Research On Head Pose Estimation Method Based On Deep Learning

With the rise of artificial intelligence technology and the rapid development of deep learning,computer vision and pattern recognition have become hot research fields,and face recognition technology has been widely used in security,education,transportation and other aspects.Because a single face recognition method cannot provide richer head information,it cannot accurately detect and recognize in the presence of occlusion,complicated light environments,etc.The biological information feature recognition combined with head pose has become a research focus in the field of pattern recognition.Head pose estimation studies the calculation of three head positions relative to the camera in a given video or image: pitch,yaw,and roll.Traditional head pose estimation methods are affected by natural environmental factors,and are prone to large errors in complex lighting and local occlusion.This paper uses deep learning and target detection methods to research and improve the head pose estimation algorithm.The main work includes the following three aspects:The principles of convolutional neural networks and deep learning,the overall process of head pose estimation,and commonly used algorithms are discussed in detail,including face detection,face alignment,and camera annotation.At the same time,it includes the application of two improved methods of target detection algorithms in face detection,and the principle of head posture solution afterwards.A multi-loss head pose estimation method is improved.This method mainly uses the deeper deep residual network RestNet101 for feature extraction,uses the AdaBound algorithm for gradient optimization in the training phase,and finally performs multiple loss calculations for fine classification to obtain a more robust algorithm.Experiments on the data set have achieved better estimation results than the classical algorithms,and can meet the real-time performance.A multi-task-based fine-grained head pose estimation method is proposed.Using the advantages of multi-task learning in image processing,The paper considers face detection,landmark regression prediction and head pose estimation as three tasks to jointly train the network model,uses VGG16 as the backbone network to extract image features,and uses fine-grained modules to spatially group the extracted features to obtain more representative features,reducing memory usage and consumption of irrelevant calculations,and finally obtains the estimated results of the head pose main task.Experiments are performed on a common data set,which shows better robustness to illumination,and the average absolute error is lower than algorithms such as FAN.It is an effective head pose estimation method.