Improving data acquisition speed and accuracy in sport using neural networks.

Area of Science:

• Biomechanics
• Sports Science
• Artificial Intelligence

Background:

• 2D video analysis is crucial for deriving kinematic variables in sports.
• Manual digitisation of body landmarks is time-consuming and prone to error.
• Automating this process can enhance efficiency and feedback in sports performance analysis.

Purpose of the Study:

• To evaluate the speed, accuracy, and reliability of a neural network (NN) for 2D body landmark digitisation.
• To compare NN digitisation with manual digitisation for the swimming glide phase.
• To assess the impact of NN digitisation on key kinematic glide variables.

Main Methods:

• A neural network (NN) was trained on 400 frames of 2D swimming glide video.
• The NN was tested against manual digitisation using data from four glide trials.
• Agreement was assessed for body marker positions (knee, hip, shoulder) and derived glide variables.

Main Results:

• The NN digitised landmarks 233 times faster than manual methods, with a root-mean-square error of ~4-5 mm.
• High accuracy and reliability were observed between NN and manual methods.
• Relative error was ≤5.4% and correlation coefficients were >0.95 for all analysed glide variables.

Conclusions:

• Neural networks offer a substantial speed improvement for kinematic analysis in sports.
• NN-based digitisation provides accurate and reliable performance measures.
• This technology can facilitate rapid feedback for athletes and coaches, optimizing training.