential road damage.
Explain the evidence. Keep inspectors in control.
Upload a road video. An AWS-hosted CPU runs OpenCV 5 + the global YOLO model, actively reinspects uncertain frames, compares optical flow and saves a traceable case to private S3. Lambda produces the cloud review report.
5Connected AI + cloud stages
Frames inspected—Recorded video frames read
AI candidate groups—Not yet human-confirmed defects
Additional vision calls—Agent-selected frame reinspection
Review candidates—Evidence-backed cases
Run an AWS-hosted inspection
1 · Secure inspector access
This is a publicly reachable AWS website; upload and inspection APIs require a demo access code to prevent abusive compute usage.
Not unlocked
Use the existing cloud report to show evidence even before uploading another video.
2 · Upload a video to AWS for inference
Road video · MP4 / MOV / AVI / MKV
Maximum 80 MB. The first 300 frames are processed by the existing RoadGuard agent (configurable server-side). Uploaded source video and evidence are stored privately in S3.
No upload submitted
Live pipeline status
Load a cloud run or upload a video to see its processing stages.
Run: none selected
Why this is agentic, not ordinary YOLO
YOLO supplies detections. The agent chooses its next observation: it selects extra frames based on confidence and previous support, compares candidate motion with surrounding-road motion, and records every request and result.
Perception → Inspect → Evidence → ReviewHuman-governed decisions
A candidate with a bounding box is an AI finding, not a proven road defect or calibrated severity estimate. The image and frame checks are available in Inspection evidence.
Inspection evidence
Real images, detection classes, model scores and temporal support from the selected cloud inspection. No simulated road-damage confirmations.
Inspection queue
No run loaded
Load a report to view evidence.
Select an event
No dataOriginal and annotated images will appear here after a cloud report loads.
Model confidence—
Nearby-frame support—
Grouped observations—
Model confidence is a detection score, not the probability that a physical defect is truly present. Multiple observations can refer to the same region.
Agent decision trace
Auditable decisions made after perception. Expand each step to see which frames the agent requested and how many supported the candidate.
Select an event to see its agent reasoning trace.
OUR INNOVATIONS · REAL IMPLEMENTATION
RoadGuard doesn't stop at detecting a box.
We built an evidence-driven inspection workflow that asks for more observations, exposes uncertainty, detects possible fixed-camera artifacts, and connects to a live AWS reporting pipeline.
AI →Evidence · Action · Audit
What makes RoadGuard different
Adaptive agentic reinspection
The agent tests additional frame offsets when confidence is low or temporal support is insufficient. Inspection requests and matching outcomes are recorded, not hidden.
Confidence-awareTool callsTemporal evidence
Motion-based artifact checks
Sparse Lucas–Kanade optical flow compares candidate-region motion with surrounding road motion. It raises a possible camera-fixed artifact signal rather than asserting a reflection is proven.
Optical flowHeuristic safety check
Auditable evidence trail
Every candidate includes source frame, box, model confidence, associated observations, reinspection tool results and unmodified/annotated images. Inspectors can challenge the AI's claim.
Explainable workflowTraceable evidence
Actual AWS-side inference
This public UI calls an API hosted on Amazon EC2. The host runs the same CPU OpenCV/ONNX inspection code; private S3 receives source video, evidence and trace; S3 events trigger Lambda to generate review summaries.
EC2 computeS3Lambda
Human-governed decisions
AI findings never become automatically verified road defects. A reviewer can confirm, reject or request further evidence; a timestamped decision history is stored in private S3.
Human-in-the-loopAudit history
Actual deployment architecture
Public Inspector HTMLBrowser upload, evidence, review
AWS EC2 / FastAPIPython API & OpenCV 5 inference
RoadGuard AgentYOLO + reinspection + optical flow
Private Amazon S3Video · image evidence · JSON
S3 → LambdaAsynchronous review summary
S3 Review AuditHuman decisions & timestamps
Measured validation (held-out multicountry images)
mAP@500.509
mAP@50–950.252
Evaluated on 3,833 held-out images. These scores do not measure UK dashcam video performance or prove real-world defect accuracy.
Responsible engineering boundaries
RoadGuard does not infer GPS location, measure crack depth, calculate repair costs or claim infrastructure safety. Temporal groups are image-space associations, not verified distinct physical defects. A camera artifact can remain undetected.
Cloud video inference is on EC2, not inside the S3-triggered Lambda function. A hosted demo is not automatically a horizontally scalable production service.
Human review & audit trail
Inspectors review the model evidence and decide whether each candidate is accepted, rejected or needs further inspection. Decisions are stored in S3, not just in the browser.
Selected inspection candidate
Select an event in Inspection evidence.
This action records a human conclusion about the visible evidence. It does not automatically schedule road repair or verify physical crack dimensions.
Persistent cloud review history
No decisions recorded for this run.
Stored as reviews/<run-id>/decisions.json in a private S3 bucket. Record includes event ID, decision, reviewer label and UTC timestamp.