Computer Vision Framework Predicts Jianzi Trajectories and Hitting Types in Tests

A computer vision and pose estimation framework combining object localization and human skeleton tracking has achieved integrated trajectory prediction and hitting strategy estimation for jianzi play. Developed by researchers Jian Huang and Wenchao Bi and published in Discover Applied Sciences, the system addresses the analytical challenge of jointly representing the irregular projected flight of the jianzi shuttlecock and an athlete’s hitting motion. The technical framework couples YOLOv8s localization with 17-keypoint skeleton sequences extracted from video footage, establishing continuous observation channels that link the projectile’s aerial path directly to the player’s physical posture.

To capture the physical dynamics of flight, the framework augments two-dimensional image-plane observations with several derived motion features, specifically principal-axis orientation, angular velocity, aspect ratio, and lateral acceleration. In addition to these visual proxies, the architecture integrates a gravity and quadratic-drag consistency term that supplies a soft, short-horizon motion prior. This regularizing mathematical formulation embeds aerodynamic constraints directly into the network’s predictive pathways, helping ground statistical feature learning in physical principles governing projectile deceleration and free fall.

The computational architecture relies on specialized deep learning modules to process trajectory paths and hitting actions simultaneously. Trajectory prediction is supported by graph attention mechanisms paired with Transformer encoding, which model spatial and temporal relationships across sequential video observations. Concurrently, a channel-attention spatiotemporal graph convolutional network processes the 17-keypoint skeleton sequences to jointly estimate the hitting type executed by the athlete and the shuttlecock’s subsequent landing position, coordinating movement recognition with outcome forecasting.

The system was evaluated on a benchmark dataset comprising 200 match clips recorded from 28 athletes, using athlete-isolated splits to ensure that test data evaluated models on previously unseen players. On this test collection, the full model achieved a nominal 20-frame average displacement error of 21.4 pixels for trajectory prediction. In hitting analysis, the architecture demonstrated a hitting type classification accuracy of 91.3 percent and a Macro-F1 score of 89.7 percent. Furthermore, the system attained a landing mean absolute error of 14.2 pixels, alongside an exact nine-zone agreement rate of 78.4 percent across target court sectors.

Component ablation experiments demonstrated the operational contribution of each structural module within the unified pipeline. When retaining the identical observation channels but replacing graph attention with average pooling, the nominal 20-frame trajectory error rose from 21.4 pixels to 23.3 pixels. Replacing the Transformer encoder with a Long Short-Term Memory network produced a further increase in trajectory error to 24.0 pixels. Isolating the physical consistency term while keeping the state proxies showed that aerodynamic regularization accounted for a modest error reduction from 21.7 to 21.4 pixels. Model optimization testing under a nominal 50 percent pruning budget retained 35.7 percent of model parameters while maintaining a 22.8-pixel trajectory error and 89.1 percent hitting type accuracy. Evaluated on edge hardware, a Jetson Xavier NX processing benchmark logged 40.7 frames per second, a processing throughput capacity that the authors note remains distinct from total observation-to-output latency.

To assess operational utility, the researchers conducted a within-participant evaluation involving 12 coaches and eight athletes comparing unassisted performance against system-cued assistance. Among the participants, mean landing judgement accuracy measured 61.3 percent without technical assistance compared to 84.6 percent when assistance was provided. For the eight participating athletes, mean footwork initiation time was recorded at 412 milliseconds without the cue and decreased to 337 milliseconds with cue presentation. Crucially, the authors emphasize that these paired observational differences strictly describe the evaluated experimental conditions and do not by themselves separate cue assistance from possible practice or condition-order effects. Overall, the reported findings support integrated trajectory and action analysis strictly within the evaluated dataset and participant sample.

Subject of Research: Computer vision and pose estimation for jianzi trajectory prediction and hitting strategy analysis

Article Title: Optimized design of a jianzi trajectory prediction and hitting strategy analysis system based on computer vision and pose estimation

Article References: Huang, J., & Bi, W. (2026). Optimized design of a jianzi trajectory prediction and hitting strategy analysis system based on computer vision and pose estimation. Discover Applied Sciences. https://doi.org/10.1007/s42452-026-09739-y

Image Credits: AI Generated

DOI: 10.1007/s42452-026-09739-y

Keywords: jianzi, computer vision, pose estimation, YOLOv8s localization, 17-keypoint skeleton sequences, trajectory prediction, hitting strategy analysis, graph attention, Transformer encoding, spatiotemporal graph convolutional network, channel attention, quadratic-drag consistency term, average displacement error, landing position estimation, model pruning, Jetson Xavier NX, footwork initiation time