基于神经网络的双目视觉下头部姿态估计

Head Pose Estimation under Binocular Vision Based on Neural Network

头部姿态估计是人体姿态检测的关键技术之一。本文基于神经网络设计一种在双目视觉下由人脸中的关键点在空间中的相对位置的变化估计头部姿态,并对头部进行定位的方法。将头部姿态分为6种,空间位置关系分为2种。利用改进SDM算法对双目视觉下的人脸关键点进行标记;标记出人脸关键点后利用的POSIT算法对头部姿态角度估计,计算出头部欧拉角;根据左右图像中对应的头部关键点位置的视差由三角测量原理算出其深度信息。并设定阈值对其进行分类。通过实验,该方法的头部姿态估计准确率高,头部空间定位精度良好。

Head pose estimation is one of the key technologies for human pose detection.Based on the neural network.A method is designed to estimate the head posture from the changes in the relative position of the key points in the face in space under binocular vision,and to locate the head.The head posture is divided into 6 kinds,and the spatial position relationship is divided into 2 kinds.Use the improved SDM algorithm to mark the key points of the face under binocular vision.After marking the key points of the face,use the POSIT algorithm to estimate the head pose angle and calculate the head Euler angle according to the corresponding head in the left and right images.The parallax of the key point position is calculated by the triangulation principle to calculate the depth information.Finally,set threshold to classify these poses.The experimental results showed that:the method has high accuracy of head pose estimation and good head space positioning accuracy.