This project demonstrates the integration of advanced AI models with resource-constrained embedded systems, specifically using the STM32H747I-DISCO microcontroller board for real-time image recognition. The system leverages Google's Teachable Machine for model training and STM32Cube.AI Developer Cloud for optimization, showcasing the potential of deploying sophisticated AI applications on low-power hardware.
| Prototype | Components | Testing |
|---|---|---|
 |
 |
 |
- Project Objectives
- Features
- Hardware Components
- Software Tools
- Installation
- Usage
- Results
- Challenges
- Future Work
- Contributors
- License
- Optimization Techniques: Develop and implement optimization strategies for image classification models to be deployed on resource-constrained STM32 microcontrollers.
- Real-Time Inference: Deploy optimized models onto STM32 microcontrollers ensuring real-time inference capabilities with adequate accuracy and frame rates.
- Real-time image recognition using STM32H747I-DISCO board.
- User-friendly model training via Teachable Machine.
- Model optimization for STM32 architecture using STM32Cube.AI Developer Cloud.
- Dual-core architecture with strong performance.
- Comprehensive peripherals including USB OTG HS, Ethernet, SAI Audio DAC, etc.
- Extension connectors compatible with multiple camera modules.
- Based on the OV5640 image sensor offering a 5-Mpixel resolution.
- Intuitive web-based tool for training machine learning models without extensive coding knowledge.
- Free online service for developing AI on ST devices, supporting tools for creation, optimization, and benchmarking.
- Integrated development environment for STM32 microcontrollers and microprocessors with debug tools, code creation, compilation, and peripheral configuration.
- STM32H747I-DISCO board
- B-CAMS-OMV camera module
- Computer with internet access
-
Train the Model:
- Use Teachable Machine to train your image classification model.
- Export the TensorFlow Lite model.
-
Optimize the Model:
- Upload the exported model to STM32Cube AI Developer Cloud.
- Optimize and download the C code tailored for the STM32H747I-DISCO board.
-
Integrate and Deploy:
- Open the downloaded C code in STM32CubeIDE.
- Configure project settings, build, compile, and flash the firmware onto the STM32H747I-DISCO board.
- Connect the B-CAMS-OMV camera module to the STM32H747I-DISCO board using a flexible flat cable (FFC).
- Power up the STM32H747I-DISCO board.
- Place objects in front of the camera within the optimal range for image recognition.
- Observe the output on the display or through connected interfaces.
| Accuracy | Loss | Confusion Matrix |
|---|---|---|
 |
 |
 |
- Successfully demonstrated real-time image recognition on STM32H747I-DISCO board.
- Achieved notable accuracy in recognizing various objects including Arduino UNO, ESP32 cam, ESP8266, and Wi-Fi expansion board with STM32.
- Frame Rate Performance: Limited to 1.7 frames per second (fps), impacting responsiveness.
- Camera Focus: Lack of auto-focus capability in the B-CAMS-OMV camera module affected image quality and classification accuracy.
- Enhance frame rate performance through advanced optimization techniques.
- Upgrade to a camera module with auto-focus capabilities.
- Expand the training dataset to improve robustness and generalizability.
- Explore different neural network architectures optimized for embedded systems.
- Improve user interface and experience by developing a comprehensive dashboard.
- Fakhrul Fauzi (me)
- Saiful Azree
- Haziq Hanafi
This project is licensed under the MIT License - see the LICENSE file for details.
Feel free to contribute to this project by opening issues or submitting pull requests. For any questions, please contact the contributors listed above.