Application And Research Of People Flow Statistics Based On Video Detection And Tracking Algorithm

In recent years,the development of image processing technology has effectively promoted the intelligent process of crowd statistics.Among them,the rapid development of artificial intelligence technology plays a important role.It can provide decision support for resource allocation and pre-warning in public places by obtaining people flow information such as airports,subways,and bus stations,and provide data support for formulating scientific and reasonable management methods,thereby preventing public safety accidents.In actual scenes,the complicated monitoring environment,interlaced motion trajectories,and different degrees of mutual occlusion of people and surrounding objects make the problem of missed and false detection of pedestrians prominent.Therefore,to carry out accurate and real-time statistics of the flow of people has high practical needs and research significance.we mainly study the pedestrian detection and tracking methods based on deep learning in this work.Aiming at the deficiencies and existing problems of the existing algorithm models,we propose improved algorithms and solutions.Specifically,the improved lightweight target detection network YOLOv3-MP(YOLOv3-Mobile Net-Person)and pedestrian re-identification enhanced detection-based multi-target tracking track-by-detection algorithm R-Deep SORT(Simple Online and Realtime Tracking)fusion,design a YOLOv3-MP-R-Deep SORT people flow statistics model based on deep learning,and finally design a two-way counting rule to realize real-time statistics of people flow.The main work and innovations of this paper are as follows:1.In view of the disadvantages of the one stage series of deep neural network models YOLOv3 and YOLOv3-tiny,we proposed a lighter with superior robustness and generalization performance based on the YOLOv3 algorithm in pedestrian detection and more suitable for complex monitoring scenarios,which was named YOLOv3-MP.The algorithm model uses a lightweight network Mobile Net instead of the inherent Darknet-53 on the backbone network,optimizes the calculation of the cross-entropy loss function according to the detection task and uses the K-means algorithm in training to cluster the coordinate boxes in the labeled pedestrian dataset.The experiments results revealed that the advanced algorithm can improve the detection effect of small-scale and occluded pedestrians.Especially,our person Average Precision(person AP)reaches87.54% and Log-average Miss Rate(LAMR)decreases by 0.03,which fully demonstrates the effectiveness of the proposed detection algorithm.2.Aiming at the problem of pedestrian ID-switch in multi-target tracking,we feed the model with pedestrian re-identification result,deepen the depth of the deep cosine softmax convolutional network,and integrate the Batch Normalization(BN)acceleration strategy after each residual network to improve the original network in the re-identification public dataset Market-1501.Subsequently,the multi-pedestrian tracking algorithm R-Deep SORT is designed by fusing the Person Re-identification(ReID)enhancement module,using the method of information association fusion measurement and increasing the weight of appearance information.Experimental results show that R-Deep SORT reduces the number of pedestrian ID-switch and reduces Mostly Lost(ML)by nearly 8%,which fully shows that the algorithm improves the performance of multi-pedestrian tracking.3.In order to realize the intelligent statistics of pedestrians,the YOLOv3-MP and R-Deep SORT algorithms are combined,two-way counting rules are designed,and the YOLOv3-MP-R-Deep SORT pedestrian statistics system model is constructed.In real-time monitoring scenarios,the statistical accuracy of the model is up to 91.67%,and the statistical speed is about 30 fps.It can be seen that the intelligent human flow statistics system designed in this paper can meet the accuracy and real-time requirements.