yinguobing/head-pose-estimation
Head Pose Estimation
This project is a computer vision tool designed to calculate the pitch, yaw, and roll of a human head in real time. It functions as a facial landmark detection library that identifies sixty-eight specific points on a face to enable precise orientation tracking from video streams.
The system utilizes a high-performance inference engine to execute pre-trained machine learning models, ensuring that geometric analysis is performed efficiently on input frames. By mapping detected facial coordinates to a three-dimensional model, the software determines the camera-relative rotation of the head through perspective projection mathematics.
The pipeline integrates standard computer vision libraries to manage video acquisition, frame processing, and the sequential execution of pose estimation algorithms. This approach allows for the extraction of spatial orientation data from moving subjects captured via connected cameras or local video files.
Features
- Head Pose Estimation - Calculates the pitch, yaw, and roll of a human head using facial landmarks and perspective algorithms.
- Facial Landmark Detection - Identifies sixty-eight specific facial points to generate the geometric data required for pose analysis.
- Head Pose Tracking - Monitors the orientation of a human head in real time to determine where a person is looking.
- Perspective-n-Point Solvers - Solves the relationship between 2D image points and 3D object points to determine head rotation.
- Facial Landmark Analysis - Maps facial landmarks to 3D space to enable geometric analysis of head orientation.
- Inference Engines - Executes pre-trained machine learning models to perform real-time geometric analysis on video streams.
- ONNX Runtime Inference - Executes pre-trained models using a cross-platform runtime for high-speed facial feature detection.
- Real-Time Video Analysis - Captures and analyzes live video frames to perform real-time pose estimation on moving subjects.
