Talk on the 31st Mediterranean Conference on Control and Automation (MED)
Date:
Title Slide:
- Title: “A CNN-based Real-time Forest Fire Detection System for Low-power Embedded Devices”
Slide 1: Overview
- Content: “Today, we’re exploring the deployment of state-of-the-art target detection algorithms on low-power computing devices like the Raspberry Pi 4B. Our focus is on real-time forest fire detection, a critical application for small UAVs. We’ve modified the YOLOv5 backbone to a lightweight network, performed channel pruning for compactness, and overclocked the Raspberry Pi’s CPU to enhance detection frame rates. Our results show significant improvements in accuracy and frame rates compared to the original YOLOv5.”
Slide 2: Comparison of Single Board Computers (SBCs)
- Content: “We’ve researched mainstream single-board computers for deploying deep learning algorithms. Our main contenders are Raspberry Pi and Nvidia Jetson series. Let’s compare their specifications and suitability for our task.”
Slide 3: Benchmarks and Selection
- Content: “We’ve benchmarked the Raspberry Pi 4B and Nvidia’s Jetson series. While the Jetson series is often preferred for deep learning applications, our focus on the Raspberry Pi 4B demonstrates its potential in a cost-effective and efficient setup for our application.”
Slide 4: Proposed Optimization Framework
- Content: “Our proposed framework includes a series of optimizations. Starting with network architecture, we’ve chosen ShuffleNetV2 as our backbone for its efficiency. Network pruning and sparse training are applied to streamline the model further, and we’ve introduced hardware acceleration techniques for enhanced performance.”
Slide 5: Dataset Preparation
- Content: “For training and testing our model, we’ve prepared a comprehensive dataset divided into training, validation, and test sets. We’ve also employed image-based data augmentation techniques to enhance the robustness of our model.”
Slide 6: Performance and Evaluation
- Content: “We’ve rigorously evaluated our model, comparing it with the original YOLOv5. The optimized version shows superior performance in terms of detection accuracy and frame rate. We also delve into the specifics of model pruning and how it impacts the performance.”
Slide 7: Deployment and Real-world Testing
- Content: “Finally, we’ve successfully deployed our optimized model on the Raspberry Pi platform. We’ll discuss the practical aspects of this deployment and show real-world testing results, including video input tests that demonstrate the model’s effectiveness in various scenarios.”
Conclusion Slide:
- Content: “In conclusion, our optimized YOLOv5 model not only achieves high accuracy but also operates efficiently on low-power devices like the Raspberry Pi 4B. This advancement significantly enhances the feasibility of deploying deep learning-based fire detection systems in real-world, resource-constrained environments.”
Questions Slide:
- Content: “Thank you for your attention. We are now open to any questions you may have.”