In-sensor 24 classes HAR under 850 Bytes

TL;DR

This paper demonstrates a novel in-sensor human activity recognition model that operates within 850 bytes, achieving 85% accuracy on 24 classes using ultra-constrained hardware, enabling real-time processing with low power consumption.

Contribution

The study introduces a scalable, ultra-compact HAR model for sensor-level deployment on ISPUs, employing innovative memory management and feature optimization techniques.

Findings

Achieved 85% accuracy on 24-class HAR task

Operates within 850-byte memory limit

Consumes only 0.5 mA power

Abstract

The year 2023 was a key year for tinyML unleashing a new age of intelligent sensors pushing intelligence from the MCU into the source of the data at the sensor level, enabling them to perform sophisticated algorithms and machine learning models in real-time. This study presents an innovative approach to Human Activity Recognition (HAR) using Intelligent Sensor Processing Units (ISPUs), demonstrating the feasibility of deploying complex machine learning models directly on ultra-constrained sensor hardware. We developed a 24-class HAR model achieving 85\% accuracy while operating within an 850-byte stack memory limit. The model processes accelerometer and gyroscope data in real time, reducing latency, enhancing data privacy, and consuming only 0.5 mA of power. To address memory constraints, we employed incremental class injection and feature optimization techniques, enabling scalability without compromising performance. This work underscores the transformative potential of on-sensor processing for applications in healthcare, predictive maintenance, and smart environments, while introducing a publicly available, diverse HAR dataset for further research. Future efforts will explore advanced compression techniques and broader IoT integration to push the boundaries of TinyML on constrained devices.