Research On Human Abnormal Behavior Recognition Method In Vault Based On Multi-source Cameras

Video surveillance is an important part of Skynet system,and also one of the important driving forces to improve the level of public security protection in recent years.With the popularity of video surveillance devices increasing year by year,the demand for intelligent monitor system is also increasing.At present,most of the domestic video surveillance systems need to be on duty manually,and arrange 24-hour shift of monitoring personnel to achieve supervision.Usually,each supervisor needs to monitor multiple displays at the same time,which has problems such as low efficiency,poor anti-interference ability and easy to be affected by the subjective consciousness of supervisors.With the deepening research of image processing,video analysis,behavior recognition and other technologies,intelligent video surveillance system has been widely used in specific occasions,such as roads,classrooms,access control and other areas,because of its high degree of intelligence,strong robustness and high efficiency.However,in the face of the bank vault and the special abnormal behavior in the vault,the general intelligent monitoring system can not achieve the desired effect.Therefore,it is of great value to study the method of abnormal bank behavior recognition based on human bank camera.In this paper,taking the bank vault as the research scene,taking mobile phone calls and illegal unpacking behavior as the recognition target,a camera scheduling algorithm based on topology analysis is designed for the huge number of camera groups in the vault.According to the characteristics of the vault environment,a moving object segmentation method based on improved Gaussian mixture model and a reflection detection operator are designed In this paper,we design a HOG-LBP feature extraction operator based on region partition and dimension reduction,and use SVM to classify.On this basis,we complete the design and experimental analysis of ui software.The main contents are as follows:(1)The overall scheme design of bank vault abnormal behavior recognition system.According to the scene distribution of bank vault and personnel working characteristics,the hardware selection of camera,recorder,server and design of hardware system are completed.Based on the requirements of abnormal behavior recognition,a multi-threaded software structure of Multilevel cameras is proposed,which can dynamically allocate resources according to the position of moving objects in the vault,and reduce the occupation of software resources on the premise of ensuring real-time performance and recognition accuracy.(2)Camera scheduling algorithm based on topology analysis.Aiming at the requirements of multi-source camera configuration and abnormal behavior recognition in bank vault,a camera scheduling algorithm based on topology analysis is designed,which can realize the opening and closing scheduling of camera group according to the pose state of moving target.On the basis of ensuring the recognition rate,the dynamic configuration of camera and computer resources is optimized.The algorithm can temporarily dispatch PTZ cameras to assist video surveillance according to the area of the target when facing the possible personnel occlusion in the bank vault.Aiming at the situation that the moving target appears in the field of vision of multiple scheduling cameras at the same time,a result voting mechanism based on distance weight is proposed.The mechanism dynamically allocates the weight according to the position of personnel in the field of vision of scheduling cameras,so as to avoid misjudgment caused by personnel too far away from cameras.(3)Moving object segmentation method based on Improved Gaussian mixture model.Aiming at the ghost and reflection problems in the bank vault,an improved Gaussian mixture model algorithm is proposed.In this algorithm,the concept of pixel speed is introduced to dynamically adjust the model learning rate of different pixels,so as to effectively solve the ghost problem in foreground extraction.At the same time,using the speed and historical information of pixels,the algorithm can effectively warn and eliminate the illumination mutation in the bank vault,and avoid interference to the target segmentation.Aiming at the reflection in the bank vault,this paper proposes a feature vector fusion of target location and weighted HSV color model,and then uses cosine similarity for correlation analysis,so as to achieve accurate target segmentation.(4)Multi dimensional feature fusion and SVM behavior recognition algorithm.Firstly,according to the camera layout in the bank vault and the dressing characteristics of the staff,a head and hand ROI screening method based on HSV and location parameters is proposed,and then the ROI of the unpacking tool is located according to the location relationship between the hand ROI and the connected domain of the unpacking tool.Then,a combined feature fusion of LBP feature and hog feature is proposed,and a feature dimension reduction method based on region partition is proposed.This feature can effectively solve the shortcomings of the classical hog features,such as ignoring diagonal gradient information,large dimension of combined features,redundant information and long extraction time.Finally,according to the classification requirements of bank vault abnormal behavior,the OVO type multi classification SVM model is used to realize the classification of abnormal behavior.Above all,according to the technical requirements of bank vault human abnormal behavior recognition,this paper studies the camera scheduling algorithm and video analysis algorithm,and designs a bank vault human abnormal behavior recognition system based on multi-source cameras.The system uses Hikvision tube camera and pan tilt camera as the video acquisition terminal,Hikvision hard disk recorder and Dell high performance server as the storage unit and platform main body to build the hardware platform;this paper is based on Visual Basic Studio software platform designed the system software,mainly including user login module,automatic operation module,manual test module,data management module,and completed the scheduling and management of each module.Through the actual scene test,it is proved that:the bank vault human abnormal behavior recognition algorithm based on multi-source camera can effectively identify abnormal behavior,the miss detection rate of two kinds of abnormal behavior is less than 0.5%,and the over detection rate is less than 5%,which achieves the performance index of abnormal behavior recognition,and the system software runs stably and meets the production requirements.