Research On Bus Passenger Statistics And Analysis Based On Deep Learning

With the implementation of smart city and smart transportation strategies,the speed of urban public transportation informatization is accelerating.Buses can easily obtain real-time location,road conditions,and other information to provide passengers with station information services,However,it is difficult to provide information on passenger congestion in the bus.there are various ways to obtain this information.Taking into account that the bus does not require a swipe card and the inside of bus is equipped with a surveillance camera,the most convenient method is to count the passengers by visual and analyze the congestion,the difficulty lies in the passenger flow statistics and analysis algorithm.Vision-based passenger flow statistics and analysis algorithms have been researched for a long time.Among them,traditional algorithms that manually design feature matching patterns have the disadvantages of difficulty in development,poor portability of the scene,and susceptibility to changes in the external lighting environment.They have been unable to meet the needs of changing scenes and stable monitoring.People need a kind of adaptable,Stable and develop simple algorithms.With the continuous breakthrough of deep learning in the field of computer vision,deep learning algorithms have the characteristics of low application difficulty,high accuracy,and strong robustness.Therefore,researchers gradually turn their attention to deep learning algorithms.This article focuses on the application of deep learning in bus passenger flow statistics and analysis scenarios,using deep learning to monitor and count the number of people getting on or off the bus in real time.The specific research content is as follows:1)A performance analysis and optimization method for deep convolutional neural networks for bus scenarios is proposed.The method locates model bottlenecks by visualizing deep learning model weights,intermediate layer feature maps,and runtime analysis,and proposes optimization solutions and experimental verifications for bottleneck modules.2)A real-time bus passenger flow statistics method based on deep convolutional neural networks is proposed.This method consists of a VGG16-SSD model with Bias bias term removed,a Kalman filter-based tracking algorithm,and a calibration line passenger flow statistical algorithm.Experiment test results on bus monitoring videos show that the accuracy of this method is far better than the passenger flow statistics system based on depth images and traditional image processing algorithms.Another experimental results of pedestrian counting tests at the entrance of the laboratory show that the real-time passenger flow statistics algorithm for buses based on deep convolutional neural networks also has a high accuracy rate in the laboratory scene,It is proved that the algorithm has learned a very robust feature expression mode,and has high value in popularization and application in other scenarios.3)A faster real-time passenger flow statistics algorithm for buses based on lightweight convolutional neural networks is proposed.The algorithm first increases the size of the input picture accepted by the lightweight network Mobile Netv3,and uses the improved Mobile Netv3 as the pre-convolutional neural network,lightweight transformation of multiple feature layers,to form a Mobile Netv3-Lite SSD model.Compared with the VGG16-SSD model,the model nearly doubles speed without a significant drop in accuracy,Using this model to replace the VGG16-SSD model in the previous algorithm,and adding an image enhancement algorithm to the passenger flow statistics system,to form a faster real-time passenger flow statistics system for bus.The analysis of the bus monitoring video found that the accuracy of the algorithm was slightly lower than the previous algorithm,but still far higher than the traditional image processing algorithm.