Real-time Fine-grained Abnormal Behavior Detection Based On Human Skeleton

Examination is a strict method of knowledge appraisal.Through the examination,students' learning ability and knowledge reserve can be assessed.In order to ensure the justice and fairness of the examination results,there must be strong discipline and restraint in the examination spots,especially the patrol staff and other examination staff to supervise the examination process,absolutely prohibit any form of cheating.Besides,cheating candidates have to bear legal and criminal responsibility.The traditional means of invigilation mainly rely on manpower,which is to set up the corresponding staff in each examination room to invigilate,patrol and deal with emergencies in the examination room.Large scale centralized examinations will consume a lot of human resources of examination institutions,and the scope of vision of human is limited,which will inevitably lead to the omission of some cheating behaviors.Although digital devices such as surveillance cameras have been widely used in various test scenes,these devices only have the function of simple video recording and playback.The recognition task still relies on a lot of manpower to complete,which is lack of intelligence.Most of the existing intelligent invigilation methods use optical flow method,hog and other methods to repeatedly process a large number of redundant background information,which has the disadvantages of low recognition accuracy and slow recognition speed.Aiming at the technical vacancy of intelligent invigilator,a real time fine-grained abnormal behavior detection method based on human posture is proposed.The contributions of our work mainly include the following aspects: firstly,we build a basic technical pipeline of intelligent invigilator by applying Open Pose technology and recurrent neural network,innovatively use the human pose estimation method to complete the invigilator task.Then we simplify the structure and compress the model of Open Pose method,and realize the real-time and parallelism of the method;Secondly,considering that there are a large number of cheating micro actions in the actual exam scene,we propose a fine-grained action recognition optimization method for recurrent neural network,which combines temporal attention information and spatial attention information to improve the action recognition method,and improve the recognition ability of the model for fine-grained abnormal behaviors such as looking around,bowing and plagiarism;On this basis,we have carried out a large number of comparative experiments to verify the effectiveness of our work.The results show that the recognition efficiency of76.8fps and the recognition accuracy of 98.5% are achieved in the GPU;Finally,we apply the above scheme to improve the basic framework,and build an intelligent invigilator application system,and then show part of the recognition results in the actual exam scene.The system uses the information exchange mode of the front and back servers of the message queue,and has the functions of marking,recording and alarming abnormal behaviors.