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

Accurate analysis of jianzi play requires the joint representation of irregular projected flight and the athlete’s hitting motion. This study presents a computer vision and pose estimation framework that combines YOLOv8s localization with 17-keypoint skeleton sequences. Principal-axis orientation, angular velocity, aspect ratio and lateral acceleration augment the image-plane observations, while a gravity and quadratic-drag consistency term provides a soft, short-horizon motion prior. Graph attention and Transformer encoding support trajectory prediction, and a channel-attention spatiotemporal graph convolutional network jointly estimates hitting type and landing position. On 200 match clips from 28 athletes with athlete-isolated splits, the full model achieves a nominal 20-frame average displacement error of 21.4 pixels, hitting type accuracy of 91.3%, Macro-F1 of 89.7% and landing mean absolute error of 14.2 pixels, with 78.4% exact nine-zone agreement. Retaining the same observation channels while replacing graph attention with average pooling or the Transformer with LSTM increases trajectory error to 23.3 and 24.0 pixels, respectively. With the state proxies retained, the consistency term reduces error from 21.7 to 21.4 pixels, identifying a modest additional regularization contribution. At a nominal pruning budget of 50%, the model retains 35.7% of its parameters and achieves 22.8-pixel trajectory error and 89.1% hitting type accuracy. The Jetson Xavier NX processing benchmark reports 40.7 frames per second; this capacity is distinct from observation-to-output latency. In a within-participant evaluation involving 12 coaches and eight athletes, mean landing judgement accuracy is 61.3% without assistance and 84.6% with assistance. Among the eight athletes, mean footwork initiation time is 412 ms without the cue and 337 ms with it. These paired differences describe the evaluated conditions and do not by themselves separate assistance from possible practice or condition-order effects. The results support integrated trajectory and action analysis within the evaluated dataset and participant sample.