Cách kiểm tra mẫu bài tập của bạn bằng AI

Pose estimation is the core technology. A neural network looks at each video frame and outputs coordinates for body landmarks, called keypoints. OpenPose, released by Carnegie Mellon researchers in 2017, made real-time multi-person pose estimation widely available. Google's MediaPipe Pose tracks 33 landmarks, and MoveNet predicts 17 keypoints in the common COCO format. Apple's Vision framework also includes body pose detection. Form apps build on models like these. They connect the keypoints into a stick figure, calculate angles at the knee, hip and elbow, and compare them against thresholds, such as whether the hip dropped below the knee. This works well for things a camera can see clearly: squat depth from the side, rep counts, tempo, left-right differences, and big errors like half-range push-ups. It works poorly for others. A single camera produces a flat image, so movement toward or away from the lens is hard to measure. Knees caving inward are nearly invisible from the side, while depth is hard to judge from the front. Small changes in spine position, bracing, breathing, pressure through the feet and grip are mostly invisible. Bar path needs separate object tracking. Loose clothing, a squat rack or plates blocking the view, poor lighting and fast movements all make keypoints less accurate. The main misconception is that an app saying 'good form' means the lift is safe. The thresholds are generic, and people's bodies differ. A lifter with long thigh bones may need more forward lean to squat well, and ankle mobility changes what works. The app does not know your injury history or how heavy the weight feels. For better results, set the camera on a stable surface at about hip height, keep your whole body in frame, and film from the side and the front. See a coach when you feel pain, when you are learning heavy barbell lifts, during rehab, or when you stop making progress.