Research On Some Key Problems Of Bus Carriage Crowding Degree

Nowadays,Bus travel has been one of the main means of travel for urban residents.To solve the problem of crowding on bus and utilize social bus resources more efficiently and reasonably,it is necessary to dispatch buses intelligently according to the real-time crowding degree of every bus.Image classification and target detection technology based on deep learning provide a new method for real-time perception of bus crowding.Our research conducts from four aspects: bus crowding image classification,passenger head target detection,deep exploration of convolutional neural networks and application of algorithms.Firstly,image classification is studied in this thesis.We make a bus crowding degree classification dataset and a bus crowding degree classification fuzzy dataset by utilizing the bus carriage picture taken by the Chongqing bus video surveillance system.The existing mainstream classification algorithms based on deep learning such as LeNet network,GoogLeNet network and ResNet network are implemented in this thesis and their classification performance on the dataset of bus crowding degree are compared.A classification model based on fuzzy classification datasets is proposed in this thesis,the model can make full use of inaccurately labeled datasets to prevent mislabeled data from competing with each other in training.Furthermore,a continuous classification model is proposed to solve the continuous classification problem.The model innovatively outputs category classes and category offsets,which provides a more fine-grained classification method.Secondly,target detection is studied in this thesis.We use the existing target detection algorithm Faster R-CNN and the common human head detection dataset SCUT_HEAD to train a universal head detection model,and use the model and image data of bus carriage to make a passenger head detection dataset.An improved target detection model is proposed in this thesis by introducing the LSTM mechanism,which can effectively deal with the recognition problem of small targets which are close to each other.Based on the results of bus passenger target detection,a classification model based on the target position is proposed,which takes targets' bounding boxes in an image as input,and outputs the crowding degree of the image.Thirdly,in the process of research,training process of deep convolutional neural networks are deeply analyzed.In this thesis,the problem of homogenization ofconvolution kernels is proposed.The similarity matrix is introduced to try to quantify this property.The change process of the weights of neural network during the training is tracked,and the curve of the high-dimensional space is used to describe the change process and the phenomenon that the direction of the convolution kernel is abruptly changed is found in this thesis.Finally,An application system for the bus carriage crowding degree sensing algorithm is implemented and deployed to cloud to provide external services.We also tried to combine the algorithm with the hardware and try to run the algorithm on terminal device.