Team siames

Overall Objectives
Scientific Foundations
Application Domains
New Results
Contracts and Grants with Industry
Other Grants and Activities

Section: New Results

Crowd simulation inside exchange areas, with levels of services characterisation.

Participants : Sébastien Paris, Stéphane Donikian [ contact ] .

Crowd simulation is an emergent problem nowadays, because of a growing interest on behalf of industries and certain organisations. Many studies have been performed, mostly using macroscopic models, or light microscopic ones (like particle based). We propose, in our model, to simulate crowd motion by association of a multitude of individual behaviours, and to analyse results in order to extrapolate a levels of services classification.

This study is carried out within the framework of an industrial thesis in collaboration with AREP, Aménagement Recherche pour les Pôles d'Echange , which is a subsidiary company of the SNCF, Société Nationale de Chemins de Fer . The goal of this study is to validate train station architectural plans with respect to the movements of people inside the station. That is made possible thanks to a tool for crowd simulation allowing data extrapolation which will then be synthesized in levels of services.

Levels of services

The concept of Levels Of Services was firstly defined by J.J. Fruin in 1971, and was reused by many researchers. But all of them made their classification only with two discriminating factors: density and flow of people. Such a classification seems well suited for security studies, but suffers a lack of information for thorough ones. What we propose to take into account with our levels of services, in addition to classical ones, is:

The last point to be approached is the fact that a level of service is not evaluated on an overall basis for the studied place, but locally at each zone of interest.

Environment abstraction

The first point to be approached for the simulation of autonomous agents is the description of their navigation environment. Our model is based on a spatial subdivision ( Fig.  27 .a ) introduced by F. Lamarche, which produces a set of convex cells by using a constraint Delaunay triangulation. The first step, called informed subdivision  ( Fig.  27 .b ), computes a topological representation of the spatial subdivision by naming cells according to their number of connexity relations: dead end for one relation, corridor for two, and crossroad for three or more. Then, a topological abstraction is performed twice. A grouping algorithm is first applied to the cells of the informed subdivision to produce groups  ( Fig.  27 .c ). Then, the same algorithm is applied to groups to produce more conceptual zones  ( Fig.  27 .d ).

Figure 27. The entire abstraction process for a simplified environment
a. Spatial subdivisionb. Informed subdivision
(c) First abstraction(d) Second abstraction

This process results in a three level hierarchical graph, which is enhanced with some preprocessing, such as potential visibility sets ( Fig.  28 ), and regular grids linked to each group in order to evaluate local densities.

Figure 28. Precalculated visibility through an entry in a test environment
a. Composed view frustab. PVS representation

Topological knowledge

In order to improve the realism of the simulation of moving entities, our model manages for each agent a topological knowledge. Such a knowledge restricts the ability of the agent to globally access to stored data on the environment. The initial knowledge of th agent can vary from nothing to perfectly known. The knowledge is updated during the simulation by an observation process, using precalculated potential visibility sets. Then, when an agent need some information relative to a not observable part of its environment, it can refer to its topological knowledge.

Path planning

The next necessary task to enable navigation inside an environment is path planning. We propose to take advantage of the hierarchical topological abstraction to perform path planning by part. The path evaluation is first performed entirely on the more conceptual layer ( zones ), then locally on the first abstraction layer to connect the current group to the second zone of the path, and finally on the informed subdivision layer to connect the current cell to the second group of the path ( Fig.  29 ).

Figure 29. Hierarchical path planning diagram

The same A* algorithm is used in the three cases, using a multi-criteria heuristic to characterise the path weight, and only taking into account known environment. This heuristic takes into account travelling distance, local densities of population and flows of people, relative and absolute direction changes, and finally passages width. Moreover, the heuristic parameters can be dynamically changed to reflect changes in the entity path planning behaviour. Finally, the path planning algorithm is reactive to events sent by the observation process, or by the rational procedure. These events result on a partial or full path revaluation.

Interaction with objects and other actors

The interaction with objects and other actors has been approached during the engineering training of Laurent Millet-Lacombe, in collaboration with AREP. The model we propose is close to the ecological theory of J.J. Gibson, describing interactions as affordances linked directly to the objects with which the interaction is possible. This model takes place as a platform called BIIO: Behavioural Interactive and Introspective Objects . What we call objects are certainly physical objects like a chair or a door, but also agents representing virtual people. BIIO enables to attach interactive behaviours to objects, in a hierarchical way: each object inherits the properties of its parent(s), including interaction potentials. Moreover, the objects have strong introspective capacities which enable to recover the whole of their properties. The interactions are classified in two categories. First, using interactions are only available to one actor at the same time, which require to manage a waiting queue for the object. Second, observation interactions are available to many actors at the same time. An interaction is composed of four parts: A rational precondition, which is a boolean expression relative to the actor and the type of the object to interact with; A local precondition, which is a boolean expression relative to the actor and the object; An effect, which may affect one or both of the actor and the object; And finally, a duration which is relative to the actor and the object.

Rational behaviour

The rational behaviour consists in linking the basic interactions provided by BIIO in order to perform a goal oriented behavioural planning. The first task consists in the creation of a behavioural graph whose root is the goal state, and the nodes are the basic interactions leading to that goal. Then, a path must be found in this graph to select each interaction according to the actor knowledge and the interactions evaluated cost. All of these processes are part of our future work.


Logo Inria