SMARTER PHONE

BUSINESS CHALLENGES

Issues	Our Solution
Less relevance of ads	Smart health assistant - Improvement of recommendations based on user’s activity
Extra hardware to detect person’s activity	Improvements of ads by categorizing ads for different physical activities
No generic API/SDK for the above	API/SDK for other developers

APPROACH

• Step 1 - Data generation
• Step 2 - Training our activity prediction model
• Step 3 - Using prediction model on smartphones (offline)

STEP 1 - DATA GENERATION

• Android application developed to generate data
• Collected data of each activity for 10-15 mins from seven android phones.

STEP 2 - TRAINING OUR PREDICTION MODEL

• Use of generated data to train our prediction model
  • Using Convolutional Neural Network (Machine Learning concept)
  • Combined 100 samples (at 50 Hz) of accelerometer data for training
  • Performed 1D Convolutional with 60x3 and 20x1 weight matrix, followed by Max Pooling
• Achieved accuracy of more than 90%

STEP 3 - USING PREDICTION MODEL ON SMARTPHONE

• Import trained model to phone to predict the activity offline
  • Used NDK* provided by android to link TensorFlow to JAVA
  • Predict activity for 100 readings
  • Taking 100 predictions (10/minute) and finding the most occurred activity
  • Build Notifications by fetching the data from local database
    *NDK - Native Development Kit

TECHNICAL ARCHITECTURE

CURRENT DATA FLOW DIAGRAM

WALKING DATA EXAMPLE