Dealing with Uncertainty in Situation Assessment: towards a Symbolic Approach

The situation assessment problem is considered, in terms of object, condition, activity, and plan recognition, based on data coming from the real­ word via various sensors. It is shown that uncer­ tainty issues are linked both to the models and to the matching algorithm. Three different types of uncertainties are identified, and within each one, the numerical and the symbolic cases are distin­ guished. The emphasis is then put on purely sym­ bolic uncertainties: it is shown that they can be dealt with within a purely symbolic framework resulting from a transposition of classical nume­ rical estimation tools. 1 SITUATION ASSESSMENT PROBLEM Let us consider the generic problem that is dealt with in the PERCEPTION project1 [BCF+98], [CCMT97): a sym­ bolic representation of what is going on in an observed environment has to be built and updated, for applications such as surveillance, intelligence, or decision-aid systems, and autonomous systems. The environment includes mo­ bile entities and is observed via various sensors (black and white, color and infrared cameras, radars). Numerical pro­ cessings take sensor data as inputs and deliver recognized and tracked objects with symbolic properties (e.g. the type of the objects: pedestrian, vehicle ... ) and numerical attribu­ tes (position, speed ... ) The symbolic level interprets these objects in terms of on-going and future activities (e.g. the pedestrian is going to take his car and leave the parking­ lot), so that the decision level (e.g. a human decision­ maker) should be informed with semantically rich data and that further relevant actions should be undertaken. Human observers may also be involved as "sensors", and their reports are direct inputs for the symbolic level. 1 http://www.cert.fr/fr/dcsd/PUB/PERCEPTION/ 2 PRINCIPLES The set of the current activities is assessed by the symbolic level thanks to plan prototypes based on the expected properties and attributes of the objects and on the expected variations of the properties and attributes with time. Three basic notions are used: • a condition prototype is an expected property a priori qualifying the objects that are likely to be observed. Condition prototypes are expressed by atomic formulas built from predefined predicates, e.g. (type x pedestrian), (speed x 4kmlh), (getting-closer x y) with x andy being variables; • an activity prototype is a set of expected conditions and constraints a priori qualifying the objects that are likely to be observed. Activity prototypes are represented by constrained cubes [TGLP88], i.e. conjunctions of atomic formulas associated with constraints, in which all the variables are assumed to be existentially quantified [CCMT97]. Ex: {(type x pedestrian), (type y vehicle) , (speed y v) {v = 0}, (getting-closer-tox y)}, withx, y and v being variables. • a plan prototype is a temporal graph of activity proto­ types; it is represented by an interpreted Petri net [DA91] whose places are associated with activity prototypes. Ex: pedestrian moving towards-vehicle pedestrian gelling into-vehicle parked-occupied-vehicle vehicle moving towards-exit parked-vehicle (typ e y vehicle) (speedyv){v=O) (type y vehicle) (speedy v){v = 0) (type y vehicle) (speedy v){v H) (moving -away-from y parking-lot) (type y vehicle) (speedy v){v H) (getting-closer-to y exit) Figure 1: vehicle-departure plan prototype Let P be the set of plan prototypes. 62 Castel, Cossart, and Tessier An observation obsn is a set of properties directly issued by the numerical processing, resulting from a numerical­ symbolic translation, or issued by a human observer at time tn. The current situation Sn at time tn is a set of plans (Pi,m,,n), defined as marked elements Pi of P; the mar­ king mi of a plan prototype Pi at time tn corresponds to the fact that some properties in obsn match the interpreta­ tion of some places (activity prototypes) in Pi. Given obsn+l and Sn, the elaboration of the current situ­ ation Sn+l at time tn+l is a prediction-verification process which is based on the following principles: i) a greater importance is given to the continuation of existing plans; ii) all the objects appearing in properties within obsn have to be explained, i.e. to belong to at least one plan; iii) the prediction of situation Sn+l from situ­ ation Sn is the set of the reachable markings mi + k of plans (Pi,m;,n); iv) the verification consists in matching the properties of obsn+l with those reachable markings; if some properties remain unmatched this way, new plans (Pj,mj,n+l) are created. As a given object may be associated with several plans, Sn+l represents the different hypothetic plans that are likely to be in progress in the observed environment. 3 UNCERTAINTY ISSUES Whatever the situations are built for (immediate or delayed warnings or actions, information collecting in an intelli­ gence context, detection of specific activities ... ), the situ­ ation assessment process has to deliver appropriate results, which means that [KSH91]: • results (assessed situations) have to agree with the global mission goal: a potentially hazardous situation has to be reported early, even if the assessment is not complete or certain; all the situations that are significant for the mission must be expected and recognized. • results have to be accurate, i.e. situations must not in­ clude a high number of different plan hypotheses. There­ fore, activity and plan prototypes, as well as the matching algorithm, have to be discriminating enough (a plan proto­ type that would model that anything can happen is of minor interest). • results have to be computed efficiently, without too nu­ merous or too complicated models. Let us now analyze the situation assessment process from the uncertainty point of view. 3.1 UNCERTAINTY AND THE MODELS The whole situation assessment process is a series of trans­ formations from sensor data into high level symbolic pro­ perties.