A Single Image Face Pose Estimation Method Based on Depth Value

Abstract

The invention relates to a method for estimating the face pose of a single image based on a depth value, comprising the following steps: obtaining a face image of a pose to be estimated; converting the face image of a pose to be estimated into a grayscale image, and performing denoising processing and standardization Processing; Introduce the principle of affine transformation invariance initialization to establish an initial model, improve the traditional active shape model method, and extract the two-dimensional values ​​of ten facial feature points; according to the principle of perspective imaging, use an iterative algorithm to obtain the ten facial features The depth value of the point; establish the face pose estimation objective function; use the four face feature points to estimate the face pose once, as the input value of the least squares optimization operation on the objective function, and output the face pose result of the second estimate; Use the M-estimation algorithm to further optimize the results obtained by the least squares optimization operation, and output the final estimation results. The invention overcomes the ill-conditioned problem of attitude estimation, and has good attitude estimation accuracy and robustness.

Description

technical field [0001] The invention relates to a single image human face attitude estimation method based on depth value. Background technique [0002] In recent years, although the development of depth camera technology and the advent of Microsoft Kinect have broken the situation of high prices of depth cameras in the past, and two new depth image databases, ETH Face Pose and Biwi Kinect, have appeared in the field of face pose estimation. In practical applications such as customs, airports, exhibition halls and other public places and public security systems for chasing criminals, it still takes time to collect and update the two-dimensional database to the three-dimensional database. Therefore, at present, multi-pose recognition for a single two-dimensional face image is still the mainstream . At present, the frontal face recognition system has achieved good recognition results, but the effective recognition of multi-pose face images is still not good. According to sta...

AI Technical Summary

Problems solved by technology

[0004] Based on the following points: (1) Multi-pose two-dimensional face image recognition is still the mainstream; (2) The geometric relationship method of facial feature points has the advantages of simplicity, short time consumption, and high efficiency; (3) Most of the existing pose estimation methods lack face features. Depth information causes the problem of ill-conditioned attitude estimation process; (4) the least squares method cannot distinguish the inner point and the outer point well, so there are problems affecting the estimation accuracy and robustness

Images