Design And Implementation Of Train Passenger Counting Software Based On TOF

With the appearance of face-swiping check-in service and autonomous-driving subway,the intelligent level of rail transit has become higher and higher.Passenger flow counting system based on machine vision is expected to become an intelligent infrastructure of rail transit,which can improve the ability of train advertising revenue,reduce the labor cost and increase the safety of passengers,by counting the number of people who board and leave the train.In order to solve the problem that accuracy and real time of the pedestrian(passenger)detection and tracking technology are not high enough in the passenger flow counting system,and to achieve a stable train passenger flow counting software system,this thesis carries out the following research work.To solve the problem that RGB images are easily affected by ambient light and most complex backgrounds are not conducive to pedestrian detection,this thesis uses a TOF camera to collect depth images for pedestrian flow counting.Compared with RGB images,depth images are not affected by ambient light and can protect the privacy of pedestrians.Removing backgrounds of depth images avoids the interference of complex backgrounds on pedestrian detection.To solve the problem that deep learning based object detection algorithms cannot run in real time on the CPU platform,this thesis constructs a fast pedestrian detection algorithm which have a low computation and storage cost by lightening the network structure and parameters.The experiments on the TVhead dataset show that the proposed detection algorithm can run in real-time speed,reach 45 FPS without GPU acceleration,and reach 95% on AP(an accuracy index),which meets the requirements of the scene.To solve the problem that the pedestrian congestion leads to the decrease of tracking accuracy,this thesis constructs a pedestrian tracking algorithm based on online learning.The tracking algorithm can learn the appearance of pedestrians in the tracking process,which increases the tracking accuracy in the pedestrian congestion scene.Due to reusing the features extracted by the proposed detection algorithm,the tracking algorithm only contains three convolutional layers with a low computation cost.The experiments on the TVhead dataset show that the proposed pedestrian tracking algorithm has only 1 time on IDs(an error rate index),and can effectively deal with the multi-person scene.Finally,this thesis uses the proposed pedestrian detection and tracking algorithm to construct the two-stage pedestrian flow counting algorithm,and realizes the complete train passenger flow counting software.The software takes the pedestrian flow counting algorithm as the core,and adds the functions of result uploading,data backup,security self-check.The experiments in the simulated train-door scene show that the software is accurate,real-time and stable.