Research On Statistics Method Of Bus Passenger Flow Based On Machine Learning

Pedestrian Counting in the bus scene can provide important data for public traffic management department,which can be used to optimize urban traffic routes and vehicle scheduling.How to get the accurate count of the passengor flow of bus scene has important research value.In this thesis,the overall framework of bus pedestrian flow counting system is divided into three main parts: pedestrian target detection,tracking,and classification trajectories.There are still many problems in using traditional image processing methods to detect passenger targets in two-dimensional images such as,illumination change,background shaking,and so on.As the basic part of the whole system,the accuracy of the pedestrian target detection will have a direct impact on the following tracking and the precision of statistics.Considering the above problems,two different detection algorithms are used in the passenger target detection.The first detection algorithm makes full use of depth camera which can provide depth stream.Then we use the translate the depth image into the world coordinates by the extrinsic parameters of the camera.Based on the the head of the single pedestrian is a local maximum area which can be located by the proposed local height maxima algorithm combined with SVM classifier for classification the head of the target.Using depth information to build 3D information of the bus scene not only effectively can reduce the influence of illumination,but also avoids the problem of mutual occlusion among passengers.It creates favorable conditions for passenger target detection.The second algorithm applies the SSD network model in deep learning to the head detection of passengers.The SSD network model can learn autonomously from the sample data and express the characteristics of the essential attributes of the target so that it can detect passenger targets more accurately.In this paper,a large number of videos recorded in bus scene to build the sample library of pedestrian's head,and the human head target models with different sample sizes are trained respectively,and the experimental results are compared.On the issue of passenger's head tracking,we use the combination method of frame matching and Kalman filtering to match and track the target and get the trajectory.Firstly,Based on the detected pedestrian's heads,frame matching method are used for the headassociation.Besides,the preliminary matching method(i.e.Kalman Filter)also is used when the target is blocked due to the missing which searches the neighborhood of predicted position within the range of continuous tracking target.Finally,according to the characteristics of passenger target trajectory and pseudo-target trajectory,we select the best feature combination to classify and count the trajectories,so as to achieve the statistics of bus numbers.Experiments have been done for the above methods,whose results show that these two methods have high statistical accuracy.The accuracy of the traditional detection algorithm is95.16%.Meanwhile,the accuracy of the method based on deep learning SSD target detection algorithm is 97.72%.The proposed methods can satisfy the needs for pedestrian counting in real bus scene.