Pose estimation
Also known as: body pose detection, skeleton tracking, keypoint detection
AI that finds the positions of body joints such as shoulders, knees and hands in video, building a skeleton that describes how a person moves.
Updated
What it means
Pose estimation locates key points on the human body in each frame, such as the head, shoulders, elbows, wrists, hips, knees and ankles, and connects them into a simple skeleton. Over time, the skeleton shows clearly how a person moves: crouching before a jump, extending in the air, landing and riding away.
It is used in sports analysis, coaching, fitness apps, animation and motion capture, and as one of the signals in action recognition. In combat sports, it can describe stances and strikes; in skiing, body angles through each turn of a run.
Pose estimation is a branch of computer vision and works from ordinary video, without any special suits, sensors or markers.
How it works in practice
Clear views of the whole body give the best results. Side angles of a jump, a fixed camera covering a sparring area, or a follow camera behind a skier all show the body well.
POV footage from a helmet or chest shows little of the rider body, so pose estimation has little to work with there. Third-person angles complement POV for any analysis that depends on body movement.
Good light, a contrasting background and a moderate distance from the subject all improve accuracy. Baggy clothing and protective gear can hide joints and lower reliability noticeably.
What to watch out for
Fast motion, heavy motion blur and very unusual positions, such as inverted tricks or tangled grappling, confuse pose models. Results may jump or flip between frames, especially at lower frame rates.
Pose estimates are approximate. They are useful for trends and highlights, less so for precise biomechanical measurements unless the system is designed and calibrated for that purpose by experts.
Multiple people overlapping, as in grappling or crowded scenes, make it hard to assign joints to the right person, so results need careful checking before you draw conclusions from them.
How RawClip handles it
RawClip uses AI analysis of motion, jumps, faces and peak action to choose moments, rather than detailed body measurements. Upload third-person and POV clips together, and the AI picks the best moments for your highlight.