Research And Application Of Abnormal Behavior Recongnition Method In Exam Based On Spatio-temporal Feature Analysis

With the rapid development of technology,artificial intelligence has been widely applied in computer vision,of which abnormal behavior detection has always been a research hotspot.However,in the practical abnormal behavior detection in videos,the existing single-frame detection algorithms have problems such as imbalance of speed and accuracy,and lack of timing dependencies.Nowadays,video behavior detection algorithms based on deep learning can simultaneously extract video frame spatial features and inter-frame temporal features with high behavior recognition accuracy,which has attracted extensive attention of researchers.Therefore,based on the limitations of single-frame detection,this paper focuses on the abnormal behavior of candidates in exam,and conducts corresponding algorithm model research and application deployment.First of all,due to the lack of related public datasets,a handcrafted normal behaviors recongnition dataset is made from scratch in this paper.The dataset in this paper selects 4 types of representative behaviors for behavior positioning and classification,and completes the labeling of 12,081 frames of images,including a total of 160,284 rectangle labels.Secondly,aiming at the problems of unbalanced speed and accuracy of and lack of timing dependencies in single-frame detector,a time-series-based video behavior detection model ST Yolov4 is proposed in this paper.Based on Yolov4,this model includes 3 innovations: 1)Replace the PANet module of Yolov4 with the spatial pyramid ASPP module,and release the binding between the model features on the premise of ensuring multi-scale detection.2)Introduce the temporal key frame mechanism to reduce the highly redundant feature calculation of the continuous frames,and balance the model inference speed and detection accuracy more flexibly.3)The 2CS-ConvLSTM module is proposed to extract the spatiotemporal features of the continuous frames to improve the accuracy of behavior detection.Experiments show that,on the examination abnormal behavior recognition dataset,ST Yolov4 compared with Yolov4,the detection accuracy mAP is increased by 1.92%,and the model speed is increased by 7.3 fps.Finally,this paper completes the deployment and application of the proposed ST Yolov4 model and algorithm,and implements a Web-based examination abnormal behavior detection system.The system can detect and display the behavior of candidates in the examination room in real time,alarm and archive the abnormal behavior of the candidates,and also provides the function of checking the abnormal behavior information after the examination,which ensures the traceability of the invigilation process.