Vehicle Detection and Classification using OpenCV and CNNs

by Ranjeet Reddy Manchana

← Projects Model: loading… View Project Code

Live demo · runs in your browser

Count and classify vehicles

Upload a traffic video and the model (YOLOv5 via ONNX Runtime Web, with OpenCV.js) detects, tracks and counts vehicles crossing the line. Nothing is uploaded to a server: your video is processed locally.

Jump to the demo ↓ GitHub repo

What it doesCount and classify traffic, entirely in the browser

EagleEye (the repo is called vehicle-counter) takes a traffic video, finds the vehicles in every frame, follows each one from frame to frame, sorts it into light, medium or heavy, and counts it once as it crosses a line. There is no server and no upload: the video plays locally, each frame is drawn to a canvas, and the neural network runs in your browser through WebAssembly.

Frame→Letterbox 640×640→YOLOv5s→Filter + NMS→Track→Classify→Count

How it worksFive stages per frame

Detect. OpenCV.js converts each frame to RGB and letterboxes it to 640×640 (aspect ratio kept, grey padding). Pixels are scaled to 0–1 and passed to a YOLOv5s model (ONNX, about 28 MB) running on ONNX Runtime Web's WebAssembly backend.
Filter. The model returns 25,200 candidate boxes. I keep those scoring at least 0.35 (objectness × class score) in the five COCO vehicle classes: bicycle, car, motorcycle, bus and truck. Non-maximum suppression at IoU 0.45 removes duplicates, and the boxes are mapped back to the original frame.
Track. A small custom tracker matches each detection to an existing track by box overlap (IoU above 0.1), falling back to centroid distance (under 80 px). A track that goes unmatched survives 10 more frames, which covers brief occlusions.
Classify. The class and the box's size relative to the frame decide the bucket (table below).
Count. A dashed line sits 60% of the way down the frame. When a tracked vehicle's centre moves from one side of the line to the other, it is counted, in either direction, and each track is counted only once. Seeking backwards resets the counters.

ClassificationType first, then size

Weight class is a heuristic, not a measurement: it combines what the model says the object is with how much of the frame its box covers.

Detected as	Rule	Bucket
Bus	Always	Heavy
Truck	Box covers more than 4% of the frame, otherwise…	Heavy / Medium
Car	Box covers more than 2.5% of the frame, otherwise…	Medium / Light
Motorcycle, bicycle	Always	Light

Engineering notesTrade-offs and limits

Why in the browser. Nothing is uploaded, which matters for traffic footage, and the whole thing is a static site with nothing to host or pay for. The cost is that inference is CPU-bound: the repo's architecture notes estimate roughly 60–170 ms per frame, so expect single digits to mid-teens frames per second on a laptop.
The tracker is deliberately simple. It has no appearance model, so a vehicle hidden for longer than the 10-frame grace period can come back with a new ID and be counted twice.
Classes depend on the camera. Because size is measured against the frame, the same car can land in a different bucket if the camera is zoomed or mounted differently. A stable camera angle gives the best results.
Next steps I've noted: a WebGPU backend, Web Workers for pre- and post-processing, a stronger tracker (ByteTrack or DeepSORT), per-direction counts, speed estimation and a quantised INT8 model.

Built with YOLOv5 by Ultralytics, ONNX Runtime Web and OpenCV.js. The model file is served from this site, so the page needs to be opened over HTTP, not as a local file.

Live demo

Try it with your own video

Upload Video

Drag & drop a traffic video here

Browse video

0:000:00

Please upload video input.

Light Medium Heavy

Vehicle Detection and Classification using OpenCV and CNNs

Count and classify vehicles

What it doesCount and classify traffic, entirely in the browser

How it worksFive stages per frame

ClassificationType first, then size

Engineering notesTrade-offs and limits

Try it with your own video

Upload Video

Counts

Total

Light

Medium

Heavy