Research And Application Of People Counting Method Under Video Surveillance Scenes

In recent years,with the rapid development of computer science and technology,especially,with great achievements in the field of computer vision based on the deep learning technology,the intelligent monitoring system based on computer vision technology has been widely used in real life.The people counting system is one of the most important applications of intelligent video surveillance system.Therefore,the research about the people counting methods in video surveillance scenes not only has an important practical significance,but also has far-reaching significance.This paper uses the human object detection method to count people,based on the Region Fully Convolutional Networks model.Through the analysis of the actual video surveillance scene,this paper designs a human object detection method based on region fully convolution network model,and to count the people in the video surveillance scene.The specific content of the method includes the following three aspects.Firstly,this paper uses the Human Omega head and shoulder model as the representation model of human body under video surveillance scene.Compared to the widely used human body model,Human Omega head and shoulder model has more stability,and the model also greatly reduce the scene of occlusion of the human object.Secondly,aiming at the problem that some human target samples are difficult to detect in the complex monitoring scene,this paper applied a kind of hard example mining algorithm to the region fully convolution network model.Through modifying the update strategy of network model,this hard example mining method not only improves the ability of detect human targets in complex scenes,but also reduces the parameters of training model.Thirdly,aiming at the problem of multi-scale detection of human body in video surveillance scene,from the perspective of network model that generates candidate detection window,this paper analysis the scale distribution and the aspect ratio distribution of the human calibration in several data set.This paper designs an improved Region Proposal Network Model,by modifying the generation rules of Anchors in the Region Proposal Network Model,to improve the localization accuracy of candidate detection box.This paper uses this method to compare with other target detection methods based on deep learning technology.The experimental results show that the designed method achieves better results in video surveillance scene.At the same time,this paper also based on the method to count the person under the video surveillance scenes.Through the comparison of the experiment,we verify that our method has a higher statistical accuracy in the monitoring scene.Finally,the method applied to the practical intelligent monitoring system,and obtained the satisfactory results.