ealpha / harX

Human activity recognition(HAR) using sensor data

harX

What's over here?

Human activity recognition(HAR) based on sensor data and pedestrian dead reckoning(PDR) in PyTorch.

• dataset readers(You need download raw data and provide the dataset path)
  • UCI-HAR
  • USC-HAD
  • Ubicomp13
• models
  • CNNwithActivityImage
  • CNNLSTM
  • CNNwithStatistics
  • StackedAutoencoders(not tested!)for this
  • NaivePeakDetector

Background(You can pass this)

Smart phones are pervasive nowadays and have become important intermediaries of our lives. There are many phone sensors, e.g., accelerometers and gyroscopes, generating data almost every second. From these raw recordings we can estimate the motion states and further promote some other down-streaming tasks such as indoor inertial navigation when GPS are unavailable and healthy cares for the elders living alone.

How to use

Just hack! A direct way to use these code is to write your own script under harX directory. example1.py is provided for reference and you can run it as easy as:

cd harX
python example1.py

You need the following packages in your python environment.

tqdm             
numpy            
scipy           
torch             
matplotlib

Some Visualization Results

You can show the data with some tools in utils.

One sensor data sample coming from Ubicomp13 Check it from another perspective.

ActivityImage proposed in Human activity recognition using wearable sensors by deep convolutional neural networks

A confusion matrix of one model's performance on UCI-HAR test dataset.