Welcome to the JetRacer Pro Autonomous Driving project! This repository showcases our team’s efforts in building an autonomous driving car using the NVIDIA Jetson Nano. Our goal is to enable the JetRacer to navigate tracks and avoid obstacles using deep learning and real-time decision-making. 🛤️🏁
This project uses the NVIDIA Jetson Nano and the JetRacer Pro AI Kit to build an autonomous driving car capable of navigating complex tracks and avoiding obstacles, all powered by deep learning. 🚙💡
jetsonNanoDemo.mp4
- Deep Learning Models: Trained and tested models like ResNet-18, ResNet-34, and SqueezeNet for real-time navigation. ResNet-18 turned out to be the champion, balancing speed and accuracy! 🥇
- Data Augmentation: Applied techniques like brightness adjustment, flipping, and darkness modification to make the models more robust across diverse conditions. 📸✨
- Edge AI with Jetson Nano: Real-time inference using the NVIDIA Jetson Nano for low-latency steering control and decision-making. 🌐💨
- NVIDIA Jetson Nano 🌟
- JetRacer Pro AI Kit: Includes a chassis, wheels, IMX219-160 camera, and servo motor. 🛠️
- WiFi Connection: Configured using mobile hotspots for reliable communication with the car. 📶
- Collected real-world driving data with the JetRacer navigating various tracks. 🛤️
- Performed data augmentation using OpenCV to simulate different environments and lighting conditions. 🌞🌚
| Dataset Type | Images |
|---|---|
| Perfect Route | 530 |
| Self-Correction | 180 |
| Obstacle Detection | 780 |
| Combined (Augmented) | 2980 |
- ResNet-18: Achieved the best performance with fast inference and accurate steering control.
- Fine-tuned Driving Parameters: Adjusted throttle and steering gain for smoother navigation.
- Model Evaluation: Assessed models based on the number of valid laps and average lap time. ResNet-18 led the pack! 🏎️
- Best Model: ResNet-18 with optimal settings completed 5 perfect laps at an average speed of 6.2 seconds per lap. 🏁
- Obstacle Avoidance: Successfully navigated 3 obstacle-free laps with an average speed of 6.5 seconds per lap. 🚧
- Encountered hardware issues like loose camera holders and Wi-Fi connectivity challenges. 🔧
- Adjusted the Frame Per Second (FPS) for smoother navigation—finding that lower FPS (around 22) resulted in better driving performance. 🎥
- Upgrade the software for compatibility with advanced models like Levit-128.
- Collect more data and explore improved augmentation techniques for greater robustness. 📊
- Enhance obstacle detection with advanced vision techniques for more dynamic environments.
This project is licensed under the MIT License. See the LICENSE file for details.
- Oluwadara Adedeji – University College Dublin
[email protected] - Sanat Thukral – Technological University Dublin
[email protected] - Madhuri Sawant – Technological University Dublin
[email protected] - Nika Gorchakovar – Dublin City University
[email protected] - Hong-Hanh Nguyen-Le – University College Dublin
[email protected] - Abigail Naa Amankwaa Abeo – Dublin City University
[email protected]
For any questions or collaboration ideas, reach out to us through the email addresses above.
Happy driving! 🏎️💨