Machine Learning Techniques for Pattern Recognition in Technical Swimming

This paper proposes an innovative, open-source and low-cost system, for data acquisition, processing and analysis, applied to the practice of technical swimming. This system allows not only to obtain swimming metrics, but also to detect errors in an athlete's stroke execution through the usage of prototype wristbands, that allow the collection of inertial signals, a mobile application that allows the control of the wristbands, and a complete data processing and analysis system that determines the swimming style. In addition, the full stack system detects possible execution errors of strokes and evaluates the swimmer's overall performance. The results obtained are based on real swimming data collected over several months from a high number of swimmers of varying skill levels. The proposed system achieves accuracy in the recognition of the different styles of 96.95%, 98.63%, 100.0% and 94.0% for crawl style, backstroke, breaststroke and butterfly, respectively, plus a stroke technique execution accuracy of 92.2%, using a support vector machines-based approach in confined waters real time performance, surpassing state-of-the-art commercial solutions.