This is a test of feature extraction performance (speed, RAM usage, ROM usage) on ARM Cortex-M3 and M4 cores.

Please cite the following paper if you use this repository:

• A. Elsts, R. McConville, X. Fafoutis, N. Twomey, R. Piechocki, R. Santos-Rodriguez and I. Craddock. On-Board Feature Extraction from Acceleration Data for Activity Recognition, EWSN 2018.

Features

Importing data:

cd import
./load_sphere_challenge_files.py

This creates a number of .c files in the data directory, each containing 15000 samples of 3-axis acceleration data. The data is expressed as 8-bit signed integers.

Running

Running natively (the same architecture as on the machine it is compiled on):

cd feature-test
make run ARCHITECTURE=native

Running on an emulator - needs Zephyr OS and Zephyr SDK to be installed:

cd feature-test
make run ARCHITECTURE=zephyr

Running on hardware (a SPHERE board) - needs Contiki OS and ARM compiler to be installed:

cd feature-test
make run ARCHITECTURE=sphere

Other supported targets are: z1 (for Zolertia Z1 with msp430), zoul (for Zolertia Zoul with CC2538) and nrf52dk (for Nordic NRF52DK with Cortex-M4F).

Features

Note: for increased performance, some of the FFT-based feature results are not normalized! To get the correct result, they should be divided either by FREQUENCY_FEATURE_WINDOW_SIZE or by FREQUENCY_FEATURE_WINDOW_SIZE^2 depending on the feature.

Generating results for comparison

Go to the top directory and run

./generate_features.py

This will generate a number of CSV files in the export directory. Each file is based on 25-minute data from a single participant and includes all of the supported features.

This will run the feature test natively. Prerequisites:

• GNU make
• GCC
• Python3

Results

• The export directory will contain the .csv files with the different features after running generate_features.py.
• The result directory contains the processing duration evaluation results on the different hardware platforms.