Research On Application Of Deep Convolutional Neural Network In Passenger Flow State Recognition In Buses

With the increasing in the number of private cars of urban residents,traffic congestion problems plague most cities.Vigorously developing public transportation to share the travel rate of urban residents is an important measure to relieve urban congestion.The state of passenger flow in the bus is the basic decision basis for timely adjustment of bus operation strategies,optimization of route planning,and whether passengers choose to take this bus.This paper uses image processing technology,convolutional neural network classification algorithm and target detection algorithm to achieve automatic detection of bus passenger flow status.It has important application value for improving the satisfaction,comfort and attractiveness of bus travel,improving the bus service level and alleviating urban traffic congestion.Firstly,a four-class bus passenger flow status classification dataset is established with a total of 12,545 pictures.Four types of convolutional neural networks are trained and tested.The batch normalization layer was added to the Vgg Net-16 model which performed well,and the fully connected layer was replaced with a 1×1 convolutional layer to improve the model accuracy to 94.3%.The Swish activation function is used to replace the original Re LU activation function for the Google Inception Net V2 model,and the first layer convolution step is adjusted to increase the accuracy of the model to 94.8%.The test accuracy of the improved two models is 0.6% higher than that of the original model.Secondly,the detection dataset of bus passenger flow state is constructed,and the model is trained by the transfer learning method,clustering the passenger head annotation frames in the dataset improves the size of the prior frames of the model and normalizes the part of the YOLOv3 loss function that calculates the predicted frame width and height,so that the model’s detection accuracy of the passenger head target is 91.4%,the test accuracy of bus passenger flow status recognition is 95.7%,and the detection time of single sheet is 25.7ms.In order to improve the detection speed of the model,improved YOLOv3-tiny network and the light-weight network Squeeze Net replacing the YOLOv3 feature extraction network are used for experimentation,so that the two models can reduce the single sheet detection speed to 8.6ms and 16.2ms under a certain detection accuracy,which allows the computing terminalto process more data at the same time.In order to solve the problem of missing detection in the original YOLOv3 for small passenger head targets and passenger head occlusion adhesion targets,a small receptive field scale fusion detection method is added to improve the model’s small target detection effect,and the detection accuracy of the model on the passenger head target is improved to 92.7%,and test accuracy of bus passenger flow status recognition is improved to 96.1%.By integrating Dense Net’s design ideas into the Darknet-53 network to improve the model’s feature extraction capability,the model’s detection accuracy of the passenger head target is increased to 93.1%,and the accuracy of the bus passenger flow status recognition is improved to 96.5%.Compared with the original model of YOLOv3,the above two algorithms have improved the detection accuracy by 1.3% and 1.7% respectively,and the test accuracy of the bus passenger flow status recognition has been improved by 0.4% and0.8%.Finally,from the two aspects of model test accuracy and performance,the two methods used in this paper to solve the problem of bus passenger flow state recognition are compared and analyzed.