Research On Bus Passenger Crowding Detection Method Based On Deep Learnin

In recent years,with the popularization of the concept of "green transportation,smart travel" and the proposal of smart city construction,various cities have invested a lot of resources in promoting the construction of new energy intelligent public transportation systems.New energy buses not only require a complete public transportation system,and needs intelligent supporting facilities,but the current bus operation scheduling and service level still need to be improved,and problems such as passenger congestion occur from time to time.As a result,it has become an urgent problem to efficiently and in real-time monitor passenger congestion on buses,as well as to dynamically dispatch each bus route and vehicle reasonably and accurately while meeting the needs of the majority of passengers.With the continuous development of computer vision and the improvement of deep learning-related algorithms,current new energy smart buses have been equipped with fatigue driving monitoring and dangerous driving monitoring,but there has been a lack of effective methods and techniques for passenger flow detection and in-vehicle passenger congestion analysis.The main goal of this thesis is to propose research of bus passenger crowding degree method based on deep learning,in which the algorithm can identify the crowding degree of bus passengers by capturing pictures inside the bus in real-time using the on-board camera,and provide the crowding degree of bus passengers while monitoring the number of passengers getting on and off the bus.The following are the main research content and innovation points:(1)The bus crowding degree is redefined and an image sample dataset is built.According to the six levels of crowding sub-categorization,the criteria of crowding definition are proposed,i.e.,comfortable,normal,crowded,and very crowded.Due to the lack of a dataset about the number of passengers on the bus,the sample dataset of bus images is established by collecting video surveillance inside the bus and collecting web galleries.(2)Propose a bus passenger congestion detection algorithm based on target detection.For regions with more obvious passenger characteristics,a deep learning-based target detection method is used to calculate the number of passengers in the video image.Since the YOLO series uses end-to-end neural networks with real-time detection speed and YOLOv5 has the best trade-off performance,the YOLOv5 s network model with high accuracy and high detection rate is chosen.The algorithm annotates the passengers in the training set of bus image samples and uses the YOLOv5s pre-training model to train the dataset.The accuracy of this algorithm for bus passenger crowding detection is verified up to 88.3% after test set experiments.(3)A crowd density estimation algorithm for crowded area scenes on buses is proposed.The model filters all objects in the original image that do not have passenger features and outputs the information features of passengers in the image.Since the bus has a certain upper limit on the number of approved passengers,there is no need to make a number prediction for excessive passenger density,so a double branch CNN(DB-CNN)is proposed to analyze the bus passenger crowding.The bus ridership estimation is performed by using a two-branch CNN to extract features and transpose convolutional layers to detect the loss of details in the density map performed by stacked pooling for regions where passenger head features are not obvious.In this paper,we combine two methods,target detection and density estimation,to evaluate the number of passengers in a bus,and predict the crowding in a bus by defining and grading the concept of crowding,and the accuracy of predicting the crowding in a bus can reach 93.2% according to the dataset used in the experiments.The experimental results show that this method solves the problem of difficult system deployment in the bus environment,improves the characteristics of traditional classification methods such as poor robustness and low accuracy,and is extremely useful for bus passenger crowding analysis.