Design And Implementation Of Smart City Traffic Statistics System Based On Machine Vision Technology

With the continuous development of machine vision technology,the coverage of surveillance cameras is becoming more and more extensive.The real-time statistics and tracking of the flow density of public places in smart cities have also been widely studied and applied.It is also possible to count the number of people in public areas with high population density such as scenic spots and parks,and accurately grasp the number of tourists in the current area,which is beneficial to avoid various adverse events such as stamping and stealing,and can also reasonably reduce the number of tourists waiting in line.Waiting time and pressure on the use of public facilities.The statistics of smart city traffic based on machine vision can better serve the masses,reduce security risks and increase management efficiency.In this paper,the design of the number system based on machine vision is divided into three parts: the detection and extraction of moving pedestrian targets based on hybrid Gaussian background modeling,the detection of head targets based on deep learning and multi-target tracking based on improved Kalman filtering.Through the detection and tracking of pedestrian head targets,the number of pedestrians entering and leaving is counted in real time.Firstly,the hybrid Gaussian background modeling method is used to model the monitoring environment background,and all the static targets and regular motion non-pedestrian targets in the monitoring environment are analyzed and unified modeling.After that,three types are compared.The advantages and disadvantages of the detection method of moving targets are selected,and the background difference method is selected to detect all moving target areas containing pedestrians.Finally,the morphological image filtering algorithm is used to further process the detected motion areas,and the motion areas are removed and reduced.The extracted moving target area reduces the workload of the head target detection and reduces the false detection rate of the pseudo pedestrian target.Then,based on deep learning feature extraction,pedestrian head detection is implemented.In order to better detect the head of the moving pedestrian,the surveillance camera must be placed in the middle.In this paper,the human head detection data set is made by real video,and the method of data annotation is combined with the pre-training model.This paper also tests and compares the target detection algorithms based on deep learning,Faster-RCNN,RFCN and SSD algorithms.Finally,the ResNet101+RFCN network structure is selected as the human head detection algorithm.Finally,a method combining Kalman filter and Mean Shift algorithm is proposed to perform motion tracking and counting on pedestrians in the monitoring scene.Using the result of Kalman filtering as the initial input value of the Mean Shift algorithm,the moving target is tracked,which effectively reduces the tracking target loss caused by short-term occlusion or interference,and improves the tracking efficiency.In the counting method,a virtual wireframe is set in the monitoring environment,and when the pedestrian passes the virtual wireframe,the counting is performed,and the direction is determined according to the order of the virtual wireframe edges.