Traffic Sense AI is a research-driven, full-stack system built for real-time multi-stream traffic analysis, powered by enhanced YOLO variants — including our custom YOLO-FDE (Feature Dynamic Enhanced) model.
Designed to deliver live vehicle detection, counting, analytics, and model benchmarking, this platform merges deep learning, computer vision, and full-stack engineering into one seamless solution.
Traffic Sense AI allows users to run simultaneous object detection across up to 4 video sources — whether live streams, YouTube links, or uploaded videos — all through an intuitive web dashboard.
It features:
- Multi-model YOLO integration (YOLOv8, YOLOv8-FDD, FDIDH, DySample, and our proposed YOLO-FDE).
- Live real-time inference and MJPEG streaming using Flask.
- Interactive performance analytics dashboard with mAP and loss comparisons.
- Fully persistent sessions with SQLite backend and RESTful API.
- Clean, reactive frontend using React + TypeScript + TailwindCSS.
| Category | Description |
|---|---|
| 🚘 Real-time Inference | Run up to 4 concurrent live or offline video detections. |
| 🎥 Video Input Support | Supports RTSP, HTTP, YouTube (via yt-dlp), or local uploads. |
| 📊 Analytics Dashboard | Displays saved sessions, model performance charts, and statistics. |
| 🧠 YOLO-FDE Architecture | Integrates FDIDH, DWR, DySample, and Post-SPPF Attention modules. |
| 💾 Session Management | Save inference results to an SQLite database via Flask backend. |
| 🧱 Model Comparison | Benchmark YOLOv8-FDE vs YOLOv8, FDIDH+DWR, DySample, etc. |
| 🧮 Lightweight Yet Accurate | Achieves 0.9242 mAP@50 at only 2.69M params and 3.5 GFLOPs. |
Figure: YOLOv8-FDE architecture with integrated FDIDH, DWR, DySample, and Post-SPPF Attention modules.
The proposed YOLO-FDE introduces dynamic feature interaction and adaptive sampling for enhanced detection accuracy and generalization.
It outperforms other YOLOv8 variants on the UA-DETRAC dataset, with the best mAP@50–95 = 0.8159 while maintaining minimal computational overhead.
For detailed research methodology and ablation studies, refer to our extended repository:
👉 YOLO-FEE: Feature Experimentation and Enhancement
| Model | Params (M) | GFLOPs | Precision | Recall | mAP@50 | mAP@50–95 |
|---|---|---|---|---|---|---|
| YOLOv8-N (Baseline) | 3.01 | 4.10 | 0.8689 | 0.8315 | 0.9090 | 0.7680 |
| YOLOv8-FDIDH + DWR | 13.57 | 12.50 | 0.9056 | 0.8319 | 0.9003 | 0.7603 |
| YOLOv8-FDIDH + DySample | 21.19 | 15.00 | 0.8676 | 0.8310 | 0.8980 | 0.7207 |
| YOLOv8-FDD | 16.56 | 40.02 | 0.8433 | 0.7919 | 0.8717 | 0.6797 |
| YOLOv8-FDE (Proposed) | 2.69 | 3.50 | 0.9077 | 0.8806 | 0.9242 | 0.8159 |
Machine Learning / CV
- PyTorch
- Ultralytics YOLOv8
- YOLO-FDE (Custom Architecture)
- Supervision
- OpenCV
- NumPy
- Matplotlib
Backend
- Flask
- Flask-CORS
- SQLAlchemy
- SQLite
- Werkzeug
yt-dlpfor YouTube video streams- Threading & OS Path Utilities
Frontend
- React + TypeScript
- TailwindCSS
- Framer Motion
- Recharts
- React Router
- Lucide Icons
Utilities
- Fetch API
- LocalStorage API
- Concurrent Development (Flask + Vite)
Traffic Sense AI is grounded in deep architectural refinements introduced in YOLO-FDE (Feature Dynamic Enhanced).
The model integrates four critical modules:
- FDIDH — Enables deformable feature interaction between classification and regression heads.
- DySample — Dynamically learns offset-based sampling for high-resolution localization.
- DWR — Uses dilation-wise residuals for efficient multi-scale context expansion.
- Post-SPPF Attention — Amplifies discriminative cues under occlusion and illumination variance.
- +6.4% mAP@50–95 improvement vs YOLOv8 baseline.
- ~33% fewer parameters with superior accuracy.
- Significant FPS gain (3.5 GFLOPs only).
- Robust under poor lighting, occlusion, and dense traffic.
- Deployed successfully in real-time (RTX 4060) .
git clone https://github.com/pnihal6/Traffic-Sense-AI.git
cd Traffic-Sense-AIcd backend
pip install -r requirements.txtcd ../frontend
npm installnpm startThis uses concurrently to start Flask + React simultaneously.
For in-depth architecture details, ablation studies, and other YOLO variants, visit the companion repository:
👉 YOLO-FEE: Feature Experimentation and Enhancement
Team Traffic Sense AI
- Priyadarshi Nihal — Full Stack ML Engineer (Team Lead)
- Dev Tailor — Full Stack Developer
- Pratyush Dubey — AI & ML Engineer
- Dattatrey — Frontend Developer
- Shivalik Mathur — Backend Developer
Supervisor:
🎓 Dr. I. Jasmine Selvakumari Jeya, Assistant Dean, VIT Bhopal University
This project is released under the MIT License.
Feel free to use, modify, and cite with attribution to the original authors and research team.
We deeply thank our mentor, Dr. Jasmine, for her guidance, and the VIT Bhopal for support.
Credits to Ultralytics, Supervision, and YOLOv8 open-source communities for their foundational contributions.
If you use this work or YOLO-FDE in your research, please cite:
@software{TrafficSenseAI2025,
title={Traffic Sense AI: Real-Time Multi-Stream Traffic Monitoring using YOLO-FDE},
author={Priyadarshi Nihal, Dev Tailor, Pratyush Dubey, Dattatrey, Shivalik Mathur},
year={2025},
institution={VIT Bhopal University},
url={https://github.com/pnihal6/Traffic-Sense-AI}
}
Traffic Sense AI represents a fusion of research-grade deep learning and production-grade engineering, enabling scalable, intelligent, and real-time traffic surveillance with YOLO-FDE at its core.
Smarter Cities, Safer Roads — Powered by Intelligent Vision.