Automatic fault detection in race walking from a smartphone camera via fine-tuning pose estimation

Automatic fault detection is an important issue in sports. In race walking, researchers have attempted it using sensors and machine learning. However, there exists problems with sensor attachments, conflicting with referee’s judgement, and the model’s interpretability. Here, we propose a fault detection system via fine-tuning pose estimation from a smartphone camera and logistic regression models trained by referee’s judgement. Using normal race-walk and walking with intentional faults, we showed that the proposed system detected faults with an average accuracy of over 90%, and the model used grounds for detecting faults according to the rules of race walking.