This program can be used to perform PCA dimensionality reduction on accelerometer data acquired from mobile smartphone clients.
The idea is that the smartphone connects to a websocket server which is used to collect the accelerometer data through a web page. Therefore, this program must be used with a mobile browser that supports retrieval of accelerometer data. Some examples:
- Firefox Nightly (Android)
- Safari (iPhone)
When the program is executed, a Qt GUI appears with several panels. The top panel allows to specify a port and start and stop the websocket server. After starting the server, the browser should be pointed to the IP address of the computer running CvPCA, with the port specified. For example, if the computer's IP address is 192.168.0.5, and the server was started on port 8080, the following URL should be provided to the phone:
http://192.168.0.5:8080
A message will be posted to the phone indicating whether the websocket was connected successfully. If there are issues accessing accelerometer data, an error message will also be displayed on the phone.
The "Connections" panel lists connected clients, which can be used to verify that the smartphone is correctly connected to the websocket server.
The "Gestures" panel contains an editable list of gestures tags. These are the instructions to the user, and can be any description of gestures that the user should record. For example, "Move up and down", or "tap 3 times a second."
On pressing "Start Recording," the phone will display a gesture instruction, and start transmitting the accelerometer data to the server, which will be stored in memory. When all gestures have been recorded, press the "Stop Recording" button, and the in-memory recording will be displayed on the screen. Press the "Save" button to save the data to a comma-separated value (csv) file.
After recording and saving data, the data will be re-loaded into the analysis portion of CvPCA. Alternatively, previously-saved files can be loaded by pressing the "Load" button in the "Data Sets" panel.
Pressing the "PCA" button will extract features from the data using a sliding window and perform Principle Component Analysis. The first two principle components will then be displayed in the visualization window on the right as a 2D scatter plot.
The features calculated are based on the accelerometer data re-sampled to 25 Hz. A sliding window is applied after a high-pass filter, extracting windows of 256 points, such that each window analyses the last 10 seconds of data. Subsequently, the log-FFT is taken giving amplitudes of frequencies present in each accelerometer axis. Finally, these axes are linearly combined to provide a feature unbiased by device orientation and gravity.
Currently, the principle components as well as the eigenvectors are additionally printed to the console.
This program uses the Qt4 build system. It also depends on libwebsockets, an excellent minimal C websocket and HTTP library. Finally, it uses OpenCV2 to perform PCA.
On Ubuntu Linux, you should install the packages libopencv-ml-dev
and libqt4-dev. Then, you must download and build libwebsockets,
and finally run Qt's qmake program.
Since libwebsockets is included as a git submodule, simply run,
git submodule init
git submodule update
to download the libwebsockets source code. Then, cd to the
libwebsockets directory and build it:
cd libwebsockets
./autogen.sh
./configure && make
Finally, build CvPCA:
cd ..
qmake-qt4
make
On success, there should be an executable called cvpca in the
current directory.