Study On Methods For State Detection Of Watchman In Surveillance Video

Along with the modernization of the Armed Police Force, the video surveillancetechnologies have become more and more important to the tasks of on duty. As one ofthe most crucial components of surveillance systems, the watcher will directlydetermine the normal operation of the duty systems and efficient treatment under theemergency situation. Therefore, the state detection and alert of the watcher has greatsignificance to improve the safety factor and prevent the various security incidentsduring the period of on duty.This paper mainly studies the surveillance video-based analysis method forwatcher’s state detection. Firstly, The position of watcher’s face is obtained based onthe analysis of surveillance video and so that we can identify the watcher is absence ornot. Then, we can locate and track the eye of the watcher based on the facial image. Byanalyzing the state of the eye, we can ultimately determine whether the watcher is tiredor not.The main contributions of this thesis are described as below:In the research of face detection, this thesis proposes a face detection method basedon the combination of skin color and improved AdaBoost algorithm.Considering thegood assembling character of the skin color in the color space, the thesis firstlyestablishes a two-dimensional Gaussian model to segment the skin regions from thedetected face image and obtains the candidate face regions. Then the exact location ofthe face regions are obtained by cascade classifier which is trained by improvedAdaBoost algorithm. Experimental results show that the proposed method can not onlyeffectively reduce the false acceptance rate and improve correct detection rate but alsobe more accelerate than classical AdaBoost algorithm.In the research of eye location, this thesis investigates the method of eye locationbased on extended Haar features and vertical integral projection. On the basis of facedetection, we can firstly reduce the search space of the eye region according the priorknowledge of the distribution of facial features. The location of the eyebrow and eyeregion can be obtained by using AdaBoost algorithm with an extended Haar feature.Finally, we can remove interference of eyebrows by using vertical integral projectionand obtain the precise position of the eye. In the research of eye tracking, this thesisstudies the real-time eye tracking algorithm combined with Kalman Filter and MeanShift. The experiment results show that the above methods can achieve a good effect oneye locating and tracking。In the research of eye state identification, this thesis proposes an eye state detectionmethod combined with template matching and projection method. Through thecalculation of the correlative coefficient between the test image and two states templates images, an absolute value of difference images can be compared with the presupposedthreshold to determine the eye state. If the correlative coefficient between test imageand two states templates is too closer, the projection method can be used to judge e yestates. Finally, the PERCLOS method combined with eye blink frequency is used to thefatigue judgment of watcher. The experimental results show that the proposed methodhas a good performance on detection and identification for different state of watcher.In the research of application, the Armed Police Force monitor states detectionsystem has been designed on the basis of this thesis. We propose the framework as wellas the software and hardware requirements of the system design. The experimentswhich implement under laboratory environments show that the proposed detectionsystem can achieve the desired results.