COOPER: the dashcam that sees it coming.

A Raspberry Pi 5 dashcam that spots people and vehicles, predicts where each one is heading, and lights a warning before one of them enters your lane.

Scroll to see

Problem

Forward-collision warning is standard on new cars. Millions of older ones have none.

Rear-end and pedestrian collisions often come down to a second or two of inattention. Getting that warning usually means buying a newer car or paying for an aftermarket system that may need professional installation. COOPER closes that gap: a Raspberry Pi, a camera and two LEDs that watch the road and warn the driver before a collision, for far less than either costs. It plugs into the car’s power and mounts on the windshield or dash, with no wiring into the car’s systems.

It doesn’t drive, brake or steer, and it doesn’t record footage — it’s a warning light, not a replacement for an attentive driver.

Detects
people & vehicles ahead
Predicts
collision risk, live
Runs
on-device, no internet needed

See

It starts with a frame: brightness values, and nothing that knows what they are.

A 5 MP OV5647 camera on a Raspberry Pi 5 looks down the road through the windshield, powered by the car, and never moves. COOPER reads it 640 × 480 pixels at a time, and every frame goes through the same loop.

Sensor
OV5647, 5 MP
Field of view
63° × 49°
Frame
640 × 480

Detect

Boxes snap onto what matters, and each one keeps its name.

YOLO11n runs on the Pi 5’s CPU at a 320 px input, with no accelerator. It finds people and vehicles, and ByteTrack gives each one an ID that survives from frame to frame. COOPER doesn’t pick a favourite: every object on the road gets its own track, its own motion model and its own predicted path.

Model
YOLO11n, 320 px
Tracker
ByteTrack
Classes
people, cars, buses, trucks, motorcycles

Predict

Watch where it’s going, not where it was.

Every object gets its own constant-velocity Kalman filter, run in image space, so motion is relative to the car. Each frame it predicts (slide the object along its velocity), then updates (blend that guess with the new box, trusting whichever is less noisy). What’s left is a smooth position and velocity for each one.

The velocity draws a path 1.5 s ahead, checked against my lane: a trapezoid over the road just in front of the car. Something coming straight at us barely moves sideways, so COOPER also watches its box grow: height over growth rate is the time to contact.

The path, sampled every 0.1 s to 1.5 s, and time to contact

Warn

Two lights, before it happens.

Yellow: a predicted path will enter my lane soon, like a car about to cut in, or something ahead is closing in under 2 s. Red: something is in my lane now, and red overrides yellow. A level lights after two frames in a row and holds half a second, so one noisy box can’t make it flicker. The dashboard shows the live video, every predicted path and the risk level, and can beep in step with the LEDs.

Yellow
path enters my lane
Red
in my lane now
Horizon
1.5 s

Build

One board, one camera, two lights.

The Pi 5 does all of it on its CPU: detection, tracking, prediction, the lane check and the video stream. The two LEDs sit on its GPIO pins, and the car powers everything. Nothing moves — COOPER started as COOP, a motorized tracking camera, and we removed the motor to point the same vision and prediction pipeline at collision warning instead.

  • OV5647 camera5 MP, 3.6 mm, 75° diagonal
  • Camera mountfixed, facing the road
  • Yellow LEDa path will enter my lane
  • Red LEDsomething is in my lane
  • Lidthe LEDs show through it
  • Raspberry Pi 58 GB, vision and risk
  • Casecar power in over USB-C

Built with: board to browser

  • Raspberry Pi 5 logoRaspberry Pi 5
  • Python logoPython
  • Picamera2 logoPicamera2
  • YOLO11 · Ultralytics logoYOLO11 · Ultralytics
  • PyTorch logoPyTorch
  • NCNN logoNCNN
  • OpenCV logoOpenCV
  • NumPy logoNumPy
  • Flask logoFlask
  • gpiozero logogpiozero
  • systemd logosystemd
  • Cloudflare logoCloudflare
  • three.js logothree.js
  • anime.js logoanime.js
  • JavaScript logoJavaScript
  • Blender logoBlender
  • pytest logopytest
  • GitHub Actions logoGitHub Actions

Privacy

Everything runs on the Pi. Nothing is recorded.

Detection, tracking, prediction and risk all run on the device, and the warnings never depend on an internet connection. Each frame is processed, sent to any open dashboard, and discarded — no video recording, no saved images. COOPER knows a person or a car is there, not who they are: it doesn’t read faces or license plates.

By default the dashboard is reachable only from devices on the same network as the Pi. Making it reachable over HTTPS from anywhere is opt-in, sits behind Cloudflare Access so only approved email addresses can see it, and only passes video while someone is actually watching.

Stored
nothing, ever
Dashboard
same network, by default
Remote view
opt-in, behind login

Next

A rehearsal tool already proves the logic. The next step is real driving.

A synthetic road already tests COOPER end to end: a car cutting in, a pedestrian crossing, the car ahead braking, and a car in the next lane that must never light an LED. Next: validate against the Waymo Open Dataset, comparing COOPER’s time-to-collision estimates against lidar ground truth; add a buzzer on the unit alongside the LEDs (the dashboard already beeps); detect the lane from road markings instead of calibrating it by hand; improve detection at night and in rain; and road-test the thresholds in real cars.

Team

Three people, built in parallel.

Everything runs on a laptop with a webcam and mocked LEDs, so the software and the hardware never had to wait for each other.

  • David Hernandez Backend Backend development including vision, tracking, and prediction.
    LinkedIn
  • Diego Avila CAD & hardware CAD and hardware. Also put together and wired the finished project.
    LinkedIn
  • Diego Tabares Presentation & design Presentation, pitch, and elements of website design.
    LinkedIn

Read the code on GitHub Devpost page