Echo state network training method and apparatus, and monitoring track determining method and apparatus

Abstract

The present invention provides an echo state network (ESN) training method and apparatus, and a monitoring track determining method and apparatus. The ESN comprises an input ESN, a chaotic ESN and a feedback classifier. The training method comprises: acquiring attractor sample data, wherein the attractor sample data comprises position information of monitoring tracks arranged in a plurality of regions of interest in a monitoring region of a monitoring camera; pre-training a connection matrix of the input ESN by means of the attractor sample data; inputting the attractor sample data into the pre-trained input ESN to obtain the states of nodes of the pre-trained input ESN; training a connection matrix of the chaotic ESN by means of the states of the nodes output by the pre-trained input ESN to obtain a trained chaotic ESN; and training a connection matrix of the feedback classifier by means of the states of the nodes output by the pre-trained input ESN to obtain a trained feedback classifier.

Description

State echo network training method, monitoring trajectory determination method and device

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to Chinese patent application No. 202311188646.2 filed in China on September 14, 2023, the entire contents of which are incorporated herein by reference. Technical Field

[0003] The embodiments of the present invention relate to the field of security monitoring technology, and in particular to a state echo network training method, a monitoring trajectory determination method and a device. Background Art

[0004] In existing security surveillance systems, surveillance cameras can be deployed to focus on areas of interest (also known as key areas) as needed. However, the camera's trajectory across these areas of interest is typically fixed, and the transition paths between these areas are often single. This approach makes the camera's trajectory easily predictable, potentially causing loss of life and property.

[0005] Summary of the Invention

[0006] The embodiments of the present invention provide a state echo network training method, a monitoring trajectory determination method and an apparatus, which are used to solve the problem that in the existing security monitoring process, the monitoring trajectory of a monitoring camera is usually fixed and easily predicted by criminals.

[0007] In order to solve the above-mentioned technical problems, the present invention is achieved as follows:

[0008] In a first aspect, an embodiment of the present invention provides a method for training a state echo network, wherein the state echo network includes: an input ESN, a chaotic ESN, and a feedback classifier, and the method includes:

[0009] Acquire attractor sample data, where the attractor sample data includes location information of monitoring tracks set in a plurality of interest areas within a monitorable area of ​​a monitoring camera;

[0010] Pre-training the connection matrix of the input ESN using the attractor sample data to obtain a pre-trained input ESN;

[0011] Inputting the attractor sample data into the pre-trained input ESN to obtain the node state of the pre-trained input ESN;

[0012] Using the states of the nodes output by the pre-trained input ESN to train the connection matrix of the chaotic ESN to obtain a trained chaotic ESN;

[0013] The connection matrix of the feedback classifier is trained using the states of the nodes output by the pre-trained input ESN to obtain a trained feedback classifier, which is used to achieve autonomous conversion between the regions of interest.

[0014] Optionally, obtaining attractor sample data includes:

[0015] Control the surveillance camera to collect images of the monitored area using a traversal method;

[0016] All collected images of the monitorable area are stitched together to obtain an overall monitorable area image;

[0017] determining a region of interest in the overall monitorable area image;

[0018] Obtaining a monitoring track set in the area of ​​interest;

[0019] The position information of the monitoring track set in the region of interest is used as the attractor sample data.

[0020] Optionally, the pre-training the connection matrix of the input ESN using the attractor sample data to obtain the pre-trained input ESN includes:

[0021] Inputting the attractor sample data into the input ESN, and obtaining the node status of the input ESN output by the input ESN;

[0022] Determining a first loss function according to a state of a node output by the input ESN and a preset target state of the node of the input ESN;

[0023] The connection matrix of the input ESN is optimized using the first loss function.

[0024] Optionally, the nodes of the input ESN participating in the training are half of all the nodes of the input ESN.

[0025] Optionally, the input ESN is expressed as:

[0026] Where t is time, τ is the proportional coefficient, and x in is the state of the node in the input ESN, g in is the scaling parameter, J in is the connection matrix of the input ESN, u in (s) is the attractor sample data.

[0027] Optionally, the first loss function is:

[0028] Among them, C 1-in is the first loss function, t is time, x s (t) is the state of the node of the input ESN at time t, is the target state of the node of the input ESN at time t, L innate is the maximum value at time t, s is the monitoring track set in the region of interest, and S is the set of monitoring tracks set in the region of interest.

[0029] Optionally, training the connection matrix of the chaotic ESN using the states of the nodes output by the pre-trained input ESN to obtain the trained chaotic ESN includes:

[0030] Inputting the state of the node output by the pre-trained input ESN into the chaotic ESN to obtain the monitoring position of the monitoring camera output by the chaotic ESN;

[0031] Determining a second loss function according to the monitoring position of the surveillance camera output by the chaotic ESN and the designated monitoring position;

[0032] The second loss function is used to optimize the connection matrix of the chaotic ESN.

[0033] Optionally, the chaotic ESN is expressed as:

[0034] Where t is time, τ is the proportional coefficient, and x ch is the state of the node in the chaotic ESN, g ch is the scaling parameter, J ch is the connection matrix of the chaotic ESN, J ic is the connection matrix between the input ESN and the chaotic ESN, x in is the status of the node in the input ESN.

[0035] Optionally, the second loss function is:

[0036] Among them, C 1-out is the second loss function, t is time, x s (t) includes the state of the node of the input ESN and the state of the node of the output ESN at time t, x s (t) = [x in ;x ch ], f s (t) is the designated monitoring position at time t, L out is the maximum value at time t, w outis the connection matrix between the chaotic ESN and the output layer, s is the monitoring trajectory set in the region of interest, and S is the set of monitoring trajectories set in the region of interest.

[0037] Optionally, the connection matrix of the feedback classifier is trained using the states of the nodes output by the pre-trained input ESN to obtain a trained feedback classifier, including:

[0038] Input the state of the node output by the pre-trained input ESN into the feedback classifier to obtain the monitoring position of the surveillance camera at the next moment output by the feedback classifier;

[0039] Determining a third loss function according to the monitoring position of the surveillance camera at the next moment output by the feedback classifier and the preset monitoring position of the surveillance camera at the next moment;

[0040] The connection matrix of the feedback classifier is optimized according to the third loss function.

[0041] Optionally, the connection matrix of the feedback classifier is trained using the states of the nodes output by the pre-trained input ESN to obtain a trained feedback classifier, including:

[0042] Input the states of the nodes output by the pre-trained input ESN and the transition probabilities between the preset regions of interest into the feedback classifier, and obtain the monitoring position of the surveillance camera at the next moment output by the feedback classifier;

[0043] Determining a fourth loss function according to the monitoring position of the surveillance camera at the next moment output by the feedback classifier and the preset monitoring position of the surveillance camera at the next moment;

[0044] The connection matrix of the feedback classifier is optimized according to the fourth loss function.

[0045] Optionally, the feedback classifier is expressed as:

[0046] Among them, f max is the feedback classifier, t is time, x(t) includes the state of the node of the input ESN and the state of the node of the output ESN at time t, x(t)=[x in ;x ch ], w is the connection matrix of the feedback classifier, s is the monitoring trajectory set in the region of interest, and S is the set of monitoring trajectories set in the region of interest.

[0047] Optionally, the third loss function is:

[0048] Wherein, C2 is the third loss function, t is time, s sper (t+Δt) is the monitoring position of the monitoring camera at the next moment output by the feedback classifier, x(t) is the state of the node of the input ESN at time t, w is the connection matrix of the feedback classifier, s is the monitoring trajectory set in the area of ​​interest, S is the set of monitoring trajectories set in the area of ​​interest, T rec The maximum value of time.

[0049] Optionally, the fourth loss function is:

[0050] Wherein, C3 is the fourth loss function, t is time, s sto (t+Δt) is the monitoring position of the surveillance camera at the next moment output by the feedback classifier according to the preset transition probability between the regions of interest, x(t) is the state of the node of the input ESN at time t, w is the connection matrix of the feedback classifier, s is the monitoring trajectory set in the region of interest, S is the set of monitoring trajectories set in the region of interest, T rec is the maximum value of time, and the transition probability is determined according to the importance of the region of interest.

[0051] In a second aspect, an embodiment of the present invention provides a monitoring trajectory determination method, including:

[0052] Get the current monitoring position of the surveillance camera;

[0053] Inputting the monitoring position at the current moment into a state echo network to obtain the monitoring position of the monitoring camera at the next moment output by the state echo network, the state echo network comprising: an input ESN, a chaotic ESN and a feedback classifier; the feedback classifier is used to realize autonomous switching between the regions of interest of the monitoring camera head;

[0054] The monitoring camera is controlled to rotate to obtain an image of the monitoring position at the next moment.

[0055] Optionally, the input ESN is expressed as:

[0056] Where t is time, τ is the proportional coefficient, and x in is the state of the node in the input ESN, g in is the scaling parameter, J in is the connection matrix of the input ESN, u in (s(t)) is the monitoring position at time t.

[0057] Optionally, the chaotic ESN is expressed as:

[0058] Where t is time, τ is the proportional coefficient, and x ch is the state of the node in the chaotic ESN, g ch is the scaling parameter, J ch is the connection matrix of the chaotic ESN, J ic is the connection matrix between the input ESN and the chaotic ESN, x in is the status of the node in the input ESN.

[0059] Optionally, the feedback classifier is expressed as:

[0060] Among them, f max is the feedback classifier, t is time, x(t) includes the state of the node of the input ESN and the state of the node of the output ESN at time t, x(t)=[x in ;x ch ], w is the connection matrix of the feedback classifier, s is the monitoring trajectory set in the region of interest, and S is the set of monitoring trajectories set in the region of interest.

[0061] Optionally, inputting the monitoring position at the current moment into the state echo network includes:

[0062] The transition probability between the monitoring position at the current moment and the region of interest of the monitoring camera is input into the state echo network.

[0063] In a third aspect, an embodiment of the present invention provides a training device for a state echo network, wherein the state echo network includes: an input ESN, a chaotic ESN, and a feedback classifier, and the device includes:

[0064] A first acquisition module is configured to acquire attractor sample data, wherein the attractor sample data includes location information of monitoring tracks set in a plurality of interest areas within a monitorable area of ​​a monitoring camera;

[0065] A first training module is configured to pre-train the connection matrix of the input ESN using the attractor sample data to obtain a pre-trained input ESN;

[0066] A second acquisition module is used to input the attractor sample data into the pre-trained input ESN to obtain the node status of the pre-trained input ESN;

[0067] A second training module is used to train the connection matrix of the chaotic ESN using the states of the nodes output by the pre-trained input ESN to obtain a trained chaotic ESN;

[0068] The third training module is used to train the connection matrix of the feedback classifier using the states of the nodes output by the pre-trained input ESN to obtain a trained feedback classifier, and the feedback classifier is used to achieve autonomous conversion between the regions of interest.

[0069] In a fourth aspect, an embodiment of the present invention provides a monitoring trajectory determination device, comprising:

[0070] An acquisition module, configured to acquire the current monitoring position of the monitoring camera;

[0071] A trajectory determination module is used to input the monitoring position at the current moment into a state echo network to obtain the monitoring position of the monitoring camera at the next moment output by the state echo network, wherein the state echo network includes: an input ESN, a chaotic ESN and a feedback classifier; the feedback classifier is used to realize autonomous switching between the regions of interest of the monitoring camera head;

[0072] The control module is used to control the monitoring camera to rotate to obtain the image of the monitoring position at the next moment.

[0073] In a fifth aspect, an embodiment of the present invention provides an electronic device, comprising: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, the steps of the state echo network training method as described in the first aspect are implemented, or when the program is executed by the processor, the steps of the monitoring trajectory determination method as described in the second aspect are implemented.

[0074] In a sixth aspect, an embodiment of the present invention provides a computer-readable storage medium, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the computer program implements the steps of the state echo network training method as described in the first aspect above; or, when the computer program is executed by the processor, the computer program implements the steps of the monitoring trajectory determination method as described in the second aspect above.

[0075] In an embodiment of the present invention, the position information of the monitoring trajectory set in the area of ​​interest in the monitorable area is used as attractor sample data to train a state echo network based on chaotic patrol. Because chaos is highly sensitive to initial values, the monitoring trajectory planned by the trained state echo network has random and unpredictable characteristics. Due to the existence of the attractor, the probability of the monitoring trajectory planned by the trained state echo network appearing near the attractor is high. Applying this state echo network to the field of security monitoring not only maintains the unplanned, random and unpredictable characteristics of chaos, but also enables the area of ​​interest to obtain a higher scanning frequency, making it impossible for criminals to effectively predict the monitoring trajectory, thereby improving the security of the security monitoring system. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0077] FIG1 is a schematic flow chart of a training method for a state echo network according to an embodiment of the present invention;

[0078] FIG2 is a schematic diagram of the structure of a state echo network according to an embodiment of the present invention;

[0079] FIG3 is a schematic diagram of a method for acquiring an image of an entire monitorable area according to an embodiment of the present invention;

[0080] FIG4 is a schematic diagram of a Markov matrix according to an embodiment of the present invention;

[0081] FIG5 is a schematic flow chart of a monitoring trajectory determination method according to an embodiment of the present invention;

[0082] FIG6 is a schematic structural diagram of a state echo network training device according to an embodiment of the present invention;

[0083] FIG7 is a schematic structural diagram of a monitoring trajectory determination device according to an embodiment of the present invention;

[0084] FIG8 is a schematic structural diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0085] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0086] The following is a brief introduction to the technical terms related to the present invention.

[0087] Chaos: Chaos refers to the presence of random, irregular motion within a deterministic system. Its behavior is uncertain, non-repeatable, and unpredictable. This is the chaotic phenomenon. Chaos is an inherent characteristic of nonlinear dynamical systems and a ubiquitous phenomenon in nonlinear systems.

[0088] Chaotic parade: It is a nonlinear phenomenon frequently observed in high-dimensional dynamical systems, characterized by chaotic parade transitions between local contraction domains (i.e., quasi-attractors).

[0089] Chaotic attractor: It has a complex stretched and twisted structure. The attractor is the product of the combined effect of the overall stability and local instability of the chaotic system. It has self-similarity and a fractal structure.

[0090] State Echo Network (ESN): is a type of recurrent neural network. The state echo network usually consists of an input layer, a storage layer (also called a reservoir, hidden layer, or hidden layer, etc.) and an output layer. Among them, the input layer’s input and the connection matrix w in The product obtained by multiplying (the connection matrix from the input layer to the storage layer) is input into the storage layer. The storage layer includes N nodes (also called neurons). Each node corresponds to a state, for example, it can be represented by x(t). The storage layer processes the input of the input layer to obtain the state vector of the storage layer. The output layer is the state vector output by the storage layer and the connection matrix w out (the connection matrix from the storage layer to the output layer) and output the product.

[0091] Referring to Figure 1, an embodiment of the present invention provides a method for training a state echo network. Referring to Figure 2, the state echo network includes an input ESN, a chaotic ESN, and a feedback classifier (not shown). The input ESN and the chaotic ESN constitute the storage layer of the state echo network, and the input ESN serves as the interface between the input layer of the state echo network and the chaotic ESN. Each of the input ESN and the chaotic ESN contains multiple nodes.

[0092] The training method of the state echo network includes:

[0093] Step 11: Obtaining attractor sample data, wherein the attractor sample data includes location information of monitoring tracks set in multiple regions of interest (ROIs) within the monitorable area of ​​the monitoring camera;

[0094] Step 12: pre-training the connection matrix of the input ESN using the attractor sample data to obtain a pre-trained input ESN;

[0095] Step 13: Input the attractor sample data into the pre-trained input ESN to obtain the node status of the pre-trained input ESN;

[0096] Step 14: using the states of the nodes output by the pre-trained input ESN to train the connection matrix of the chaotic ESN to obtain a trained chaotic ESN;

[0097] Step 15: Using the states of the nodes output by the pre-trained input ESN to train the connection matrix of the feedback classifier, a trained feedback classifier is obtained. The feedback classifier is used to achieve autonomous conversion between the regions of interest.

[0098] Among them, the pre-trained input ESN, the trained chaotic ESN and the trained feedback classifier constitute the trained state echo network.

[0099] In an embodiment of the present invention, the position information of the monitoring trajectory set in the area of ​​interest in the monitorable area is used as attractor sample data to train a state echo network based on chaotic patrol. Because chaos is highly sensitive to initial values, the monitoring trajectory planned by the trained state echo network has random and unpredictable characteristics. Due to the existence of the attractor, the probability of the monitoring trajectory planned by the trained state echo network appearing near the attractor is high. Applying this state echo network to the field of security monitoring not only maintains the unplanned, random and unpredictable characteristics of chaos, but also enables the area of ​​interest to obtain a higher scanning frequency, making it impossible for criminals to effectively predict the monitoring trajectory, thereby improving the security of the security monitoring system.

[0100] In the embodiment of the present invention, optionally, obtaining attractor sample data includes:

[0101] Step 111: Control the surveillance camera to collect images of the monitorable area using a traversal method;

[0102] Optionally, the monitorable area of ​​the surveillance camera is determined according to the movable range of the surveillance camera, and images of the monitorable area are collected in sequence using a traversal method.

[0103] Step 112: stitch all the collected images of the monitorable area to obtain an overall monitorable area image; please refer to FIG3 .

[0104] Step 113: determining an area of ​​interest in the overall monitorable area image;

[0105] In the embodiment of the present invention, the region of interest may be selected manually.

[0106] Step 114: Obtaining a monitoring track set in the region of interest;

[0107] In the embodiment of the present invention, the monitoring trajectory within the region of interest may be set manually or automatically and randomly.

[0108] In the embodiment of the present invention, different types of monitoring tracks can be set according to the size of the region of interest.

[0109] The monitoring track may be a circle, a square, or an irregular line (such as a Lisa curve).

[0110] Step 115: Using the location information of the monitoring track set in the region of interest as the attractor sample data.

[0111] In an embodiment of the present invention, optionally, the pre-training the connection matrix of the input ESN using the attractor sample data to obtain the pre-trained input ESN includes:

[0112] Step 121: Input the attractor sample data into the input ESN, and obtain the node status of the input ESN output by the input ESN;

[0113] Step 122: determining a first loss function according to the state of the node output by the input ESN and a preset target state of the node of the input ESN;

[0114] Step 123: Optimize the connection matrix of the input ESN using the first loss function.

[0115] In an embodiment of the present invention, optionally, the nodes of the input ESN participating in the training are half of all the nodes of the input ESN to reduce the redundancy of the connection matrix of the input ESN. Of course, in some other embodiments of the present invention, the nodes of the input ESN participating in the training may also be all the nodes of the input ESN.

[0116] In the embodiment of the present invention, optionally, the input ESN (having N in nodes) is represented as:

[0117] Where t is time, τ is the proportional coefficient, which can affect the update rate, x in is the state of the node in the input ESN, g in is the scaling parameter (i.e., the connection matrix J in Scaling, i.e. normalization operation), J in is the connection matrix of the input ESN, J in The elements of u follow the normal distribution, in (s) is the attractor sample data. tanh() is the hyperbolic tangent function, which is an activation function.

[0118] In this embodiment of the present invention, optionally, the first loss function is:

[0119] Among them, C 1-in is the first loss function, t is time, x s (t) is the state of the node of the input ESN at time t, is the target state of the node of the input ESN at time t, L innate is the maximum value at time t, s is the monitoring track set in the region of interest, and S is the set of monitoring tracks set in the region of interest.

[0120] Of course, in some other embodiments of the present invention, the first loss function may also adopt other loss functions, and the present invention is not limited thereto.

[0121] In an embodiment of the present invention, optionally, the connection matrix of the chaotic ESN is trained using the states of the nodes output by the pre-trained input ESN to obtain the trained chaotic ESN, including:

[0122] Step 141: inputting the node status output by the pre-trained input ESN into the chaotic ESN to obtain the monitoring position of the monitoring camera output by the chaotic ESN;

[0123] Step 142: determining a second loss function according to the monitoring position of the monitoring camera output by the chaotic ESN and the designated monitoring position;

[0124] Step 143: Optimize the connection matrix of the chaotic ESN using the second loss function. In the embodiment of the present invention, optionally, the chaotic ESN (with N ch nodes) is represented as:

[0125] Where t is time, τ is the proportional coefficient, and x ch is the state of the node in the chaotic ESN, g ch is the scaling parameter (i.e., the connection matrix J ch Scaling, i.e. normalization operation), J ch is the connection matrix of the chaotic ESN, J ic is the connection matrix between the input ESN and the chaotic ESN, x in is the status of the node in the input ESN.

[0126] In some embodiments of the present invention, τ = 10.0, g in =0.9,g ch=1.5 makes the input ESN non-chaotic and the chaotic ESN chaotic.

[0127] In this embodiment of the present invention, optionally, the second loss function is:

[0128] Among them, C 1-out is the second loss function, t is time, x s (t) includes the state of the node of the input ESN and the state of the node of the output ESN at time t, x s (t) = [x in ;x ch ], f s (t) is the designated monitoring position at time t, L out is the maximum value at time t, w out is the connection matrix between the chaotic ESN and the output layer, s is the monitoring trajectory set in the region of interest, and S is the set of monitoring trajectories set in the region of interest.

[0129] Of course, in some other embodiments of the present invention, the second loss function may also adopt other loss functions, and the present invention is not limited thereto.

[0130] In embodiments of the present invention, the input ESN and / or output ESN training process can utilize a first-order-reduced and controlled-error learning method. This training method does not involve matrix inversion or extensive convolution calculations, thus optimizing the computational approach and increasing computational speed.

[0131] In the embodiment of the present invention, in order to prevent the chaotic ESN from changing to a non-chaotic state, during the training process, the link between the input ESN and the chaotic ESN is trained to be transiently stable with a convergence to zero.

[0132] In order to realize the navigation transformation between attractors, a feedback link is set between the storage layer and the input layer, that is, a feedback classifier is added to the trained state echo network (the actual output around the set attractor trajectory is fed back to the input link to realize state jump. For example, the output of the storage layer is from state A to state B, to state C, and feedback is added to allow ABC to proceed in sequence) to autonomously generate specific symbolic dynamics.

[0133] In some embodiments of the present invention, optionally, the connection matrix of the feedback classifier is trained using the states of the nodes output by the pre-trained input ESN to obtain the trained feedback classifier, including:

[0134] Step 151: inputting the state of the node output by the pre-trained input ESN into the feedback classifier, and obtaining the monitoring position of the monitoring camera at the next moment output by the feedback classifier;

[0135] Step 152: determining a third loss function according to the monitoring position of the monitoring camera at the next moment output by the feedback classifier and the preset monitoring position of the monitoring camera at the next moment;

[0136] Step 153: Optimize the connection matrix of the feedback classifier according to the third loss function.

[0137] In the embodiment of the present invention, optionally, the feedback classifier is expressed as:

[0138] Among them, f max is the feedback classifier, t is time, x(t) includes the state of the node of the input ESN and the state of the node of the output ESN at time t, x(t)=[x in ;x ch ], w is the connection matrix of the feedback classifier, s is the monitoring trajectory set in the region of interest, and S is the set of monitoring trajectories set in the region of interest.

[0139] In this embodiment of the present invention, optionally, the third loss function is:

[0140] Wherein, C2 is the third loss function, t is time, s sper (t+Δt) is the monitoring position of the monitoring camera at the next moment output by the feedback classifier, x(t) is the state of the node of the input ESN at time t, w is the connection matrix of the feedback classifier, s is the monitoring trajectory set in the area of ​​interest, S is the set of monitoring trajectories set in the area of ​​interest, T rec The maximum value of time.

[0141] Among them, in the formula, 1{s sper (t+Δt)=s} means: when s sper (t+Δt)=s, which is 1 in this case and 0 otherwise.

[0142] In some other embodiments of the present invention, optionally, the connection matrix of the feedback classifier is trained using the states of the nodes output by the pre-trained input ESN to obtain the trained feedback classifier, including:

[0143] Step 151 ′: inputting the node status output by the pre-trained input ESN and the transition probability between the preset regions of interest into the feedback classifier, and obtaining the monitoring position of the surveillance camera at the next moment output by the feedback classifier;

[0144] Step 152 ′: determining a fourth loss function according to the monitoring position of the monitoring camera at the next moment output by the feedback classifier and the preset monitoring position of the monitoring camera at the next moment;

[0145] Step 153 ′: Optimize the connection matrix of the feedback classifier according to the fourth loss function.

[0146] In the embodiment of the present invention, optionally, the feedback classifier is expressed as:

[0147] Among them, f max is the feedback classifier, t is time, x(t) includes the state of the node of the input ESN and the state of the node of the output ESN at time t, x(t)=[x in ;x ch ], w is the connection matrix of the feedback classifier, s is the monitoring trajectory set in the region of interest, and S is the set of monitoring trajectories set in the region of interest.

[0148] In this embodiment of the present invention, optionally, the fourth loss function is:

[0149] Wherein, C3 is the fourth loss function, t is time, s sto (t+Δt) is the monitoring position of the surveillance camera at the next moment output by the feedback classifier according to the preset transition probability between the regions of interest, x(t) is the state of the node of the input ESN at time t, w is the connection matrix of the feedback classifier, s is the monitoring trajectory set in the region of interest, S is the set of monitoring trajectories set in the region of interest, T rec is the maximum value of time, and the transition probability is determined according to the importance of the region of interest.

[0150] In the formula of the fourth loss function described above, 1{s sto (t+Δt)=s} means: when s sto (t+Δt)=s, which is 1, otherwise it is 0.

[0151] In the embodiment of the present invention, the transition probability between regions of interest includes: the transition probability between different regions of interest and the transition probability between regions of interest themselves.

[0152] In this embodiment of the present invention, the transition probability between ROIs can be determined based on the importance of the ROI. Referring to Figure 4, assuming that the importance of ROI1 is 20, the importance of ROI2 is 50, and the importance of ROI3 is 30, the transition probabilities between the various ROIs are shown in Figure 4. These transition probabilities form a Markov matrix. In this embodiment of the present invention, the importance of a ROI can be determined based on its danger level. For example, if a particular ROI is crowded, the attention it receives should be increased. A higher danger level increases the probability of transitioning from other ROIs to that ROI.

[0153] Referring to FIG5 , an embodiment of the present invention further provides a monitoring trajectory determination method, including:

[0154] Step 51: Obtain the current monitoring position of the surveillance camera;

[0155] Step 52: Input the monitoring position at the current moment into a state echo network to obtain the monitoring position of the monitoring camera at the next moment output by the state echo network, wherein the state echo network includes: an input ESN, a chaotic ESN, and a feedback classifier; the feedback classifier is used to realize autonomous switching between the regions of interest of the monitoring camera head;

[0156] Step 53: Control the monitoring camera to rotate to obtain an image of the monitoring position at the next moment.

[0157] In the embodiment of the present invention, optionally, the input ESN is expressed as:

[0158] Where t is time, τ is the proportional coefficient, and x in is the state of the node in the input ESN, g in is the scaling parameter, J in is the connection matrix of the input ESN, u in (s(t)) is the monitoring position at time t.

[0159] In the embodiment of the present invention, optionally, the chaotic ESN is expressed as:

[0160] Where t is time, τ is the proportional coefficient, and x ch is the state of the node in the chaotic ESN, g ch is the scaling parameter, J ch is the connection matrix of the chaotic ESN, J ic is the connection matrix between the input ESN and the chaotic ESN, x in is the status of the node in the input ESN.

[0161] In the embodiment of the present invention, optionally, the feedback classifier is expressed as:

[0162] Among them, f max is the feedback classifier, t is time, x(t) includes the state of the node of the input ESN and the state of the node of the output ESN at time t, x(t)=[x in ;x ch ], w is the connection matrix of the feedback classifier, s is the monitoring trajectory set in the region of interest, and S is the set of monitoring trajectories set in the region of interest.

[0163] In the embodiment of the present invention, optionally, inputting the monitoring position at the current moment into the state echo network includes: inputting the transition probability between the monitoring position at the current moment and the region of interest of the monitoring camera into the state echo network.

[0164] 6 , an embodiment of the present invention further provides a state echo network training device 60, wherein the state echo network includes: an input ESN, a chaotic ESN, and a feedback classifier, and the device 60 includes:

[0165] A first acquisition module 61 is configured to acquire attractor sample data, wherein the attractor sample data includes location information of monitoring tracks set in a plurality of interest areas within a monitorable area of ​​a monitoring camera;

[0166] A first training module 62 is configured to pre-train the connection matrix of the input ESN using the attractor sample data to obtain a pre-trained input ESN;

[0167] A second acquisition module 63 is configured to input the attractor sample data into the pre-trained input ESN to obtain the node status of the pre-trained input ESN;

[0168] A second training module 64 is configured to train the connection matrix of the chaotic ESN using the states of the nodes output by the pre-trained input ESN to obtain a trained chaotic ESN;

[0169] The third training module 65 is used to train the connection matrix of the feedback classifier using the states of the nodes output by the pre-trained input ESN to obtain a trained feedback classifier, and the feedback classifier is used to realize the autonomous conversion between the regions of interest.

[0170] In an embodiment of the present invention, optionally, the first acquisition module 61 is used to control the monitoring camera to collect images of the monitorable area using a traversal method; perform image stitching on all collected images of the monitorable area to obtain an overall monitorable area image; determine an area of ​​interest in the overall monitorable area image; obtain a monitoring track set in the area of ​​interest; and use the location information of the monitoring track set in the area of ​​interest as the attractor sample data.

[0171] In an embodiment of the present invention, optionally, the first training module 62 is used to input the attractor sample data into the input ESN to obtain the state of the node of the input ESN output by the input ESN; determine a first loss function based on the state of the node output by the input ESN and the preset target state of the node of the input ESN; and use the first loss function to optimize the connection matrix of the input ESN.

[0172] In the embodiment of the present invention, optionally, the nodes of the input ESN participating in the training are half of all the nodes of the input ESN.

[0173] In the embodiment of the present invention, optionally, the input ESN is expressed as:

[0174] Where t is time, τ is the proportional coefficient, and x in is the state of the node in the input ESN, g in is the scaling parameter, J in is the connection matrix of the input ESN, u in (s) is the attractor sample data.

[0175] In this embodiment of the present invention, optionally, the first loss function is:

[0176] Among them, C 1-in is the first loss function, t is time, x s (t) is the state of the node of the input ESN at time t, is the target state of the node of the input ESN at time t, L innate is the maximum value at time t, s is the monitoring track set in the region of interest, and S is the set of monitoring tracks set in the region of interest.

[0177] In an embodiment of the present invention, optionally, the second training module 64 is used to input the state of the node output by the pre-trained input ESN into the chaotic ESN to obtain the monitoring position of the monitoring camera output by the chaotic ESN; determine the second loss function based on the monitoring position of the monitoring camera output by the chaotic ESN and the specified monitoring position; and use the second loss function to optimize the connection matrix of the chaotic ESN.

[0178] In the embodiment of the present invention, optionally, the chaotic ESN is expressed as:

[0179] Where t is time, τ is the proportional coefficient, and x ch is the state of the node in the chaotic ESN, g ch is the scaling parameter, J ch is the connection matrix of the chaotic ESN, J in is the connection matrix between the input ESN and the chaotic ESN, x in is the status of the node in the input ESN.

[0180] In this embodiment of the present invention, optionally, the second loss function is:

[0181] Among them, C 1-out is the second loss function, t is time, x s (t) includes the state of the node of the input ESN and the state of the node of the output ESN at time t, x s (t) = [x in ;x ch ], f s (t) is the designated monitoring position at time t, L out is the maximum value at time t, w out is the connection matrix between the chaotic ESN and the output layer, s is the monitoring trajectory set in the region of interest, and S is the set of monitoring trajectories set in the region of interest.

[0182] In an embodiment of the present invention, optionally, the third training module 65 is used to input the state of the node output by the pre-trained input ESN into the feedback classifier to obtain the monitoring position of the monitoring camera at the next moment output by the feedback classifier; determine the third loss function based on the monitoring position of the monitoring camera at the next moment output by the feedback classifier and the preset monitoring position of the monitoring camera at the next moment; and optimize the connection matrix of the feedback classifier according to the third loss function.

[0183] In an embodiment of the present invention, optionally, the third training module 65 is used to input the status of the node output by the pre-trained input ESN and the transition probability between the preset regions of interest into the feedback classifier to obtain the monitoring position of the monitoring camera at the next moment output by the feedback classifier; determine the fourth loss function based on the monitoring position of the monitoring camera at the next moment output by the feedback classifier and the preset monitoring position of the monitoring camera at the next moment; and optimize the connection matrix of the feedback classifier according to the fourth loss function.

[0184] In the embodiment of the present invention, optionally, the feedback classifier is expressed as:

[0185] Among them, f max is the feedback classifier, t is time, x(t) includes the state of the node of the input ESN and the state of the node of the output ESN at time t, x(t)=[x in ;x ch ], w is the connection matrix of the feedback classifier, s is the monitoring trajectory set in the region of interest, and S is the set of monitoring trajectories set in the region of interest.

[0186] In this embodiment of the present invention, optionally, the third loss function is:

[0187] Wherein, C2 is the third loss function, t is time, s sper (t+Δt) is the monitoring position of the monitoring camera at the next moment output by the feedback classifier, x(t) is the state of the node of the input ESN at time t, w is the connection matrix of the feedback classifier, s is the monitoring trajectory set in the area of ​​interest, S is the set of monitoring trajectories set in the area of ​​interest, T rec The maximum value of time.

[0188] Among them, in the formula of the third loss function above, 1{s sper (t+Δt)=s} means: when s sper (t+Δt)=s, which is 1 in this case and 0 otherwise.

[0189] In this embodiment of the present invention, optionally, the fourth loss function is:

[0190] Wherein, C3 is the fourth loss function, t is time, s sto(t+Δt) is the monitoring position of the surveillance camera at the next moment output by the feedback classifier according to the preset transition probability between the regions of interest, x(t) is the state of the node of the input ESN at time t, w is the connection matrix of the feedback classifier, s is the monitoring trajectory set in the region of interest, S is the set of monitoring trajectories set in the region of interest, T rec is the maximum value of time, and the transition probability is determined according to the importance of the region of interest.

[0191] In the formula of the fourth loss function described above, 1{s sto (t+Δt)=s} means: when s sto (t+Δt)=s, which is 1, otherwise it is 0.

[0192] Referring to FIG. 7 , an embodiment of the present invention further provides a monitoring trajectory determination device 70 , comprising:

[0193] An acquisition module 71 is used to obtain the monitoring position of the monitoring camera at the current moment;

[0194] A trajectory determination module 72 is configured to input the current monitoring position into a state echo network to obtain the monitoring position of the surveillance camera at the next moment output by the state echo network. The state echo network includes: an input ESN, a chaotic ESN, and a feedback classifier; the feedback classifier is configured to achieve autonomous switching between regions of interest of the surveillance camera head;

[0195] The control module 73 is used to control the monitoring camera to rotate to obtain the image of the monitoring position at the next moment.

[0196] Optionally, the input ESN is expressed as:

[0197] Where t is time, τ is the proportional coefficient, and x in is the state of the node in the input ESN, g in is the scaling parameter, J in is the connection matrix of the input ESN, u in (s(t)) is the monitoring position at time t.

[0198] Optionally, the chaotic ESN is expressed as:

[0199] Where t is time, τ is the proportional coefficient, and x ch is the state of the node in the chaotic ESN, g ch is the scaling parameter, J ch is the connection matrix of the chaotic ESN, J icis the connection matrix between the input ESN and the chaotic ESN, x in is the status of the node in the input ESN.

[0200] Optionally, the feedback classifier is expressed as:

[0201] Among them, f max is the feedback classifier, t is time, x(t) includes the state of the node of the input ESN and the state of the node of the output ESN at time t, x(t)=[x in ;x ch ], w is the connection matrix of the feedback classifier, s is the monitoring trajectory set in the region of interest, and S is the set of monitoring trajectories set in the region of interest.

[0202] Optionally, the trajectory determination module 72 is configured to input the transition probability between the monitoring position at the current moment and the region of interest of the monitoring camera into a state echo network.

[0203] Referring to Figure 8 , an embodiment of the present invention further provides an electronic device 80, including a processor 81, a memory 82, and a computer program stored in the memory 82 and executable on the processor 81. When executed by the processor 81, the computer program implements the various processes of the above-mentioned state echo network training method or monitoring trajectory determination method embodiment, and can achieve the same technical effects. To avoid repetition, they are not further described here.

[0204] An embodiment of the present invention further provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the various processes of the aforementioned state echo network training method or monitoring trajectory determination method embodiments, achieving the same technical effects. To avoid repetition, the details are omitted here. The computer-readable storage medium may be, for example, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0205] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0206] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present invention.

[0207] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are protected by the present invention.

Claims

1. A training method for a state echo network, the state echo network comprising: Inputting ESN, chaotic ESN and feedback classifier, the method comprises: Acquire attractor sample data, wherein the attractor sample data includes location information of monitoring tracks set in a plurality of interest areas within a monitorable area of ​​a monitoring camera; Pre-training the connection matrix of the input ESN using the attractor sample data to obtain a pre-trained input ESN; Inputting the attractor sample data into the pre-trained input ESN to obtain the node status of the pre-trained input ESN; Using the states of the nodes output by the pre-trained input ESN to train the connection matrix of the chaotic ESN to obtain a trained chaotic ESN; The connection matrix of the feedback classifier is trained using the states of the nodes output by the pre-trained input ESN to obtain a trained feedback classifier, and the feedback classifier is used to realize the autonomous conversion between the regions of interest.

2. The method according to claim 1, wherein: The obtaining of attractor sample data comprises: Control the monitoring camera to collect images of the monitorable area using the traversal method; All the collected images of the monitorable area are stitched together to obtain the overall monitorable area image; Determining a region of interest in the overall monitorable area image; Acquire the monitoring track set in the area of ​​interest; The position information of the monitoring track set in the region of interest is used as the attractor sample data.

3. The method according to claim 1, wherein: The method of pre-training the connection matrix of the input ESN using the attractor sample data to obtain the pre-trained input ESN includes: Input the attractor sample data into the input ESN to obtain the state of the node of the input ESN output by the input ESN; Determine a first loss function according to the state of the node output by the input ESN and the preset target state of the node of the input ESN; The connection matrix of the input ESN is optimized using the first loss function.

4. The method according to claim 1 or 3, wherein: The nodes of the input ESN participating in the training are half of all the nodes of the input ESN.

5. The method according to claim 1, wherein: The input ESN is represented as: Where t is time, τ is the proportionality coefficient, and x in is the state of the node in the input ESN, g in is the scaling parameter, J in is the connection matrix of the input ESN, u in (s) is the attractor sample data.

6. The method according to claim 3, wherein: The first loss function is: Among them, C 1-in is the first loss function, t is time, x s (t) is the state of the node of the input ESN at time t, is the target state of the node of the input ESN at time t, L innate is the maximum value at time t, s is the monitoring track set in the region of interest, and S is the set of monitoring tracks set in the region of interest.

7. The method according to claim 1, wherein: The connection matrix of the chaotic ESN is trained using the state of the node output by the pre-trained input ESN to obtain the trained chaotic ESN, including: Inputting the state of the node output by the pre-trained input ESN into the chaotic ESN to obtain the monitoring position of the monitoring camera output by the chaotic ESN; Determine a second loss function according to the monitoring position of the monitoring camera output by the chaotic ESN and the designated monitoring position; The second loss function is used to optimize the connection matrix of the chaotic ESN.

8. The method according to claim 1 or 7, wherein: The chaotic ESN is expressed as: Where t is time, τ is the proportionality coefficient, and x ch is the state of the node in the chaotic ESN, g ch is the scaling parameter, J ch is the connection matrix of the chaotic ESN, J ic is the connection matrix between the input ESN and the chaotic ESN, x in is the status of the node in the input ESN.

9. The method according to claim 7, wherein: The second loss function is: Among them, C 1-out is the second loss function, t is time, x s (t) includes the state of the node of the input ESN and the state of the node of the output ESN at time t, x s (t) = [x in ;x ch ], f s (t) is the designated monitoring position at time t, L out is the maximum value at time t, w out is the connection matrix between the chaotic ESN and the output layer, s is the monitoring track set in the region of interest, and S is the set of monitoring tracks set in the region of interest.

10. The method according to claim 1, wherein: The connection matrix of the feedback classifier is trained using the state of the node output by the pre-trained input ESN to obtain the trained feedback classifier, including: Input the state of the node output by the pre-trained input ESN into the feedback classifier to obtain the monitoring position of the monitoring camera at the next moment output by the feedback classifier; Determine a third loss function according to the monitoring position of the monitoring camera at the next moment output by the feedback classifier and the preset monitoring position of the monitoring camera at the next moment; The connection matrix of the feedback classifier is optimized according to the third loss function.

11. The method according to claim 1, wherein: The connection matrix of the feedback classifier is trained using the state of the node output by the pre-trained input ESN to obtain the trained feedback classifier, including: Input the states of the nodes output by the pre-trained input ESN and the transition probabilities between the preset regions of interest into the feedback classifier, and obtain the monitoring position of the monitoring camera at the next moment output by the feedback classifier; Determine a fourth loss function according to the monitoring position of the monitoring camera at the next moment output by the feedback classifier and the preset monitoring position of the monitoring camera at the next moment; The connection matrix of the feedback classifier is optimized according to the fourth loss function.

12. The method according to claim 1, 10 or 11, wherein: The feedback classifier is expressed as: Among them, f max is the feedback classifier, t is time, x(t) includes the state of the node of the input ESN and the state of the node of the output ESN at time t, x(t)=[x in ;x ch ], w is the feedback classification The connection matrix of the device, s is the monitoring track set in the area of ​​interest, and S is the set of monitoring tracks set in the area of ​​interest.

13. The method according to claim 10, wherein: The third loss function is: Wherein, C2 is the third loss function, t is time, s sper (t+Δt) is the monitoring position of the monitoring camera at the next moment output by the feedback classifier, x(t) is the state of the node of the input ESN at time t, w is the connection matrix of the feedback classifier, s is the monitoring trajectory set in the area of ​​interest, S is the set of monitoring trajectories set in the area of ​​interest, T rec is the maximum value of time.

14. The method according to claim 11, wherein: The fourth loss function is: Wherein, C3 is the fourth loss function, t is time, s sto (t+Δt) is the monitoring position of the monitoring camera at the next moment output by the feedback classifier according to the preset transition probability between the regions of interest, x(t) is the state of the node of the input ESN at time t, w is the connection matrix of the feedback classifier, s is the monitoring trajectory set in the region of interest, S is the set of monitoring trajectories set in the region of interest, T rec is the maximum value of time, and the conversion probability is determined according to the importance of the region of interest.

15. A monitoring trajectory determination method, comprising: Get the current monitoring position of the surveillance camera; The monitoring position at the current moment is input into the state echo network to obtain the monitoring position of the monitoring camera at the next moment output by the state echo network, wherein the state echo network includes: an input ESN, a chaotic ESN and a feedback classifier; the feedback classifier is used to realize the autonomous conversion between the regions of interest of the monitoring camera head; The monitoring camera is controlled to rotate to obtain an image of the monitoring position at the next moment.

16. The method according to claim 15, wherein: The input ESN is represented as: Where t is time, τ is the proportionality coefficient, and x in is the state of the node in the input ESN, g in for Scaling parameter, J in is the connection matrix of the input ESN, u in (s(t)) is the monitoring position at time t.

17. The method according to claim 15, wherein: The chaotic ESN is expressed as: Where t is time, τ is the proportionality coefficient, and x ch is the state of the node in the chaotic ESN, g ch is the scaling parameter, J ch is the connection matrix of the chaotic ESN, J ic is the connection matrix between the input ESN and the chaotic ESN, x in is the status of the node in the input ESN.

18. The method according to claim 15, wherein: The feedback classifier is expressed as: Among them, f max is the feedback classifier, t is time, x(t) includes the state of the node of the input ESN and the state of the node of the output ESN at time t, x(t)=[x in ;x ch ], w is the connection matrix of the feedback classifier, s is the monitoring trajectory set in the region of interest, and S is the set of monitoring trajectories set in the region of interest.

19. The method according to claim 15, wherein: Inputting the monitoring position at the current moment into the state echo network includes: The transition probability between the monitoring position at the current moment and the region of interest of the monitoring camera is input into the state echo network.

20. A training device for a state echo network, the state echo network comprising: Input ESN, chaotic ESN and feedback classifier, the device comprises: A first acquisition module is used to acquire attractor sample data, wherein the attractor sample data includes location information of monitoring tracks set in multiple interest areas within a monitorable area of ​​a monitoring camera; A first training module is used to pre-train the connection matrix of the input ESN using the attractor sample data to obtain a pre-trained input ESN; A second acquisition module is used to input the attractor sample data into the pre-trained input ESN to obtain the node status of the pre-trained input ESN; A second training module is used to train the connection matrix of the chaotic ESN using the states of the nodes output by the pre-trained input ESN to obtain a trained chaotic ESN; The third training module is used to train the connection matrix of the feedback classifier using the state of the node output by the pre-trained input ESN to obtain the trained feedback classifier. The device is used to realize autonomous switching between the regions of interest.

21. A monitoring trajectory determination device, comprising: An acquisition module, used to acquire the monitoring position of the monitoring camera at the current moment; A trajectory determination module is used to input the monitoring position at the current moment into a state echo network to obtain the monitoring position of the monitoring camera at the next moment output by the state echo network, wherein the state echo network includes: an input ESN, a chaotic ESN and a feedback classifier; the feedback classifier is used to realize the autonomous conversion between the regions of interest of the monitoring camera head; The control module is used to control the monitoring camera to rotate to obtain the image of the monitoring position at the next moment.

22. An electronic device, comprising: A processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, the program implements the steps of the state echo network training method as described in any one of claims 1 to 14, or, when the program is executed by the processor, the program implements the steps of the monitoring trajectory determination method as described in any one of claims 15 to 19.

23. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the computer program implements the steps of the state echo network training method as described in any one of claims 1 to 14; or, when the computer program is executed by a processor, the computer program implements the steps of the monitoring trajectory determination method as described in any one of claims 15 to 19.

Images