Project · Jan – Apr 2025 · Team of 4
AI Traffic Management System
Fixed timers give an empty road the same green as a full one. This system watches each road with a camera, counts the vehicles with OpenCV, and shares the green time by demand. When a microphone picks up an ambulance siren, it clears the way.
- Presented at Universiti Sains Islam Malaysia, Kuala Lumpur
- Python · OpenCV · NumPy · FastAPI
- My part: the signal logic and the pitch
A working rebuild of our competition project: a Python package, a test suite, and the web API this page calls.
See it run
Output of python -m traffic junction on four approach videos and one microphone recording. Each tile is a camera: green boxes are vehicles OpenCV found, the yellow line is where they are counted, and the inset is the foreground mask. The bar along the bottom is the signal controller and the siren detector.
Signal plan, cycle by cycle
Each cycle's counts set the next cycle's greens. This run used a 40-second demo cycle (5 to 20 seconds per road) so several cycles fit in the clip.
| Cycle | Starts | Green N · E · S · W (s) | Counted N · E · S · W |
|---|
Counting accuracy on the sample footage
Vehicles heading to the junction that crossed the line, against what the cameras counted.
| Road | Actual | Counted | Accuracy |
|---|
Misses are mostly dark vehicles on dark asphalt and vehicles too close to separate. That is the main limit of background subtraction.
Try the API
Checking the API…These call the same Python code running on the server. Use the sample files or your own.
1 · Count vehicles in a video
A fixed camera looking along a road, under 4 MB. The first 30 seconds are processed.
2 · Turn counts into green times
Vehicles counted on each road last cycle. 120-second cycle, 10 to 60 seconds each.
3 · Listen for a siren
A WAV recording from the roadside, under 4 MB. An FFT compares the siren band (600 to 1,600 Hz) with the rest of the sound.
How the code works
Four small modules, each tested on its own, joined by junction.py.
-
Find what moves
traffic/detector.py· MOG2 learns the empty road and marks anything new. Shadows come back grey, so a threshold drops them. Opening removes specks, dilation joins broken shapes, and contours outline each vehicle.gray = cv2.GaussianBlur(cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY), (5, 5), 0) fg = self.mog.apply(gray) _, fg = cv2.threshold(fg, 200, 255, cv2.THRESH_BINARY) # shadows are 127 fg = cv2.morphologyEx(fg, cv2.MORPH_OPEN, self.k_open) fg = cv2.dilate(fg, self.k_dilate) contours, _ = cv2.findContours(fg, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE) found = [cv2.boundingRect(c) for c in contours if cv2.contourArea(c) >= self.cfg.min_area] -
Count each vehicle once
traffic/tracker.py· Each centre is matched to the nearest track from the last frame. A vehicle counts when its centre crosses the line in the direction of the junction, and only once.pairs.sort() # closest matches first for d, ti, ci in pairs: ... prev_y = tr.y tr.x, tr.y = centres[ci] if not tr.counted and prev_y < self.line_y <= tr.y: self.down += 1 # crossed, heading to the junction tr.counted = True -
Share the green
traffic/signal.py· Next cycle, each road's green is its share of the count, kept between a minimum and a maximum. The controller then runs green, yellow and all-red for each road in turn.def green_times(counts, cycle=120, min_green=10, max_green=60): total = sum(counts) if total == 0: return [round(min(max_green, max(min_green, cycle / len(counts))))] * len(counts) return [round(min(max_green, max(min_green, n / total * cycle))) for n in counts] green_times([24, 14, 10, 2]) # [58, 34, 24, 10] -
Hear the siren
traffic/siren.py· Every quarter second, an FFT measures power inside the siren band against power outside it. Above 4× for 0.8 seconds counts as a siren, which ends the current green safely and holds green on that road.power = np.abs(np.fft.rfft(window * np.hanning(len(window)))) ** 2 freqs = np.fft.rfftfreq(len(window), 1 / rate) inside = (freqs >= 600) & (freqs <= 1600) ratio = power[inside].mean() / power[~inside].mean() # ratio >= 4.0 for 0.8 s in a row -> controller.emergency(road)
Project layout
traffic/
config.py every tunable number
detector.py OpenCV background subtraction
tracker.py centroid tracking, line counting
signal.py green times, phases, emergencies
siren.py FFT siren detection
junction.py runs it all on four videos
cli.py python -m traffic …
api/index.py FastAPI, served on Vercel
tests/ 21 pytest tests
scripts/make_samples.py
Run it yourself
pip install -r requirements.txt
python scripts/make_samples.py
python -m traffic plan 24 14 10 2
python -m traffic count samples/north.mp4 --show
python -m traffic junction \
--north samples/north.mp4 --east samples/east.mp4 \
--south samples/south.mp4 --west samples/west.mp4 \
--siren samples/siren_east.wav --siren-road east \
--out output/junction.mp4 --json output/results.json
pytest
Limits and next steps
- Background subtraction struggles at night, in rain, and with shadows that move.
- Vehicles bumper to bumper merge into one shape, so jams are undercounted, which is when counts matter most.
- Counting passing vehicles measures flow, not how long the queue is or how long people have waited.
- One junction on its own; it does not coordinate with the next signal.
Next: a trained detector such as YOLO for jams and night, queue length as the input, and linked junctions. A pilot would compare average wait and ambulance clearance time against the fixed timer at one junction.