Physio AI Companion. In Data Science & Engineering Master of Advanced Study (DSE MAS) Capstone Projects

Physical Therapists (PT) and Kinesiologists recommend a series of exercises but often face challenges in continuously monitoring individuals performing exercises to ensure correct postures and prevent injury aggravation. This research attempts to address this issue by building a product designed to automate the detection of the incorrect exercises and provide users with timely feedback. The research effort began with a set of curated exercise videos, a set of biomechanical standards as well as developing a core model to analyze a single exercise - overhead squat. The core model approach consists of three main steps: preprocessing to standardize the videos, creating a 3D body model and calculating incorrectness scores for each repetition and aggregation based on measured joint angles. The work uses state-of-the-art computer vision models and computational algorithms for a customized solution. The results from the core model are used to provide feedback to both practitioners and users through visual overlays on the exercise video and graphical presentation of biomechanical measures captured during the exercise.

物理治疗师（Physical Therapist，PT）与运动机能学家（Kinesiologist）通常会为训练者推荐一系列训练动作，但往往难以持续监控练习者的动作姿态是否规范，亦无法及时防止运动损伤加重。本研究旨在解决这一痛点，开发一款可自动识别错误训练动作并为用户提供及时反馈的产品。研究启动之初，团队首先构建了精选训练动作视频数据集、生物力学标准库，并针对单种训练动作——过头深蹲（overhead squat）开发了核心分析模型。该核心模型的实现流程包含三大主要步骤：对输入视频进行标准化预处理、构建三维人体模型，以及基于实测关节角度计算每组动作的错误评分并完成聚合汇总。本研究采用当前最先进的计算机视觉模型与计算算法，打造定制化解决方案。核心模型的输出结果将通过在训练视频上叠加可视化标注，以及以图表形式展示运动过程中采集的生物力学指标，为从业者与练习者双方提供针对性反馈。