Facility surveillance systems and methods
Systems and methods for surveillance of a facility including facility elements. The system includes a central computing unit providing a digital model of the facility providing topological or logical or functional relationships of the facility elements, surveillance sensors adapted for surveillance of a plurality of the facility elements and for generation of surveillance data, communication means for transmitting data from the surveillance sensors to the central computing unit, and state derivation means configured to analyse the surveillance data and derive a state of a respective facility element. The central computing unit is configured to record a state pattern by combining states of at least one facility element based on at least one relationship of the facility element provided by the facility model, provide a state pattern critical-noncritical classification model which considers relationships provided by the facility model, and perform a criticality-classification based on the relationship.
1 . An automated surveillance system for an automated detection of states at a facility,
the system comprising multiple surveillance-sensors of which at least two are operating in different modalities, and
wherein the automated detection of states in course of an automated surveillance of the facility is derived from a combination of those multiple surveillance-sensors of which at least two are operating in different modalities in course of said automated surveillance of the facility,
wherein,
the combination of multiple surveillance-sensors of which at least two are operating in different modalities is provided by a machine learned system comprising a context-adaptive model for data of different modalities, which is trained on training data comprising a contextual information of the facility in a training phase in a plurality of contexts,
whereby in course of the training of the machine learned system contexts are learned and extracted and an optimal combination of surveillance-sensors of which at least two are operating in different modalities for the automated detection of states is learned for these extracted contexts,
deriving a parallel and/or hierarchical evaluation structure for the data from the surveillance sensors based on the machine learning in the plurality of contexts, whereby an optimal combination of the surveillance-sensors in hierarchical and/or parallel structures is learned automatically from data for each of the extracted contexts,
wherein the combining is established by a hierarchical and/or parallel structure for merging the at least two surveillance sensors, wherein a weighting of the different sensing modalities is derived by the machine learning and dependent on the contextual information,
and, after the training phase, using the learned optimal combination of surveillance-sensors of which at least two are operating in different modalities for the automated detection of states in course of said automated surveillance of the facility.
2 . The surveillance system according to claim 1 , wherein the combination is chosen from at least a subset of the surveillance-sensors by a machine learned weighting function for the modalities, in particular wherein the weighting function comprises weighting factors for the modalities which are depending on the contextual information.
3 . The surveillance system according to claim 2 , wherein the weighting factors are classified by a classifier that is machine learned on the training data comprising the contextual information.
4 . The surveillance system according to claim 3 , wherein the classifier or a detector for the detection of states is at least partially trained on training data which is at least partially synthetically generated and derived from a virtual model.
5 . The surveillance system according to claim 1 , wherein the contextual information is at least partially derived from the surveillance-sensors, in particular wherein the contextual information which is comprised in a weighting of one modality is from another modality than the one modality which the surveillance-sensor is operating on.
6 . The surveillance system according to claim 1 , wherein the contextual information is at least partially derived from an environmental sensor.
7 . The surveillance system according to claim 1 , wherein the contextual information comprises one or more of a temperature context and/or a spatial context.
8 . The surveillance system according to claim 1 , wherein the contextual information is partitioned/segmented into classes by machine learning of a machine learned classifier for filtering the surveillance-sensors dependent on the contextual information.
9 . The surveillance system according to claim 1 , wherein the modalities comprise at least a visual modality, an infrared modality and a depth modality, in particular also a spatial modality and an audio modality, which modalities are derived by different sensors.
10 . The surveillance system according to claim 9 , wherein multiple of the modalities are combined into a single dataset, by which dataset the machine learned system is trained.
11 . The surveillance system according to claim 10 , wherein a combination of different modalities into a single data set comprises geometric transformation of the data of at least one of the modalities in such a way that the combined modalities are referenced to single common coordinate system of the single dataset, in particular combined with a pixel-to pixel-correspondence of the modalities in a multi-channel image.
12 . An automated surveillance method for a machine learned detection of a critical state of a facility, the method comprising the steps of:
deriving multiple surveillance data from at least two surveillance sensors for the facility, which surveillance sensors are operating in at least two modalities in course of the automated surveillance of the facility,
deriving at least one contextual information at the facility, in particular an environmental information by a context-sensing means,
combining data from one or more of the surveillance sensors of which at least two are operating in different modalities for the detection of the critical state in course of the automated surveillance of the facility with a machine learned automated information filter comprising a context-adaptive model for data of different modalities, which is established by a training on training data which comprises the contextual information in a training phase in a plurality of extracted contexts,
whereby in course of the training of the machine learned automated information filter contexts are learned and an optimal combining of surveillance-sensors of different modalities for the detection of a critical state is learned for these extracted contexts,
deriving a parallel and/or hierarchical evaluation structure for the data from the surveillance sensors based on the machine learning in the plurality of contexts, whereby an optimal combination of the surveillance-sensors in hierarchical and/or parallel structures is learned automatically from data for each of the extracted contexts,
wherein the combining is established by a hierarchical and/or parallel structure for merging the at least two surveillance sensors, wherein a weighting of the different sensing modalities is derived by the machine learning and dependent on the contextual information,
and, after the training phase, using the learned optimal combining of surveillance-sensors of which at least two are operating in different modalities for the automated detection of states in course of said automated surveillance of the facility.
13 . The surveillance method according to claim 12 , comprising a weighting or parameterization of the parallel and/or hierarchical evaluation structure by machine learning of context dependent weighting factors.
14 . The surveillance method according to claim 12 , wherein the machine learned automated information filter is configured by machine learning for a contextual information dependent combining of the at least two modalities to derive the detection of a critical state.
15 . The method according to claim 13 , wherein the weighting factors are classified by a classifier that is machine learned on the training data comprising the contextual information.
16 . The method according to claim 15 , wherein the classifier or a detector for the detection of states is at least partially trained on training data which is at least partially synthetically generated and derived from a virtual model.
17 . The method according to claim 16 , wherein the contextual information is at least partially derived from the surveillance-sensors, in particular wherein the contextual information which is comprised in the weighting of one modality is from another modality than the one modality which the surveillance-sensor is operating on.
18 . A method for deriving a machine learned automated surveillance system embodied in a computation unit, in particular for a surveillance system according to claim 1 , which comprises an automatic classifier and/or detector for security issues comprising a context-adaptive model for data of at least two different modalities based on at least two surveillance sensors operating in at least two different sensing modalities for an automated surveillance of a facility,
wherein a training of the automatic classifier and/or detector comprises providing training data, in particular real world training data from the surveillance sensors of which at least two are operating in different modalities, which is at least partially comprising a contextual information for the training data, and
whereby a combining of the at least two different sensing modalities is machine learned in a training phase in a plurality of extracted contexts, which combining is at least partially segmented according to the contextual information,
whereby contexts are learned and extracted in course of the training of the automatic classifier and/or detector and an optimal combining of surveillance-sensors of different modalities for the automated surveillance is learned for these extracted contexts,
deriving a parallel and/or hierarchical evaluation structure for the data from the surveillance sensors based on the machine learning in the plurality of contexts, whereby an optimal combination of the surveillance-sensors in hierarchical and/or parallel structures is learned automatically from data for each of the extracted contexts,
wherein the combining is established by a hierarchical and/or parallel structure for merging the at least two surveillance sensors, wherein the structure and/or a weighting of the different sensing modalities is derived by the machine learning and dependent on the contextual information,
the learned optimal combining of surveillance-sensors of which at least two are operating in different modalities being adapted for the automated detection of states, after the training phase, in course of an automated surveillance of the facility with the machine learned automated surveillance system.
19 . A computer program product comprising program code which is stored on a non-transitory machine-readable medium, and having computer-executable instructions for performing the steps of the method according to claim 12 , when run on a computing unit of a surveillance system.