µT-Kernel AI Interpreter (Proof of concept)

Combining AI and RTOS for resource constrained embedded systems.

Introduction

The rapid expansion of the AIoT underscores the increasing significance of embedded systems and their fundamental real-time operating systems. Moreover, in today's societal drive toward carbon neutrality, the need for energy-efficient embedded systems is increasing rapidly.

A machine learning model is utilized by an AI interpreter to process data and generate outcomes. µT-Kernel is a real-time OS for small embedded systems. It can run on embedded systems with a very small amount of ROM and RAM. Employing a µT-Kernel AI Interpreter enhances flexibility and scalability compared to using a bare-metal AI interpreter. With this setup, each RTOS task can be fine-tuned in terms of priority and periodicity, enabling developers to meet specific real-time and energy-efficiency requirements effectively.

To showcase the capabilities of the µT-Kernel AI Interpreter, a person detection example application is deployed on an MCU-based board. This proof of concept would aptly demonstrate the capability of the µT-Kernel AI Interpreter development.

Demonstration

The person detection example application demonstrates how an application detects the presence of a person. The example application creates and synchronizes 3 µT-Kernel tasks:

• person_detection_input: This task loads images stored in ROM.
• person_detection: This task detects the presence of a person in the loaded image.
• person_detection_output: This task responses to the result of detection.

The below images are stored in ROM and loaded alternatively.

no-person	person

To run the person detection example app, import the below projects:

• tflm: AI interpreter.
• tflm_person_detection_mtk3: Example application.

Build and run the project tflm_person_detection_mtk3.

Specifications

• Refer here.

Getting started

Setup the development environment

• Refer here.

Clone the repository

Run the command git clone.

git clone https://github.com/tu-thai/tflm-ekra8m1-mtk3.git --recurse-submodules

If you downloaded the repository without using the --recurse-submodules argument, you should run the below command to acquire the submodules.

git submodule update --init --recursive

Import projects to e² studio

(1) Open e² studio.

(2) Choose a workspace and click Launch.

(3) File > Import... > General > Existing Projects into WorkSpace.

(4) Click Browse... and choose the projects folder. Click Select Folder.

(5) Select the projects you want to import.

(6) Ensure that Copy projects into workspace is not selected.

(7) Click Finish to import the projects.

Build and debug a project with e² studio

(1) Select a project to be debugged on the Project Explorer view. Click Project > Build Project to build the project.

(2) Ensure that debugger connection is established. Click Run > Run Debug Configurations... Select the corresponding debug configuration under the Renesas GDB Hardware Debugging group. Click Debug to start debugging the project.

Setup debugging

(1) Connect the host computer to the USB Debug Port (J10) on EK-RA8M1.

(2) Launch e² studio.

(3) Select a project to be debugged on the Project Explorer view. Click Project > Build Project to build the project.

(4) Click Run > Run Debug Configurations.... Select the corresponding debug configuration under the Renesas GDB Hardware Debugging group. Click Debug to start debugging the project.

(5) Launch J-Link RTT Viewer. Configure the connection as follow.

Item	Contents
Connection to J-Link	USB
Specify Target Device	R7FA8M1AH
Target Interface	SWD
Speed	4000 kHz
RTT Control Block	Search Range
	Address range: 0x22000000 0xE0000