Squad B (Vision) Roadmap

Catarman LGU Climate-Resilient SMART Traffic Light 8-week sprint · Oct 19 – Dec 14, 2026 · 3 members

Key point: The stock YOLOv8 model has no tricycle or jeepney class and struggles with small toy vehicles seen from above. We must train our own model on photos of the exact toy vehicles used in the diorama. Counting, PCU scoring and MQTT are standard once detection works.


Roles

Member 1: Dataset & Model Lead

Owns detection accuracy. If the model can't tell a tricycle from a motorcycle, nothing downstream works.

Member 2: Vision Pipeline & Integration Lead

Owns the live system: camera → detection → counts → PCU → ESP32.

Member 3: Violations, Data & Dashboard Bridge

Owns everything the dashboard and judges see from vision.


Hardware

Item Purpose If we're lacking
Laptop (ideally NVIDIA GPU, 8GB+ RAM) Runs YOLO, MQTT, servers Intel CPU only: OpenVINO export at imgsz 416. Any CPU: ONNX Runtime + process every 2nd frame.
Smartphone, 1080p camera Overhead camera USB webcam works; 720p is enough at diorama scale. Bring a second phone as backup.
Overhead mount Holds camera fixed above diorama Gooseneck phone holder, ring-light stand, or DIY boom from PVC pipe/wood clamped to a table. Selfie stick taped to a chair works for early data collection. It must not move, or zones go out of alignment.
USB cables (phone + ESP32) DroidCam over USB, serial fallback DroidCam over USB instead of Wi-Fi = less lag, avoids venue interference.
Toy vehicles (several per class, same scale) Training data and demo Tricycle toys are hard to find: add a cardboard or 3D-printed sidecar to a toy motorcycle. Jeepney: toy van with a painted roof rack. Use the exact same toys for training and demo.
Desk lamp or LED strip Consistent lighting Any bright, steady lamp. Lighting changes between training and demo are a major cause of accuracy drops.
Pocket Wi-Fi or spare phone hotspot Private local network Don't rely on school or venue Wi-Fi.
Laptop cooling pad Prevents thermal throttling Raise the laptop so air flows underneath; keep it plugged in.

Software

Tool Purpose Alternatives
Python 3.11 + VS Code Main language/editor PyCharm
ultralytics YOLOv8 training and inference —
opencv-python Camera feed, image processing —
supervision (Roboflow) Ready-made zone counting, line crossing, box drawing Write your own with OpenCV
paho-mqtt + Mosquitto Messaging with ESP32 and dashboard —
MQTT Explorer View MQTT messages live while debugging mosquitto_sub in terminal
pyserial USB serial fallback —
Flask or FastAPI MJPEG stream + JSON API for Squad C WebSocket with base64 frames
sqlite3 (built in) + pandas Logging and CSV export Plain CSV files
Roboflow (free tier) Labeling, augmentation, YOLO export CVAT, Label Studio, labelImg (fully offline)
Google Colab Free GPU training Kaggle Notebooks (~30 free GPU hrs/week)
OpenVINO / ONNX Runtime Faster inference on CPU —
Git + GitHub Shared repo (/edge_vision) —
DroidCam or IP Webcam app Phone as camera Iriun Webcam

MQTT schema (draft, must be agreed with Squads A & C in Week 1)

topic: traffic/vision/approach
{"approach":"N","pcu":4.5,"counts":{"tricycle":3,"motorcycle":2,"car":2},"confidence":0.82,"ts":1729312000}

topic: traffic/vision/heartbeat   (every 1 second)
{"status":"ok","fps":17.3}

Suggested PCU weights (confirm with Squad A): car 1.0 · tricycle 0.5 · motorcycle 0.3 · jeepney 2.0 · van 1.5


Weekly to-do

Week 0: Oct 19–20 (before hardware freeze)

Week 1: Oct 20–26 (Setup & baseline)

Week 2: Oct 27–Nov 2 (First custom model)

Week 3: Nov 3–9 (Connect to ESP32)

Week 4: Nov 10–16 (Diorama retraining)

Week 5: Nov 17–23 (Fail-safes)

Week 6: Nov 24–30 (Stress testing, all squads)

Week 7: Dec 1–7 (Code freeze)

Week 8: Dec 8–14 (Rehearsal)


Watch out