The assignment we need to do is to classify standing, walking or running, given the data collected by sensor fusion app. We need to clean the data first, in order to be used for training.
- pandas
- sklearn
- seaborn (for visualization)
- matplotlib (for visualization)
In the data collecting stage, we first stand for a while to save the log only for standing; then stop the logging on the app, and start walking while starting to save another log file only for waling; then do the same for running. In this way we have three log files with the ground truth labels.
In the data folder, the sensorLogr_ indicates it is for running, the sensorLogs_ indicates it is for standing, the sensorLogw_ indicates it is for walking.
run read_data.py with sys.argv[1] as the file path and sys.arvg[2] as the activity (s/w/r), as in standing/walking/running.
Then you have three .csv file with time, activity, acc_x, acc_y, acc_z, gyr_x, gyr_y, gyr_z.
To be mentioned, the log file saved by the sensor fusion app, saves the data in such a way, for every time stamp, it saves either acc or gyro, not both. So in the raw data, for the same time stamp, we cannot have both acc and gyro. So in read_data.py, I simply copy the acc/gyro at the last stamp as the acc/gyro for the curent time stamp, since the difference between the two adjunt time stamps are relatively small, I think it is reasonable.
Run activity-recognition.ipynb cell by cell, allows you build the dataset for standing/walking/running activity recognition, and train with random forest in sklearn, with a high prediction accuracy on the test set.
Run t-sne dimensionality reduction and visualization.ipynb to visualize the data