Overview

Application overview

Traditional human-machine interfaces (HMI) rely on physical buttons or touchscreens which are impractical in wet, gloved, or hygiene-sensitive environments. Rule-based gesture recognition struggles to account for natural variation in user movements. Cloud-based solutions introduce latency and privacy concerns for real-time control.

Our edge AI solution demonstrates practical always-on gesture detection implementations for battery-constrained wearables, smart watches, health bands, and industrial touchless controls with 94.86% accuracy and 9.07 ms inference latency. Our solution runs a compact convolutional neural network (CNN) directly on the MSPM0G5187 microcontroller, leveraging our TinyEngine™ NPU. The solution uses 8-bit integer (INT8) quantization to classify four dynamic hand gestures in real time from a from 3-axis accelerometer.

Starting evaluation

Data collection

We capture our 3-axis accelerometer data though serial peripheral interface (SPI) from our edge AI sensor BoosterPack at 100 Hz. The incoming data is windowed into 256-sample frames with 25% overlap and range-normalized to [-1, 1] before inference. The dataset, hand_gesture_dataset, contains labeled recordings of four classes and one idle state: clockwise circle, counterclockwise circle, wave, tap, and non-gesture motion.

Data quality assessment

Use integrated capture tools in Edge AI Studio to record site-specific gesture samples and customize datasets for your target application. In this use case our goodness of fit (GoF) evaluation model verifies that your gesture dataset encompasses distinct accelerometer patterns for each motion class. Our GoF evaluation model was trained and validated using the TIDA-010997 BoosterPack reference design; for the best results follow that design.

Training the model

Quickly train and evaluate hand gesture recognition models with our Edge AI Studio's model composer. Our hand gesture recognition example project provides an annotated dataset of hand motions for circle, wave, and tap gestures captured from the accelerometer on TIDA-010997. 

For advanced users looking to further optimize the training pipeline a command line interface (CLI) is available to fine tune the training process and achieve the best performance.

Find the right model for your needs

Access our library of time-series classification models optimized for gesture recognition on MSPM0 devices. This use case uses the CLS_55k_NPU model in the TI tinyML model zoo on GitHub. It provides 94.86% accuracy in 67 KB flash with 9.07ms inference latency

Deploying the model

The Edge AI Studio model composer provides an end-to-end workflow for deploying trained gesture recognition models. For developers seeking deeper customization, refer to the deployment guides for a comprehensive framework of integrating edge AI technology into embedded applications. We provide two options for deploying machine learning models:
1. Edge AI studio graphical user interface (GUI) tools
2. Command line interface (CLI) tools

ProductProcessing coreNPU availableNILM Benchmarks
   Latency (ms)Flash (kB)SRAM (kB)
MSPM0G5187Arm® Cortex®-M0+ CoreYes9.07675.4
SIMILAR USE CASES
INDUSTRIAL | VISION

Object detection in vision systems with 25x lower latency

Find and localize specific objects, people in real-time at high framerate with AI-accelerated processors and industry-standard software


INDUSTRIAL | VISION

Pixel-level Segmentation for navigation and perception with 40x lower latency

Detect fine-grain obstacles and pathways from imagery in real-time using C7™ NPU equipped processors