A large-scale analysis of athletes’ cumulative race time in running events

Abstract

Action recognition models and cumulative race time (CRT) are practical tools in sports analytics, providing insights into athlete performance, training, and strategy. Measuring CRT allows for identifying areas for improvement, such as specific sections of a racecourse or the effectiveness of different strategies. Human action recognition (HAR) algorithms can help to optimize performance, with machine learning and artificial intelligence providing real-time feedback to athletes. This paper presents a comparative study of HAR algorithms for CRT regression, examining two important factors: the frame rate and the regressor selection. Our results indicate that our proposal exhibits outstanding performance for short input footage, achieving a mean absolute error of 11 min when estimating CRT for runners that have been on the course for durations ranging from 8 to 20 h.