Concept
Pose Estimation
Computer vision technology that detects the position of a person's body joints - enabling AI to understand human posture, movement, and gesture from video or images.
Added May 18, 2026
Pose estimation identifies the spatial locations of key points on a human body - joints like shoulders, elbows, wrists, hips, knees, and ankles - from images or video. The output is a skeleton: a set of 2D or 3D coordinates for each joint, connected into the anatomical structure of the body. From this skeleton, downstream systems can analyse posture, recognise gestures, track movement, and understand physical activity.
The task is challenging because humans appear in enormous variety: different body sizes, clothing, lighting conditions, occlusions (parts hidden behind objects or other people), and camera angles. A pose estimation model must generalise across all of this while maintaining precise joint localisation.
Modern pose estimation approaches are mostly deep learning-based. A convolutional or transformer backbone processes the image and produces feature maps. These features are used to generate heatmaps - one per joint - where the peak of each heatmap indicates the most likely location of that joint. The final joint coordinates are extracted from these heatmaps.
Two main paradigms exist. Top-down approaches first detect each person using an object detector, then estimate the pose of each detected person independently. This is accurate but slow when many people are present. Bottom-up approaches detect all joints in the image simultaneously and then group them into individuals. This scales better to crowded scenes but is more complex.
3D pose estimation extends the task to recover full 3D joint positions from 2D image observations. This is substantially harder because a 3D pose is not uniquely determined by a 2D image - depth is ambiguous. 3D pose models use body priors (knowledge of what anatomically plausible 3D poses look like) to resolve this ambiguity.
Applications span sports analytics (detailed movement analysis for training), animation (capturing actor motion for game characters or VFX), healthcare (gait analysis for rehabilitation, fall detection for elderly care), fitness apps (form correction during exercise), and human-computer interaction (gesture control interfaces).
Analogy
The way motion capture suits work in film production - detecting the precise position of every joint in an actor's body to drive a digital character. Pose estimation does the same from regular camera footage, without requiring the actor to wear any special equipment.
Real-world example
Google's Pose Estimation API (MediaPipe Pose) runs on mobile devices in real time and has enabled a wave of fitness apps: apps that analyse your squat form as you exercise, count repetitions correctly, and warn you when your posture puts you at risk of injury. The entire pipeline runs on the phone camera, with no server required.
Why it matters
Pose estimation is what allows AI to understand human physical presence and action - not just recognise that a person is present, but understand what they are doing with their body. This is the perceptual foundation for sports analysis, human-robot interaction, rehabilitation technology, and motion capture workflows that would otherwise require expensive specialised equipment.
In the news
No recent coverage - check back later.
Related concepts