The Action Logger
Smartphone multimodal sensor-based activity recognition

Abstract View
Feature Ext.

Methodology > Abstract View
Abstract view of Action Logger

Action Logger System Architecture

  • The Action Logger provides technology to detect userí»s activity and to calculate consumed calories.
  • It uses accelerometer and GPS embedded in smartphone to recognize 5 different activities which are stay, walking, jogging, taking a bus and subway.
  • The orientation and the position (Top/bottom pocket, bag, hand, etc) of the smartphone is irrelevant.
  • With action logger, you may review your daily activity and moved path with consumed calorie which could be used widely in healthcare, fitness, wellness, and entertainment services.
  • Action Logger has two functions. One is feature extractionand the other one is machine learning algorithm(GMM).
  • Figure above shows system architecture of Action Logger.

Feature Extraction


  • Action Logger is using machine learning Technics. There are two types feature to recognize physical Activities (stay, walking and jogging) and vehicles (subway and bus).
  • Physical Activites are recognized by SVM (Signal Vector Magnitude) features (average, standard devision and each X,Y,Z correlations) to recognize Activities for position independantly.
  • Vehicle Activities are recognized by FFT (Fast Fourier Transform) features. Because it is hard to recognize stay betwenn vehicles using only SVM features.
  • So we proposed new feature extraction technics through difference number of vibrations by accelerometer sensor.
  • Figure above shows Action Logger's feature extraction techniques.



  • Action Logger is using GMM(Gaussian Mixture Model) machine learning algorithm.
  • GMM Algorithm is very light with other machin learning algorithms like SVM(Support Vector Machine), Baysian Network , etc.
  • So we select GMM algorithm for porting on a smartphone. And we build optimised Activity Model using GMM.
  • Figure above shows how to build GMM Model.