"Interaction Analysis with A Bayesian Trajectory Model"

Human behavior recognition is one of the most important and challenging objectives performed by intelligent vision systems. Several issues must be faced in this domain ranging from the extraction of reliable and meaningful features from video streams, to the accurate definition of models able to characterize in a discriminative but, at the same time, general enough way people actions. In this work to cope with these problems a Dynamic Bayesian Network (DBN) model is proposed to recognize people interactions using their patterns of movement. Couples of observed interacting trajectories are processed by the Instantaneous Topological Map (ITM) algorithm to automatically create a topological map used to define the states of the DBN. The trajectories are encoded into the model where conditional probability densities are learned in order to statistically describe interactions. Finally, an online cumulative measure is defined to recognize interactions both in simulated and real-world environments. In this work instead of considering the trajectories independently one from another, we focus on the analysis of motion dependencies between couple of persons to recognize interactions. This choice is motivated by the fact that often human actions are interpretable only if the context is also taken into consideration. For example, a meeting situation can be detected only by identifying that a person is still because he is talking with someone in front of him/her. A statistical representation of interactions based on Dynamic Bayesian Networks (DBNs) is proposed to encode causal relationship in a probabilistic framework and classify unknown couples of trajectories. The Instantaneous Topological Map (ITM) algorithm is used to define the state of the DBN in an online way by creating a topological map based on trajectory observations. The states of the DBN are motion events defined in the map that are used to represent interactions in terms of cause/effect relationships. The proposed model presents the following properties: 1) the trajectory representation is not dependent to the starting point; 2) it is able to represent trajectories sampled with different number of points; 3) it is robust to tracking error and local track loss; 4) it makes possible to classify interactions in an online way; 5) it is easily adaptable to a new scenario. The paper is organized as follows: in Section II an overview of previous works on trajectory modeling and interaction recognition is provided. In Section III the bio-inspired model of interaction and its representation into a Dynamic Bayesian Networks framework are proposed with the aim to model and classify interactive trajectories. In Section IV the discrimination capabilities of the proposed representation for different types of interactive trajectories are shown and finally, in Section V conclusions are drawn.