The goal of this project was to design and deploy an AI-powered traffic sign recognition system for autonomous driving applications.
The system had to be accurate, fast, and lightweight enough to run on edge hardware in real time.
- Gathered and labeled 1,400+ images across 6 traffic sign classes.
- Used Roboflow for data preprocessing, augmentation, and annotation.
- Prepared the dataset for training with YOLOv8.
- Trained a YOLOv8 object detection model on the custom dataset.
- Experimented with hyperparameters to balance accuracy vs. speed.
- Validated performance using test images with diverse lighting and backgrounds.
- Deployed the optimized model on a Jetson Nano connected to an OAK-D depth camera.
- Integrated real-time video feed for live detection and classification of traffic signs.
- Applied optimizations to ensure inference stability under hardware constraints.
- Conducted tests in simulated road environments.
- Measured frame rate, detection accuracy, and latency.
- Compared results against project expectations for real-time performance.
- Successfully achieved real-time traffic sign recognition at ~14 FPS on Jetson Nano.
- Model detected and classified 6 types of traffic signs with strong accuracy.
- System remained stable under constrained edge-computing resources.
- Identified challenges in distinguishing visually similar signs under poor lighting — valuable insight for future improvement.
- Complete end-to-end computer vision pipeline: data collection → training → evaluation → deployment.
- Hands-on experience in deep learning (YOLOv8), computer vision, and edge AI deployment.
- Clear demonstration of the trade-offs between model complexity, accuracy, and hardware efficiency.
Due to academic restrictions, the full report and source code cannot be shared publicly.
However, the complete project report is available on request.
Maryam Jahangir
Master’s in Automation & IT (Data Science)
Machine Learning | Computer Vision | Edge AI | Optimization