Research On The Key Technology Of Terminal Passenger Intelligence Analysis System Based On Deep Learning

With the development of civil aviation,there is a growing demand for computer vision technology in the field of civil aviation.Passenger management in airport terminals that are the most crowded scenes in the civil aviation industry is an essential guarantee of aviation safety.At present,the monitoring and management of passengers in the terminal mainly rely on manual,which is labor-intensive and poses safety hazards as staffs are easily fatigued due to the high intensity of work.To improve the intelligent analysis of passenger surveillance video in the terminal,this paper investigates and improves the key technologies of the passenger intelligence analysis system.The intelligent passenger analysis system is mainly based on computer vision,which involves key technologies such as crowd density estimation algorithms,pedestrian detection algorithms,and human posture estimation algorithms in crowded scenes.The main work of this paper includes:1.A multi-task learning for pedestrian detection algorithm is proposed.The algorithm utilizes knowledge of crowd density estimation tasks to assist pedestrian detection tasks in crowded scenes.The performance of pedestrian detection is improved while reducing missed detection.In addition,a method for generating pseudo-labels for crowd density estimation with pedestrian detection annotation is proposed,which enables the crowd density estimation task to be free from reliance on labels and reduces a large amount of annotation work for crowd density estimation.Compared to the traditional approach of using two networks separately for density estimation and pedestrian detection,this paper successfully fuses the two tasks into a network,which not only improves the performance of pedestrian detection and the accuracy of crowd density estimation in dense scenes,but also greatly reduces the computational consumption of the traveler intelligence analysis system.2.An improved human pose estimation method based on a high-resolution network is proposed.Firstly,a heatmap-based limb orientated loss function is designed to solve the lack of explicit constraints on the human structure in current pose estimation methods.The limb loss function and the traditional heatmap loss function jointly supervise the training of the network model to reduce the prediction of bizarre poses by the network.Secondly,a small refined network is added with essentially no change to the vanilla network,which utilizes the shallow features of the vanilla network model to refine the rough predicted poses.Finally,comparative experiments demonstrate that the method in this paper greatly improves the accuracy of pose estimation with negligible 1M extra parameters.The method in this paper can be used not only for highresolution networks,but can also be transferred without modification to other pose estimation networks.3.An intelligent analysis system for passenger monitoring and management is designed.The terminal passenger intelligent analysis system is built based on the human pose estimation algorithm,crowd density estimation,and pedestrian detection algorithm proposed in this paper.Furthermore,chaotic queue detection and passenger behavior recognition are implemented after obtaining the crowd density map and human pose estimation.