Simple sensor based human action recognition end-to-end pipeline using Bi-LSTM model. We collect the dataset using our own smartwatch wear os sensor data collecter as you can check here. Our dataset consists of accelerometer and gyroscope from smartwatch data with the action of Walking, Standing, Jumping, and Falling, you can make and use your own dataset as well.
- Tensorflow 1.14
Human action recognition for wearable sensor data is conduct by using bidirectional long-short term memory (Bi-LSTM) to capture the long-term dependencies of hand movement and automate feature extraction from raw sensor inputs and multilayer perceptron as the classifier of each activity classes. In order to train our Bi-LSTM model, we do not perform any hand-crafted feature pre-processing and directly split each collected data into a number of windows. Through the experiment of different window sizes, in our case, we find the optimal window size is about 120 Hz with a step size of 20Hz. The Bi-LSTM learns to map and predict each window sensor data to an activity as shown in figure above.
You can save the model as frozen protobuf file and use our action recognition wear os application here
.