/Human-Activity-Recognition-on-Android

Primary LanguageJupyter Notebook

Human-Activity-Recognition-on-Android

Note:This repository is for training neural network classifier.

Required packages

  • pip3 install tensorflow==2.3.0

or

  • pip3 install tensorflow-gpu==2.3.0
  • pip3 install keras==2.4.3
  • pip3 install pandas
  • pip3 install numpy==1.19.1
  • pip3 install pyunpack
  • pip3 install patool

0. I've made a video too with animated explanation and code for people who are lazy enough to read all text below.

yYoutube Video

1. About the dataset

  • The dataset contains sensor values from Acceletometer, Linear Acceleration Sensor and Gyroscope sensor.

  • For each timestamp the dataset contains activities as labels corresponding to each data point.

  • The dataset has been recored on following activities:

    • Walking
    • Standing
    • Jogging
    • Sitting
    • Biking
    • Upstairs
    • Downstairs

  • Download the dataset here

2. Visualization of sensors data

  • The visualizations below are plotted values of X-Y and Z axis with respect to activities.

Note: Since the dataset was large, I took only 100 samples from starting for each activity to demonstrate.

during walking

during jogging

during standing

during biking

3. Data Preprocessing

Step 1: Since the dataset was divided into 10 files which was collected by 10 different people by using their devices, all those files concatenated into one single file

Step 2: Split the dataset into Train and Validation distribution.

Step 3: Encode the labels into numbers.

Step 4: Since the data collected by sensors on each timestamp are continous in nature, it's better to convert them into Time Series data so that it can fit as a sequence on a timestamp in LSTMs or GRUs (continous models).

Note: We do not use any discrete modeling because the prediction cannot be made by just observing a single datapoint.

4: About Model

  • The architecture of model is following: Layer 1: The first layer is a LSTM layer for learning from sequence of 100 points at each timestamp are returns the sequence mapping as well. Layer 2: Flatten layer is used to 2 dimensional (Number of timestamps, Number of features) output from above LSTM layer and convert it into 1-d vector Layer 3: From this layer onwards the classifier part is starting which is a Dense layer that takes flatten output from above layer and pass it to Layer 4. Layer 4: This is softmax layer taking input from Layer 3 and predict the probability corresponding to each activities.

  • The model used Categorical Crossentropy for loss measure and Adam or Adaptive Momemtum as a gradient descent optimizer.

  • It has been trained for 5 epochs with ModelCheckpoint callback for saving best model while training.

5. Export to Proto Buffer file

  • The reason behind exporting Hadoop (h5) format file to Proto Buffer file (.pb) is just because it is lightweight and supported by Tensorflow Mobile for deployment in Android devices.

6. Deployment

  • The model has been deployed on and Android Device using Tensorflow Mobile for tracking human activities in real-time.

  • If you want deployment of the model in Android. See this repository

Don't forget to leave a star if you liked it.

PRs are welcome.