Show HN: Real-time face recognition on low power microcontroller

3 months ago 10

A comprehensive embedded face recognition system implementing real-time face detection and recognition on STM32N6570-DK development board using STEdgeAI.

Real-time Face Detection using CenterFace neural network model
Face Recognition with MobileFaceNet embeddings and cosine similarity
Multi-face Tracking with embedding bank and similarity voting
Hardware Acceleration via STM32N6 NPU (Neural Processing Unit)
Live Camera Input with ISP processing pipeline
PC Streaming Interface for remote monitoring and control
LCD Display Output with face detection visualization
Optimized Performance for embedded deployment

STM32CubeIDE - Primary development environment
STM32EdgeAI - Model conversion tool (stedgeai)
STM32CubeProgrammer - Firmware flashing tool
ARM GCC Toolchain - Cross-compilation toolchain (included in STM32CubeIDE)

Configure tool paths Edit stm32_tools_config.json and set the correct paths for your installation
Build Process

# Convert AI models ./scripts/compile_model.sh face_detection ./input_models/centerface.tflite ./scripts/compile_model.sh face_recognition ./input_models/mobilefacenet_int8_faces.onnx # Build in STM32CubeIDE, then sign and flash ./scripts/sign_binary.sh ./embedded/STM32CubeIDE/Debug/Project.bin ./scripts/flash_firmware.sh all

STM32N6570-DK development board
Camera Module (IMX335, VD55G1, or VD66GY supported)
LCD Display (800x480 resolution)
USB Connection for programming and debugging
PC with STM32CubeIDE or ARM GCC toolchain

STM32N6-FaceRecognition/ ├── stm32_tools_config.json # Tool configuration ├── scripts/ # Build and deployment scripts │ ├── compile_model.sh # AI model conversion │ ├── sign_binary.sh # Binary signing │ └── flash_firmware.sh # Firmware flashing ├── input_models/ # Input model files (.onnx, .tflite) ├── converted_models/ # Generated model code and binaries └── embedded/ # STM32 embedded project ├── STM32CubeIDE/ # IDE project files ├── Src/ # Application source code ├── Inc/ # Header files ├── Models/ # Generated model C files ├── Binary/ # Final binaries for flashing ├── Makefile # Alternative build system ├── Middlewares/ # STM32 middleware └── STM32Cube_FW_N6/ # STM32 firmware library

STM32CubeIDE (recommended) or ARM GCC toolchain
STM32CubeProgrammer for flashing
STM32EdgeAI for model conversion

Configure and convert models:

# Configure tool paths in stm32_tools_config.json # Convert AI models ./scripts/compile_model.sh face_detection ./input_models/centerface.tflite ./scripts/compile_model.sh face_recognition ./input_models/mobilefacenet_int8_faces.onnx
Build the embedded project:

# Using Makefile (from embedded directory) cd embedded && make clean && make -j$(nproc) # Or using STM32CubeIDE # Open embedded/STM32CubeIDE project, clean and rebuild
Flash the complete system:

The STM32N6 requires flashing four separate components in this order:

# 1. Flash FSBL (First Stage Boot Loader) at 0x70000000 STM32_Programmer_CLI -c port=SWD mode=HOTPLUG -el MX66UW1G45G_STM32N6570-DK.stldr -w Binary/ai_fsbl.hex # 2. Flash Face Detection Model at 0x71000000 STM32_Programmer_CLI -c port=SWD mode=HOTPLUG -el MX66UW1G45G_STM32N6570-DK.stldr -w Binary/det_network_data.hex # 3. Flash Face Recognition Model at 0x72000000 STM32_Programmer_CLI -c port=SWD mode=HOTPLUG -el MX66UW1G45G_STM32N6570-DK.stldr -w Binary/rec_network_data.hex # 4. Flash signed application at 0x70100000 STM32_Programmer_CLI -c port=SWD mode=HOTPLUG -el MX66UW1G45G_STM32N6570-DK.stldr -w Binary/STM32N6_GettingStarted_ObjectDetection_signed.hex

or use the provided flashing script!

Important Notes:
- Set BOOT1 switch to right position (dev mode) before flashing
- All .hex files are pre-built and available in the Binary/ directory
- The external loader MX66UW1G45G_STM32N6570-DK.stldr is required for external flash programming
- After flashing, set BOOT1 switch to left position (boot from flash) and power cycle
Connect and run:
- Connect camera module and LCD
- Power on the board
- System will start face detection automatically

Camera → ISP → Face Detection (NPU) → Face Cropping → Face Recognition (NPU) → Display/Stream ↓ ↓ ↓ ↓ ↓ Image Buffer Detection Results Cropped Faces Embeddings UI Output

Face Detection: CenterFace model (128x128 input, INT8 quantized)
Face Recognition: MobileFaceNet model (112x112 input, INT8 quantized)
Image Processing: Hardware-accelerated cropping, resizing, format conversion
Embedding Management: Multi-face tracking with similarity-based voting
Communication: UART protocol for PC interface

Architecture: CenterFace
Input: 128x128 RGB
Output: Face bounding boxes + keypoints
Quantization: INT8
Performance: ~9ms inference time

Architecture: MobileFaceNet
Input: 112x112 RGB aligned faces
Output: 128-dimensional embeddings
Quantization: INT8
Performance: ~120ms inference time

Key Configuration Options (Inc/app_config.h)

// Input source selection #define INPUT_SRC_MODE INPUT_SRC_CAMERA // or INPUT_SRC_PC // Display settings #define LCD_FG_WIDTH 800 #define LCD_FG_HEIGHT 480 // AI model parameters #define AI_PD_MODEL_PP_CONF_THRESHOLD (0.5f) // Detection confidence #define AI_PD_MODEL_PP_MAX_BOXES_LIMIT (10) // Max detected faces // Camera settings #define CAMERA_FLIP CMW_MIRRORFLIP_NONE #define CAPTURE_FORMAT DCMIPP_PIXEL_PACKER_FORMAT_RGB565_1

├── Src/ # Application source code ├── Inc/ # Header files ├── Models/ # AI model files (C code) ├── Exercises/ # Implementation examples and tutorials ├── STM32Cube_FW_N6/ # STM32 HAL drivers and BSP ├── Middlewares/ # AI runtime and camera middleware ├── python_tools/ # PC-side Python utilities ├── Binary/ # Pre-built firmware binaries └── Doc/ # Additional documentation

The system supports PC connectivity for enhanced functionality:

cd python_tools pip install -r requirements.txt python run_ui.py # Launch GUI interface