kropsposeestimation

Kropsposeestimation, in English body pose estimation, is the task of determining the configuration of the human body in images or video. It typically involves locating a standardized set of joints—such as the head, shoulders, elbows, wrists, hips, knees, and ankles—and estimating their coordinates in two or three dimensions, along with confidence scores for each joint. Outputs may be 2D coordinates in the image plane or 3D coordinates in a human-centered or camera coordinate system.

Modern methods rely on deep learning. Approaches are commonly categorized as top-down, where a detector finds each person before a pose is estimated, and bottom-up, where joints are detected first and individuals are grouped afterwards. Notable architectures include OpenPose, HRNet, MoveNet, and AlphaPose, among others. 2D models are widely used, while 3D pose estimation aims to recover depth from monocular images, multi-view setups, or depth sensors, often requiring temporal information for video.

Key datasets include COCO Keypoints, MPII for 2D pose, PoseTrack for video, and 3D datasets such as Human3.6M and 3DPW. Common evaluation metrics include PCK (percentage of correct keypoints) and PCKh for 2D, and MPJPE (mean per-joint position error) for 3D pose, as well as PCKA, AP-based measures, and PA-MPJPE for calibrated depth.

Applications span sports analytics, animation, virtual/augmented reality, human-computer interaction, and ergonomics. Challenges include occlusion, fast motion, clothing variability, crowd scenes, viewpoint changes, and privacy concerns in public spaces. Ongoing trends emphasize real-time performance, robust 3D reconstruction from monocular video, and efficient models for mobile devices.