All code (but not the dataset) necessary to reproduce the machine learning results from the master's thesis: Drews, Haakon. “Classification of Error Types in Physiotherapy Exercises,” 2017. https://www.duo.uio.no/handle/10852/56891.
What follows is an overview of functionality and data flow. For more details please refer to the usage instructions of the different components by passing the --help argument. See the examples folder for examples of the file structure layout. Note: the example csv files do not contain real exercises or error types.
- Record the sensor data for each sample and save it to a CSV file.
- Calculate the distance between each sample using DTW
- Generate confusion matrices
- Visualize the recorded motion data in a very simple way (real-time and playback of csv files)
Tested on Linux only.
Main requirements are
Message bus
- message_bus.py
Sensor clients
- myo_client.py
- readLP.c
Data consumers
- capture_motion_client.py
- visualizer.py
Data flow
- A message bus is run which the other components use to interface.
- For each type of sensor used, a client is run which forwards the motion data from the sensor to the message bus. Contains clients for the Myo and LPMS-B sensors.
- To record the sensor data, a recording client is connected to the message bus which receives all sensor data and saves it to a CSV file in an appropriate folder structure.
- The visualizer can be connected to the message bus to make sure the motion data is received correctly in real-time.
- Having recorded all samples, calc_dtw.py is used to calculate all DTW distances between the samples. A pickle containing the processed data set is saved to a file.
./calc_dtw.py --help
usage: calc_dtw.py [-h] [-d {quaternion,euclidean}] [-t DATA_TYPE] [-V]
[-i SENSOR_ID_CATEGORY]
[-r [SENSORS_REQUIRED [SENSORS_REQUIRED ...]]] [-j JOBS]
Calculate all distances between samples in the current directory
optional arguments:
-h, --help show this help message and exit
-d {quaternion,euclidean}, --dist-type {quaternion,euclidean}
Either 'quaternion' (default) or 'euclidean'.Note that
the quaternion distance function will likely throw
weird errors for anything but quaternions.
-t DATA_TYPE, --data-type DATA_TYPE
Must be the same as the column label in the data files
used (for example 'quat')
-V, --verbose Be verbose
-i SENSOR_ID_CATEGORY, --sensor-id SENSOR_ID_CATEGORY
The name of the csv column in the CSV input file
containing the sensor id. Default is 'id'
-r [SENSORS_REQUIRED [SENSORS_REQUIRED ...]], --required-sensors [SENSORS_REQUIRED [SENSORS_REQUIRED ...]]
List of required sensors. If specified, will abort if
a required sensor is missing from input data
-j JOBS, --jobs JOBS Number of jobs to run concurrently
- Using the exported data set file, data_tester.py is used to classify each sample using the k-nearest-neighbor algorithm and generate confusion matrices.
./data_tester.py --help
usage: data_tester.py [-h] [-s [SENSORS [SENSORS ...]]] [-k KNEIGHBORS]
[--class-type {exercise,mode}] [-t DATA_TYPE]
[-l {all,exact}] [-d DPI] [-V] [--save] [--show]
inputfile
Generate confusion matrices and classify using k-NN
positional arguments:
inputfile Pickle file generated by calc_all_dtw.py containing an
ExerciseRecordingDataSet
optional arguments:
-h, --help show this help message and exit
-s [SENSORS [SENSORS ...]], --sensors [SENSORS [SENSORS ...]]
List of required sensors to combine for kNN matching.
If not specified, combine all available.
-k KNEIGHBORS, --kneighbors KNEIGHBORS
Number of neighbors to use for majority voting matches
--class-type {exercise,mode}
'exercise' for classification by exercise or 'mode'
for classification by error type
-t DATA_TYPE, --data-type DATA_TYPE
Data type to use
-l {all,exact}, --leave-me-out {all,exact}
set leaveMeOut scheme. Must be 'all' (leave out all
with same tsID) or 'exact' (leave out all with same
tsID and same exercise)
-d DPI, --dpi DPI Set DPI for the saved figure
-V, --verbose enable verbose mode
--save save figure to file
--show show figure
