一种基于特征点的快速人脸朝向检测算法
2021-12-24高泠溦李志扬邓蕾杨泽信邹颖
高泠溦 李志扬 邓蕾 杨泽信 邹颖



摘要:人脸朝向是判断学生是否认真听课的重要线索,而其中所涉及的物体姿态检测技术一般需要6轴传感器、KINECT等专业设备。本文提出了一种采用单摄像头的便捷人脸朝向检测算法,该算法在基准照片和转动后的照片中找出人脸的5对匹配特征点,然后根据光学成像原理,通过粒子群算法找出这些匹配特征点所对应的旋转矩阵,从而求出旋转角度。实验表明,对刚性物体该算法的测量精度可以达到2度左右。该方法可以广泛用于人脸朝向检测,或其他物体的姿态检测。
关键词:人脸检测;人脸朝向;特征点;姿态估计;粒子群算法
中图分类号:TP391 文献标识码:A
文章编号:1009-3044(2021)29-0105-03
Fast Face Orientation Detection Method Based on Feature Points
GAO Ling-wei, LI Zhi-yang, DENG Lei, YANG Ze-xin, ZOU Ying
(College of Physical Science and Technology, Central China Normal University, Wuhan 430079, China)
Abstract: Face orientation is an important clue to see whether a student is listening carefully, in which the ect posture detection usually requires professional equipment such as 6-axis sensor and KINECT. The paper proposed a convenient face orientation detection method using only a single camera. It first finds 5 pairs of matched face feature points from the reference photo and the rotated photo. Then it calculates by means of the particle swarm algorithm the rotation matrix and next the rotation angle of these pairs of feature points from the equations set up following geometrical optics. As the experiments show, a measurement accuracy of about 2 degrees can be reached for rigid ects. Therefor it can be widely used for face orientation detection and posture detection of other ects.
Key words:face detection; face orientation; feature point detection; attitude estimation;particle swarm algorithm
1引言
在全面防控疫情的背景下,各級各类学校积极开展线上教学[1]。网络教学具有很多优势,例如,资源共享以及学习形式趋向自由化。但是网络教学也带来一些不便,例如难以实时监测和把控学生的学习状态。通过人体姿态,特别是人脸朝向,可以判断学生是否认真听课。而对人脸或其他物体的姿态检测一般需要专业硬件设备或辅助装置,例如6 轴传感器[2]、KI? NECT[3]、双目摄像头[4]以及辅助棋盘格[5]等。人体姿态估计也一直是计算机视觉中一个备受关注的研究热点[6]。传统的姿态估计算法基于结构模型,往往应用在二维图场景中,受制于几何模糊性,其精度一般较低;随着深度学习的流行,近年来提出大量基于深度学习的检测算法,例如R-CNN[7]、SSD[8]等,估算精度得到大幅提升。……
