cmu-perceptual-computing-lab/openpose

OpenPose is a real-time computer vision library developed at Carnegie Mellon University that automatically detects and tracks the positions of human body parts in images and videos. Given a photo or a video frame, OpenPose identifies where a person's joints are, shoulders, elbows, wrists, knees, ankles, and many more, by drawing a "skeleton" of keypoints over each person in the scene. It handles multiple people simultaneously and was among the first systems to do this in real time. The library detects up to 135 keypoints across four categories: 25 points for the body and feet (covering the major joints), 21 points per hand (for detailed finger tracking), and 70 points for the face (for precise facial landmark positions). There is also a 3D reconstruction module that, when multiple camera views are available, can estimate where those joints are in three-dimensional space rather than just on a flat image. The technical approach uses deep learning models trained on large human pose datasets. A convolutional neural network (CNN) processes the image and produces confidence maps indicating where each type of joint is likely to be, then a separate algorithm assembles those detections into coherent skeletons for each individual in the scene. The body detection runtime stays constant regardless of how many people appear in the frame, which is what makes it fast enough for real-time use. You would use OpenPose for motion capture without special suits or equipment, sign language recognition, human-computer interaction systems, sports biomechanics analysis, fitness apps, animation, robotics, and any research requiring detailed understanding of how people move. It is written in C++ with Python bindings, supports CUDA-accelerated GPUs for maximum speed, and can also run on CPU-only machines at reduced speed. It is free for academic and non-commercial use.