Facial Feature Detection And Its Application In Video Surveillance

As information technology thrives, intelligent multi-media is widely applied in the fields of security and detective work. Facial feature detection(face alignment) always draws a lot of attention from researchers, which is regarded as a key method of face recognition and intelligent surveillance. This paper mainly discusses about the facial feature detection in the aspects of algorithm and application. Facial feature detection research in this paper mainly includes shape model with classifier and global to local explicit shape regression for facial feature detection. Meanwhile, application based on facial feature detection in this paper consists of detection and tracking in videos.1. This paper proposes two methods based on shape model with classifier. One is the combination of Active Shape Model(ASM) and Random Forest. The other is the combination of ASM and Adaboost. The experiments compare the two proposed methods to ASM and the corresponding classifier, afterwards; validate the advantages of detection speed and accuracy of the two proposed methods. Additionally, the experiments find out how the searching space affects accuracy of Adaboost. As a result, the optimization of searching space is available, namely, the parameters are able to be tuned for optimization.2. This paper applies explicit shape regression in local shape detection as a method of local explicit shape regression, which improves the effectiveness of facial feature detection on facial details. Furthermore, we analyze global explicit shape regression and local explicit shape regression, and propose global to local explicit shape regression for facial feature detection. This method inherits the advantages of global explicit shape regression in the aspect of shape constrain and the advantages of local explicit shape regression in the aspect of local detail. The experiments analyze the specificity of each method among global explicit shape regression, local explicit shape regression and our proposed method in the perspective of regressive estimator. Meanwhile, in the perspective of the numbers of global regressors and local regressors, the experiments analyze the importance of global shape restrain and local detail in global to local shape regression. Finally, the analysis provides evidence for parameter tuning.3. This paper also discusses about the application of facial feature detection and tracking in videos. Firstly, the status quo of visual tracking research is analyzed. We analyze the current problems of tracking algorithms by comparison between Learning a Deep Compact Image Representation for Visual Tracking and High-Speed Tracking with Kernelized Correlation Filters in experiments. Secondly, we develop a video surveillance system using pedestrian detection algorithm based on Hog and SVM, face detection based on Adaboost and facial feature detection based on ASM with Random Forest, which illustrates the key points of facial feature detection. We draw a conclusion that High-Speed Tracking with Kernelized Correlation Filters performs better in tracking in complex background. Meanwhile, Learning a Deep Compact Image Representation for Visual Tracking is more suitable for tracking in simple background. Moreover, High-Speed Tracking with Kernelized Correlation Filters is faster than Learning a Deep Compact Image Representation for Visual Tracking.