Vision-based movement recognition reveals badminton player footwork using deep learning and binocular positioning

Abstract

Coordinating dynamic interceptive actions in sports like badminton requires skilled performance in getting the racket into the right place at the right time. For this reason, the strategic movement and placement of one's feet, or footwork, is an important part of competitive performance. Developing an automated, efficient, and economical method to record individual movement characteristics of players is critical and can benefit athletes and motor control specialists. Here, we propose new methods for recording data on the footwork of individual badminton players, in which deep learning is used to obtain image coordinates (2D) of their shoes and binocular positioning to reconstruct the 3D coordinates of the shoes. Results show that the final positioning accuracy is 74.7%. Using the proposed methods, we revealed inter-individual adaptations in the footwork of several participants during competitive performance. The data provided insights on how individual participants coordinated footwork to intercept the projectile, by varying the distance traveled on court and jump height. Compared with visual observations by biomechanists and motor control specialists, the proposed methods can obtain quantitative data, provide analysis and evaluation of each participant's performance, revealing personal characteristics that could be targeted to shape the individualized training programs of players to refine their badminton footwork.

Keywords: Badminton player trajectories; Binocular positioning; Computer vision; Coordination; Deep learning.

Figure 1

Workflow of the proposed methodology.

Figure 2

The VGG16 network architecture (adapted from [37]).

Figure 3

Camera setup and definition of the world coordinate system.

Figure 4

Shoe identification results: (a) both the shoe location and category are accurately identified; (b) only the shoe location is accurately identified; and (c) neither the location nor category of the shoe is accurately identified.

Figure 5

Footwork trajectory of a player.

Figure 6

Results of movement distance and maximum bounce height among participants (line a: Average movement distance of males, 137.24 m; line b: Average maximum bounce height of males, 0.237 m; line c: Average movement distance of females, 133.73 m; line d: Average maximum bounce height of females, 0.21 m).

Figure 7

The relationship between distance moved for each shot and winning points.