Goal of this assignment was to recognize Activities of Daily Living (ADLs) from sensor data.
The sensor used was the Shimmer 3 (or similar). It is capable of measuring acceleration along three axes.
For data collection, each student performed four activities:
- Brushing teeth (20 seconds)
- Binding shoes (normal speed)
- Drinking water (three movements)
- Writing (a rather long sentence)
The labeled data can be found in
data/raw_from_matlab/data2018.mat. This file therefore contains a number of measurements, each representing a single activity.
In order to train a classifier on this data, we extracted features from each measurement, such as:
- The mean acceleration along each axis (
gx,gy,gz) - The standard deviation and skewness of the overall acceleration (
std,skewness) - The 25. and 75. percentile of the Fourier transform of the measurements
(
f25,f75)
We then trained a support vector machine (SVM) on on these features. As can be
seen in Assignment.ipynb, the collected data can be perfectly separated with
just two features and a linear classifier.
In a realistic setting, the task would be very different: The data would consist of one continuous stream of accelerometer data, and each point in time then has to be classified. We therefore can not directly apply the above method.
The dataset that was provided to us can be found in `data/raw_from_matlab/testData.mat**. By using sliding windows, we generate chunks of data, for which we can then again extract the features and use those to classify this chunk. Finally, the different predicted labels have to be combined in order to generate a single label for each point in time. Majority voting worked well for this task.
Install all necessary packages, ideally in a virtual environment
pip install -r requirements.txtThe jupyter notebook Assignment.ipynb provides all the function calls and goes
through the described approach step by step.
Open jupyter:
jupyter notebookThen go through Assignment.ipynb in your browser.