RoadLens AI

Inspiration

Traffic cameras are everywhere, but most of them mainly record video. When something important happens, someone still has to review footage, find the right vehicle, understand what happened, and collect the useful information manually.

We wanted to explore a different approach: what if an ordinary camera could understand traffic activity in real time and automatically turn important events into structured, reviewable reports?

That idea became RoadLens AI.

What it does

RoadLens AI turns a camera-enabled device into an AI-powered traffic intelligence system.

The application can:

• Detect vehicles, pedestrians, bicycles, motorcycles, buses, and trucks in real time
• Track the same vehicle as it moves through the scene
• Estimate vehicle speed when the camera is fixed and calibrated
• Compare measured speed against a configured speed limit
• Automatically generate reports when sustained speeding is detected
• Analyze readable license plates
• Compare plate readings across multiple frames instead of trusting a single OCR result
• Include vehicle information, speed, plate information, and evidence in reports
• Allow reports and traffic activity to be viewed remotely from another device
• Export traffic reports for additional review

The main goal is to move from raw video monitoring to event-based traffic intelligence.

Instead of reviewing every frame, operators can focus on the events that actually matter.

How we built it

RoadLens is built as a web-based traffic monitoring platform.

The frontend uses React, TypeScript, and Vite for the camera interface, traffic visualization, reports, and remote viewer.

Computer vision is used for real-time vehicle and pedestrian detection. RoadLens also maintains persistent vehicle tracking so the same vehicle can be followed across multiple frames.

For speed analysis, a fixed and calibrated camera maps vehicle movement in the video to real-world road coordinates. RoadLens then evaluates the measured speed against a configurable speed policy.

License plate recognition uses a separate plate detection and OCR pipeline. Instead of relying on one frame, RoadLens can collect multiple plate observations and use consensus to improve the final result.

The system also includes a temporary reporting layer that combines vehicle information, speed data, plate information when readable, and supporting evidence into one event report.

RoadLens can run traffic detection through a local AI worker, while also supporting browser-based inference as a fallback.

Challenges we ran into

One of the biggest challenges was maintaining the identity of the same vehicle across changing frames. Vehicle classifications and bounding boxes can change slightly over time, which can cause tracking IDs to switch.

Speed estimation was another major challenge. A moving handheld camera can create misleading speed measurements, so we designed RoadLens to only provide absolute vehicle speed when the camera is stationary and calibrated.

License plate recognition was also difficult because plates can be small, blurry, angled, or partially hidden. We improved this by analyzing higher-resolution vehicle crops and comparing multiple OCR readings instead of trusting a single frame.

Another challenge was keeping the live camera experience smooth while running AI analysis in parallel.

Accomplishments that we're proud of

We built an end-to-end traffic intelligence workflow rather than only an object detection demo.

RoadLens can connect detection, tracking, speed analysis, license plate recognition, evidence capture, and reporting into one system.

We are especially proud of the automatic speeding workflow:

Vehicle detected → vehicle tracked → speed analyzed → threshold exceeded → traffic report created

We also designed RoadLens so that reports and session information are temporary by default instead of building a permanent database of traffic activity.

What we learned

This project showed us that computer vision is only one part of building a useful real-world AI system.

Detecting a car is relatively easy compared with maintaining its identity, measuring its movement correctly, deciding when an event matters, handling uncertain OCR results, and presenting everything in a form that an operator can actually use.

We also learned how important it is to design AI systems that can say when they are uncertain instead of always returning an answer.

What's next for RoadLens AI

Our next goal is to expand RoadLens from traffic observation into broader traffic-event intelligence.

Possible next steps include:

• Collision and abnormal-event detection
• Wrong-way driving detection
• Near-miss and pedestrian safety analysis
• Traffic flow and congestion analytics
• Multi-camera traffic intelligence
• Integration with traffic operations and public-safety workflows
• Incident evidence packages for authorized insurance or claims workflows
• Deployment with compatible existing city camera infrastructure

The long-term vision is simple:

Make existing cameras intelligent instead of replacing the entire infrastructure.

Built With

• computer
• cuda
• express.js
• fast-plate-ocr
• html5
• javascript
• learning
• machine
• node.js
• ocr
• onnx
• opencv
• python
• pytorch
• react
• render
• runtime
• typescript
• vercel
• vision
• vite
• web
• websockets
• yolo
• zod