Research On The Detection Method Of Candidates In The Examination Room Based On SSD

Teaching,learning and management based on artificial intelligence is becoming an important symbol of modern education in China.Examination is not only a common form of checking students' learning,but also an important means of talent assessment and selection.Although the current examination classrooms has installed a lot of monitoring equipment in accordance with relevant regulations,it has not been able to automatically identify and understand the behavior of candidates with the help of advanced computer vision technology.So,the invigilation method still takes up a lot of power and time.In this dissertation,the object detection in standardized examination classroom is taken as the research object.On this basis,we study the examinee position detection method with deep learning,which serves for the analysis and understanding of candidates' behavior in the subsequent intelligent examination classroom.The main innovative work in this thesis includes:(1)Taking the examinees in the standardized examination classroom as the research object,the image dataset of "Examinee Position Detection Dataset" is constructed.According to the position distribution and morphological characteristics of examinees and invigilators in standardized examination classroom,the appropriate video frame is selected.After manual annotation with annotation tool LabelIing,the dataset of examinee position detection "Examinee Position Detection Dataset" is constructed.On this basis,a method of examinee position detection based on SSD(Single Shot multibox Detector)is proposed to verify the dataset.The experimental results show that the proposed method is accurate and effective,and the dataset improves the accuracy of examinee position detection.(2)In order to improve the detection accuracy of examinees in standardized examination classroom,a method of examinee position detection based on SSD-DU is proposed according to the scene characteristics of monitoring equipment acquisition and imaging in standardized test room.SSD is improved in the following aspects:Firstly,the adaptive threshold of Intersection-over-Union is used for dividing the samples to increase the proportion of the positive samples in the training process;secondly,the feature up-sampling is introduced into the prediction layer of SSD to improve the extraction ability for small and weak objects;finally,the above two detection results are combined to balance the detection effect of different scale targets.The experimental results show that SSD-DU can effectively detect the object position in the real test scene,and the mAP(mean Average Precision)reaches 86.05%,which is 15.15%higher than the original SSD.(3)A detection method of real-time video moving object based on SSD512 is introduced after EPD dataset was expanded.In this method,SSD512 is used to detect the examinee position in the surveillance video of examination classroom for obtaining the position information of examinees in each frame.Then,the difference degree of the same object position between frames is calculated to determine whether the position has changed or not.Finally,the real-time detection of moving object in consecutive frames is carried out.The experimental results show that the proposed method can effectively detect the positions of all candidates and invigilators in the video,and find the object whose position changed in real-time in the monitoring video,which can reduce the visual fatigue of invigilators.