Research On Multi-view Face Detection Technology In Video

Social security has been a major problem in people's livelihood.As for the densely populated and highly mobility city,the public security problem is more serious.For the development of electronic technology,a variety of social security occasions have been installed the camera for monitoring,which alleviate the pressure on public security.After the incident,the video data can provide clues about the important information.Even so,the face information collection of video data still needs human resources.The computer detection can automate this process very well,but for the detection technology,some special situation is still difficult to handle: such as detection for partial target;the position of the camera may be artificially movement,cause rotation and occlusion problem will interfere effect of detection.In the past,the detection method is better in the up-right face detection,but it is not good for the detection results of the rotation and partial occlusion.For this reason,we have trained a detector which has strong adaptability in rotation and occlusion.using the video attribution of the spatio-temporal consistency to improve the detection effect in video,also trained a convolution neural network to estimate the pose of human faces in different viewpoints.Specific contributions are as follows:Firstly,train a face detector based on the convolution neural network.Optimize the network model design,reduce the number of parameters to learn.Using data augmentation and improving sample distribution to improve learning effect.The partial face is introduced as weak positive sample for solving the occlusion problem.For the problem that the weak positive samples and the positive samples have the same score,cause the failure of the non-maximum suppression algorithm which dependent on the score and the overlap rate,propose to use the strategy of imbalanced samples to enhance the positive samples' score.Secondly,There are motion blur,occlusion,morphological changes,illumination changes in the video,and only the use of image detection methods can not achieve good results.the detector is used in the video,and the performance of the detector is compared with that of the video detector without the use of the video features such as the spatio-temporalconsistency,the motion field and other information.Furthermore,for the multi-view problem,different view-point face has different pitch,yaw and roll.The traditional pose estimation methods usually get some of the key points position of the face by using the regression method,such as the eyes,nose,mouth,and then use those key points to match as much as possible with a 3D face model ‘s 2D projection,in order to get the pose of face.In this paper,a convolutional neural network is trained to obtain the pose of the human face,avoiding the process of making 3D model and matching with the 3D model.