Research On Public Traffic Passenger Flow Detection System Based On Deep Learning

Generally,the passenger flow statistics of buses can provide the full passenger flow data of operating vehicles.The vehicle dispatching management personnel use this data for statistics,analysis and decision-making to achieve effective allocation of vehicles and personnel,thereby highlighting the advantages of intelligent dispatching.In recent years,with the speeding up of urban public transport infrastructure construction and vehicle equipment renovation and updating in various parts of China,public transport has been increasing in terms of vehicle ownership,operating line length,operating mileage,etc,which has led to a large amount of operation and management workload and incapable of real-time monitoring of public transport The condition of the vehicle makes operation inefficient and passenger satisfaction poor.Aiming at problems such as the current situation of unstable public passenger traffic data acquisition,outdated processing methods,and incapability of real-time dispatching,this paper conducts research on a public transit passenger flow detection system based on deep learning.The system combines image processing and deep learning methods,transplants a TensorFlow deep learning framework in Raspberry Pi,and builds a model of a convolutional neural network CNN to extract passenger head features in the passenger compartment.The learning rate is optimized by a comprehensive gradient descent algorithm.The three-channel data fusion technology is used to judge the congestion in the car.Compared with traditional algorithms,the system not only improves the recognition accuracy but also accelerates the convergence rate.It is especially suitable for real-time detection.Real-time congestion judgment can be more practical.The main work of the paper is summarized as follows:(1)The overall function analysis and scheme construction of the bus passenger flow detection system.First of all,it researches from the four aspects of the development process of deep learning,scene application,mainstream model analysis and learning tools,and selects a suitable deep learning network model.Then,based on the design principles and the realization of basic functions,the system requirements for the in-car image acquisition module and the in-car image recognition and discrimination module are constructed,and the overall system design is completed.(2)Algorithm analysis and design.In order to verify the effectiveness of deep learning models,the traditional passenger flow counting algorithm,detection-based passenger flow algorithm and regression-based passenger flow algorithm are studied first,and some feature extraction methods,classifiers,and regression image verification are performed,and then deep learning-based The convolutional neural network architecture is researched,and finally,a suitable algorithm is selected for passenger flow statistics in combination with objective scene factors.(3)Hardware analysis and design of bus passenger flow system.According to the actual application scenario and functional requirements of the system,the main controller module,the anti-vibration in-car camera,the video decoding chip and the communication display module were selected and designed,and the in-car image acquisition module was completed.Secondly,the communication display module is designed for 4G module,GPS positioning module and power module,and completes the interaction with the PC.(4)Software analysis and design of the bus passenger flow system.By transplanting the Raspbian operating system OpenCV library transplantation and TensorFlow framework transplantation in Raspberry,the software design of the car image acquisition module,the car image recognition and discrimination module,and the communication display module are completed.Among them,the in-car image acquisition module collects through the camera and transmits the image through the serial port.The transmitted image is the input of the in-car image recognition and discrimination module.After the image preprocessing is completed,it is input into the built convolutional neural network model,and a comprehensive The gradient descent algorithm replaces the traditional gradient descent algorithm,which solves the problem that the computing power is too weak and overfitting and local loss function are easy to occur.The experiment proves that the recognition accuracy is 87.23%,and the convergence rate is increased by 20.92%.This paper researches and designs a public transit passenger flow detection system based on deep learning that can realize the functions of passenger collection,identification and communication display in the car.This system makes up for some of the shortcomings of the current bus passenger flow system in vehicle-mounted embedded detection.At the same time,the paper mainly proposes a convolutional neural network model of a comprehensive gradient descent algorithm.Experiments show that the algorithm not only improves the recognition accuracy but also accelerates the convergence rate,which is particularly suitable for real-time detection.