Research On Statistics Algorithm For Participants In Video Conference

Considering the problems of occlusion,people moving with a small range,multi-venue polling of the multi-venue polling videos.Traditional people counting means have low detection accuracy and high false positive on video conference.HOG operator is widely used in pedestrian because of Optical and Geometric invariance,but in video conference scene,but people are occluded and background has a serious influence on detection,so that HOG operator is not best for people counting in video conference scene.Face recognition algorithm generally includes three frames such as face detection,feature extraction and feature contrast,face recognitions has high detection rate on simple scenes,but for video conference scene with occlusions,face recognition has low detection accuracy.The human head detection algorithm is used in video surveillance with hanging top camera,so it is not suitable for video conference scene.Then we improved an algorithm of head-shoulder model detection.Firstly,we use dualhybrid Gaussian background model to extract the foreground.Then the Hu moment invariant features of foreground image counters can be calculated.The moment of change is matched to the minimum distance.Finally,we use skin color detection to eliminate the interference from the human target.But the detection accuracy still does not meet the engineering requirements.For these reasons,this paper proposed two improvement.1)Firstly,the two-mixed Gaussian background model is used to extract the foreground,and compared with the higher hybrid background,the model uses different updates speeds,Gaussian model updating background can extract the foreground of video conference scene better with smaller motion range.2)There are so much occlusion phenomena in video conference,which causes the error of statistics,in view of this,this paper adopts a method of connected domain multiplex detection to partially solve the population statistic error caused by occlusion problem,and when the foreground is extracted,a corrosion operation of the foreground pixel is formed into the connected domain,This paper makes a threshold judgment for the pixel energy statistics of the connected domain,and determines whether the connected domain is the influence of the people occlusion or noise and small objects,so as to make a retest of the population statistics and improve the accuracy of the statistics.3)This paper improved and proposed an adaptive energy model algorithm,first,through the two-mixed Gaussian background model to extract the required foreground information,and then using adaptive energy to construct a human's occupied pixels are not with the distance between the camera and the change of energy model to deal with the foreground pixel information,By using the angle information of the vertex and inflection point of the foreground connected domain,the occlusion degree of the connected domain is close to the reference person,and the number of participants in the statistical video conference scene is improved significantly compared with the accuracy rate of the traditional method.In summary,the improved algorithm can be used for real-time statistics of video conferencing scene,and it has high statistic precision,and has good robustness to the occlusion and bow of the meeting scene.