Research Of Student Position Detection And Number Estimation System Based On Face Recognition In Video Streaming

Checking attendance is the basic class management tool.The most common method of checking attendance is to call and answer from a list of registered students,but this original manual method not only occupies the teaching time,but also does not rule out the phenomenon of late arrival or replacement.Therefore,the checking attendance method using the human biological characteristics has been widely used.Among them,the face feature is the most natural and intuitive biometric recognition method because the subject does not need to touch a device.However,the most common method based face recognition is the face scanning method,which requires the subject to actively approach the device,and the device can read the face image with a clear fixed posture.The development of deep learning in the field of face recognition makes it possible to apply face detection and recognition in an unconstrained environment.In this thesis,the face detection and recognition technology based on deep learning is applied to checking class attendance.The combination of a panoramic camera and a PTZ(Pan,Tilt and Zoom)camera is used to complete the HD face image acquisition for each student,and realize the student number estimation function.The main work of this thesis is as follows:1.According to the problem causing by face occlusion,posture and face image size in the classroom environment,this thesis divides the classroom attendance problem into two parts:position detection and face recognition.In this thesis,moving target detection is used to assist face detection to compensate for the position detection of students when face information is missing.A detection box fusion algorithm based on multiple detection box weights is proposed to obtain non-repetitive effective student position coordinates.The coordinates are modeled on the PTZ imaging coordinates,and the PTZ is automatically controlled to focus on the students at each position,and the face image set with less occlusion and ideal size is obtained for face recognition.2.For the problem that the lack of local data set and that the face classifier cannot be trained well,this thesis introduces the unsupervised learning method into the checking attendance system to complete the number estimation function.Through a period of focused face image collection,this thesis establishes a HD face dataset,uses FaceNet network with improved loss function to perform face verification on the dataset,and face similarity to clear face dataset.Clustering is performed to finally output face images of all different classes.3.To improved testing result of the detection algorithm and the face recognition algorithm proposed in this thesis,this thesis collects and produces the image dataset containing the student's face in the actual classroom environment.As a result,the detection rate reaches 93%,especially in the case that the face is occluded seriously,the algorithm is 23.3%higher than the FDNet detection rate.By combining the improved Loss function,the accuracy of the face recognition is up to 98.6%.