Head pose estimation
This repo shows how to estimate human head pose from videos using TensorFlow and OpenCV.
The code is tested under Ubuntu 16.04.
How it works
There are three major steps:
Face detection. A face detector is adopted to provide a face box containing a human face. Then the face box is expanded and transformed to a square to suit the needs of later steps.
Facial landmark detection. A custom trained facial landmark detector based on TensorFlow is responsible for output 68 facial landmarks.
Pose estimation. Once we got the 68 facial landmarks, a mutual PnP algorithms is adopted to calculate the pose.
The marks is detected frame by frame, which result in small variance between adjacent frames. This makes the pose unstaible. A Kalman filter is used to solve this problem, you can draw the original pose to observe the difference.
The 3D model of face comes from OpenFace, you can find the original file here.
The build in face detector comes from OpenCV. https://github.com/opencv/opencv/tree/master/samples/dnn/face_detector
The code is licensed under the MIT license. However, the pre-trained TensorFlow model file is trained with various public data sets which have their own licenses. Please refer to them before using this code.
(Currently we are trying to empower embeded devices to be more powerful with deep learning, please get in touch if you are interested and would like to join us.)