Research On Detection Algorithm Of Crowding Degree In Bus Carriage And Passenger Count In Platform

Improving bus travel comfort is one of the important tasks of bus dispatching.How to use deep learning technology to realize the intelligent dispatch of buses is a research hotspot in this field.Based on the convolutional neural network,this thesis focuses on the detection of crowding degree in bus carriage and the number of people waiting on the platform.The research work and results of this thesis are as follows:(1)Based on the on-board video monitoring system of the bus,we have produced a data set of classification of crowdedness in the bus compartment and a data set of head marking of waiting passengers on the bus platform respectively.(2)For the detection of the degree of congestion in the bus compartment,we propose a detection algorithm based on the spatial attention module of the adaptive receptive field.Through experiments on the mainstream classification networks Res Net,Google Net,and Dense Net,and it’s found that the Res Net network performs best.According to the problems of adjacent category classification errors and cross-category classification errors in the detection results,this thesis combines the characteristics of the degree of congestion in the carriage,and proposes a spatial attention module to enable the network to strengthen the distribution characteristics of passengers on both sides of the aisle,and at the same time suppress features that have nothing to do with the degree of congestion.In addition,an attention module of adaptive receptive field is built,so that the network can adjust the size of the receptive field to obtain more context information and reduce the interference of local features on the judgment of the degree of congestion in the compartment.Finally,the two modules are combined and embedded in the Rse Net network.The experimental results show that the proposed algorithm has a great improvement in the performance of the congestion degree detection.(3)For the counting the number of people waiting on a bus platform,we propose a algorithm for counting the number of people waiting on the platform based on adaptive spatial feature fusion.First,the head of passengers was detected based on the YOLOv3 network,and it’s found that there was a problem that it was difficult to detect when the platform was crowded with passengers.In response to this problem,this thesis proposes enhancing data to improve the relative resolution of the passenger head in the image;the feature fusion module in the YOLOv3 network is improved,and an adaptive feature fusion module is constructed to deal with the conflicts information that occur during the fusion of multi-scale features is filtered to improve the expression ability of shallow features in fusion features,thereby optimizing the network’s ability to detect small targets;the loss function is improved,and the modulating factor is used to solve the problem of low confidence of the bounding box in detection.Experimental results show that the proposed algorithm is effective and feasible for small-scale passenger head detection.We deploy the detection algorithm of passenger congestion in the bus compartment and the number of people waiting on the platform to the server,and adds an intelligent detection module to the existing bus dispatch system to provide data support for bus dispatch.