We are engaged on a development of a system of swimming performance and technique evaluation with the aim of enhancing the swimming level of competitive swimmers. It evaluates performances of swimming motions (diving, strokes and turns) of a swimmer by using an underwater wireless wearable sensor attached on swimmer's waist. This system gives feedback to a swimmer/coach at the timings: (A) real-time, (B) end of swimming and (C) after training. To make these feedbacks by analyzing measured acceleration data, it requires an automatic swimming style classification. For the feedback at the timing (A), it has to classify the swimming style at each one stroke from corresponding part of data, and for (B) it has to classify from whole data. We examined the availability of the method of constructing a decision tree by dividing a feature space which feature quantities of each stroke are distributed in. After constructing the decision tree from training data and confirming the reliability of it, we assessed the availability of our method by F-measures of the result from test data. In our experiment, the styles freestyle and backstroke were well classified with high F-measure (greater than 0.95), but the styles breaststroke and butterfly were not well classified with not so high F-measures (greater than 0.72). This result means the evaluation for the feedback at (A) is still difficult to realize. However, it also means that this method is enough available for the feedback at (B) because the classification a this timing decides on a majority vote.
