Head Pose Estimation Based On Monocular Recognition Technology

The era of artificial intelligence is a data-driven era.With data,you can get the rules of the real world and help people make better decisions.In the field of image recognition and machine vision technology,vectors containing a large amount of data such as images and videos are also important data sources in the era of artificial intelligence.Different from ordinary tabular data,the representation and processing methods of these data directly determine how much useful information people can extract from it.Therefore,image recognition technology is a very important part of the field of artificial intelligence development.Head pose estimation is a hot research topic in image processing and machine vision.The detection process is non-contact and realtime,and can also detect and analyze existing image and video data,which has high practical value.The head pose estimation studied in this thesis is to process and analyze the twodimensional images acquired by the camera in monocular recognition technology to obtain the head posture.The steps are divided into face detection,face alignment,face tracking and pose calculation.Face detection and face alignment are used for face recognition and feature point localization of single images.Face tracking is used for accurate and fast face localization between frames in video.Pose calculation uses iterative algorithm to map the face model from two-dimensional face model into a three-dimensional face model.Among them,face detection consumes a lot of computing resources,resulting in a decrease in video real-time performance.Illumination variation and image noise affect the accuracy of tracking.The process of head pose estimation cannot balance real-time efficiency and accuracy.In view of the above problems,this thesis makes research in the following aspects:(1)Using face detection based on skin color segmentation and adaptive boosting and face alignment based on ensemble of regression trees to accurately get the location of face feature points,and experimentally test the real-time and robustness of the positioning effect.It was proved that the method flow had high accuracy of face detection and good effect of face alignment;(2)In view of the decline of real-time efficiency and failure of face detection,the adaptive particle filter tracking method with multi-feature fusion is introduced.Through the screening of multiple feature methods,central symmetric local binary mode is selected as the texture feature,and combined with skin color features,adaptively adjust their respective influence on the tracking results,which makes the algorithm more stable and accurate;(3)In the head pose calculating stage,the inter-frame difference method is used to determine whether the face motion between two frames exceeds the threshold.In the two frames that do not exceed the threshold,the optimized parameter matrix of the previous frame will become the initial parameter matrix of iterative algorithm for the next frame.The experiment proved that this optimized method not only stabilized the accuracy of the algorithm,but also enhanced real-time performance.