Research On Real-time Face Detection And Alignment

With the rapid development of the modern science and technology,applications of computer vision are more and more widely used in social life,face detection and alignment technology is one of the hottest research topics,which is the basis of many applications related to face like 3Dmodel reconstruction of a face,facial expression recognition,face recognition and so on.Face detection has been able to get a good performance of faces of frontal view,however,there are still many challenges in face detection with complex background.HOG+SVM algorithm is a classic algorithm which is used widely in face detection.But it still has some limitations.It is not fast enough and only use a single kind of feature.This paper proposes a multi-feature cascade SVM classifier using parallel threads with a faster extraction of HOG feature and a faster image pyramid building and scanning way,which improves the precision of face size detection and efficiency,while guarantees a certain accuracy at the same time.Most of today's face alignment algorithm are based on single images,however,the practicability,accuracy and stability of feature point have some certain limitations while used in the video.First is the precision of face alignment of rolling faces.Second is the uncertainty of the face got from a face tracker.Last is the unstable feature points got from face alignment in video caused by the changed environment and light.This paper proposes three improvements to enhance the performance of face alignment in videos,including feature points stability,alignment of rolling face and the certainty of the tracked face.Extensive experiments have been carried out and the results show that our strategies are effective.This paper aims to improve the robustness and practicability of face detection and alignment in video,first proposes a multi-feature cascade SVM classifier using parallel threads whose efficiency is 2.6 times of the HOG+SVM algorithm,and then proposes three improvements to enhance the performance of face alignment in videos.