Domain-Knowledge Enhanced Machine Learning for Table Tennis Stroke Characterization

Abstract

Machine learning in wearable applications characterizing body movements is particularly challenging due to constraints in energy and computation, as well as due to a large space of possible usage scenarios. In this work, we are pushing these boundaries with a focus on table tennis stroke recognition and present a novel domain-specific approach at various abstraction levels, ranging from low-level signal corrections, context-dependent sensor fusion, stroke type characteristics, to a valid gameplay state machine. Our implementation runs entirely on a player’s smartwatch and shows that ML specialization is critical to achieving robustness for practical use. Our pipeline and classifier cascade is based on LSTM and successfully distinguishes between strokes and non-stroke actions, as well as eight types of strokes (drive, loop, block, and push in forehand and backhand). Training and validation were controlled in an environment where a table tennis robot served different players thousands of precisely controlled shots that required specific stroke responses. The final ML classifier pipeline is only 82.5 KB in size, achieves a maximum F1-score of 97.27 % for stroke classification and a maximum F1-score of 91.75 % in distinguishing strokes from non-stroke actions. Another novel feature is that our stroke type classifiers can self-correct based on the gameplay model that excludes humanly impossible transitions of moves.