Project: Evaluating Edge AI Platforms on STM32

This academic team project aimed to explore the field of embedded AI (Edge AI) by evaluating and comparing the performance of two new hardware platforms from STMicroelectronics.

We developed and ported three practical Machine Learning applications (computer vision and audio processing) to measure the effectiveness of hardware acceleration via integrated Neural Processing Units (NPUs).

1. Context and Problem

Deploying AI models is often limited to cloud infrastructure, which poses challenges regarding latency, power consumption, and data privacy. Edge AI (embedded AI) addresses these issues by executing models directly on local devices, such as microcontrollers (MCUs) and microprocessors (MPUs).

Our challenge was to concretely evaluate the suitability of the new STM32 platforms for these tasks.

2. Project Objectives

• Evaluate the performance of the STM32N6570-DK (MCU) and STM32MP257F-DK2 (MPU) platforms.
• Compare the efficiency of the MCU architecture (geared towards real-time, low-power) versus the MPU architecture (geared towards processing power).
• Leverage the hardware acceleration of the integrated Neural Processing Units (NPUs).
• Optimize ML models (via quantization, etc.) for deployment on constrained resources.
• Demonstrate the feasibility of AI applications for security, industrial, and communication uses.

3. Applications Developed

To conduct our comparative study, we implemented three distinct use cases:

1. Access Control via Face Detection

   • Field: Computer Vision (Security)
   • Objective: Detect a face in real-time from a camera stream.
   • Metrics: Detection accuracy (target: 90%) and latency (target: < 200ms).

2. Gesture Recognition

   • Field: Computer Vision (Industry 4.0)
   • Objective: Enable touchless human-machine interaction for equipment control.
   • Metrics: Recognition accuracy (target: 85%) and latency (target: < 250ms).

3. Audio Enhancement in Noisy Environments

   • Field: Audio Signal Processing
   • Objective: Isolate or enhance human speech in a noisy environment.
   • Metrics: Signal-to-Noise Ratio (SNR) improvement (target: +6 dB) and latency (target: < 100ms).

4. Technical Environment (Tech Stack)

• Hardware:
  • STM32N6570-DK Development Board (MCU)
  • STM32MP257F-DK2 Development Board (MPU)
• STMicroelectronics Ecosystem:
  • STM32Cube.AI (for model conversion and optimization)
  • STM32CubeIDE (Development IDE)
  • STM32 ModelZoo (Bank of pre-trained models)
  • OpenSTLinux (Embedded Linux distribution for the MPU)
• AI & Machine Learning:
  • Python (3.7-3.9)
  • TensorFlow / Keras / TensorFlow Lite
  • OpenCV, Librosa, NumPy, scikit-learn
• Project Management:
  • Agile Methodology (Scrum)
  • GitHub (Version Control)
  • Jira (Task Tracking)

Links

• GitHub: [Coming soon]
• Documentation: [Project Wiki]