Research On Key Technology Of Perception And Recognition For Intelligent Security

Intelligent security system refers to a comprehensive system that provides intrusion alarm services for important places through a variety of security devices.It mainly consists of three subsystems:access control subsystem,surveillance subsystem and intrusion alarm subsystem.Intelligent sensing and recognition technology bears the responsibility of information acquisition and intelligent analysis in intelligent security application,and provides technical support for various intelligent applications such as video surveillance.The identification system based on fingerprint identification integrates convenience,security and interaction,and is most commonly used in the access control subsystem.At the same time,the security requirements of the access control subsystem pose great challenges to the recognition accuracy and false acceptance rate of fingerprint identification.In the surveillance subsystem,monitoring based on single camera always has a blind spot;monitoring of multiple cameras has the disadvantage that the scene is disorderly distributed and is difficult to observe comprehensively.The current static camera video stitching algorithm needs to be improved in terms of algorithm speed and stitching effect.In the intrusion alarm subsystem,various alarm detectors need to be equipped with intelligent detection and identification technology to achieve automatic and timely alarm functions.The detection algorithm not only needs to have high-precision detection effect in the common scene,but also needs to adapt to uncertain factors such as weather,illumination and occlusion existing in the surveillance scene.In view of the problems and challenges in the security scenarios,this topic is aimed at intelligent security applications,focusing on the key technologies of visual perception and recognition.The main research contents and innovations include:1.Fingerprint identification technology based on fingerprint orientation field regularization is studied to improve the security of the access control subsystem.The initial fingerprint orientation field obtained by the gradient-based algorithm usually contains a large amount of noise.The cracks,stains and blurs are summarized as structural noise,block noise,gaussian noise and mixed noise.In view of data shortage,this paper expands the training data by adding noise to meet the training requirements of neural network,and then constructs the network model and loss function by regression.Finally,the accuracy of fingerprint identification is used to verify the effectiveness of our algorithm.Comparison experiments are carried out on FVC2002,FVC2004 and FVC2006 data sets.High accuracy is achieved by our algorithm.Fingerprint identification technology with high precision and low false acceptance rate is the basic requirement of the security of smart lock.The fingerprint orientation field regularization technology proposed in this paper can effectively improve the accuracy on multiple indicators and make the entrance guard more reliable.2.A fast and natural video stitching algorithm is proposed,including an image stitching algorithm based on structure preservation and a seam searching algorithm based on 3d graph cuts,which can provide large-angle security monitor for monitoring subsystem.Perspective transformation is easy to produce perspective distortion and the grid deformation method is easy to cause structure distortion.Thus,this paper puts forward an image alignment algorithm with feature point alignment term,local similarity item,global similarity item and line collinearity term as energy terms of optimization based on similarity transformation.In the subsequent stitching,the generated mask and a seam searching algorithm based on 3d graph cuts are used to obtain a smooth and ghost-free video.The data set is constructed by means of network collection and autonomous collection for comparative experiments.Great results are obtained in both image experiments and user experiments.The use of fast and natural video stitching algorithm to combine the scene content captured by multiple cameras can effectively solve the situation that the monitoring scene is disorderly and difficult to comprehensively observe,providing a more convenient user experience for the customers.3.This paper presents a top-down human pose estimation network based on atrous convolution,which can provide technical support for intrusion alarm system.Firstly,atrous convolution is used to replace the downsampling convolution module in the backbone,which can expand the receptive field while keeping the resolution of the feature map.Secondly,multi-scale features are extracted through the atrous spatial pyramid pooling module to enrich the scale information.Finally,output heatmaps are constructed by replacing the upsampling and convolution operation with deconvolution.In this paper,experiments are carried out on COCO 2017,SPID and SHPD data sets.The experimental results show that the accuracy of COCO test-dev data sets with different input sizes is 1.1%and 0.5%higher than the existing algorithms.Moreover,the proposed algorithm has good robustness in the monitoring scene,and can adapt to various angles of view,weather,illumination and occlusion.Efficient human pose estimation technology combined with large scene monitoring technology can identify intruders and issue alarms in time to effectively protect important sites.