Unsupervised People Counting From Multiple Views

People counting is an important topic in computer vision field.As monitoring systems rapidly expand recently,analysis and recognition by systems themselves require more demanding techniques.While most proposed methods make use of only images or videos of the scene captured by a single camera from a certain angle,considering the 3D complexity such as occlusion in the scene of interest,the information in these 2D data from a single camera is often limited;therefore,external information is often adopted such as the training set of the objects.In this paper,we propose to monitor the scene of interest from different angles simultaneously with multiple cameras.By integrating these images or videos with additional information on the geometric features of the scene as well as the calibrated cameras,we propose an efficient method to reconstruct the 3D structure.Based on the information in the reconstructed 3D scene,people counting can be accomplished.After extract the foreground in advance using the Gaussian mixture model,at the step of 3D reconstruction,we segment the scene into cube units.Through the proposed index called filling rate,information can be borrowed between different cameras,furthermore to determine the cube units to which the foreground pixels in the data hold spatial correspondence.By properly choose the size of the cube units,we can improve the algorithm efficiency without compromising precision required.After projecting the information in the reconstructed 3D scene to the horizontal plane according to certain rules,we can perform clustering algorithms such as contour tracing to get the counting results.Simulation shows that even in quite crowd scenes can the proposed method give precise counting numbers.Applied to real data,our method also shows results of some practical values even the monitoring system's settings hamper the method.