Research On The Algorithm Of Face Real-time Tracking Based On Deep Learning

With the enhancement of computer computing capabilities and the support of big data,deep learning has demonstrated its powerful feature learning capabilities.As one of the important Combination of deep learning and target tracking,face tracking has become a hot topic in the industry.In video sequences,face tracking algorithms based on detection methods are disturbed by factors such as ambient light changes,human posture changes,affecting the tracking effect.In order to achieve a robust and accurate face tracking,through deep learning,we extract two-frame image features and learning to obtain the face position information of the next frame,then position facial feature points in the face window.Therefore,this paper focuses on three aspects:face detection,face tracking and feature point location.The face detection part is the initialization of face tracking,providing tracking target position information and generating a face initialization window.This paper first introduces the principle of Haar-like cascading classifier based on AdaBoost for face detection.The analysis shows that this algorithm is trained by the gray image of human face,and it is easy to make mistakes in the background area that resembles the features of human face.Testing.In order to solve this problem,this paper introduces the skin color detection based on YCbCr color space in the traditional face algorithm,removes the non-skin color area and improves the accuracy of face detection by performing a series of noise reduction operations such as skin color extraction and erosion expansion reduce the false detection rate.Experiments show that this algorithm can effectively avoid the false detection of some non-faces and improve the stability of face detection.After the face detection,the face is tracked.With the change of light in the complex scene,the occlusion of the object,and the change of the human posture,a robust face tracking algorithm is required.This paper firstly introduces the process and principle of GOTURN target tracking algorithm based on deep learning,next analyzes the advantages and disadvantages of using it for face tracking.Finally,experimental tests show that,it was found that GOTURN tracking algorithm has a scale deviation in face tracking.Aiming at this problem,this paper proposes a GOTURN face tracking algorithm based on a correction network.A correction network is added on the basis of the feature extracted from the initialization window to perform feature extraction on the face of the training set,and the face scale is reduced in the tracking process.deviation.The experimental results show that the optimized GOTURN algorithm significantly improves the tracking performance of the face tracking and improves the tracking stability.Finally face window obtained by face tracking continues to locate feature points.This paper firstly introduces the structure of cascaded convolutional neural network model proposed by Tang's research group.The algorithm uses deep learning to extract image features,then uses a cascade structure to enable accurate positioning of facial feature points.In view of the problem of network model's instability in feature point location,this paper uses the first-level parallel structure of Tang's algorithm model structure,refers to the AlexNet model,and introduces batch normalized layer and image color channel information in the network to increase the parallel model prediction data.reliability.Experiments show that the optimized algorithm can reduce the error while reducing the number of training iterations,and further improve the model's prediction accuracy.