Abnormal Vessel Behavior Detection in Port Areas Based on Dynamic Bayesian Networks Francesco Castaldo and Francesco A.N. Palmieri DIII Seconda Universit` a degli Studi di Napoli Aversa (CE), Italy {francesco.castaldo, francesco.palmieri}@unina2.it Vahid Bastani, Lucio Marcenaro and Carlo Regazzoni DITEN Universit` a degli Studi di Genova Genova, Italy vahid.bastani@ginevra.dibe.unige.it, {lucio.marcenaro, carlo.regazzoni}@unige.it Abstract—Automatic recognition of abnormal situations in harbor environments is approached in this paper with a system based on Dynamic Bayesian Networks. The area under surveil- lance is partitioned in zones of different sizes and shapes by means of an Instantaneous Topological Map, on which events are detected and inference is carried out. The model is trained with synthetic normal trajectories of ships and vessels mooring in the port, and each time a new trajectory is presented to the system, comparisons with the normal behaviors stored in the network are performed. If no match is found, an abnormal situation is declared and countermeasures can be taken. The algorithm has been tested in a real port with simulated data in order to evaluate the false alarm rate and the abnormal detection capabilities of the proposed approach. I. I NTRODUCTION In recent years the design of systems that guarantee an high level of security in highly-crowded and critical areas such as harbors, airports, busy streets, etc., have become an important research subject for both academia and industry. Dangerous situations and accidents (as the one happened in 2013 in Genoa, Italy [1] in which the control tower in the harbor was pulled down by a ship with a malfunctioning engine) are increasingly perceived as intolerable, given the quality and the quantity of the sensory equipment that nowadays surrounds and monitors these areas. These situations seem to suggest the need of new mechanisms to merge low level information in such a way that a behavioral analysis of what is happening in the area can be pursued. Nowadays, many kinds of sensors such as radar, video and infrared cameras and Automatic Identification System (AIS) are widely deployed for monitoring maritime environments. However, the effectiveness of this broad usage is limited, mainly because of the involvement of human operator that is required to analyze the enormous amount of information generated by the sensors and decide accordingly. Thus, in the past years, considerable research effort has been devoted to developing efficient automatic detection, tracking and classi- fication of targets and sensor fusion to help users of such systems. Yet, there is still the need for a system which is able to analyze the environment, its elements and the relation between events in order to provide a state of situation awareness. The key to situation awareness is the ability to analyze the behavior of the actors in order to understand abnormal events. In the maritime domain actors’ (ships, boats, etc.) behavior appears in their movement and trajectories. Thus, the next step after detection and tracking is motion and journey analysis for detecting unexpected and abnormal trajectories in order to warn the operator, or activate an automatic action mechanism. In this paper we propose a system, based on the probabilis- tic graphical model (PGM) [2] of Dynamic Bayesian Networks (DBN) [3] [4], in which causal relationships regarding the normal movements of a ship in a port area are encoded in a probabilistic model based on events. Behaviors that signifi- cantly differ from the ones defined as normal are likely to be dangerous and have to be properly managed. For such systems low alarm rates and rapid evaluation of dangerous situations are mandatory in order to start appropriate security policies (alarm siren, evacuation of the area, etc.) in time. The approach proposed in this paper is inspired by the work presented in [5] [6] in which methods for modeling and analyzing trajectories using topological representation of the environment and dynamic Bayesian network learning and inference are introduced. The approach shows to be promising not only for detection of abnormal interactions [5] but also for analyzing the operator’s interaction with the environment [6]. Unlike other classic modeling techniques which mostly rely on trajectory clustering and learning [7], DBNs due to their inherited dynamics are able to easily learn a wide variety of spatio-temporal models and provide statistically optimum inferences. Compared to [5] [6], we enhanced the method by allowing the time difference between events to exist as a separated random variable in the DBN. In addition, we modified the cumulative measure to achieve bounded value in order to be able to apply fixed abnormality detection threshold. More details on the algorithmic implementation can be found in Section III. The main contribution of this paper is to combine the Bayesian approach with reduction state space techniques (ad- dressed in Section III) in order to design a system capable of recognizing normal/abnormal maritime situations in real time, giving to the port authorities the time to intervene if the acknowledged situation is labeled as dangerous. Unlike other methods (detailed in Section II), our algorithm is devised for an online learning of behaviors: at first any unseen trajectory is regarded as abnormal, but with the intervention of the users (e.g. security operators in port), it is possible to add the new detected trajectory into the normality dictionary. In this way, as the algorithm works, its model of the environment becomes more comprehensive.