Research On People Counting Algorithm Based On Binocular Stereo Vision In Crowd Scene

People counting algorithm based on binocular stereo vision has the advantages of high efficiency,accuracy and low cost,and is widely used in shopping malls,scenic spots,subway and other crowded places.However,under the condition of dense crowds,due to the effect of pedestrian occlusion,uneven illumination and small head size,the algorithm of binocular people counting is difficult to meet the accuracy requirements.In order to improve the accuracy of the binocular people counting algorithm in the crowded scene.The thesis introduced the following work.(1)A two-stream SSD pedestrian detection algorithm based on feature fusion was proposed to improve the precision of pedestrian detection under illumination and occlusion conditions.Firstly,in order to improve the impact of pedestrian occlusion on the algorithm,the head information of pedestrians was obtained by taking vertically downward.Then,a semi-global stereo matching algorithm based on TADG-FCensus was used to get the disparity image,and a high-quality depth image was obtained according to the calculation relationship between disparity and depth.Secondly,in the two-stream SSD network structure,the Conv4＿3 and FC7 layers in the color image channel were combined with the Conv4＿3 and FC7 layers in the depth image channel in different fusion methods to obtain the corresponding two-stream SSD network model,and selected the optimal network model from it.Finally,the network of Conv4＿3＿Fuse layer and Conv10＿2＿Fuse and Conv11＿2＿Fusion layer were fused to fully learn the pedestrian head features.Experiments on the self-made dataset show that the improved algorithm has a detection accuracy of96.49%,which is increased by 11.97% compared with SSD algorithm.The problem of missing detection caused by illumination and occlusion is solved,and the recognition ability of pedestrian head is enhanced.Experiments on TVHeads dataset show that the detection accuracy is increased by 13.16% compared with SSD algorithm,which makes the algorithm can be applied to multiple scenes,and proves the effectiveness and applicability of the improved algorithm.(2)A pedestrian tracking algorithm based on improved Deep SORT was proposed to overcome the mis-tracking problem caused by similar pedestrian head information in the pedestrian tracking process.By adding spatial distance information to the correlation function,the pedestrian movement area was predicted,and the number of pedestrian head ID exchanges was reduced.Secondly,512-dimension feature vectors of the improved SSD network were extracted as appearance similar features to avoid secondary calculation of the feature extraction network.Finally,the double-line counting algorithm was used to statistics the pedestrians’ entry and exit,which reduced the situation of false counts caused by pedestrians wandering in the video.Experiments on the two video sequences show that the improved tracking algorithm has a tracking precision of 86.2%,which is increased by 4.6%compared with Deep SORT algorithm.The mean recognition rate of people counting is96.45%,and the miss rate is 3.55%,and the mistake rate is 1.7%.The high-precision people counting algorithm can be obtained.