Research On Methods For Video Surveillance Based Indoor Occupant Counting

Using of existing surveillance cameras in the building for people zoning counting has the advantages of no additional sensor network,rich data,high estimation accuracy,etc.This method has become an important field of computer vision research,and the acquisition of people distribution within the building controls the building intelligence.It is also of great significance.The research on indoor occupant for building energy conservation has become a research hotspot in academia and industry.Therefore,this thesis has carried out some research on the estimation of people distribution in buildings for different application scenarios,and proposed two indoor occupant counting methods based on video surveillance and a staff counting correction algorithm based on binomial distribution.The content and results of this thesis mainly have the following three aspects.Firstly,based on the characteristics of the indoor environment of the building,a scene counting method based on image analysis is proposed for the scene where the indoor effective area image can be acquired.The single-frame image often cannot accurately detect the head region of the image,and may be occluded.,false detection,missed detection,etc.,proposed a population correction method based on cluster analysis.The method first evaluates the indoor environment characteristics,extracts the ROI(region of interest)region,and then uses the cascaded Adaboost(adaptive boosting)classifier and the convolutional neural network and the SVM(support vector machine)classifier to The head target is tested to achieve staff count.Aiming at the possible missed detection and false detection in single-frame image detection,a method based on DBSCAN algorithm is proposed.The experimental results show that the indoor occupant counting has good robustness,accuracy and real-time.Secondly,considering the occlusion,incomplete coverage and the lack of images in the area where some indoor people are located based on the finite image analysis method,an indoor number counting method based on the statistics of the entry and exit of the regional boundary personnel is proposed.The method adopts an effective detection area at the boundary of the region,adopts a target detection based on the Gaussian mixture model and the background difference method for the movement of the person,uses the Kalman filter algorithm to track the people,and judges the direction of the person entering and exiting through the change of the center of gravity of the person,and finally realizes indoor occupant counting.In this thesis,the algorithm is verified in the actual monitoring scenario,and the accuracy rate can reach more than 90%.Finally,considering the cumulative error problem of the people counting method based on the boundary in and out over time,a staff counting correction algorithm based on the error model is proposed.In this thesis,the causes and effects of cumulative error based on the entry and exit statistics of boundary people are analyzed in detail.The characteristics of people movement in the boundary region are analyzed.A cumulative error model based on binomial distribution is proposed.The error model is used to judge the sensor statistics.When the counting people should be corrected.Finally,based on the cumulative error model,a correction method based on the error model is proposed.The error model is used to judge the correction time,and the cumulative error of the statistical population is corrected.The experimental results show that compared with the simple method of calculating the number of people entering and leaving by statisticians,after using the counting correction method of this thesis,the performance of people counting has been greatly improved,and the effect of correcting the cumulative error is obvious.