Head Pose Estimation Based On Computer Vision

According to incomplete statistics,China's annual death toll from traffic accidents exceeds 100,000,accounting for one-fifth of the death toll in global traffic accidents,ranking first in the world.With the rapid development of artificial intelligence,it is becoming more and more common to use artificial intelligence algorithms to solve daily tasks.In this context,it is possible to accurately and accurately judge the driver's head posture using computer vision technology,thereby reflecting the driver's mental state.Based on this,this thesis mainly studies how to design a large-scale head pose estimation algorithm with high precision and timeliness through a single image.During the driving process of the car,the algorithm estimates the mental state and the state of the driver by detecting the driver's head posture.The purpose is to effectively reduce the incidence of car accidents caused by fatigue driving and watching the mobile phone.The head pose estimation algorithm is the basis and focus of implementing this detection system.By consulting a large number of head pose estimation algorithm literatures,this paper designs two head pose estimation algorithms for the shortcomings of the head pose estimation algorithm.Experiments show that the head pose estimation algorithm can accurately and quickly estimate the head posture of the detected person.The algorithm in this paper can also be used in related fields such as living body detection and human-computer interaction.This paper mainly completes the following work:(1)Optimize face detection speed.The real-time and accuracy of the head pose estimation algorithm requires the face detection algorithm to be accurate and fast.In this paper,the face search strategy is improved based on the original face detection.Firstly,the foreground area is extracted by the background elimination model,the face is detected in the foreground area,the face coordinate value is obtained,and the face position information between the two frames is used to predict the face.Position,effectively reduce the algorithm search area and optimize face detection time.(2)Optimize face tracking accuracy.There are two problems with the existing target tracking algorithm: the tracking cannot be resumed when the tracking fails;the target cannot be repeated after the drift occurs.In view of these two problems,based on the Kernelized Correlation Filter(KCF)and Camshift algorithm,this paper proposes an improved method that can be re-tracked when the tracking fails.When tracking video targets,the algorithm calls the KCF and Camshift tracking algorithms in conjunction with background elimination modeling to design the "voting" module.Experiments show that the proposed algorithm can effectively resist occlusion interference,especially in the case of missing targets,it can re-initialize the exact position of the target and improve the robustness of tracking.(3)Propose a face tracking algorithm based on feature points.The coordinate position of the face feature points is detected,and the feature of the face position "degree of association" between the two frames is designed by using the face coordinate position.The feature is combined with the support vector machine classifier to design a support vector machine(Support).Vector Machine,SVM)face tracking algorithm.This algorithm can achieve multi-target tracking.This tracking algorithm has good specificity,high robustness and good practicability.(4)SVM-based head pose estimation algorithm is proposed.Using the coordinate position of 68 face feature points,an artificial feature describing the head pose is designed,and the extracted features are optimized.The optimized feature vector is combined with the SVM classifier to train the feature vector to obtain a head pose.Estimation algorithm.Experimental results show that this head pose estimation algorithm has higher accuracy and real-time performance.(5)Head pose estimation algorithm based on convolutional neural network is proposed.Parameter estimation is the core of the head pose estimation algorithm.A head pose estimation algorithm based on convolutional neural network is proposed.The neural network layer,width,depth and fully connected layer have been carefully analyzed and adjusted.A convolutional neural network structure is designed for head pose estimation.Touching glasses,hair,etc.are added during the self-built training set.Action,the test shows strong robustness.It can be seen from the experimental results that the head attitude estimation algorithm has strong anti-interference ability and high real-time performance.Finally,the entire head pose estimation algorithm has undergone a lot of tests,and the accuracy,robustness and real-time performance of the algorithm meet the requirements,which can be used in living body detection,vehicle assisted driving and human-computer interaction.