Research On Bus Passenger Counting System Based On Binocular Camera

As the basic data of the bus system,the accuracy of bus passenger flow counting and analysis has an important impact on the efficiency of the system and the effective decision of the management department.Therefore,the research of bus passenger counting system has important application value and economic benefits.Most of the bus passenger count methods are based on monocular cameras or infrared sensors.The influence of various disturbances in the bus scene makes the detection accuracy of the existing detection method unsatisfactory.Binocular camera can get three-dimensional information of the scene,meanwhile not affected by lighting.Besides,its installation and maintenance costs are low and it can effectively reduce the influence of factors such as occlusion,adhesion and pseudo-target on the detection accuracy.Therefore,it is of great significance to research the bus passenger counting system based on binocular camera.In this paper,an improved Census stereo matching algorithm is proposed for the problems existing in Census stereo matching algorithm,which can get a more complete and dense depth map compared with the original algorithm.At the same time,two depth map filling algorithms are designed to fill the invalid matching points.Finally,the multi-feature fusion filling algorithm is selected to effectively fill the invalid matching points of the depth map in the low texture and discontinuous region.Based on the post-fill depth map,this paper establishes the world coordinate system and use the vanishing point calibration to convert the depth map to the top view of the world coordinate system.A robust algorithm for head locking is designed by searching the local depth maximum algorithm for the initial locking of the head target,extracting the characteristics of head target and use the extracted features to remove part of the pseudo target.On the basis of the extracted target characteristics,the feature sample library is established.In this paper,the SVM classifier is constructed by training and testing for the sample to select the appropriate kernel function and the corresponding parameters,and the decision function is used to identify the head target.Then the Kalman filter is used to predict the coarse positioning which serves as the center to achieve the target tracking in the specified smaler matching window using the block matching method for full search.Finally the number of passengers is counted through the on-off counting test line.A high-precision and efficient passenger number statistics algorithm is designed and a number of passengers statistics system for the intelligent bus is set up.In the real bus scene,test results show that the algorithm can achieve more than 90% of the counting accuracy.Its overall performance is better than the monocular statistics and can thus meet the practical application requirements.