Crowd Monitoring Using Image Sequence Processing
VANNOORENBERGHE P, MOTAMED C, BLOSSEVILLE J M and POSTAIRE J G, Universite du Littoral, France
Pedestrians motion parameters are required by traffic engineers for the design and managing of urban areas subject to crowd movements, such as road crossings or business centres. In-site pedestrians flow measurements by human observers, which require a la
Pedestrians motion parameters are required by traffic engineers for the design and managing of urban areas subject to crowd movements, such as road crossings or business centres. In-site pedestrians flow measurements by human observers, which require a large manual effort, are very difficult to apply in crowded conditions. During the last years, a significant reduction of video hardware costs has led to the installation of a great number of CCTV systems in urban areas. Methods used in crowd monitoring applications must take into account some specific properties of pedestrians motion
* their movements are unpredictable, changing in speed and direction, especially in crowded conditions.
* the shape of a person in the image is highly variable since it derives from a complex combination of individual image velocities created by the different parts of the body.
* people are not always walking continuously, they can stop and remain completely stationary for a significant period of time.
This last characteristic is of major importance : the motion-based segmentation method must cope with image sequences showing areas that can remain fixed during several images. An object that stops under the camera should still be considered as mobile during a controlled time lag. However, a direct segmentation of image sequences showing pedestrian flows would be a very complex task. Assuming that pedestrians are moving in the scene, motion has been suggested as an important cue for the segmentation of crowd images. In addition to motion estimation, several studies [Rourke 1994, Velastin 1994] show that edge detection can play an important role in automatic crowd monitoring processes.
Motion analysis is a crucial problem in computer vision, for which many different approaches are available [Nagel 1983]. Estimating motion parameters from the analysis of a dynamic scene is one of the basic problems that must be solved. Motion reconstruction in the original 3D scene is not always possible since it involves a resolution of the inverse projection problem which is ill-posed. Thus, many approaches limit themselves to the estimation of image velocities --or optical flow-- which are the projections of 3D velocities on the image plane. The methods which have been proposed in order to extract motion information from a sequence of video images, can be divided into two main approaches [Ullmann 1981]. Object- oriented methods [Kories 1984, Sethi 1988] are based on two successive steps : an extraction of features describing the objects and a matching of these features detected in the successive images of the sequence. On the other hand, motion-oriented methods analyse the spatio- temporal variations of grey levels in order to estimate the local motion [Horn 1981].
We present, in this paper, a video-based system used to collect pedestrian traffic data. It is composed of two different processes. The first process is a motion detection algorithm, which can cope with deformable moving bodies like pedestrians [Vannoorenberghe 1996]. Its originality consists in the on line construction of a reference image, containing all static edges in the scene. It is particularly robust towards illumination changes. A tracking process is then used to extract pedestrians traffic parameters like occupation rate, crossing time. This second process is based on a feature matching algorithm and works in real time. The system is adapted to crowd monitoring and used as an intelligent sensor in the project ~, Intelligent Cross-roads >> developed by I.N.R.E.T.S. (National Research Institute of Transportation and Security).
Association for European Transport