Efficient human activity recognition on edge devices using DeepConv LSTM architectures.

Area of Science:

• Artificial Intelligence
• Embedded Systems
• Machine Learning

Background:

• The Internet of Things (IoT) necessitates deploying deep learning on resource-constrained hardware.
• TinyML offers a solution for efficient machine learning on edge devices.
• Human Activity Recognition (HAR) is crucial for real-time applications.

Purpose of the Study:

• To deploy lightweight deep learning models for HAR using TinyML on edge devices.
• To evaluate and compare the performance of different deep learning architectures for HAR.
• To optimize a model for resource-constrained deployment.

Main Methods:

• Designed and evaluated 2D CNN, 1D CNN, and DeepConv LSTM models for HAR.
• Applied full integer quantization to the best-performing model.
• Deployed the quantized model on an Arduino Nano 33 BLE Sense Rev2 using the Edge Impulse platform.

Main Results:

• DeepConv LSTM achieved 98.24% accuracy and 98.23% F1 score before quantization.
• Quantization reduced model size from 513.23 KB to 136.51 KB.
• The deployed model maintained 97% accuracy and 97% F1 score with low memory (29.1 KB) and flash (189.6 KB) usage.

Conclusions:

• TinyML facilitates efficient and low-latency HAR systems on edge devices.
• The DeepConv LSTM model, after quantization, is suitable for real-time HAR applications.
• This study demonstrates the practical feasibility of deploying advanced AI models on microcontrollers.