Research On Indoor Vacancy Detection Algorithm Based On Surveillance Videos

With the improvement of living standards,people's awareness of security is also greatly increasing.For some special places,such as the fire control room,duty room,and frontier guard post,if there is vacancy,it is likely to bring huge losses to people's lives and property.Therefore,vacancy detection is of great significance.Most of the early vacancy detection methods were carried out in a human supervision way.Afterwards,with the installation of a large number of cameras in public places,many organizations began to allocate special personnel to view the surveillance videos.All these traditional detection methods have high labor costs.Besides,long-term attention to the display screen may easily make people ignore some critical video clips,unable to gain optimal results.With the continuous development of computer vision technology,the vacancy detection algorithm based on intelligent video surveillance has been proposed.It takes the effective detection of indoor humans as the most critical step,then analyzes the status changes and finally determines whether there is vacancy,thus eliminating potential safety hazards.Most of the existing human detection methods are aimed at the distant upright pedestrians outside,but it is different for the nearby humans in a surveillance video.Specifically,due to the fixed location of the camera,humans are often in a multi-view state with various deformations and occlusions,which brings difficulties and challenges to the detection task.In view of the above situation,this thesis proposes a neural network model of indoor human detection algorithm based on surveillance videos,consisting of the image preprocessing part and the principal detection network part.For the former,it refers to the region proposal network combined with the multi-view model,which draws on the design of the RPN model in the two-stage object detection network Faster R-CNN.And it can obtain a higher recall rate and a higher detection rate for considering multiple views,including the frontal,back and side views.For the latter,it integrates sub-modules,including the feature extraction,deformation processing,visibility reasoning and classification module.Among them,the deformation processing and visibility reasoning module are based on the design idea of the deformable part model algorithm.By fully considering the relationship among the components as well as different modules of the model,the detection performance has been improved under the conditions of multiple deformations and occlusions.What's more,given the temporal feature of human behavior in videos,the vacancy judgement criteria have been put forward,conforming to actual needs.By collecting a large number of related surveillance videos in several typical indoor scenes,this thesis establishes an indoor vacancy detection dataset,containing samples of multiple views,deformations and occlusions.Based on this established dataset,this thesis has conducted the testing experiment of this proposed algorithm model in different scenes,and the comparative experiment with other classical human detection algorithms,which fully verifies the effectiveness and robustness of the proposed algorithm model.In general,the proposed algorithm can provide reference for related research fields.