Real-time traffic detection, tracking and counting — AIMS Senegal CV 2026
# 🇸🇳 National Traffic Intelligence System
**AIMS Senegal — Computer Vision Project 2026**
> Real-time detection, tracking, and counting of road-traffic objects across multiple video scenes, built for the Senegalese urban context and deployed publicly on Fly.io.
🌐 **Live Demo:**
aims-traffic-monitor-marble…
📁 **GitHub:**
github.com
---
## 👥 Authors
- Charlotte Chepngeno Sigei
- Fatou Binetou Mbaye
- Akoto Juducael Eric
**Supervisor:** Jordan Felicien Masakuna
**Institution:** African Institute for Mathematical Sciences (AIMS) Senegal
**Academic Year:** 2025–2026
---
## 📌 Overview
This system processes traffic video footage frame by frame, assigns persistent unique IDs to every detected object using ByteTrack, and exposes results through a Flask web interface with live MJPEG streaming and a Chart.js analytics dashboard.
The model was fine-tuned on the VisDrone2019-DET dataset using YOLOv11m on a Kaggle Tesla T4 GPU across two training stages (10 + 20 epochs), covering 10 traffic object classes including tricycles (tuktuks) common in West Africa.
---
## 🚀 Key Features
| Feature | Detail |
|---------|--------|
| **Detection** | YOLOv11m fine-tuned on VisDrone — 10 classes |
| **Tracking** | ByteTrack via Ultralytics — persistent unique IDs |
| **Counting** | Unique object counts, not per-frame detections |
| **Tuktuk logic** | Tricycle classes relabelled as "tuktuk" for Senegal context |
| **Live streaming** | MJPEG stream with colour-coded bounding boxes and track IDs |
| **Video upload** | Drag-and-drop or browse to upload MP4/AVI/MOV |
| **Class filter** | Toggle which object types to track per session |
| **Confidence slider** | Adjustable detection threshold (0.10 – 0.90) |
| **CSV logging** | 19-column structured logs saved after each session |
| **Analytics dashboard** | Cross-scene charts, timeline, class breakdown (Chart.js) |
| **Public deployment** | Hosted on Fly.io — ac …