Rail Transit Crowd Counting Method And Its Embedded Implementation

In recent years,with the continuous development and improvement of urban rail transit,subway has become the first choice for more and more people.In order to better assist the fine management of passenger flow in subway operation and eliminate safety risks,the crowd density analysis system based on monitoring video can be used to estimate the crowd density in each area of the subway station in real time.The existing crowd count estimation methods are often inadaptable and difficult to be applied in the scene of rail transit.The research work of this paper focuses on the design of superior performance of the crowd density analysis algorithm based on convolutional neural network and the embedded design of intelligent camera with crowd counting function under the subway platform scene.The main tasks include the following:a.Considering the difference of crowd counting area and the diversity of camera shooting angles,the counting algorithm based on pedestrian detection and the existing openscene-crowd dataset can hardly meet the needs of crowd counting in subway platform scene.This paper makes an experimental analysis of the above problems,and builds a test and training dataset of rail transit scene based on crowd density map on the basis of collecting a large number of video data of typical actual scenes.b.Metro Net,a crowd counting model based on convolutional neural network,is proposed for the monitoring scene of low Angle and high occlusion in subway.The model uses the improved VGG16 network structure as the front-end network to extract the shallow feature map,then use the Inception network architecture combined with void convolution as the back-end network.The weighted loss function is used to train with the head loss and density map loss.In addition,the platform scene dataset and the open dataset are compared with several models with outstanding performance in the same method.The results show that Metro Net model has high accuracy.c.The Metro Net model was ported to the Hi3559CV100 embedded AI chip.The embedded software prototype design of smart camera supporting crowd counting is realized by using Hayes multimedia processing platform MPP and hayes NNIE deep neural network framework.The actual effect is tested and results show that the system can meet the accuracy requirement of practical application.At last,the thesis summarizes the whole thesis and looks forward to the problems worthy of further study in the future.