Motion Classification

• This project utilizes Digital Signal Processing and Machine Learning techniques to classify motion patterns: in this specific use case: digits!
• For a more detailed overview my system design please see APPENDIX.md
• It utilizes the ARM-Based STM32 Microcontroller and its on-board accelerometer
  • This board had 2 megabytes of Flash memory with 786 Kilobytes of SRAM
• This project is written in C and utilizes the STM Cube IDE
• For an example console output through model training and classification please see Console_Output.txt

System Design:

Navigating the Code:

• The code I wrote is in main.c and embedded_ML.c utilizing the STM32 example design and an Embedded Machine Learning Library written by TA Charles Zaloom

main.c:

Contains the code for:

• STM32 setup code (Clock frequency, HAL Library, Instruction Cache, etc)
• The instantiation of the Neural Network with appropriate parameters and random number generation
• Infinite loop to start training and classification

embedded_ML.c:

Contains the code for:

• Digital signal processing (velocity calculations, sampling, and filtering)
• Feature extraction and training loops for machine learning
• Classification calculations taking max softmax output
• Helper functions to provide LED output to help user with performing motions
• Machine learning library code for gradient descent and back propagation

Project Results and Evaluation

• Sampling results across fifty different classifications this system predicted 47 of them to be the correct digit: 94% Accuracy!
  • The wrong predictions were likely to have been introduced by human error in data collection
• This project is a dynamic system that can train on changing data depending on what we want to classify in the field and Embedded ML is very flexible and dynamic
  • In this case we are simply detecting digits and can recognize different letters and digits
• I enjoyed this project because Internet of Things (IoT) requires many disciplines of engineering including: Transducers and Sensors, Digital Signal Processing, Embedded Systems, and Machine Learning!

Product use and Application: Education and Accessibilty

• This product can be used to teach young children how to correctly draw numbers and provide feedback
• This product can also be used to aid those who do not have the dexterity to draw numbers but can move a box to draw digits

Thank you for reading this

• I appreciate your time and thank you for reading this!