Method For Acquiring Fundamental Diagram Of Pedestrian Movement Based On Deep Learning And Optical Method

In the study of pedestrian dynamics,the study of pedestrian motion law can provide a strong theoretical support for the design of building facilities and the compilation of emergency plan for activities,and ensure public safety.The fundamental diagram is an effective tool in pedestrian dynamics analysis.In recent years,deep learning has developed rapidly,and many achievements have been made in the field of computer vision.Although pedestrian dynamics have been extensively studied,there is still a lack of research methods and practical applications combined with deep learning.This paper proposes a deep fundamental diagram network structure.The deep fundamental diagram network consists of a multi-scale recursive convolutional neural network(MSR-Net)and an optical flow module.MSR-Net has learned the mapping relationship between the video frame image and the pedestrian density map,and the density map reflects the pedestrian spatial information.The optical flow module obtains the pedestrian speed map through the sparse optical flow algorithm,and the speed map reflects the pedestrian's motion vector.Through the spatial correspondence between the speed map and the density map,a fundamental diagram of pedestrian movement can be obtained.We choose straight channel experiments and bottleneck experiments to verify our method.The accuracy of pedestrian density and speed obtained through the deep fundamental diagram network is high,and the robustness of the network is good.The trend of pedestrian movement reflected by the fundamental diagram is consistent with the fundamental diagram obtained by traditional methods.The deep fundamental diagram network runs fast,can get pedestrian dynamics information frame by frame,and can quickly obtain fundamental diagram and perform pedestrian dynamics analysis.Using deep fundamental diagram networks,this paper proposes an anomaly detection method using pedestrian motion information.It avoids the problems of difficult to distinguish foreground and background,difficult to detect quickly,and difficult to label samples in pedestrian anomaly detection.This paper also applies the pedestrian motion information output from the deep fundamental diagram network to the pedestrian crossing counting task,which solves the problem that the target tracking method in high-density crowd is difficult to distinguish individual pedestrians,and the problem of poor robustness and counting jumps of the timing slicing method.Pedestrian dynamics experimental verification shows that compared with commonly used target tracking methods,our line crossing technology method has higher accuracy.Good results have been obtained in the detection of pedestrian anomalies and pedestrian crossing counts in the deep fundamental diagram network,which proves that our network has the ability to adapt to different tasks,and has achieved the research and application of the combination of pedestrian dynamics research and deep learning.