Automatic shooting detection in archery from acceleration data for score prediction

In archery, avoiding postural tremor while aiming is one of the most important factors improving performance. Thus, efforts have been made to automatically detect shooting events and use basic statistics of tremor during the aiming period to predict the total score the archer will achieve. However, although they require many shots, previous shooting detectors are not accurate enough because they use simple classification algorithms. This makes them arguably unsuitable for use in practice. Therefore, in this study, we compared seven commonly used machine learning methods on these tasks. In our experiment, first, the winning model, random forest, outperformed the conventional method by 26% in terms of the F-measure in shooting detection. Second, random forest even showed higher performance in predicting the score from the tremor during the aiming periods of detected shots than that of actual shots identified from video recordings. This is an interesting result in that random forest was trained using manual annotations as the ground truth for shooting detection in a cross-validation manner but outperformed the annotations in score prediction. In addition, we found that random forest can halve the number of shots required for both shooting detection and score prediction compared to that in the literature.