Method for extracting spatial information based on spiking neural network

Abstract

The invention discloses a method for extracting spatial information based on a spiking neural network, and relates to the technical field of 3D point cloud processing, and the method comprises the specific steps: firstly collecting 3D point cloud original data, and obtaining effective data through denoising and coordinate standardization; inputting a pulse space position coding layer, and generating pulse features containing local details by using an algorithm; dividing a feature map according to spatial sub-regions and maintaining a state vector; inputting a dynamic event coding layer, detecting correlation, triggering connection, and adjusting weight to realize global fusion; and finally, outputting by an output layer, providing spatial information for subsequent tasks, and completing extraction. According to the method, the 3D point space position is accurately represented through pulse coding, local topology details are mined, and accurate local information support is provided; and through a mutual information entropy dynamic weight rendering mechanism, global space context efficient transmission fusion is realized, local and global conditions are considered, the method adapts to complex scenes, comprehensive and reliable support is provided for space calculation tasks, and the processing effect is improved.

Description

Technical Field

[0001] This invention relates to the field of 3D point cloud processing technology, specifically a method for extracting spatial information based on a spiking neural network. Background Technology

[0002] With the rapid development of 3D perception technology, 3D point clouds, as a data source capable of accurately representing the spatial structure of target scenes, play a crucial role in many fields. Spatial information extraction, as the core link in 3D point cloud processing, directly affects the effectiveness of various subsequent spatial computing tasks. Spiking neural networks, with their characteristic of simulating the pulse firing mechanism of biological neurons, exhibit unique advantages in spatiotemporal information processing. They can carry rich information through the time and frequency characteristics of pulse sequences and have gradually become an important research direction in the field of 3D point cloud spatial information extraction. Currently, various industries have continuously increased their requirements for the accuracy and comprehensiveness of spatial information extraction, requiring a technical solution that can deeply mine the local details and global correlations of 3D point clouds to adapt to the application needs in complex scenarios and promote the implementation and development of related technologies. Whether it is environmental perception of intelligent terminals or 3D inspection in industrial scenarios, there is an urgent need for efficient and accurate 3D point cloud spatial information extraction technology, providing broad development space for related technological innovation.

[0003] Traditional 3D point cloud spatial information extraction methods have limitations in their encoding methods. They often employ single-dimensional encoding logic, which makes it difficult to fully capture the three-dimensional spatial characteristics of 3D point clouds. This results in generated features that cannot accurately reflect local spatial details. In the feature fusion stage, most methods rely on fixed connection patterns or static weight allocation, failing to dynamically adjust the information transmission method based on the correlation of impulse activity in different spatial sub-regions. This leads to insufficient fusion of local and global features, easily resulting in information loss or redundancy. Furthermore, the network architecture design of some technologies lacks coordination, and the connection between data transmission and processing between modules is not smooth enough, making it difficult to balance processing efficiency and result accuracy. This fails to meet the real-time and reliability requirements of spatial information extraction in complex scenarios. At the same time, traditional methods are not precise enough in dividing and maintaining spatial sub-regions when dealing with complex scenarios, making it difficult to track the impulse activity status of each region in real time. This further affects the overall quality of spatial information extraction and limits its application in demanding fields. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for spatial information extraction based on a spiking neural network. This method acquires effective data through point cloud acquisition and preprocessing, generates spiking features containing local details through spatial location spiking encoding, divides feature maps into spatial sub-regions and maintains state vectors, and then uses a mutual information entropy dynamic weight evolution algorithm to achieve global dynamic fusion across feature maps. Finally, it outputs spiking features that fuse local details and global relationships. The efficient collaboration of various modules in the spatially enhanced spiking neural network ensures the stability and accuracy of data transmission and processing, providing reliable support for subsequent spatial computing tasks such as target detection, segmentation, and recognition. It balances extraction accuracy and processing efficiency, adapting to the needs of complex application scenarios.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for extracting spatial information based on a spiking neural network, the specific steps of which are as follows: S1, Point Cloud Acquisition Preprocessing: Acquire raw 3D point cloud data, denoise the raw data, and then perform coordinate standardization on the denoised data to obtain effective 3D point cloud data; S2, Spatial Position Pulse Encoding: Input effective 3D point cloud data into the pulse spatial position encoding layer, use a three-dimensional coupled pulse frequency encoding algorithm to obtain the pulse firing frequency, combine spatial sorting rules to encode the pulse firing time, and generate pulse features containing local spatial details; S3, Feature map partitioning and maintenance: Divide the pulse features according to the spatial sub-region distribution of the 3D point cloud to obtain multiple feature maps corresponding to different spatial sub-regions. Each feature map maintains a state vector representing its own pulse activity intensity in real time. S4, Cross-Graph Global Dynamic Fusion: Input all feature maps into the pulse cross-feature map dynamic event encoding layer, use the mutual information entropy dynamic weight evolution algorithm to detect the pulse activity correlation pattern of any two feature maps, dynamically trigger temporary high-bandwidth connection channels based on the correlation determination results and adjust the connection weights in real time, realize the global spatial context transfer and fusion of pulse features between different feature maps through the connection weights, and output fused pulse features containing local details and global dynamic spatial relationships. S5, fusion feature output completed: The fusion pulse features are output through the output layer of the spatially enhanced spiking neural network, providing spatial information support for subsequent spatial computing-related target detection, segmentation and recognition tasks, and completing the process of extracting spatial information from 3D point cloud; The spatially enhanced spiking neural network includes an input layer, a spiking spatial location encoding layer, a feature map partitioning and vector maintenance module, a spiking cross-feature map dynamic event encoding layer, and an output layer connected in series. The input layer receives valid 3D point cloud data, adopts a parallel receiving architecture, supports multi-channel data input, and the single-channel data receiving rate is no less than 300MB / s; The input data is transmitted to the pulse spatial position coding layer after format conversion; The pulse spatial location coding layer and the feature map partitioning and vector maintenance module are connected through a high-speed data bus with a bus bandwidth of no less than 1GB / s; The feature map partitioning and vector maintenance module and the pulse cross-feature map dynamic event coding layer are connected by a distributed interface, which supports batch transmission according to feature map identifiers, with a transmission delay of no more than 5ms. The output of the pulse cross-feature map dynamic event coding layer is directly connected to the output layer. The fused pulse feature matrix is ​​processed by the format normalization unit of the output layer and then output according to the preset data protocol. The output layer has a built-in data verification unit that verifies the integrity and accuracy of the output data in real time. The network adopts a 32-bit floating-point computing architecture, supports batch data parallel processing, and can process no less than 100,000 effective 3D point cloud points in a single batch. The network's power consumption is controlled within 50W, and it can support continuous 72 hours of trouble-free operation.

[0006] Further, in S1, the point cloud acquisition preprocessing, the acquisition method of the original 3D point cloud data is as follows: an acquisition device with three-dimensional coordinate detection function is used, the acquisition device including a rotating lidar, a three-dimensional laser scanner and a structured light three-dimensional camera. Before acquisition, the internal and external parameters of the device are calibrated, including lens distortion correction, distance measurement accuracy calibration and coordinate system one; during the acquisition process, the device operates according to a preset scanning mode, the number of scanning lines of the rotating lidar is set to 16 to 128 lines, and the scanning frequency is 10Hz to 30Hz; the point cloud density of the three-dimensional laser scanner is set to not less than 50 points per square meter, and the scanning resolution is 0.1 mm to 1 mm; the shooting frame rate of the structured light three-dimensional camera is 15fps to 60fps, and the field of view is 60 degrees to 120 degrees; the device captures the spatial position information of each point in the target scene in real time, and synchronously records the x, y, z three-dimensional coordinate data, acquisition timestamp, and device attitude parameters. Each frame of data contains three-dimensional coordinate information of not less than 5000 points.

[0007] Furthermore, in step S2, the spatial position pulse coding layer includes an input module, a frequency coding module, a time coding module, and an output module. The input module receives standardized three-dimensional coordinate data of valid 3D point clouds and transmits it sequentially according to the point cloud sequence number using a serial reading method, with a data transmission rate of not less than 200MB / s. The frequency coding module has a built-in three-dimensional coupled pulse frequency coding algorithm operation unit, which performs frequency calculation using 32-bit floating-point arithmetic precision and supports parallel processing of coordinate data of multiple points. The time coding module stores preset spatial sorting rules, performs point cloud sorting by coordinate weighted sum, and encodes the pulse emission time of each point. The output module integrates the binary pulse sequence of each point according to the sorting number, generates a pulse feature matrix, and outputs it to step S3.

[0008] Furthermore, in S2, the mathematical expression for the three-dimensional coupled pulse frequency coding algorithm in spatial position pulse coding is: ,in For pulse delivery frequency, , , To standardize the three-dimensional coordinates of effective 3D point cloud data, Based on the pulse frequency, The hyperbolic tangent nonlinear activation function is... For nonlinear gain coefficients, , , For independent weights of three-dimensional coordinates, The three-dimensional coordinate coupling coefficient, This is a tiny offset.

[0009] Furthermore, in S2, the specific process of generating pulse features containing local spatial details by combining spatial sorting rules with the pulse firing time in spatial location pulse coding is as follows: weighted independently according to three-dimensional coordinates... , , With standardized three-dimensional coordinates , , Calculate the weighted sum for each point , It is a weighted sum of the coordinates of each 3D point, applied to all valid points by weighted summation. The data is globally sorted from smallest to largest and assigned a unique sort number, starting from 1 and increasing sequentially. Based on an encoding period of 100 milliseconds and a time resolution of 1 millisecond, the sort numbers are... The first pulse emission time point is set to The pulse emission time points are shifted sequentially according to the pulse emission frequency obtained by the three-dimensional coupled pulse frequency encoding algorithm. The interval between each pulse is 1000 milliseconds divided by the pulse emission frequency. Within a 100-millisecond encoding period, each point generates a binary pulse sequence. The position corresponding to the emission time point in the sequence is marked as 1, and the other positions are marked as 0. The binary pulse sequences of all points are concatenated in sorted order to form a pulse feature matrix with a dimension of the number of effective point cloud points multiplied by 100 time steps. The pulse feature matrix represents the spatial positional relationship of different points through the difference in emission time points, and together with the difference in pulse emission frequency, it carries local spatial detail information.

[0010] Furthermore, in S3, the specific process of feature map partitioning and maintenance is as follows: using standardized three-dimensional coordinates of effective 3D point cloud data... , , Based on the common value range [0, 1], the meshing operation is performed using a three-dimensional equidistant meshing method. First, set... , , The number of divisions in the three dimensions are respectively , , The calculated interval lengths of individual sub-regions in each dimension are as follows: , , Based on this, the coordinate intervals of all cube sub-regions are determined, where any cube sub-region... coordinate range , coordinate range , coordinate range , The value range is 0 to , The value range is 0 to , The value range is 0 to Each cube sub-region is assigned a unique feature map identifier; all pulse features generated in step S2 are traversed, and the normalized 3D point cloud coordinates corresponding to each pulse feature are extracted. By comparing these coordinates with the coordinate ranges of each cube sub-region, the cube sub-region to which the pulse feature belongs is determined, and the pulse feature is assigned to the feature map of the corresponding identifier; after all pulse features have been assigned, a total of [number missing] pulse features are generated. The feature map contains a complete pulse sequence of all points within the corresponding cubic sub-region. The pulse sequences are arranged in ascending order of the weighted sum of the coordinates of the corresponding points within the feature map.

[0011] Furthermore, in step S4, the cross-map global dynamic fusion, the pulse cross-feature map dynamic event encoding layer includes an input module, a correlation calculation module, a connection management module, a feature transfer fusion module, and an output module. The input module receives all feature maps and corresponding pulse activity state vectors output in step S3, and reads data sequentially according to feature map identifiers using a parallel interface, with a data transmission rate of not less than 500MB / s. The correlation calculation module has a built-in mutual information entropy calculation unit, which groups the feature maps according to pairwise pairing rules, and calculates the mutual information entropy of the state vectors of each group independently. The calculation process uses 32-bit floating-point arithmetic precision and supports simultaneous processing of not less than 500MB / s. The system pairs 1000 feature maps; the connection management module maintains a dynamic connection channel list, recording the channel-associated feature map identifier, current connection weight, and bandwidth allocation. The channel status is updated at 1ms intervals, supporting real-time triggering, weight adjustment, and closing / releasing of channels; the feature transfer and fusion module includes a weighted operation unit and a summation and integration unit. The weighted operation unit performs time-step weighted processing on the pulse features according to the dynamic connection weights, and the summation and integration unit performs element-wise weighted summation of its own original features and associated transferred features; the output module formats and splices the fusion results of all feature maps to generate a global fused pulse feature matrix and caches it. The mutual information entropy calculation module executes a mutual information entropy calculation formula that includes a formula for calculating the information entropy of a single impulse activity state vector and a formula for calculating the joint information entropy of two impulse activity state vectors. The formula for calculating the information entropy of a single impulse activity state vector is as follows: ,in The information entropy represents the activity state vector of a single impulse. This represents the first state vector. One element, Represents the logarithmic function with base 2. This represents the summation operation on all elements in the state vector; the formula for calculating the mutual information entropy of two impulse activity state vectors is: ,in express Time-of-flight feature map The corresponding impulse activity state vector With feature map The corresponding impulse activity state vector Mutual information entropy, Represents the state vector Information entropy Represents the state vector Information entropy Represents the state vector With state vector Joint information entropy, joint information entropy The probability distribution of the two state vectors is calculated, and then obtained using the following formula: ,in Represents the state vector The element With state vector The element The joint probability of simultaneous occurrence; The pairwise pairing rule is as follows: First, all feature maps input to the pulse cross-feature map dynamic event coding layer are assigned globally unique identifiers. The feature maps are then sorted according to the three-dimensional coordinate intervals of their corresponding spatial sub-regions. Each feature map is assigned a continuous and non-repeating natural number, ranging from 1 to Q, where Q is the total number of feature maps. Pairing is performed using a combination logic that avoids repetition and omissions. Only two feature maps with distinct numbers are selected to form a pairing group. For any two feature maps with numbers a and b, pairing is only performed when the value of a is less than the value of b. Reverse pairing groups where the value of a is greater than the value of b are not generated, and pairing groups with completely different values ​​are excluded. The feature maps with the same number are paired up. When pairing, the feature maps are traversed in ascending order of their numbers. First, the feature map with number 1 is used as the base and paired with the feature maps with numbers 2, 3, ... Q one by one. Then, the feature map with number 2 is used as the base and paired with the feature maps with numbers 3, 4, ... Q one by one. This process continues until the feature map with number Q-1 is paired with the feature map with number Q. The total number of pairing groups generated is Q×(Q-1) / 2. All pairing groups are arranged in ascending order of the base feature map number and the pairing feature map number to form a pairing list, which is read and processed by the correlation calculation module in the order of the list. The summation and integration unit performs an element-wise weighted summation of its original features and associated transitive features, with the following weighted formula: ,in, Represents the feature map at time t The value of the element in the i-th row and j-th column of the fusion feature matrix. This is the weighting coefficient for its own original features, with a value of 1. Represents the feature map at time t The value of the element in the i-th row and j-th column of its own original pulse feature matrix; Representing time t and the feature map The set of all feature maps with valid associations. Represents the feature map at time t Related feature maps Dynamic connection weights between them Represents the feature map at time t After being processed by the weighted arithmetic unit, it is passed to the feature map. The value of the element in the i-th row and j-th column of the correlated pulse feature matrix. Refers to the target feature map currently undergoing fusion processing. Reference and target feature map Other feature graphs with valid associations exist.

[0012] Furthermore, in S4, the mathematical expression for the mutual information entropy dynamic weight evolution algorithm in cross-graph global dynamic fusion is: ,in for Time-of-flight feature map and Dynamic connection weights, For correlation gain factor, for The mutual information entropy of the pulse sequences of two feature maps at time 1. To update the weighting percentage, for Connection weights at time intervals, The dynamic attenuation coefficient, For time step, The maximum value of mutual information entropy. The correlation trigger threshold, It is an exponential function. The basic attenuation coefficient.

[0013] Furthermore, in S4, the operation based on the correlation determination result in cross-graph global dynamic fusion specifically involves: when the mutual information entropy of two feature maps... When a temporary high-bandwidth connection channel is immediately triggered between the two, the correlation gain factor and dynamic attenuation coefficient are calculated according to the mutual information entropy dynamic weight evolution algorithm to update the connection weight; when At that time, only the existing connection weights are attenuated until the weight is less than 0.05, at which point the corresponding channel is closed.

[0014] Furthermore, in S4, the specific process of pulse feature transfer and fusion between different feature maps through connection weights in cross-graph global dynamic fusion is as follows: Each feature map, through a triggered temporary high-bandwidth connection channel, weights its own pulse feature matrix according to the corresponding dynamic connection weights, and then transfers the weighted pulse features to all feature maps with related relationships; after each feature map collects the weighted pulse features transferred from all related feature maps, it performs a weighted summation operation on these transferred features and its own original pulse features, where the weight of its own original pulse features is set to 1, and the weight of the transferred features from the related feature maps is the corresponding dynamic connection weight; after all feature maps have completed the weighted summation of their own original features and related transferred features, the fusion features of all feature maps are spliced ​​and integrated according to the three-dimensional position order of the corresponding spatial sub-regions of each feature map, and finally form a global fusion pulse feature matrix containing local spatial details of each point and global dynamic spatial relationships between different sub-regions.

[0015] Furthermore, in S5, during the completion of feature fusion output, the fused pulse feature is a matrix data with a dimension equal to the number of effective 3D point cloud points multiplied by 100 time steps. Each element in the matrix is ​​a binary value of 0 or 1, where 0 indicates no pulse emission at the corresponding time step and 1 indicates pulse emission at the corresponding time step. Each row in this feature matrix corresponds to the pulse sequence of a single 3D point, carrying the local spatial topology details and coordinate association information of that point. Each column corresponds to the pulse emission status of all 3D points at the same time step, reflecting the synchronicity of pulse activity in different spatial sub-regions. The data type of the fused pulse feature is 32-bit floating-point, with a numerical range limited to 0 to 1, and it adopts a row-major storage method.

[0016] Compared with existing technologies, this method for extracting spatial information based on spiking neural networks has the following advantages: I. This invention generates pulse features that carry local spatial details by performing spatial position pulse coding on effective 3D point cloud data, combined with frequency coding of three-dimensional coupling and time coding of spatial sorting rules. This achieves accurate representation of the spatial positional relationship of each 3D point. By dividing the pulse features into spatial sub-regions and maintaining corresponding state vectors, the pulse activity intensity of different regions can be tracked in real time, ensuring the independence and integrity of local spatial information. This design can deeply mine the local topological details of 3D point clouds, avoid information confusion, and make the extracted spatial information more consistent with the actual distribution of the target scene. It provides accurate local information support for subsequent target detection, segmentation, recognition and other tasks, improves the processing accuracy and reliability of related tasks, solves the problem of insufficient capture of local spatial details in traditional methods, and makes spatial information extraction more targeted.

[0017] Second, this invention utilizes a dynamic weight evolution mechanism based on mutual information entropy in the dynamic event coding layer across feature maps to accurately detect the correlation of pulse activity between different feature maps. It dynamically triggers temporary connection channels and adjusts weights in real time, achieving efficient transmission and fusion of global spatial context. Combined with the optimized architecture design of the spatially enhanced spiking neural network, the modules collaborate efficiently, balancing the preservation of local details with the capture of global dynamic relationships. This allows the extracted fused features to possess both detailed local information and comprehensive global correlations. This dynamic fusion method can adapt to changes in complex spatial scenes, avoiding the limitations of information transmission caused by fixed connections. At the same time, the network's efficient processing architecture ensures the timeliness and stability of information extraction, providing comprehensive and reliable spatial information support for subsequent spatial computing tasks and improving the overall adaptability and efficiency of task processing.

[0018] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0020] Figure 1 This is a flowchart of a 3D point cloud spatial information extraction method based on a spiking neural network. Figure 2 Input-output relationship diagram for each step of 3D point cloud spatial information extraction; Figure 3 This is a flowchart of the dynamic event encoding layer for pulses across feature maps. Detailed Implementation

[0021] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below. Example 1:

[0022] Extraction of three-dimensional spatial information of urban roads.

[0023] This embodiment is applied to the acquisition and analysis of three-dimensional spatial information of roads in urban intelligent transportation systems. By extracting the spatial relationships and detailed features of roads and their surrounding environment, it provides accurate spatial information support for target detection, path planning, and traffic flow statistics of autonomous vehicles. The acquisition equipment uses a rotating LiDAR, which is suitable for the large-scale, high-dynamic spatial data acquisition needs of open urban road scenarios. The specific steps are as follows: Figure 1 As shown.

[0024] S1, Point Cloud Acquisition Preprocessing: A rotating lidar with 3D coordinate detection capability was used as the data acquisition device. Before acquisition, comprehensive internal and external parameter calibration was performed, including lens distortion correction, distance measurement accuracy calibration, and coordinate system unification, to ensure the consistency and accuracy of the acquired data. During acquisition, the device operated according to a preset scanning mode, setting the lidar's scan line count to 64 lines and the scanning frequency to 20Hz to balance data acquisition density and transmission efficiency. The device captured the spatial location information of various points in the urban road scene in real time, simultaneously recording the x, y, and z 3D coordinate data of each point, as well as the acquisition timestamp and device attitude parameters. Each frame of data contained 3D coordinate information of no less than 5000 points, fully covering targets such as road surface, side facilities, and passing vehicles. After acquisition, the raw 3D point cloud data underwent denoising processing to remove outliers caused by environmental interference and equipment errors. Then, coordinate standardization processing was performed on the denoised data to map the 3D coordinates of all points to a unified value range, ultimately obtaining effective 3D point cloud data, providing a high-quality data foundation for subsequent spatial location pulse coding.

[0025] S2, Spatial Position Pulse Code: The valid 3D point cloud data obtained in step S1 is input into the pulse spatial position encoding layer of the spatially enhanced spiking neural network. The input module of this encoding layer uses a serial reading method, transmitting standardized 3D coordinate data sequentially according to the point cloud sequence number, maintaining a data transmission rate above 200MB / s to ensure data transmission continuity. The frequency encoding module incorporates a 3D coupled pulse frequency encoding algorithm computation unit. The mathematical expression of the 3D coupled pulse frequency encoding algorithm is: ,in For pulse delivery frequency, , , To standardize the three-dimensional coordinates of effective 3D point cloud data, Based on the pulse frequency, The hyperbolic tangent nonlinear activation function is... For nonlinear gain coefficients, , , For independent weights of three-dimensional coordinates, The three-dimensional coordinate coupling coefficient, For minute offsets, coordinate data from multiple points are processed in parallel with 32-bit floating-point precision to calculate the pulse firing frequency for each point, accurately distinguishing point cloud features at different spatial locations. The time-coding module stores preset spatial sorting rules, first calculating a weighted sum for each point based on independent weights of the 3D coordinates and standardized 3D coordinates, using the following formula: All valid points are globally sorted by weighted sum from smallest to largest and assigned a unique sorting number, starting from 1 and increasing sequentially. With an encoding period of 100 milliseconds and a time resolution of 1 millisecond, the first pulse emission time of the point with sorting number k is set to k milliseconds. Subsequent pulse emission time points are shifted sequentially according to the calculated pulse emission frequency, with the interval between each pulse determined by the pulse emission frequency. The output module integrates the binary pulse sequences of each point according to the sorting number, generating a pulse feature matrix with dimensions equal to the number of valid point cloud points multiplied by 100 time steps. This matrix represents the spatial relationship between different points through differences in emission time points, and together with differences in pulse emission frequency, it carries local spatial detail information, clearly presenting the spatial distribution characteristics of road elements. The pulse feature matrix is ​​then output to the next step.

[0026] S3, Feature map partitioning and maintenance: Standardized 3D coordinates of effective 3D point cloud data , , Based on the common value range of 0 and 1, a three-dimensional equidistant grid division method is used to divide the pulse features generated in step S2. First, set... , , The number of divisions in each of the three dimensions is MNP. The interval length of a single sub-region in each dimension is calculated, and the coordinate interval of all cubic sub-regions is determined accordingly. Each cubic sub-region is assigned a unique feature map identifier. All pulse features generated in step S2 are traversed, and the normalized 3D point cloud coordinates corresponding to each pulse feature are extracted. By comparing the coordinates with the coordinate intervals of each cubic sub-region, the cubic sub-region to which the pulse feature belongs is determined, and the pulse feature is assigned to the feature map with the corresponding identifier. After all pulse features have been assigned, a total of M×N×P feature maps are formed. Each feature map contains a complete pulse sequence of all points in the corresponding cubic sub-region. The pulse sequence is arranged in ascending order of the weighted sum of the coordinates of the corresponding points in the feature map. At the same time, each feature map maintains a state vector representing the intensity of its own pulse activity in real time, intuitively reflecting the spatial feature activity of each sub-region, and providing a clear feature basis for cross-map global dynamic fusion.

[0027] S4, cross-graph global dynamic fusion: All feature maps are input into the pulse cross-feature map dynamic event encoding layer. The input module of this encoding layer adopts a parallel interface, sequentially reading the feature maps and corresponding pulse activity state vectors according to the feature map identifiers. The data transmission rate is no less than 500MB / s, meeting the high-efficiency transmission requirements of large-scale feature data. The correlation calculation module has a built-in mutual information entropy calculation unit. First, all feature maps are assigned globally unique identifiers. They are then assigned continuous and non-repeating natural numbers according to the sorting results of the three-dimensional coordinate intervals of the corresponding spatial sub-regions of the feature maps. The number range is from 1 to Q, where Q is the total number of feature maps. Then, the feature maps are grouped according to the pairwise pairing rule without repetition or omission. Only feature maps with different numbers and where the previous number is less than the next number are selected to form pairing groups. Finally, Q×(Q-1) / 2 pairing groups are generated. The mutual information entropy of the two feature map state vectors is calculated independently for each group. The calculation process includes calculating the information entropy of a single impulse activity state vector and calculating the joint information entropy of the two state vectors. The formula for calculating the information entropy of a single impulse activity state vector is as follows: ,in The information entropy represents the activity state vector of a single impulse. This represents the first state vector. One element, Represents the logarithmic function with base 2. This represents the summation operation on all elements in the state vector; the formula for calculating the mutual information entropy of two impulse activity state vectors is: ,in express Time-of-flight feature map The corresponding impulse activity state vector With feature map The corresponding impulse activity state vector Mutual information entropy, Represents the state vector Information entropy Represents the state vector Information entropy Represents the state vector With state vector Joint information entropy, joint information entropy The probability distribution of the two state vectors is calculated, and then obtained using the following formula: ,in Represents the state vector The element With state vector The element Simultaneous joint probabilities accurately detect association patterns of features in different sub-regions. The connection management module maintains a dynamic list of connection channels, recording channel association feature map identifiers, current connection weights, and bandwidth allocation. Channel status is updated at 1ms intervals. When the mutual information entropy of two feature maps is greater than or equal to the correlation trigger threshold Upon activation, a temporary high-bandwidth connection channel is immediately triggered between the two entities. The correlation gain factor and dynamic attenuation coefficient are calculated using the mutual information entropy dynamic weight evolution algorithm to update the connection weights, thereby strengthening the transmission of strong correlation features. The mathematical expression for the mutual information entropy dynamic weight evolution algorithm is as follows: ,in for Time-of-flight feature map and Dynamic connection weights, For correlation gain factor, for The mutual information entropy of the pulse sequences of two feature maps at time 1. To update the weighting percentage, for Connection weights at time intervals, The dynamic attenuation coefficient, For time step, The maximum value of mutual information entropy. The correlation trigger threshold, It is an exponential function. The basic attenuation coefficient; when the mutual information entropy is less than the correlation trigger threshold. At this time, only existing connection weights are attenuated until the weight is less than 0.05, at which point the corresponding channel is closed to avoid invalid connections consuming resources. The feature transfer fusion module includes a weighting operation unit and a summation and integration unit. The weighting operation unit performs time-step weighting processing on the pulse features according to the dynamic connection weights. The summation and integration unit performs element-wise weighted summation on its own original features and the associated transferred features. The weight of its own original features is set to 1, and the weight of the transferred features in the associated feature map is the corresponding dynamic connection weight, realizing deep fusion of local details and global associations. The output module formats and splices the fusion results of all feature maps to generate a global fused pulse feature matrix and caches it. This matrix completely preserves the local spatial details of each point in the road scene and the global dynamic spatial relationships between different sub-regions, providing comprehensive spatial information for subsequent object detection, such as... Figure 3 As shown.

[0028] S5, Feature fusion output complete: The global fused pulse feature matrix output from the pulse cross-feature map dynamic event encoding layer is transmitted to the output layer of the spatially enhanced spiking neural network. The format normalization unit of the output layer processes the fused pulse feature matrix according to a preset data protocol, and the built-in data verification unit performs real-time verification of the integrity and accuracy of the output data to ensure data reliability. The final output fused pulse feature is a 32-bit floating-point matrix, with a dimension equal to the number of effective 3D point cloud points multiplied by 100 time steps. Each element in the matrix is ​​a binary value of 0 or 1, where 0 indicates no pulse firing at the corresponding time step and 1 indicates pulse firing at the corresponding time step. Each row corresponds to the pulse sequence of a single 3D point, carrying the local spatial topological details and coordinate association information of that point. Each column corresponds to the pulse firing status of all 3D points at the same time step, reflecting the synchronicity of pulse activity in different spatial sub-regions. The data is stored in a row-major order. This fused pulse feature provides accurate spatial information support for subsequent urban road target detection, segmentation, and recognition tasks, fully completing the extraction process of spatial information from the 3D point cloud of urban roads. The entire network adopts a 32-bit floating-point computing architecture, supports batch data parallel processing, and can process no less than 100,000 effective 3D point cloud points in a single batch. The network's power consumption is controlled within 50W, and it can support continuous 72 hours of fault-free operation, meeting the continuous operation requirements of urban intelligent transportation systems.

[0029] In summary, this embodiment targets urban road scenarios, employing a rotating LiDAR to acquire data. Preprocessing removes interference and unifies coordinates, laying a reliable foundation for subsequent processes. A 3D coupled pulse frequency encoding algorithm and spatial sorting rules within the pulse spatial position encoding layer transform the point cloud into a pulse feature matrix carrying local details. A feature map is then obtained through 3D equally spaced grid partitioning, and a state vector is maintained, providing a clear feature basis for fusion. Global feature fusion is achieved using mutual information entropy calculation and a mutual information entropy dynamic weight evolution algorithm within the pulse cross-feature map dynamic event encoding layer, ultimately outputting verified fused features. The entire process follows preset parameters and algorithms, efficiently extracting road spatial information, meeting the precise spatial data requirements of autonomous driving tasks, and ensuring stable network operation and controllable power consumption. Example 2:

[0030] 3D inspection of industrial parts.

[0031] This embodiment is applied to the 3D inspection scenario of industrial parts in the field of precision machinery manufacturing. By extracting the 3D spatial information, topological structure, and spatial features related to minute defects on the surface of the parts, it provides data support for the dimensional accuracy verification, defect detection, and quality control of the parts. The acquisition device selected is a structured light 3D camera, which is adapted to the needs of close-range, high-precision surface data acquisition of parts in industrial scenarios. The specific steps are as follows: Figure 2 As shown.

[0032] S1, Point Cloud Acquisition Preprocessing: A structured light 3D camera with 3D coordinate detection capability was used as the acquisition device. Before acquisition, the device underwent internal and external parameter calibration, including lens distortion correction, distance measurement accuracy calibration, and coordinate system unification, ensuring that the accuracy of the acquired data met industrial testing standards. During acquisition, the device operated in a preset mode, setting the shooting frame rate to 30fps and the field of view to 90 degrees to ensure complete coverage of the component surface and capture of minute details. The device captured the spatial position information of each point on the surface of the industrial component in real time, synchronously recording the x, y, and z 3D coordinate data of each point, as well as the acquisition timestamp and device attitude parameters. Each frame of data contained 3D coordinate information of no less than 5000 points, accurately reproducing the shape structure and surface condition of the component. After acquisition, the raw 3D point cloud data was denoised to remove invalid points caused by oil stains on the component surface and interference from ambient light. Then, coordinate standardization was performed on the denoised data to uniformly map the 3D coordinates of all points to the range of 0 and 1, obtaining valid 3D point cloud data, providing a clean data source for subsequent high-precision encoding.

[0033] S2, Spatial Position Pulse Code: The valid 3D point cloud data obtained in step S1 is input into the pulse spatial location encoding layer of the spatially enhanced spiking neural network. The input module uses a serial reading method, transmitting standardized 3D coordinate data sequentially according to the point cloud sequence number, with a data transmission rate of no less than 200MB / s to ensure data transmission stability. The frequency encoding module has a built-in 3D coupled pulse frequency encoding algorithm operation unit, which processes the coordinate data of multiple points in parallel with 32-bit floating-point arithmetic precision, accurately calculating the pulse firing frequency of each point, fully reflecting the spatial differences of different positions on the surface of the component, especially the characteristic differences between small defect areas and normal areas. The time encoding module stores preset spatial sorting rules. First, it calculates the weighted sum of each point according to the independent weights of the 3D coordinates and the standardized 3D coordinates. All valid points are globally sorted according to the weighted sum from smallest to largest and assigned a unique sorting number. The sorting number starts from 1 and increases sequentially. Then, with an encoding period of 100 milliseconds and a time resolution of 1 millisecond, the first pulse firing time of the point with sorting number k is set to k milliseconds. The subsequent pulse firing time points are shifted sequentially according to the calculated pulse firing frequency. The interval time of each pulse is determined by the pulse firing frequency. The output module integrates the binary pulse sequence of each point according to the sorted number to generate a pulse feature matrix with a dimension of the number of effective point cloud points multiplied by 100 time steps. This matrix characterizes the spatial positional relationship of different points on the surface of the component by the difference in the emission time point, and carries the local spatial details of the component surface by combining the difference in pulse emission frequency, clearly presenting the position and shape information of tiny protrusions and depressions. Then the pulse feature matrix is ​​output to the next step.

[0034] S3, Feature map partitioning and maintenance: Standardized 3D coordinates of effective 3D point cloud data , , Based on the common value range of 0 and 1, a three-dimensional equidistant grid division method is used to divide the pulse features generated in step S2. First, set... , , The number of divisions in each of the three dimensions is MNP. The interval length of a single sub-region in each dimension is calculated, and the coordinate range of all cubic sub-regions is determined accordingly. Each cubic sub-region is assigned a unique feature map identifier to ensure that the spatial range of each sub-region is accurate and controllable. All pulse features generated in step S2 are traversed, and the 3D point cloud standardized coordinates corresponding to each pulse feature are extracted. By comparing the coordinates with the coordinate range of each cubic sub-region, the cubic sub-region to which the pulse feature belongs is determined, and the pulse feature is assigned to the feature map with the corresponding identifier. After all pulse features have been assigned, a total of M×N×P feature maps are formed. Each feature map contains a complete pulse sequence of all points in the corresponding cubic sub-region. The pulse sequence is arranged in ascending order of the weighted sum of the coordinates of the corresponding points in the feature map. At the same time, each feature map maintains a state vector representing the intensity of its own pulse activity in real time, dynamically reflecting the spatial feature activity in the sub-region, which helps to accurately identify the sub-region where the defect is located.

[0035] S4, cross-graph global dynamic fusion: All feature maps are input into the pulse cross-feature map dynamic event encoding layer. The input module of this encoding layer adopts a parallel interface, sequentially reading the feature map and the corresponding pulse activity state vector according to the feature map identifier. The data transmission rate is maintained above 500MB / s, meeting the high-efficiency processing requirements of industrial detection. The correlation calculation module has a built-in mutual information entropy calculation unit. First, all feature maps are assigned globally unique identifiers. Then, they are assigned continuous and non-repeating natural numbers based on the sorting results of the three-dimensional coordinate intervals of the corresponding spatial sub-regions of the feature maps. The number range is from 1 to Q, where Q is the total number of feature maps. Next, the feature maps are grouped according to the pairwise pairing rule without repetition or omission. Only feature maps with different numbers and where the previous number is less than the next number are selected to form pairing groups. Finally, Q×(Q-1) / 2 pairing groups are generated. The mutual information entropy of the state vectors of the two feature maps is calculated independently for each group. The calculation process includes the information entropy calculation of a single impulse activity state vector and the joint information entropy calculation of two state vectors. Both use 32-bit floating-point arithmetic precision and support the simultaneous processing of no less than 1000 groups of feature map pairings to ensure the comprehensiveness and accuracy of the correlation analysis and accurately capture the association between defective areas and surrounding normal areas. The connection management module maintains a list of dynamic connection channels, recording the channel association feature map identifier, current connection weight, and bandwidth allocation. Channel status is updated every 1ms. When the mutual information entropy of two feature maps is greater than or equal to the correlation trigger threshold, it indicates a strong correlation between the spatial features of the corresponding sub-regions, potentially corresponding to continuous structures or defect areas of components. A temporary high-bandwidth connection channel is immediately triggered between the two, and the association gain factor and dynamic attenuation coefficient are calculated using the mutual information entropy dynamic weight evolution algorithm to update the connection weight. When the mutual information entropy is less than the correlation trigger threshold, only the existing connection weight is attenuated until the weight is less than 0.05, at which point the corresponding channel is closed to avoid invalid connections consuming resources. The feature transfer and fusion module includes a weighted operation unit and a summation and integration unit. The weighted operation unit performs time-step weighted processing on the pulse features according to the dynamic connection weights. The summation and integration unit performs element-wise weighted summation of its own original features and the associated transferred features. The weight of its own original features is set to 1, and the weight of the associated feature map transferred features is the corresponding dynamic connection weight, fully integrating local details and global association information to highlight the spatial features of defect areas. The output module formats and stitches the fusion results of all feature maps to generate a global fusion pulse feature matrix and caches it. This matrix fully preserves the local spatial details of the component surface and the global dynamic spatial relationship between different sub-regions, providing key support for defect detection.

[0036] S5, Feature fusion output complete: The global fused pulse feature matrix output from the pulse cross-feature map dynamic event encoding layer is transmitted to the output layer of the spatially enhanced spiking neural network. The format normalization unit of the output layer processes the fused pulse feature matrix according to a preset industrial data protocol. The built-in data verification unit performs real-time verification of the integrity and accuracy of the output data, ensuring that the data meets the stringent requirements of industrial inspection. The final output fused pulse feature is a 32-bit floating-point matrix, with a dimension equal to the number of effective 3D point cloud points multiplied by 100 time steps. Each element in the matrix is ​​a binary value of 0 or 1, where 0 indicates no pulse emission at the corresponding time step and 1 indicates a pulse emission at the corresponding time step. Each row corresponds to the pulse sequence of a single 3D point, carrying the local spatial topological details of that point and its coordinate association information with surrounding points, accurately reflecting the location of minute defects. Each column corresponds to the pulse emission status of all 3D points at the same time step, reflecting the spatial structural synchronicity of different areas of the component, aiding in dimensional accuracy verification. The data is stored in a row-major order. This fused pulse feature provides high-precision spatial information support for subsequent industrial component defect detection, dimensional measurement, and quality assessment tasks, completing the extraction process of 3D point cloud spatial information for industrial components. The entire network adopts a 32-bit floating-point computing architecture, supports batch data parallel processing, and can process no less than 100,000 effective 3D point cloud points in a single batch. The network's power consumption is controlled within 50W, which can meet the continuous operation requirements of industrial production lines and support continuous 72 hours of fault-free operation, ensuring efficient and stable industrial testing.

[0037] In summary, this embodiment focuses on the inspection of industrial components, employing a structured light 3D camera to acquire high-precision data. Invalid points are removed and coordinates are standardized through preprocessing to ensure data purity. The core algorithm and sorting rules of the pulse spatial position coding layer accurately capture surface details and minute defect features of the components. A 3D equidistant grid is used to generate feature maps and maintain state vectors in real time, aiding in defect area localization. The correlation detection and dynamic fusion mechanism of the pulse cross-feature map dynamic event coding layer deeply integrates local and global information, highlighting defect correlation features. The final output is a rigorously validated fused feature that meets industrial inspection standards. The entire process strictly adheres to document parameters and algorithm requirements, efficiently extracting spatial information of components, providing strong support for dimensional verification and defect detection, and adapting to the continuous operation requirements of industrial production lines.

[0038] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for extracting spatial information based on a spiking neural network, characterized in that, The specific steps of this method are as follows: S1, Point Cloud Acquisition Preprocessing: Acquire raw 3D point cloud data, denoise the raw data, and then perform coordinate standardization on the denoised data to obtain effective 3D point cloud data; S2, Spatial Position Pulse Encoding: Input effective 3D point cloud data into the pulse spatial position encoding layer, use a three-dimensional coupled pulse frequency encoding algorithm to obtain the pulse firing frequency, combine spatial sorting rules to encode the pulse firing time, and generate pulse features containing local spatial details; S3, Feature map partitioning and maintenance: Divide the pulse features according to the spatial sub-region distribution of the 3D point cloud to obtain multiple feature maps corresponding to different spatial sub-regions. Each feature map maintains a state vector representing its own pulse activity intensity in real time. S4, Cross-Graph Global Dynamic Fusion: Input all feature maps into the pulse cross-feature map dynamic event encoding layer, use the mutual information entropy dynamic weight evolution algorithm to detect the pulse activity correlation pattern of any two feature maps, dynamically trigger temporary high-bandwidth connection channels based on the correlation determination results and adjust the connection weights in real time, realize the global spatial context transfer and fusion of pulse features between different feature maps through the connection weights, and output fused pulse features containing local details and global dynamic spatial relationships. S5, fusion feature output complete: The fused pulse features are output through the output layer of the spatially enhanced spiking neural network, providing spatial information support for subsequent spatial computing-related target detection, segmentation and recognition tasks, and completing the extraction process of 3D point cloud spatial information.

2. The method for extracting spatial information based on a spiking neural network according to claim 1, characterized in that, In S2, the spatial position pulse coding layer includes an input module, a frequency coding module, a time coding module, and an output module. The input module receives standardized three-dimensional coordinate data of valid 3D point clouds and transmits them sequentially according to the point cloud sequence number using a serial reading method. The data transmission rate is not less than 200MB / s. The frequency encoding module has a built-in three-dimensional coupled pulse frequency encoding algorithm operation unit, which performs frequency calculations with 32-bit floating-point precision and supports parallel processing of coordinate data of multiple points; the time encoding module stores preset spatial sorting rules, sorts the point cloud by coordinate weighted sum and encodes the pulse emission time of each point. The output module integrates the binary pulse sequence of each point according to the sorted number, generates a pulse feature matrix, and outputs it to step S3.

3. The method for extracting spatial information based on a spiking neural network according to claim 1, characterized in that, In S2, the mathematical expression for the three-dimensional coupled pulse frequency coding algorithm in spatial position pulse coding is: ， in For pulse delivery frequency, , , To standardize the three-dimensional coordinates of effective 3D point cloud data, Based on the pulse frequency, The hyperbolic tangent nonlinear activation function is... For nonlinear gain coefficients, , , For independent weights of three-dimensional coordinates, The three-dimensional coordinate coupling coefficient, This is a tiny offset.

4. The method for extracting spatial information based on a spiking neural network according to claim 1, characterized in that, In S2, the specific process of generating pulse features containing local spatial details by combining spatial sorting rules with the pulse firing time encoding is as follows: weighted independently according to three-dimensional coordinates... , , With standardized three-dimensional coordinates , , Calculate the weighted sum for each point , It is a weighted sum of the coordinates of each 3D point, applied to all valid points by weighted summation. The data is globally sorted from smallest to largest and assigned a unique sort number, starting from 1 and increasing sequentially. Based on an encoding period of 100 milliseconds and a time resolution of 1 millisecond, the sort numbers are... The first pulse emission time point is set to The pulse emission time points are shifted sequentially according to the pulse emission frequency obtained by the three-dimensional coupled pulse frequency encoding algorithm. The interval between each pulse is 1000 milliseconds divided by the pulse emission frequency. Within a 100-millisecond encoding period, each point generates a binary pulse sequence. The position corresponding to the emission time point in the sequence is marked as 1, and the other positions are marked as 0. The binary pulse sequences of all points are concatenated in sorted order to form a pulse feature matrix with a dimension of the number of effective point cloud points multiplied by 100 time steps. The pulse feature matrix represents the spatial positional relationship of different points through the difference in emission time points, and together with the difference in pulse emission frequency, it carries local spatial detail information.

5. The method for extracting spatial information based on a spiking neural network according to claim 1, characterized in that, In S3, during feature map partitioning and maintenance, the specific process of feature map partitioning is as follows: using standardized three-dimensional coordinates of effective 3D point cloud data... , , Based on the common value range [0, 1], the meshing operation is performed using a three-dimensional equidistant meshing method. First, set... , , The number of divisions in the three dimensions are respectively , , The calculated interval lengths of individual sub-regions in each dimension are as follows: , , Based on this, the coordinate intervals of all cube sub-regions are determined, where any cube sub-region... coordinate range , coordinate range , coordinate range , The value range is 0 to , The value range is 0 to , The value range is 0 to Each cube sub-region is assigned a unique feature map identifier; all pulse features generated in step S2 are traversed, and the normalized 3D point cloud coordinates corresponding to each pulse feature are extracted. By comparing these coordinates with the coordinate ranges of each cube sub-region, the cube sub-region to which the pulse feature belongs is determined, and the pulse feature is assigned to the feature map of the corresponding identifier; after all pulse features have been assigned, a total of [number missing] pulse features are generated. The feature map contains a complete pulse sequence of all points within the corresponding cubic sub-region. The pulse sequences are arranged in ascending order of the weighted sum of the coordinates of the corresponding points within the feature map.

6. The method for extracting spatial information based on a spiking neural network according to claim 1, characterized in that, In step S4, the cross-map global dynamic fusion process includes an input module, a correlation calculation module, a connection management module, a feature transfer and fusion module, and an output module. The input module receives all feature maps and corresponding pulse activity state vectors output in step S3, and reads data sequentially according to feature map identifiers using a parallel interface. The data transmission rate is not less than 500MB / s. The correlation calculation module has a built-in mutual information entropy calculation unit. It groups the feature maps according to pairwise pairing rules, and calculates the mutual information entropy of the state vector for each group independently. The calculation process uses 32-bit floating-point arithmetic precision and supports the simultaneous processing of no less than 1000 feature map pairings. The connection management module maintains a list of dynamic connection channels, recording the channel-associated feature map identifiers, current connection weights, and bandwidth allocation. The channel status is updated at 1ms intervals, and it supports real-time triggering, weight adjustment, and closing and releasing of channels. The feature transfer and fusion module includes a weighted operation unit and a summation and integration unit. The weighted operation unit performs time-step weighted processing on the pulse features according to the dynamic connection weights, and the summation and integration unit performs element-wise weighted summation of its own original features and associated transferred features. The output module formats and concatenates the fusion results of all feature maps to generate a global fused pulse feature matrix and caches it.

7. The method for extracting spatial information based on a spiking neural network according to claim 1, characterized in that, In S4, during cross-graph global dynamic fusion, the mathematical expression for the mutual information entropy dynamic weight evolution algorithm is: ,in for Time-of-flight feature map and Dynamic connection weights, For correlation gain factor, for The mutual information entropy of the pulse sequences of two feature maps at time 1. To update the weighting percentage, for Connection weights at time intervals, The dynamic attenuation coefficient, For time step, The maximum value of mutual information entropy. The correlation trigger threshold, It is an exponential function. The basic attenuation coefficient.

8. The method for extracting spatial information based on a spiking neural network according to claim 1, characterized in that, S4, the operation based on the correlation determination result in cross-graph global dynamic fusion, specifically involves: when the mutual information entropy of two feature maps... When a temporary high-bandwidth connection channel is immediately triggered between the two, the correlation gain factor and dynamic attenuation coefficient are calculated according to the mutual information entropy dynamic weight evolution algorithm to update the connection weight; when At that time, only the existing connection weights are attenuated until the weight is less than 0.05, at which point the corresponding channel is closed.

9. The method for extracting spatial information based on a spiking neural network according to claim 1, characterized in that, S4, the specific process of pulse feature transfer and fusion between different feature maps in cross-graph global dynamic fusion through connection weights is as follows: Each feature map, through a triggered temporary high-bandwidth connection channel, weights its own pulse feature matrix according to the corresponding dynamic connection weights, and then transfers the weighted pulse features to all feature maps with related relationships; after each feature map collects the weighted pulse features transferred from all related feature maps, it performs a weighted summation operation on these transferred features and its own original pulse features, where the weight of its own original pulse features is set to 1, and the weight of the transferred features from related feature maps is the corresponding dynamic connection weight; after all feature maps have completed the weighted summation of their own original features and related transferred features, the fusion features of all feature maps are spliced ​​and integrated according to the three-dimensional position order of the corresponding spatial sub-regions of each feature map, and finally form a global fusion pulse feature matrix containing local spatial details of each point and global dynamic spatial relationships between different sub-regions.

10. The method for extracting spatial information based on a spiking neural network according to claim 1, characterized in that, In step S5, the fusion feature output is completed. The fusion pulse feature is a matrix data with a dimension equal to the number of effective 3D point cloud points multiplied by 100 time steps. Each element in the matrix is ​​a binary value of 0 or 1. 0 indicates that no pulse is emitted at the corresponding time step, and 1 indicates that a pulse is emitted at the corresponding time step. Each row in the feature matrix corresponds to the pulse sequence of a single 3D point, carrying the local spatial topology details and coordinate association information of that point. Each column corresponds to the pulse firing status of all 3D points at the same time step, reflecting the synchronicity of pulse activity in different spatial sub-regions. The data type of the fused pulse features is 32-bit floating point, with a numerical range limited to 0 to 1, and a row-major storage method is adopted.

Images