Research On Head Pose Estimation Method Based On Deep Learning

Head pose estimation algorithm is a method to obtain head orientation information in 3D space through computer vision technology.With the rapid development of artificial intelligence technology,head pose estimation has been widely used in the fields of humancomputer interaction,driver monitoring and medical detection.However,the current head pose estimation algorithms still face many challenges.For example,(1)insufficient feature extraction of the algorithm may lead to problems such as low recognition accuracy and insufficient generalisation ability of the algorithm;(2)while improving the recognition accuracy of the algorithm,it will increase the complexity of the model due to increasing the depth of the network,etc.In this paper,the following improvements are mainly made to address the above problems:(1)A three-branch head pose estimation algorithm with multi-stage feature fusion is proposed to address the existing head pose estimation algorithms in terms of insufficient feature extraction and low recognition rate of the algorithm.The algorithm has a multi-stage output structure,using three different types of tributary networks to extract features from the input image respectively,and there are three stages on each tributary,each stage only needs to refine the features of the previous stage,and the feature maps extracted from the same stages are passed through the feature fusion module to generate feature maps,which effectively avoids the feature loss problem;the feature extraction module selects the Ghost module as the feature extraction The feature extraction module chooses the Ghost module as the feature extraction network and uses model compression to make it reduce network parameters and computation while ensuring network accuracy;to extract effective features of greater importance,the efficient channel attention module ECA-Net is introduced,thus improving the accuracy of head pose estimation.The algorithm showed excellent performance with a reduced MAE of 4.68 and 3.59 on the AFLW2000 dataset and BIWI dataset respectively,and a model size of only 0.55 MB.(2)A lightweight head pose estimation algorithm based on binary neural networks is proposed to address the problems of large number of network parameters and low recognition accuracy of existing head pose estimation algorithms,which make it difficult to maintain a balance between the two.Firstly,the binary neural network can effectively reduce the number of parameters of the algorithm,and the binary neural network is chosen to replace the normal convolutional neural network in the feature extraction module.The Sim AM attention mechanism is then introduced in the feature fusion module,which can improve the accuracy of the algorithm by focusing on the fusion of robust features without increasing the number of parameters.The algorithm has a MAE of 3.65 and 4.21 on the BIWI and AFLW test sets respectively,and the algorithm model is small at 0.30 MB,effectively lightweighting the model.(3)Combining head pose estimation technology and face simulation technology to design and implement an AR fun face changing real-time interactive system.The system uses the lightweight head pose estimation algorithm proposed in this paper to perform fast and accurate head pose estimation for the user;when the user’s head pose is recognised,either automatic or manual mode can be selected,and in automatic mode the corresponding face transformation can be performed for the user’s face deflected in different directions,while in manual mode the face rendering can be selected according to the user’s own wishes.Experiments have proven that the system not only achieves high recognition accuracy in head pose estimation,but also allows for fast and accurate face transformation in the system,creating a new interactive experience for the user.