,

Human Activity Recognition 

Fork on GitHub

Human Activity Recognition (HAR) is a time series classification task identifying the specific movement or action of a person based on sensor data.  Movements can be activities performed indoors, such as walking, standing, sitting or outdoors such as driving, biking. The demo runs on a small form factor board Sensor Tile that comes along with a smartphone application connected through Bluetooth Low Energy. 

Approach

– Exploits 3-axis accelerometer data
– 5 classes: stationary, walking, running, biking, driving
– Pre/Post-processing: filtering gravity, reference rotation, temporal filter 

The main model is a ST Convolutional Neural Network model, but several other models are proposed within our function packs FP-AI-SENSING1 and FP-AI-MONITOR1, another CNN and a SVC model. 

Sensor

Vision: 3D Accelerometer (reference: LSM6DSM)

Data

Data format:
3D-accelerometer acquired @ 26Hz  
5 activities / 185 minutes per activity 
Sensor held in various places (backpack, wrist, in hand, …) 

Results

Model: ST Convolutional Neural Network  

Input size: 24×3

Memory footprint:
12 KB Flash for weights
1.8 KB RAM for activations

Performance on STM32L476 (Low Power) @ 80 MHz 
Use case: 1 classification/sec 
Pre/Post-processing: 0.02 MHz
NN processing: 0.35 MHz 
Power consumption (1.8 V) 
– System: ~ 580 uA (with optimization BLE) 
– STM32: ~ 510 uA