Real-time human face attitude estimation method based on depth video streaming

Abstract

The invention discloses a real-time human face attitude estimation method based on depth video streaming. The real-time human face attitude estimation method comprises two stages of sampling and training and estimating in real time. The sampling and training stage comprises the following steps of obtaining filed depth images of all human face angles; then conducting random sampling on the filed depth images of all angles and obtaining a training sample set; and using a supervised learning method to conduct training, and obtaining a classifier. The real-time estimation stage comprises the following steps of firstly extracting real-time human face filed depth images in the depth video streaming output by collection equipment and converting the real-time human face filed depth images into integral images; conducting random slice sampling on the integral images, utilizing the classifier obtained by training to conduct classification on samples, and obtaining a plurality of estimation results; and removing abnormal results for the results and conducting weighted average, and obtaining final human face attitude results. The real-time human face attitude estimation method based on the depth video streaming avoids influences of factors like illumination on the final results, and is good in instantaneity and precision.

Description

technical field [0001] The present invention relates to a recognition method, in particular to a real-time face pose estimation method based on depth video stream. Background technique [0002] At present, the user interacts with the computer mainly through the keyboard, mouse and touch screen, all of which rely on specific hardware input devices. Natural human-computer interaction methods have become the focus of current research, such as: human body posture, facial posture, facial expression analysis, etc. In addition, in face recognition, face pose estimation also has a very important application. After the face pose is estimated, the photo can be deformed according to the face pose before recognition, and then the recognition can be performed, which can greatly improve the accuracy of face recognition. [0003] The existing face pose estimation methods are all based on two-dimensional images and videos, and such methods still have the following problems: [0004] 1) T...

AI Technical Summary

Problems solved by technology

[0004] 1) The collected data is susceptible to noise interference

Traditional two-dimensional images and videos are easily affected by lighting, which will completely change the gray value and texture information of certain areas in the picture, making it impossible for the sample set in the training phase to take care of all possible situations, so that the accuracy rate in the generalization phase linear decrease

[0005] 2) The accuracy rate drops seriously under the condition of missing features

[0006] 3) Manual intervention is required during system operation

This type of method needs to initialize the face position when the system starts running, and it is easy to lose the target when the face moves quickly or encounters an occluder

Images