Large-scale space brain-like navigation method based on high-dimensional cognitive coding

Abstract

The invention discloses a large-scale space brain-like navigation method based on high-dimensional cognitive coding, which comprises the following steps: constructing grid cells by using an oscillation interference network with negative feedback, and carrying out path integration on three-axis speed of self-motion information to realize high-frequency rough estimation of the position of an unmanned system; a continuous attractor neural network is utilized to construct position cells, a large-scale space is efficiently represented through a cyclic mapping mechanism, and self-adaptive fusion of self-motion information and environment information is achieved; and the cognitive map is quickly decoded by using an interest iteration mechanism, so that the memory and reasoning functions of the unmanned system on the environment information are realized. According to the invention, self-motion information provided by an inertial navigation system and external environment information provided by a global navigation satellite system and a camera are fused, and an efficient and reliable large-scale space brain-like navigation method is provided for an unmanned system.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent navigation, and particularly relates to a large-scale spatial brain-like navigation method based on high-dimensional cognitive coding. Background Art

[0002] The applications of unmanned systems are rapidly expanding to various fields and have become an important driving force for the development of modern technology. With the continuous progress of technology, unmanned systems are widely used in multiple scenarios such as environmental monitoring, military reconnaissance, search and rescue, etc. These applications usually require high-precision and high-reliability navigation and positioning capabilities. Especially in large-scale spaces, unmanned systems face complex and changeable environments and geographical features, so the reliability and accuracy of their navigation technologies have become core challenges.

[0003] Existing multi-source information fusion methods mainly include filtering and optimization techniques. Filtering-based methods rely on covariance matrices for state estimation, which assume first-order Markov properties and Gaussian noise, resulting in performance degradation in complex environments. Optimization-based methods are prone to falling into local optimal solutions and have high computational complexity, making it difficult to meet the real-time application requirements in large-scale complex scenarios. When the sensors are interfered, the accuracy of error modeling will decrease in the above methods, thus directly affecting the accuracy of the fusion results.

[0004] In contrast, mammals such as bats exhibit stable and reliable navigation capabilities in large-scale complex environments. Their navigation does not rely on external precise modeling, but rather realizes navigation through high-dimensional semantic cognition of the external environment and adaptive fusion of multi-modal information. In particular, grid cells, place cells, etc. in the hippocampus-entorhinal cortex loop play a core role in spatial navigation and memory. Through the collaborative action of these cells, a cognitive map can be constructed through self-motion perception and environmental cognition to achieve reliable and stable navigation.

[0005] Therefore, researching brain-like navigation methods based on high-dimensional cognitive coding to improve the robustness, intelligence, and autonomy of unmanned system navigation in large-scale complex spaces has important scientific and application values. Summary of the Invention

[0006] Object of the Invention: The present invention proposes a large-scale spatial brain-like navigation method based on high-dimensional cognitive coding, which solves the problems that existing navigation methods rely on precise mathematical models in large-scale spaces and cannot fully utilize high-dimensional cognitive information in the externally perceived environment, and improves the reliability, stability, and intelligence of the navigation system of unmanned systems in large-scale complex environments.

[0007] Technical Solution: A large-scale spatial brain-like navigation method based on high-dimensional cognitive coding described in the present invention specifically includes the following steps:

[0008] (1) Obtain self-motion information and external perception information based on an unmanned system;

[0009] (2) Construct a cognitive map model, a grid cell model, and a place cell model; the cognitive map is used to provide an internal representation of external cognitive information; grid cells are used to simulate the path integration function in the brain; place cells are used to fuse self-motion information and external perception information;

[0010] (3) Extract semantic pointers from the images in the external perception information and perform similarity measurement with the cognitive map; if a similar scene is found, execute step (4); otherwise, execute step (5);

[0011] (4) Perform interest iterative cognitive map decoding on the spatial semantic pointers of the similar scene in step (3), output the positioning result of the unmanned system, and at the same time update the initial value of grid cell path integration, and execute step (8);

[0012] (5) Use grid cells to encode the three-axis velocity in the self-motion information, achieve path integration through active iterative update, and obtain a relatively rough high-frequency position estimate of the unmanned system by decoding the firing results of the grid cell model;

[0013] (6) Use place cells to encode, actively update, and decode the estimated information decoded in step (5) and the position information provided in the external perception information, output the position information of the unmanned system after fusing multi-source information, bind it to the spatial semantic pointer extracted in step (3), and memorize it in the cognitive map;

[0014] (7) If the wave packet moves close to the edge of the place cell representation range, use the loop mapping mechanism to remap the place cells;

[0015] (8) Return to step (3) and loop until navigation ends.

[0016] Furthermore, the self-motion information in step (1) includes three-axis velocity information provided by an inertial sensor; the external perception information includes images provided by a camera, visual odometer information, and position information provided by a global navigation satellite system.

[0017] Furthermore, the implementation process of constructing the cognitive map model in step (2) is as follows:

[0018] To achieve a high-dimensional representation of continuous spatial information, the mapping relationship from the Euclidean space position vector to the cognitive map This mapping is encoded in the following way:

[0019] Ω(x)=f -1 {e iAx}

[0020] Where x is the position vector in Euclidean space, is the encoding matrix, f -1 is the inverse Fourier transform. In order to ensure that the vector Ω in the cognitive map is real-valued, it is necessary to select such that e iAx A matrix with conjugate symmetry:

[0021]

[0022] in, Indicates a bundling operation. Represents the Hadamard product; through one or more binding operations, the continuous movement of the unmanned system in space is achieved.

[0023] Furthermore, the implementation process of the grid cell model in step (2) is as follows:

[0024] By using the velocity information provided by the motion information in the Euclidean space Transformed into the derivative of spatial semantic pointers in cognitive maps with respect to time

[0025]

[0026] The transformation matrix A is composed of multiple sets of row vectors, and each set of vectors k must satisfy:

[0027]

[0028] Through the mutual oscillation and interference of multiple groups of vectors with a phase difference of 120°, the final grid cell discharge pattern will present a grid shape along the Euclidean space.

[0029] Furthermore, the process of constructing the cell model in step (2) is as follows: PC The (m,t) update process is as follows:

[0030]

[0031] Among them, τ PC is the synaptic time constant of place cells, r PC (n,t) represents the firing rate of place cells, is the current magnitude encoded by the position of the unmanned system, γ PC is the global inhibition constant, J PC (m,n) is the connection weight between place cells, ρ PC is the representation density of the place cell network, m is the number of the current place cell, and n is the number of other place cells;

[0032] The connection weights between place cells are determined by a standard deviation of α PC The Gaussian function determines:

[0033]

[0034] The firing rate r of place cells PC (n,t) is calculated as follows:

[0035]

[0036] where k PC is the discharge adjustment coefficient.

[0037] Furthermore, the implementation process of step (3) is as follows:

[0038] During the process of extracting the semantic pointer of the image, the convolutional neural network is used to extract the features of the current frame of the picture, and the network output is normalized to a b-dimensional vector to obtain the feature semantic pointer At the same time, other semantic information perceived by this frame is characterized by the orthonormal vectors of the b-dimensional hypersphere:

[0039]

[0040] where is the Euclidean norm, and δ ij is the Kronecker symbol; then the joint representation calculation of the semantic pointer of this key frame is:

[0041]

[0042] During the comparison with the cognitive map M, the external perception information of this frame is inversely calculated to obtain the comparison result:

[0043]

[0044] The result is judged. If ‖p‖2 < ε, it can be considered that there is no relevant memory in the currently stored cognitive map. Otherwise, it is considered that a similar scene is found and the spatial semantic pointer p0 in the scene is obtained, and further decoding is required to obtain the position of the unmanned system in the Euclidean space.

[0045] Furthermore, the implementation process of step (4) is as follows:

[0046] For the large-scale space S0 with length c and width d, to decode the target spatial semantic pointer p0, it needs to be divided into r×s regions of interest, and it is required to satisfy:

[0047]

[0048] where r, s ∈ N and t is a positive constant; then each region of interest is encoded to obtain its corresponding semantic feature vector Next, perform a cosine similarity calculation with the target space semantic pointer p0 to further identify the cognitive space region that is most similar to the target position:

[0049]

[0050] At this time, the decoding result is the position with the maximum cosine similarity to the target space semantic pointer p0, which is the output of the first decoding; continue to perform iterative update within the identified region of interest S1, take the result of the w-th iteration, and calculate:

[0051]

[0052] Among them, C is the inference threshold. If the above formula is satisfied, the position information of this scene can be inferred.

[0053] Furthermore, the implementation process of step (5) is as follows:

[0054] Regarding the generation of active changes in grid cells, its update and iteration process is as follows:

[0055]

[0056] Among them, j is the Fourier component, r j =|f{Ω(x)} j | is the feedback factor, is the time-varying frequency on the unit circle; this model realizes grid-like firing on the neural engineering framework and realizes the path integration function through the interference of oscillations with different phase information.

[0057] Furthermore, the implementation process of step (6) is as follows:

[0058] Regarding the position cell network with a representation range of [c0, d0], the position information obtained by the path integration of grid cells and the position information of external perception information will be encoded using radial basis functions:

[0059]

[0060] Among them, is the total input after multi-sensor encoding, represents the absolute difference between the position represented by the position cell and the position measurement information of the i-th sensor, is the encoding weight of the i-th sensor, is the standard deviation of the encoding of the i-th sensor;

[0061] After the network is iteratively updated and stabilized, perform decoding on its position cell fusion information:

[0062]

[0063] Among them, is the position represented by the cell at the i-th position, l is the spatial length represented by the place cell, and P PC (t) is the position information of the unmanned system decoded at time t;

[0064] During the binding process of the spatial semantic pointer, for the spatial semantic pointer p of the cognitive map M and the current scene sem , the memory process can be described as:

[0065]

[0066] After binding the current spatial semantic pointer with the position of the unmanned system through the bundling operation, it is memorized in the cognitive map.

[0067] Furthermore, the implementation process of step (7) is as follows:

[0068] During the update process of the place cell based on cyclic mapping, the change in the membrane potential of the place cell is:

[0069] u PC (m,t) = u PC ((m - h + N) mod N,t)

[0070] where h is the number of cells for mapping translation, and N is the total number of place cells; the spatial range represented by the place cell also changes as follows:

[0071]

[0072] where c n and d n are the spatial ranges represented by the place cells after the n-th decoding.

[0073] Advantageous effects: Compared with the prior art, the advantageous effects of the present invention are as follows: Based on the navigation and cognition mechanism of mammals such as bats, the present invention uses high-dimensional cognitive coding to efficiently utilize external perception information, and has the advantages of high computational efficiency, strong autonomy, and high intelligence level; uses spatial semantic pointers to construct a cognitive map, which can assist in correcting the self-motion estimation of the current unmanned system through historical memory scenes, improving the robustness and reliability of the navigation system; uses a recall mechanism based on interest iteration to improve the efficiency and accuracy of cognitive map decoding through rapid iterative update of the cognitive map; realizes the representation of a large-scale space by a small number of neurons with unchanged representation resolution through cyclic mapping of place cells; the present invention does not require linearized perception information and Gaussian noise assumptions, fully utilizes high-dimensional perception information in the external environment, and has the advantages of strong universality of the navigation architecture and stable and reliable navigation results, and can be used for intelligent autonomous navigation of unmanned systems in large-scale spaces. Brief Description of the Drawings

[0074] Figure 1 is the flowchart of the present invention;

[0075] Figure 2 is the schematic diagram of the cognitive map decoding based on the interest iteration mechanism;

[0076] Figure 3 is the schematic diagram of the iterative update of place cells based on cyclic mapping;

[0077] Figure 4 is the comparison chart of experimental results. Detailed implementation manners

[0078] The present invention will be further described in detail below with reference to the accompanying drawings.

[0079] As Figure 1 shown, the present invention proposes a large-scale spatial brain-like navigation method based on high-dimensional cognitive coding. By performing high-dimensional cognitive coding on external perceptual information and intelligently fusing it with self-motion information through place cells, if the same scene is encountered again, the position can be directly decoded from the memory of the cognitive map to obtain a stable and reliable positioning result. The specific steps are as follows:

[0080] Step 1: The sensors of the unmanned system include inertial sensors, cameras, and a global navigation satellite system. Among them, the self-motion information includes the three-axis velocity information provided by the inertial sensors, and the external perceptual information includes the images provided by the cameras, visual odometer information, and the position information provided by the global navigation satellite system.

[0081] Step 2: Construct a cognitive map model, a grid cell model, and a place cell model; the cognitive map is used to provide an internal representation of external cognitive information; the grid cells are used to simulate the path integration function in the brain; the place cells are used to fuse self-motion information and external perceptual information.

[0082] During the construction of the cognitive map model, to achieve the high-dimensional representation of continuous spatial information, the mapping relationship from the Euclidean space position vector to the cognitive map This mapping can be encoded in the following way:

[0083]

[0084] where x is the position vector in the Euclidean space, is the encoding matrix, f -1 is the inverse Fourier transform. To ensure that the vector Ω in the cognitive map is real-valued, a matrix must be selected such that e iAx has conjugate symmetry:

[0085]

[0086] where, Indicates the defined bundling operation, Represents the Hadamard product. Through this operation, the vectors in the cognitive map can be directly updated without repeated decoding. Through one or more binding operations, the unmanned system can move continuously in space.

[0087] In the process of building the grid cell model, the speed information provided by the motion information in the Euclidean space is Transformed into the derivative of spatial semantic pointers in cognitive maps with respect to time

[0088]

[0089] The transformation matrix A is composed of multiple sets of row vectors, and each set of vectors k must satisfy:

[0090]

[0091] Through the mutual oscillation and interference of multiple groups of vectors with a phase difference of 120°, the final grid cell discharge pattern will present a grid shape along the Euclidean space.

[0092] During the construction of the place cell model, the membrane potential u PC The (m,t) update process is as follows:

[0093]

[0094] Among them, τ PC is the synaptic time constant of place cells, r PC (n,t) represents the firing rate of place cells, is the current magnitude encoded by the position of the unmanned system, γ PC is the global inhibition constant, J PC (m,n) is the connection weight between place cells, ρ PC is the representation density of the place cell network, m is the number of the current place cell, and n is the number of other place cells.

[0095] The connection weights between place cells are determined by a standard deviation of α PC The Gaussian function determines:

[0096]

[0097] The firing rate of place cells PC The calculation method of (n,t) is as follows:

[0098]

[0099] Among them, k PC is the adjustment coefficient.

[0100] Step 3: Extract semantic pointers from the images in the externally perceived information and measure the similarity with the cognitive map; if a similar scene is found, go to Step 4; otherwise, go to Step 5;

[0101] During the process of extracting the image semantic pointer, use a convolutional neural network to extract the features of the current frame of the picture, and normalize the network output into a b-dimensional vector to obtain the feature semantic pointer At the same time, other semantic information (such as color objects etc.) perceived in this frame is represented by the orthonormal vectors of a b-dimensional hypersphere:

[0102]

[0103] where is the Euclidean norm, and δ ij is the Kronecker symbol. Then the semantic pointer of this key frame can be jointly represented and calculated as:

[0104]

[0105] During the comparison with the cognitive map M, the inverse operation is performed on the externally perceived information of this frame to obtain the comparison result:

[0106]

[0107] Judge the result. If ‖p‖2 < ε (ε is a constant), it can be considered that there is no relevant memory in the currently stored cognitive map. Otherwise, it is considered that a similar scene is found and the spatial semantic pointer p0 in the scene is obtained, and the position of the unmanned system in the Euclidean space needs to be further decoded.

[0108] Step 4: Perform interest-based iterative cognitive map decoding on the spatial semantic pointer of the similar scene in Step 3, output the positioning result of the unmanned system, and at the same time update the initial value of the grid cell path integration, then go to Step 8;

[0109] During the process of interest-based iterative cognitive map similar scene decoding, for a large-scale space S0 with length c and width d, when decoding the target spatial semantic pointer p0, it needs to be divided into r×s interest regions, which should satisfy:

[0110]

[0111] where r, s ∈ N and t is a positive constant. Then encode each interest point to obtain its corresponding semantic feature vector Next, calculate the cosine similarity with the target position p0, and then identify the cognitive space region most similar to the target position. This process is achieved through the following formula:

[0112]

[0113] At this time, the decoding result is the position with the maximum cosine similarity to the target position p0, which is the output of the first decoding. During this process, the decoding error is controlled within t / 2. As the region of interest is gradually refined, it is not necessary to compare one by one in the entire space, but the iterative update can be continued only within the identified region of interest S1. Take the result of the w-th iteration and calculate:

[0114]

[0115] where C is the inference threshold. If the above formula is satisfied, the position information of the scene can be inferred.

[0116] Step 5: Use grid cells to encode the triaxial velocity in the self-motion information, and implement path integration through active iterative update; decode the firing results of the grid cell model to obtain a relatively rough high-frequency position estimate of the unmanned system.

[0117] The iterative process of the grid cell generating active changes is as follows:

[0118]

[0119] where j is the Fourier component, r j =|f{Ω(x)} j | is the feedback factor, is the time-varying frequency on the unit circle. The implementation of this model on the neural engineering framework can show grid-like firing, and the path integration function is realized through the interference of oscillations with different phase information.

[0120] The decoding process of the grid cell is the same as the spatial semantic pointer decoding method in Step 4 above.

[0121] Step 6: Use place cells to encode, update the activity, and decode the position information of the unmanned system after fusing multi-source information from the estimated information decoded in Step 5 and the position information provided in the external perception information, and bind it to the spatial semantic pointer extracted in Step 3 and memorize it in the cognitive map.

[0122] For the place cell network with a representation range of [c0, d0], the position information obtained from the grid cell path integration and the position information of the external perception information will be encoded using the radial basis function:

[0123]

[0124] where, is the total input after multi-sensor encoding, Denote the absolute difference between the position represented by the place cell and the position measurement information of the i-th sensor. Is the coding weight of the i-th sensor. Is the standard deviation of the encoding of the i-th sensor.

[0125] After the network is iteratively updated and stabilized, perform decoding on the fused information of the place cells:

[0126]

[0127] Among them, Is the position represented by the i-th place cell, l is the spatial length represented by the place cell, and P PC (t) is the position information of the unmanned system decoded at time t.

[0128] During the binding process of the spatial semantic pointer, for the cognitive map M and the spatial semantic pointer p of the current scene sem , then the memory process can be described as:

[0129]

[0130] After binding the current spatial semantic pointer with the position of the unmanned system through the bundling operation, it is memorized into the cognitive map.

[0131] Step 7: If the wave packet moves close to the edge of the range represented by the place cell, use the cyclic mapping mechanism to remap the place cells.

[0132] The change in the membrane potential of the place cell is:

[0133] u PC (m,t) = u PC ((m - h + N) mod N,t)

[0134] Among them, h is the number of cells for mapping translation, and N is the total number of place cells. Through this operation, the spatial range represented by the place cells also changes as follows:

[0135]

[0136] Among them, c n and d n Are the spatial ranges represented by the place cells after the n-th decoding.

[0137] Step 8: Continue to loop the above steps from Step 3 until the navigation ends.

[0138] Figure 2 Is a schematic diagram of the decoding process of the cognitive map based on interest iteration, where a and b represent the size of the cognitive map, p0 is the high-dimensional semantic pointer to be decoded in the cognitive map, and S0 is the overall area of the cognitive map. It is the (1, 1) area of the nth iteration of the region of interest, and the rapid and accurate decoding of the large-scale spatial cognitive map is realized through interest iteration. Figure 3 It is a schematic diagram of the position cell update process based on cyclic mapping. When it is detected that the wave packet is approaching the edge of the representation range, by remapping the membrane potential of the position cell, the wave packet is mapped to the middle area of the representation range, and at the same time, the spatial representation range is updated to achieve the efficient representation of the large-scale space and the accurate fusion of multi-sensor information. Figure 4 It is a comparison of the experimental results of different model algorithms. Among them, the present invention is used for multiple repeated tests, as shown in the red part of the figure. It can be clearly seen from the figure that the present method has good stability and accuracy in large-scale spatial navigation and positioning.

[0139] The present invention has been described in detail above in combination with specific embodiments, but these descriptions should not be construed as limiting the present invention. Those skilled in the art understand that without departing from the spirit and scope of the present invention, various equivalent substitutions, modifications or improvements can be made to the technical solutions and their implementation manners of the present invention, and these all fall within the scope of the present invention. The protection scope of the present invention is subject to the appended claims.

Claims

1. A large-scale spatial brain-like navigation method based on high-dimensional cognitive coding, characterized in that The following steps are involved: (1) Obtaining self-motion information and external perception information based on the unmanned system; (2) constructing a cognitive map model, a grid cell model, and a place cell model; the cognitive map is used to provide an internal representation of external cognitive information; the grid cells are used to simulate the path integration function in the brain; and the place cells are used to fuse self-motion information and external perception information; (3) Extracting semantic pointers from the images in the external perception information and measuring similarity with the cognitive map; if similar scenes are found, executing step (4); otherwise, executing step (5); (4) Decoding the spatial semantic pointers of similar scenes in step (3) based on the interest-based iterative cognitive map, outputting the positioning result of the unmanned system, and updating the initial value of the grid cell path integral, and executing step (8); (5) Use grid cells to encode the three-axis velocity in the self-motion information, implement path integration through iterative activity updates, and obtain a rough high-frequency position estimate of the unmanned system by decoding the discharge results of the grid cell model; (6) using place cells to encode the estimated information decoded in step (5) and the location information provided by the external perception information, update the activity, and decode and output the location information of the unmanned system after fusing the multi-source information, bind it with the spatial semantic pointer extracted in step (3), and memorize it in the cognitive map; (7) If the wave packet moves close to the edge of the place cell representation range, the place cell is remapped using the cyclic mapping mechanism; (8) Return to step (3) and loop until navigation ends.

2. The large-scale spatial brain-inspired navigation method based on high-dimensional cognitive coding according to claim 1, wherein, The self-motion information described in step (1) includes three-axis speed information provided by an inertial sensor; the external perception information includes images provided by a camera, visual odometer information, and position information provided by a global navigation satellite system.

3. A large-scale spatial brain-like navigation method based on high-dimensional cognitive coding according to claim 1, characterized in that The implementation process of building the cognitive map model in step (2) is as follows: To achieve a high-dimensional representation of continuous spatial information, the mapping relationship from Euclidean space position vectors to cognitive maps This mapping is encoded in the following way: Ω(x) = f -1 {e iAx} where x is a position vector in Euclidean space, is the encoding matrix, and f -1 is the inverse Fourier transform. To ensure that the vector Ω in the cognitive map is real-valued, one must choose a matrix such that e iAx has conjugate symmetry: Among them, represents a bundling operation, represents a Hadamard product; through one or more bundling operations, the continuous movement of the unmanned system in space is achieved.

4. A large-scale spatial brain-like navigation method based on high-dimensional cognitive coding according to claim 1, characterized in that The implementation process of constructing the grid cell model in step (2) is as follows: By converting the velocity information provided by the motion information in the Euclidean space into the derivative of the spatial semantic pointer with respect to time in the cognitive map The transformation matrix A is composed of multiple sets of row vectors, and each set of vectors k must satisfy: Through the mutual oscillation and interference of multiple groups of vectors with a phase difference of 120°, the final grid cell discharge pattern will present a grid shape along the Euclidean space.

5. The large-scale spatial brain-inspired navigation method based on high-dimensional cognitive coding according to claim 1, wherein, The process of constructing the cell model in step (2) is as follows: Membrane potential u PC (m, t) is updated as follows: where τ PC is the synaptic time constant of place cells, r PC (n,t) represents the firing rate of place cells, is the magnitude of the current encoded by the position of the unmanned system, γ PC is the global inhibition constant, J PC (m,n) is the connection weight between place cells, ρ PC is the representation density of the place cell network, m is the number of the current place cell, and n is the number of other place cells; The connection weights between place cells are determined by a Gaussian function with a standard deviation of α PC : The firing rate r of place cells PC (n,t) is calculated as follows: where k PC is the release adjustment coefficient.

6. The large-scale spatial brain-inspired navigation method based on high-dimensional cognitive coding according to claim 1, wherein The implementation process of step (3) is as follows: During the process of extracting the semantic pointer of an image, the convolutional neural network is used to extract the features of the current frame image, and the network output, which is a b-dimensional vector, is normalized to obtain the feature semantic pointer. Meanwhile, other semantic information perceived in this frame is represented by the orthonormal vectors on the b-dimensional hypersphere: where is the Euclidean norm, and δ ij is the Kronecker symbol; then the joint representation calculation of the semantic pointer of this key frame is as follows: In the process of comparing with the cognitive map M, the external perception information of the frame is inverted to obtain the comparison result: The result is judged. If ‖p‖2<ε, it is considered that there is no relevant memory in the currently stored cognitive map. Otherwise, it is considered that a similar scene is found and the spatial semantic pointer p0 in the scene is obtained, and further decoding is required to obtain the position of the unmanned system in the Euclidean space.

7. A large-scale spatial brain-inspired navigation method based on high-dimensional cognitive coding according to claim 1, characterized in that The implementation process of step (4) is as follows: For a large-scale space S0 with a length of c and a width of d, to decode the target space semantic pointer p0, it needs to be divided into r×s regions of interest, which must meet the following requirements: where r, s ∈ N and t is a positive constant; then each point of interest is encoded to obtain its corresponding semantic feature vector Next, the cosine similarity is calculated with the semantic pointer p0 of the target space, and then the cognitive space region most similar to the target position is identified: At this time, the decoding result is the position with the largest cosine similarity with the target space semantic pointer p0, which is the output of the first decoding; continue to iterate and update in the identified interest area S1, take the result of the wth iteration, and calculate: Among them, C is the inference threshold. If the above formula is satisfied, the position information of the scene can be obtained.

8. A large-scale spatial brain-inspired navigation method based on high-dimensional cognitive coding according to claim 1, characterized in that The implementation process of step (5) is as follows: Regarding the generation of activity changes in grid cells, its update and iteration process is as follows: where j is the Fourier component and r j = |f{Ω(x)} j | is the feedback factor, is the time-varying frequency on the unit circle; this model realizes grid-like spiking on the neural engineering framework and achieves the path integration function through the interference of oscillations with different phase information.

9. A large-scale spatial brain-like navigation method based on high-dimensional cognitive coding according to claim 1, characterized in that The implementation process of step (6) is as follows: For the position cell network with a representation range of [c0, d0], the position information obtained by the grid cell path integration and the position information of the external perception information will be encoded using radial basis functions: Among them, is the total input after multi-sensor encoding, represents the absolute difference between the position represented by the place cells and the position measurement information of the i-th sensor, is the encoding weight of the i-th sensor, is the standard deviation of the encoding of the i-th sensor; After the network iteration is updated and stabilized, its position cell fusion information is decoded: Among them, is the position represented by the cell at the i-th position, l is the spatial length represented by the place cell, P PC (t) is the position information of the unmanned system decoded at time t; During the process of binding spatial semantic pointers, for the cognitive map M and the spatial semantic pointer p of the current scene sem , the memory process is described as: Through the bundling operation, the current spatial semantic pointer is combined with the position of the unmanned system and memorized in the cognitive map.

10. A large-scale spatial brain-like navigation method based on high-dimensional cognitive coding according to claim 1, characterized in that The implementation process of step (7) is as follows: During the update process of position cells based on cyclic mapping, the change in the membrane potential of position cells is: u PC (m, t) = u PC ((m - h + N) mod N, t) Among them, h is the number of cells for mapping translation, and N is the total number of position cells; the spatial range represented by the position cells also changes as follows: where c n and d n are the spatial ranges represented by the place cells after the nth decoding.

Images