Progress in Machines and Systems Hathairat Ketmaneechairat 15 1 2026 https://doi.org/10.6025/pms/2026/15/1/1-18 https://www.dline.info/pms/fulltext/v15n1/pmsv15n1_1.pdf Accurate acquisition of athletes' motion trajectories under high-speed, complex environmental conditions remains a major challenge in sports biomechanics and intelligent training analysis. This study proposes a framework for capturing Games movement trajectories using the mean shift algorithm to address limitations of conventional tracking methods in dynamic training environments. A skeletal representation consisting of 16 joints and 51 degrees of freedom is constructed to model human motion during Games training. Quaternionbased joint representation and logarithmic mapping are employed to reduce rotational singularities and enable dimensionality reduction within a linear time-invariant system. To reduce environmental parameter dependency, the probability density function in the gradient iteration framework is integrated with kernel density estimation, while colour histograms are used as stable visual features for target localisation. Additionally, Difference of Gaussian (DoG)-based feature detection and SIFT-assisted fusion tracking are introduced to enhance robustness under illumination variation and motion blur. Experimental validation demonstrates improved detection accuracy, reduced false alarm rates, and superior clustering efficiency compared with conventional estimation-based tracking approaches. The proposed framework provides a reliable, adaptive, and computationally efficient solution for acquiring Games motion trajectories without requiring external parameter tuning