Finding an appropriate representation of a person's activities is probably the most challenging part of activity recognition. The behavior needs to be represented with simple and general features so that the model using these features will also be general and work well on behaviors different from those in the learning set.

In fact, it is not difficult to design features specific to the captured observations in a training set; such features would work well on them. However, as the training set captures only a part of the whole range of human behavior, overly specific features would likely fail on general behavior:

Let's see how this is implemented in MyRunsDataCollector. When the application is started, a method called onSensorChanged() gets a triple of accelerometer sensor readings (x, y, and z) with a specific timestamp and calculates the magnitude from the sensor readings. The methods buffers up to 64 consecutive magnitudes marked before computing the FFT coefficients.

Now, let's move on to the actual data collection.