OpenPose — распознавание поз тела в реальном времени

★ 34.4K

OpenPose is an open-source library for real-time multi-person pose estimation: given a photo, video, webcam, or IP camera stream, it detects up to 135 keypoints covering body, hands, face, and feet, even when multiple people appear simultaneously in the frame. Reach for it when you need to analyze athletes' movement technique without motion-capture suits, annotate a pose dataset for model training, build gesture-based controls, or assess a rehabilitation patient's posture automatically. The core is written in C++ with Python bindings; it supports CUDA (Nvidia), OpenCL (AMD), and CPU-only modes, runs via CLI or API, and includes a 3D reconstruction module for multi-camera setups plus a Unity plugin. Output formats include PNG, JPG, AVI for visualization and JSON, XML, YML for keypoint coordinates. Runs on Linux, Windows, macOS, and Nvidia TX2. The direction is detecting poses of people already captured on camera — not generating motion, not identifying individuals by face. A discrete GPU is recommended for real-time performance; the CMU license restricts commercial use, and a separate agreement with the university is required for product deployment.