Research And Realization Of The Face Detection System Based On Video

Face detection in video stream has received significant attention and has a widerange of applications, such as video surveillance, image retrieval, human-computerinteraction, and so on. However，the main challenge remains the real-time detection offace with different poses.In this thesis, a novel face detection method for human faces in different angles inthe video stream is introduced. And in the process the scheme combining thefirst-detection and post-validation is showed and the detecting algorithm based onAdaboost is improved. First of all, the development and technologies of face detectionin video is introduced; Second, the face detection system design and implementation isshowed in details; Third, the system is tested in the standard databases and then theresult is analyzed.The most important part of this paper is the design and realization of the detectionsystem, which is divided into four sub-modules: sub-module1for firstly detecting theactive object, sub-module2for detecting human faces, sub-module3for validatinghuman faces, sub-module4for tracking human faces.Innovation of our algorithm is that firstly, as a result of sub-module1, which isused to decrease the detecting area and simplify the environment, the system is speededup greatly while keeping lower false alarm rate and secondly, due to extending haar-likefeatures, the system can detect the faces in different angles successfully and thirdly,because of the combination skin color model with temple-matching, the system canvalidate and locate the faces with a success.The main goal of sub-module1is to obtain the active object containing facespossibly. Firstly, according to the idea of the background subtraction GMM model isused to separate the areas of interest from the background in the video, and the sizes andlocations of them are computed by means of the connected-component algorithm.And Sub-module2based on Adaboost is used to detect and obtain faces ascandidates in the area with a interest. This paper mainly demonstrates the trainingprocess of the algorithm and then the detector prototype is constructed, and finally the fact that the module can rapidly detect faces with95.35%hit rate in video is approvedby the result of experiment.Sub-module3embedding a novel cascade scheme, i.e. two-validation combiningthe skin color model in the HSV and YCbCr space and template-matching, is utilized tovalidate the candidates and then obtain real human faces. After the validating, the falsealarm rate can be greatly reduced to only0.48%, while maintaining the original hit rate.Sub-module4makes use of the effective Camshift algorithm to track the facedetected and validated.The final testing demonstrates that the detector is able to detect and track faces indifferent orientations with a great success and can be used with a wide range.