This is a 'friendly' version of the work done on anomalous activity explanation. For a more detailed version of the discourse, please see the CVPR 2005 paper.

Anomalous Activity Explanation

If we look up the word anomaly in a dictionary, we would find descriptions such as "deviation from common or regular", or "something different, or strange".

Of course these descriptions don't really mean anything without first understanding the notion of what is meant by regular. Can something common be described as regular, or do we need a set of rules to define regular more precisely? Once the idea of regular has been established, we need to answer the question of what do we mean by being different. Moreover, if we do consider being different from regular, as a measure of being anomalous, how different is different enough to be considered an anomaly? And finally, in what ways something anomalous, is different from something typical or regular? We try to tackle these questions in the backdrop of understanding everyday activities, and how we can explain anomalies in such situations.

Most of the previous attempts to tackle the problem of Anomaly Detection have focused on model based anomaly recognition, where a particular type of activity is pre-defined as being anomalous, is modeled in some way, and is then searched for [1][2]. While such an approach could prove to be useful for cases where the variance between different anomalous instances is not significantly large [3], for any reasonably unconstrained situation, anomalies are hard to completely define a priori. Since this is particularly true for everyday activities, we define an anomaly as "something different from regular", with the hope of being able to model something regular more efficiently. We consider the term regular from a Bayesian perspective, i.e. we regard things that are common as regular. This in turn leads us to our final definition of an anomaly as "something rare and different from regular".

Interestingly enough, there are studies done in Cognitive Science which show evidence that humans also learn about anomalies by considering the "distance" of a new piece of information from the mental model of the class which they believe the new information belongs to [4][5]. Thus the decision about the class membership of a new instance is made first, and then the decision about it being a normal or an anomalous class member is made.

More precisely, we divide our problem at hand into the following parts:

1- Representation of an activity.

2- The notion of similarity between activities.

3- Clustering activities into disjunctive groups.

4- Finding the typical (best representative) member of a cluster.

5- Classifying a new activity into one of the classes.

6- Detecting if this new class member is anomalous or regular.

7- In case the new class member is anomalous, explaining away in what respects is it different from the typical member of the membership class.

 

 

[1] S. Hongeng and R. Nevatia, "Multi-agent event recognition, " in ICCV, 2001, pp. II: 84--91.

[2] Koral Ilgun, R. A. Kemmerer, Phlip A. Porras, "State Transition Analysis: A Rule-Based Intrusion Detection Approach, " in IEEE Transaction on software engineering, 1995, pp. 188-199.

[3] Mei Chen, Takeo Kanade, Dean Pomerleau, Henry A. Rowley: Anomaly detection through registration. Pattern Recognition 32(1): 113-128 (1999)

[4] Chi, M. T. H. (1992). Conceptual change within and across ontological categories. Cognitive models of Science: Minnesota Studies in Philosophy of Science, pp: 129-186.

[5] Grdenfors et. al. (1995). Concept Formation in Dimensional Spaces. Lund University Cognitive Studies 26.