Computer-vision traffic counter
Vehicle Counting Agent
Counts vehicles in traffic footage with YOLO11 detection and multi-object tracking, so each vehicle is counted exactly once as it crosses a virtual line.
- Status
- Repository
- Type
- Experiment
- Year
- 2026
- Role
- Solo project
01Overview
A small agent that perceives a traffic scene, remembers each tracked vehicle, decides whether a genuine line crossing happened, and acts by updating the count and the video overlay.
02Goal
Count reliably — detection alone has no memory between frames, so a car on screen for three seconds produces 75 unrelated detections.
03Approach
- YOLO11 detects vehicles in one forward pass; only car, motorcycle, bus and truck classes are kept.
- BoT-SORT tracking gives every vehicle a stable ID, which is what makes counting once possible.
- A dead band either side of the line stops jittery boxes on the line from flipping sides and double-counting.
- A track must survive a minimum number of frames before it can count, filtering out flicker.
04Result on the sample clip
On the included 48-second, 1280×720 CCTV clip it counts 32 cars, 3 buses and 1 motorcycle — 36 in total. The single truck in frame is already past the line and moving away, so a count of zero trucks is correct.
05Outcome
Command-line tool with configurable model, line position, confidence, dead band and tracker; runs on video files or a webcam.
06Stack
- Python
- YOLO11
- Ultralytics
- BoT-SORT
Contact
Have a project or a role in mind? Let’s talk.
Send a message through the contact form, or email me directly.