Research On Face Detection In Crowd Gathering

At present,people mainly recognize the crowded faces by watching videos,which can not find effective targets in the crowd gathering scene in time.The accuracy and speed of face detection in clustered crowd are the primary conditions to ensure face recognition.Due to the complex environmental impacts such as occlusion,angle change,illumination change and size change,there is still some work to be done in the accuracy and speed of face detection for clustered crowd.In this paper,based on the deep residual 101 layer network model(ResNet101),face detection of clustered crowd is studied,and the following results are obtained:(1)Aiming at the problem of the accuracy of face detection for clustered crowd,the 101-layer network model of deep residual is combined with the selected nearest neighbor interpolation,bilinear interpolation and bicubic interpolation algorithm to detect and analyze the face of clustered crowd,which achieves satisfactory detection accuracy.The main ideas are as follows: firstly different interpolation algorithms are selected to scale the multi-scale image sequence,then the ResNet101 is used to detect the face area of the multi-scale image,and finally the non-maximum suppression is used to fuse image.The experimental results show that the average detection accuracy of the combination of ResNet101 and bicubic interpolation algorithm is 97.71% and the average false detection rate is 2.28% on the WIDER FACE data set collected by the Chinese University of Hong Kong.(2)Six image sharpness evaluation algorithms,including Average Gradient,EAV,Edge Intensity,Entropy,PSNR and Spatial Frequency,are implemented.The experimental results are used to analyze the differences of each evaluation algorithms.(3)Aiming at the problem of the speed of face detection in clustered crowd,six image sharpness evaluation algorithms,including Average Gradient,EAV,Edge Intensity,Entropy,PSNR and Spatial Frequency are selected to analyze the multiscale image sequence.It is found that the sequence scaling factor is between 1.0075 and 1.0125,and the multi-scale image sequence with stable image quality can be obtained.The scaling factor of the sequence is 1.0075 to 1.0125.And the scaling factor effectively reduces the number of multi-scale image sequences.Combining the scaling factor of interval sequence,bicubic interpolation algorithm and the ResNet101,the detection speed is improved without affecting the detection accuracy.The experimental results show that: The average detection speed of this method is 20.941 s and the average accuracy is 97.73% on the Aggregating Crowd data set WIDER FACE of the Chinese University of Hong Kong.The average detection speed of Hu algorithm is 23.894 s.The detection speed of this method is slightly higher than that of Hu algorithm.The face image data collected from crowd gathering scenes such as auditorium,military training venue,conference room,classroom,shopping mall and station with large illumination change,face scale change and posture change show that the average detection speed of this method is 24.017 s,the average accuracy is 95.08%,and the average detection speed of Hu algorithm is 27.883 s.The detection speed of this method is also slightly higher than that of Hu algorithm.