Research On Early Warning System Of City Appearance Violations And People Flow In Streets Based On Intelligent Video Analysis

City appearance violations and crowd safety are important issues for urban management in street management.With the continuous increase of the monitoring scale in urban streets,urban management departments relying solely on manual management are inefficient and there are hidden dangers of omissions.Based on the intelligent video analysis technology,this paper researches and implements the early warning system of street appearance violation and pedestrian flow.The main work includes the following three aspects:(1)A city appearance violations detection algorithm based on improved Gaussian Mixed background Modeling(GMM)is proposed.First,the post-processing feedback algorithm and the learning rate dynamic adjustment algorithm are used to improve the GMM so that it can extract the static violation target area,and at the same time can receive external feedback information to restore the background of the false alarm area in time.Then a fast false alarm suppression algorithm is proposed based on the object closure criterion,edge connectivity matching criterion and uniform illumination change criterion,which can effectively suppress false alarms caused by ghosts,light shadows,and so on.Finally,based on the twin neural network,an accurate false alarm suppression algorithm is designed,which can suppress false alarms that are difficult to filter by traditional algorithms.Compared with the combined algorithms of "background modeling + keypoint matching" and "background modeling + classifier",the proposed algorithm has better performance and can be run in real time on a mobile portable computer.(2)A crowd localization and counting network based on perspective aware is proposed.The network includes the perspective information prediction branch,crowd localization branch,and crowd density estimation branch.Utilizing the particularity of urban street scenes can obtain perspective information,thereby adding contextual information to the positioning of key points on the head,which can improve the performance of crowd positioning.In addition,integrating the crowd localization branch with the crowd density estimation branch to increase the accuracy of the estimation of the counting people.The experimental results show that compared with other algorithms,the proposed algorithm has a competitive advantage in both localization and counting tasks.(3)A video intelligent early warning system of city appearance violations and people flow is designed and implemented.The system includes three function modules: early warning of violations,over-traffic and visual analysis.In software implementation,distributed processing is used to transfer the main computing tasks to the cloud server.The client runs on a personal computer and is mainly responsible for light work such as information visualization and recording.