Detecting And Tracking Human In Examination Room Based On Deep Learning

With the development of human society,it is the most important means of ability evaluation to select social talents through examination.In recent years,in order to ensure the fair selection of talents and maintain the authority of the examination,the funds invested by examination centers at all levels and relevant employers in the field of invigilation have increased year by year.At present,most of them still use the methods of manual invigilation and manual inspection,which can not effectively avoid the occurrence of abnormal situations in the examination room.In order to improve the level of invigilation,many research institutions and technology companies began to apply intelligent monitoring technology to the examination scene,and obtained a large number of examination room video monitoring data.However,these monitoring data still rely on manual access and analysis,which is boring and inefficient.There is an urgent need for an automatic invigilator system to liberate manpower.In the automatic invigilator system,the detection of examination room personnel behavior events is the most important part of the intelligent invigilator system.The behavior events of the examination room personnel refer to the events such as the change of the posture and position of the personnel in the examination room during the period of answering questions in the examination room,This paper studies the detection algorithm of examination staff behavior events based on target tracking technology,which can detect and record the examination staff behavior events in the examination room,and provide corresponding technical accumulation for the implementation of intelligent invigilator system.The specific research work is as follows:(1)Establish relevant data sets.According to the actual research needs,the test monitoring video data is collected and recorded.The video clips from the beginning of the test to the end of the test are mainly extracted.The key image frames are manually labeled and processed into data sets in batch,which provides data support for the performance of training model and test model.(2)The monitoring video in the actual examination scene is a kind of wide-angle overhead video monitoring scene,which has two characteristics:Firstly,the redundancy of surveillance video is high,and only foreground people in the video image will cause events.Second,the proportion of people in the examination room in the image is small,so it is difficult to carry out reliable target detection.In this paper,Yolo algorithm is improved by adding pre differential filter and uncertainty measurement cost of detection frame,which effectively reduces the detection cost and improves the detection speed and success rate.(3)According to the demand of automatic invigilation,a tracking model of the staff in the examination field based on the deep learning target detection algorithm is designed.The feature extraction and target detection are carried out based on Yolo model.The target tracking is carried out based on the traditional filtering algorithm and even graph maximum weight matching algorithm.It can not only effectively solve the problem of missed detection caused by the occlusion of the personnel,but also improve the detection personnel’s performance It is the reliability of the event,and can realize the whole process personnel event detection according to the time limit characteristics of the test scene,and generate the event log of the test site with the ID of the person in the test site as the index.