Head Pose Estimation Algorithm Study And Implementation Base On Deep Neural Learning

Head pose estimation is the process of inferring the orientation of a human head from a digital imagery using computer vision and pattern recognition technology.It requires a series of processing steps to convert pixel-based representation of a head into high-level concepts of direction.As the main research content of attention direction and human head movement analysis,it has a wide application prospect in the field of computer vision.For example,in some driving assistance systems,it is necessary to estimate the driver's head posture for monitoring the driver status;or in some online teaching systems,it matters to judge the learning state of students in the same way;and it helps to improve the accuracy of human face alignment.At present,more and more researches on head posture tend to use deep learning method,and have also achieved certain results.However,there are still some problems that affect the accuracy of head pose estimation:(1)the method of using Euler angle to represent head pose makes the label not unique;(2)the model is easy to be affected by lighting,occlusion and other factors,cause generalization performance of the model is poor;(3)the angle change is not obvious in face imaging,make the feature are difficult to refine.In addition,we are also faced the challenges of data annotation and improving model performance.In view of the above problems,this paper mainly proposes a method based on face orientation vector to solve the problem that data labels are not unique due to the singularity of Euler angle;proposes a soft label to solve the problem that angle features are difficult to refine,making the model training more stable and easy in continuity regression task;at the same time,develops a tagging tool based on Unity3d for data tagging.Finally,building a real-time face detection and head pose estimation simulation system.The above scheme has been preliminarily verified in the simulation driving environmentThe research work of this paper is mainly reflected in the following aspects(1)In order to solve the singular problem of Euler angle,a label representation method based on face orientation is proposed in this paper.By studying and analyzing the rotation theory of objects in three-dimensional space,transfer Euler angle into rotation matrix to obtain the face orientation vector.This effectively avoids the case that the label of Euler angle is not unique in the large angle condition(2)In order to improve the accuracy of vector coordinate prediction,this paper proposes a classification method based on soft label to solve the problem of ordinal regression.The ordered and continuous change of the angle can be directly reflected in the continuous change of the image In this paper,the continuous coordinate change is first classified into coarse-grained categories,and then the soft labels of the samples are generated by measuring the distance between the categories The experimental results show that the proposed soft tag can make the model discover hidden relationships between classes in the learning process,which is more conducive to model training(3)In terms of improving the generalization ability of the model,this paper proposes and develops a head pose annotation tool based on Unity3D.By using Unity3D engine to manipulate the 3D head model,the quaternions of rotation is recorded by keeping the same head pose as that in the two-dimensional image.In this paper,this tool is used to annotate the non-public infrared data And enhances the data through flipping,random sampling,blurring,gray processing,etc..The diversity of model training data and the generalization performance of the model is greatly increased(4)In order to improve the performance of the model,this paper improves and experiments the model with several different basic networks.Through the comparison of model accuracy,model file size,FLOPS(floating-point operations per second)and the FPS(Frames Per Second),the most suitable network structure for terminals integration is obtained(5)Build a complete head pose estimation simulation system.For real-time head pose estimation tasks,it is necessary to quickly detect where the face is.Therefore,this paper optimizes and compresses the SSD detection network and builds a lightweight face detection network.Based on the face detection network and the head pose estimation network proposed in this paper,a real-time head pose estimation simulation system is designedIn addition to the main research content mentioned above,a large number of comparative experiments are carried out in this paper.The experimental results show that the algorithm model based on face orientation vector can effectively estimate the pose of human head in two-dimensional image,and effectively avoid the problem of large angle error caused by Euler angle singularity Secondly,the soft label proposed in this paper links the relationship between classes directly with the distance between classes,so that the model can be optimized easier than the one-hot label in the training process.Finally,by comparing several experiments,the most suitable model for real-time estimation of head posture and the model with the highest accuracy are obtained.