摘要：

在滑雪教学过程中, 由于学员的迅速移动和姿态的大幅变化, 导致以短时智能图像变化特征为主的姿态跟踪算法失效或不稳定, 特别是在雪地条件恶劣或光线不足等复杂环境下, 跟踪效果会受到较大的影响。为此, 提出基于长时跟踪的滑雪教学姿态辅助矫正方法。使用递归最小二乘法(RLS)分类器训练得到学员姿态位置核相关滤波器(KCF)。计算最大的KCF响应值, 精确检测学员的姿态位置。如果KCF结果低于经验阈值, 表明目标丢失, 启动再检测模块。利用光流法在前一帧学员姿态位置附近寻找当前帧中的姿态位置, 以获取一个大致位置。在此位置重新应用跟踪器, 获得精确的学员姿态位置, 实现长时跟踪。依据得到的姿态位置数据, 构建滑雪学员姿态误差补偿模型, 提取学员身体的运动参数和姿态误差。通过计算身体运动参数, 并结合KCF构建滑雪学员身体姿态辅助矫正模型, 从而完成滑雪教学的姿态辅助矫正。实验结果表明, 该方法在长时跟踪中展现出高度的有效性、可靠性和稳定性, 其平均PCK指标达到92.3%, 同时在目标跟踪效率上, 参数量和计算量分别为30.24 MB和9.26 GFLOPs, 速度达到142帧/s, 实现了高效实时的跟踪, 验证了该方法在滑雪教学姿态辅助矫正中的可行性。

关键词: 长时跟踪, 滑雪教学, 姿态辅助矫正, 相关滤波器, 误差补偿, 运动参数

Abstract:

In skiing teaching, the rapid movement and significant changes in the posture of students result in the failure or instability of posture tracking algorithms, which are mainly based on short-term intelligent image change features. The tracking performance can be greatly affected, especially in complex environments such as harsh snow conditions or insufficient lighting. To this end, a skiing-teaching posture correction method based on long-term tracking is proposed. The Recursive Least Squares(RLS) classifier is trained to obtain Kernelized Correlation Filters(KCFs) for student pose and position. Subsequently, the maximum KCF response value is calculated, and the posture position of the student is accurately detected. If the KCF result is below the empirical threshold, it indicates that the target is lost and the re-detection module is activated. Using the optical flow method, the pose position in the current frame becomes close to the position of the student's pose in the previous frame to obtain an approximate position. The tracker is re-applied at this location to obtain accurate student posture positions and achieve long-term tracking. Based on the obtained posture data, a skiing-student posture error compensation model is constructed, and the motion parameters and posture errors of the student's body are extracted. The body motion parameters are calculated and combined with the KCF to construct a skiing-student body posture-assisted correction model, thereby completing posture-assisted correction in skiing teaching. The experimental results show that this method is highly effective, reliable, and stable for long-term tracking. The average PCK index reaches 92.3%, and in terms of target tracking efficiency, the parameters and computational complexity are 30.24 MB and 9.26 GFLOPs, respectively. The speed reaches 142 frame/s, enabling efficient real-time tracking and confirming the feasibility of this method for posture assistance correction in skiing teaching.

Key words: long-term tracking, ski teaching, attitude assisted correction, correlation filter, error compensation, motion parameter