Improving data acquisition speed and accuracy in sport using neural networks

Video analysis is used in sport to derive kinematic variables of interest but often relies on time-consuming tracking operations. The purpose of this study was to determine speed, accuracy and reliability of 2D body landmark digitisation by a neural network (NN), compared with manual digitisation, for the glide phase in swimming. Glide variables including glide factor; instantaneous hip angles, trunk inclines and horizontal velocities were selected as they influence performance and are susceptible to digitisation propagation error. The NN was trained on 400 frames of 2D glide video from a sample of eight elite swimmers. Four glide trials of another swimmer were used to test agreement between the NN and a manual operator for body marker position data of the knee, hip and shoulder, and the effect of digitisation on glide variables. The NN digitised body landmarks 233 times faster than the manual operator, with digitising root-mean-square-error of similar to 4-5 mm. High accuracy and reliability was found between body position and glide variable data between the two methods with relative error <= 5.4% and correlation coefficients >0.95 for all variables. NNs could be applied to greatly reduce the time of kinematic analysis in sports and facilitate rapid feedback of performance measures.

Automated summary

This study compared the speed, accuracy, and reliability of 2D body landmark digitization using a neural network with manual digitization in the glide phase of swimming. The neural network digitized body landmarks 233 times faster than manual digitization, with a root-mean-square-error of around 4-5 mm, showing high accuracy and reliability. Results demonstrated strong agreement and correlation between body position and glide variable data obtained from the two methods.