A Novel Improved YOLOv8 Approach for Micro‐Motion Trajectory Modeling and Real‐Time Athlete Motion State Estimation

ABSTRACT There are problems of insufficient accuracy in capturing small motion trajectories and delay in real‐time motion state estimation in the training of track and field athletes. Based on the improved YOLOv8 framework, this paper introduces 3D convolution to construct a spatiotemporal feature pyramid (ST‐FPN) and embeds an optical flow‐guided cross‐frame feature alignment layer for motion compensation. A dual‐branch detection head is designed: the main branch uses the Anchor‐Free mechanism to detect the overall contour, and the auxiliary branch locates the joint points through key point heat map regression and uses the attention mechanism to fuse features. An LSTM‐Transformer hybrid trajectory prediction model is constructed to encode the motion parameters of historical frames, combine the self‐attention mechanism to capture long‐range dependencies, and incorporate Kalman filtering for kinematic constraints. Experiments show that the detection accuracy ([email protected]) reaches 89.6%, the average accuracy of joint positioning is 93.75%, the trajectory continuity retention rate is 96.4%, and the end‐to‐end latency is 25.8 ms on Jetson AGX Xavier, providing real‐time data support for scientific training.