Abstract: Intelligent motion coaching can provide automated and personalized support for sports training, motor skill instruction, and rehabilitation, and therefore has significant practical value. Existing studies mainly focus on Action Quality Assessment (AQA), error identification, or textual feedback, while current motion correction generation methods remain limited in correcting complex continuous motions. To address this issue, this paper proposes an edit-region-guided text-driven motion correction generation method. First, a motion editing selector is constructed to generate an editing prediction signal that characterizes the editing importance of different joints at different time steps based on the learner's original motion and expert textual feedback. Subsequently, the editing prediction signal is combined with joint-level body-region masks to construct edit-region guidance weights, which are then transformed into continuous attention biases and incorporated into a diffusion denoising network. This mechanism enhances the diffusion denoising network to focus more feedback-relevant regions during generation while reducing unnecessary modifications to unrelated motion regions. Comparative experiments are conducted on an expert-feedback paired motion-correction dataset constructed from Ego-Exo4D. The proposed method achieves R@1, R@2, and R@3 scores of 3.13%, 5.76%, and 8.59%, respectively, for Generated-to-Target Retrieval (E2T), with a Median Rank (MedR) of 49. Edit-region Mean Per Joint Position Error (Edit-MPJPE), non-edit-region preservation Mean Per Joint Position Error (Keep-MPJPE), Mean Per Joint Position Error (MPJPE), and Procrustes-Aligned Mean Per Joint Position Error (PA-MPJPE) are 17.06 cm, 21.17 cm, 16.75 cm, and 8.86 cm, respectively. Compared with TMED (Text-based Motion Editing Diffusion), SimMotionEdit (text-based human Motion Editing with motion Similarity prediction), and 3D Pose Based Feedback for Physical Exercises (abbreviated as 3DPBF in this paper), the proposed method achieves the best E2T retrieval performance. Specifically, R@1, R@2, and R@3 improve by 0.81, 0.91, and 1.82 percentage points, respectively, over the second-best method, SimMotionEdit, while MedR improves from 66 to 49. In addition, on the Exercise Correction in 3D (EC3D) dataset, the proposed method achieves a correction success rate of 88.14%, outperforming 3DPBF by 11.87 percentage points. Experimental results demonstrate that the proposed method enables targeted motion correction and achieves favorable correction performance across different motion correction datasets, providing effective support for generating corrective motion demonstrations in intelligent motion coaching systems.

Key words: motion correction generation, diffusion model, text-driven motion editing, intelligent motion coaching, attention mechanism

摘要： 智能动作指导为体育训练、动作教学及康复指导提供自动化与个性化支持，具有重要应用价值。现有研究多侧重于动作质量评分、错误标签或文本反馈，已有纠错动作生成方法在复杂连续动作的修正方面仍存在不足。为此，提出一种编辑区域引导的文本驱动动作纠错生成方法。该方法首先构建运动编辑选择器，根据学习者原始动作与专家文本反馈生成编辑预测信号，用于刻画不同关节在不同时刻的编辑重要程度；随后结合关节级身体区域掩码构建编辑区域引导权重，并将其转换为连续注意力偏置引入扩散去噪网络，使该网络在生成过程中增强对反馈相关区域关注的同时减少对无关动作区域的非必要修改。在基于Ego-Exo4D构建的专家反馈动作纠错配对数据集上的对比实验中，本文方法Generated-to-Target Retrieval(E2T)的R@1、R@2和R@3分别达到3.13%、5.76%和8.59%，Median Rank(MedR)为49；编辑区域平均关节位置误差(Edit-MPJPE)、非编辑区域保持平均关节位置误差(Keep-MPJPE)、平均关节位置误差(Mean Per Joint Position Error，MPJPE)和Procrustes对齐平均关节位置误差(Procrustes-Aligned Mean Per Joint Position Error，PA-MPJPE)分别为17.06 cm、21.17 cm、16.75 cm和8.86 cm。与TMED (Text-based Motion Editing Diffusion)、SimMotionEdit (text-based human Motion Editing with motion Similarity prediction)和3D Pose Based Feedback for Physical Exercises(本文简称3DPBF)相比，本文方法在E2T检索指标上取得最优结果，其中R@1、R@2和R@3较次优方法SimMotionEdit分别提高0.81、0.91和1.82个百分点，中位检索排名MedR由66前移至49。此外，在Exercise Correction in 3D(EC3D)数据集上，本文方法的纠错成功率达到88.14%，较3DPBF的76.27%提高11.87个百分点。实验结果表明，本文方法能够实现针对性的动作纠错，并在不同动作纠错数据集上均取得了较好的实验结果，可为智能动作指导中的纠正动作示范生成提供有效支持。

关键词: 动作纠错生成, 扩散模型, 文本驱动动作编辑, 智能动作指导, 注意力机制