Part B | 22.4
510 Part B Autonomous Ocean Vehicles, Subsystems and Control
22.4 Maritime Surveillance
Marine traffic is one of the backbones of the world’s
economy. Together with fishing and industrial exploration and recreation activities, an enormous amount
of uncorrelated movements take place. The amount of
data is tremendous; sensors are erroneous and cannot always been trusted. For example, AIS (automatic
identification system) is a locating signal which all
large commercial ships must carry, though many vessels
transporting illicit cargo turn off their AIS transmitters
in order to avoid cooperatively reporting their status.
Turning off the AIS transmitter however is an anomalous behavior which can be detected.
For information gathering tasks, the evaluation of
sampling strategies depends on the ability to produce
quantitative metrics for system performance, which in
turn depends on having a clear sense of what aspects
of the observed field are important. Much recent effort
has been directed toward anomaly detection. The concept of an anomaly or anomalous behavior is dependent
on baseline statistics that are problematic when regimes
can only be sampled sparsely. In such cases, the role of
feedback will be even more critical with perhaps active
probing (sending an UAV (unmanned aerial vehicle))
necessary to elicit a response. False positives (the false
alarm problem) and their associated costs are typically
the most serious limitation of a highly sensitive system.
When risk is included as a metric, state variables may
have different associated utility and expected values.
In the case of systems for anomaly detection, consider
a large sensor network including UAVs, AUVs, ships,
radars, etc., for the detection of smuggling, or otherwise anomalous behavior in coastal waters. The system
would not only detect anomalous behavior, but it could
also access databases about vessels, their cargo and crew
in order to estimate risk for the particular observation.
This imposes high levels of uncertainty in the system as
public databases might be inaccessible, or UAV assets
not available at the time. The system has to constantly
reconfigure in order to deduce the information relevant
to the operators. An adaptive architecture is necessary
for the coordination scheme to cope with this variability. The implementation of autonomous ocean vehicles
could potentially be a game changer.
In maritime surveillance, there are many innocent
players who are following the rules but run into strange
situations (by equipment failure or environmental conditions). Surveillance of all these players is not possible.
Practically, the target always risks being under surveillance (the covert receiver is the general case). Instead
of trying to be invisible (which is not possible), the objective of the target is to make only its illegal activities
invisible.
The aim of maritime surveillance is to check the sequence of decisions made by the target to see whether
this sequence is designed to be maximally independent. That is an anomaly because decisions of the target
should be normally following another strategy (the one
written in the AIS log or inherently given like bringing
goods from A to B, finding fish, etc.). The strategy for
the surveillance team is to gather as much data as possible for this task. Coverage is an attribute for the search
phase. A quick reaction capability of requesting more
data if an anomaly has been detected is an attribute
for Phase 2 (verification and hold). More than in the
multistatic scenario, the maritime surveillance scenario
will be demand based in order to avoid the transmission
of all collected data. Demand-based data sharing might
look like: Something is strange here. Do you have more
information? An implementation of this network should
build upon the features of a publish and subscribe architecture.
In this case, we are not only gathering information
about positions, but about processes. Position information is still important because it can be, for example, the
start position of a process or a constraint for a process,
but in general we are looking for a measured sequence
of positions and activities which is then used to estimate
the target’s intention. A decision on whether the target’s
behavior is inside or outside normal bounds is called an
anomaly detector.
The target’s goal is to disguise the illegal intention
as much as possible for a sufficiently long time. The target is modeled to have a small number of mental states,
as before, and determines how much probability of letting the surveillance detect its behavior is acceptable.
The coordination architecture is again a mix between
centralized synchronization and distributed decision
making and the coordination method is synchronized
covertness (letting the target know that you might be
watching, but not whether you are really there or not).
This can be visualized as a separation in the related
probabilistic graphical model.
510 Part B Autonomous Ocean Vehicles, Subsystems and Control
22.4 Maritime Surveillance
Marine traffic is one of the backbones of the world’s
economy. Together with fishing and industrial exploration and recreation activities, an enormous amount
of uncorrelated movements take place. The amount of
data is tremendous; sensors are erroneous and cannot always been trusted. For example, AIS (automatic
identification system) is a locating signal which all
large commercial ships must carry, though many vessels
transporting illicit cargo turn off their AIS transmitters
in order to avoid cooperatively reporting their status.
Turning off the AIS transmitter however is an anomalous behavior which can be detected.
For information gathering tasks, the evaluation of
sampling strategies depends on the ability to produce
quantitative metrics for system performance, which in
turn depends on having a clear sense of what aspects
of the observed field are important. Much recent effort
has been directed toward anomaly detection. The concept of an anomaly or anomalous behavior is dependent
on baseline statistics that are problematic when regimes
can only be sampled sparsely. In such cases, the role of
feedback will be even more critical with perhaps active
probing (sending an UAV (unmanned aerial vehicle))
necessary to elicit a response. False positives (the false
alarm problem) and their associated costs are typically
the most serious limitation of a highly sensitive system.
When risk is included as a metric, state variables may
have different associated utility and expected values.
In the case of systems for anomaly detection, consider
a large sensor network including UAVs, AUVs, ships,
radars, etc., for the detection of smuggling, or otherwise anomalous behavior in coastal waters. The system
would not only detect anomalous behavior, but it could
also access databases about vessels, their cargo and crew
in order to estimate risk for the particular observation.
This imposes high levels of uncertainty in the system as
public databases might be inaccessible, or UAV assets
not available at the time. The system has to constantly
reconfigure in order to deduce the information relevant
to the operators. An adaptive architecture is necessary
for the coordination scheme to cope with this variability. The implementation of autonomous ocean vehicles
could potentially be a game changer.
In maritime surveillance, there are many innocent
players who are following the rules but run into strange
situations (by equipment failure or environmental conditions). Surveillance of all these players is not possible.
Practically, the target always risks being under surveillance (the covert receiver is the general case). Instead
of trying to be invisible (which is not possible), the objective of the target is to make only its illegal activities
invisible.
The aim of maritime surveillance is to check the sequence of decisions made by the target to see whether
this sequence is designed to be maximally independent. That is an anomaly because decisions of the target
should be normally following another strategy (the one
written in the AIS log or inherently given like bringing
goods from A to B, finding fish, etc.). The strategy for
the surveillance team is to gather as much data as possible for this task. Coverage is an attribute for the search
phase. A quick reaction capability of requesting more
data if an anomaly has been detected is an attribute
for Phase 2 (verification and hold). More than in the
multistatic scenario, the maritime surveillance scenario
will be demand based in order to avoid the transmission
of all collected data. Demand-based data sharing might
look like: Something is strange here. Do you have more
information? An implementation of this network should
build upon the features of a publish and subscribe architecture.
In this case, we are not only gathering information
about positions, but about processes. Position information is still important because it can be, for example, the
start position of a process or a constraint for a process,
but in general we are looking for a measured sequence
of positions and activities which is then used to estimate
the target’s intention. A decision on whether the target’s
behavior is inside or outside normal bounds is called an
anomaly detector.
The target’s goal is to disguise the illegal intention
as much as possible for a sufficiently long time. The target is modeled to have a small number of mental states,
as before, and determines how much probability of letting the surveillance detect its behavior is acceptable.
The coordination architecture is again a mix between
centralized synchronization and distributed decision
making and the coordination method is synchronized
covertness (letting the target know that you might be
watching, but not whether you are really there or not).
This can be visualized as a separation in the related
probabilistic graphical model.