Posture detection achieved with a microcontroller-based AI system: STMicroelectronics NanoEdge AI Studio/VL53L8CX

①Collect learning data and verification data

The training data and verification data for building the AI learning model were collected using the following procedure.

• (1) With your back straight, sway your upper body slightly back and forth and side to side to collect data.
  In order to obtain sufficient data patterns, data was collected for 3 minutes.
  (Data is dumped in CSV format)
• (2) With your back bent and your posture hunched, data will be collected for three minutes while you sway your upper body slightly back and forth and side to side.
  (Data is dumped in CSV format)
• (3) Repeat steps 1 and 2 above twice.
  These serve as training data and validation data, respectively.
• (4) Convert the collected data into a data format that can be read by NanoEdge AI Studio.

In the end, we created the following dataset:

• Training Dataset
  (Two files: "Straight back" and "Humpback")
• Confirmation dataset
  (Two files: "Straight back" and "Humpback")

• 

  Stand straight

• 

  hunchback

② Launch NanoEdge AI Studio and select a learning model

Launch NanoEdge AI Studio (see the image on the left).
The "+ NEW PROJECT" button will appear in the upper left corner of the screen, so click on it.

The following four options for building an AI learning model are displayed (see the left image).
This time I selected "Detect Anomalies."

• Detect Anomalies
• Classify (multivalued classification)
• Detect outliers
• Extrapolate

The following five steps will be displayed (see the image on the left).
Click on each icon to go to the respective step.

• 1. Project Settings
• 2. Signals
• 3. Benchmark
• 4. Validation
• 5. Deployment

The following explains the tasks performed in each step.

③ 1. Project Setting: Select the device and microcontroller

Here, we will select the ToF ranging sensor VL53L8CX and the microcontroller STM32F401RE that we will use this time.

④ 2. Setting: Set the learning data

Load the training dataset prepared in ①.

⑤ 3. Benchmark: Building AI learning models

Build an AI learning model. NanoEdge AI Studio automatically tests multiple libraries to build an AI learning model, and lists the AI learning models in a ranking format based on the performance evaluation results.
Detailed performance information for each AI learning model can be viewed as a report on the GUI.
Below is an example report display of the AI learning model with the highest performance (Best Library).

⑥ 4. Validation: Evaluate the AI learning model

Among the AI learning models constructed in (5), several AI learning models will be selected, mainly from the Best Library, and evaluated using the confirmation dataset prepared in (1).
Based on the evaluation results, the AI learning model to be used is finally selected.
Below are the evaluation results of the AI learning model selected this time.
[Evaluation results using training dataset]
Balanced Accuracy: 87.2%
[Evaluation results based on confirmation dataset]
Balanced Accuracy: 87.7%

⑦ 5. Deployment: Output of library for embedding C source

The AI learning model selected in ⑥ will be output as a library that can be incorporated into C source.

8. The library is implemented on a microcontroller, completing a demo machine for the edge AI system.

The library generated in ⑦ was implemented on the P-NUCLEO-53L8A1 microcontroller STM32F401RE.
It takes about 3 hours to complete steps ① to ⑧.
Below you can see the demo unit in action.
The difference in the state of the back muscles is clearly detected.

• 

  Stand straight

• 

  hunchback