Target identification method based on time-space separation pulse Transform, electronic equipment and storage medium

Abstract

The invention discloses a time-space separation pulse Transform-based target recognition method, electronic equipment and a storage medium. The method comprises the steps of obtaining event stream data which is output by a target neuromorphic visual sensor and is used for target recognition; inputting the event stream data into a pre-trained target pulse neural network, performing pulse coding on the event stream data through a coding module, and outputting pulse coding features; performing local space-time feature extraction on the pulse coding features through a feature extraction module, and outputting local space-time pulse features; performing global spatio-temporal feature extraction on the local spatio-temporal pulse features through a spatio-temporal separation pulse Transform module, and outputting global spatio-temporal enhancement features; and performing linear mapping on the global space-time enhancement features through a task output module, and outputting a target recognition result. According to the invention, while event-driven sparse calculation and low power consumption characteristics are maintained, the discrimination stability of weak, small, sparse and slow moving targets is significantly improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of neuromorphic computing and machine vision, and particularly relates to a target recognition method based on a space-time separation pulse Transformer, an electronic device and a storage medium. BACKGROUND

[0002] In an autonomous driving system, a vehicle relies on high-precision sensors for environment perception. Traditional vision sensors perform poorly in low-light and dynamic scenes. Event cameras, which can capture microsecond-level temporal changes and have low power consumption, have become one of the ideal sensors in autonomous driving. However, in order to effectively process the high-frequency dynamic data generated by event cameras, traditional neural network methods often need to process a large amount of static data, and cannot fully utilize the advantages of event cameras.

[0003] Spiking Neural Networks (SNNs) are the third generation of neural networks inspired by biological neural systems, and are particularly suitable for processing such efficient and sparse event-driven data. SNNs use discrete spikes as information carriers and simulate neuron activation and synaptic connections, enabling fine coding of information through spike timing intervals and asynchronous processing of inputs, greatly reducing computational and communication energy consumption. SNN neurons typically incorporate biological regulatory mechanisms such as leakage and refractory periods, enhancing non-linear representation and short-term memory capabilities, and have good temporal pulse processing characteristics. Based on the event-driven computing paradigm, SNNs only operate on activated neurons, with inactive neurons in a resting state, significantly reducing computational and communication energy consumption during inference, and having low power consumption advantages.

[0004] Existing SNNs are mostly constructed with convolutional layers and pulse neurons, extracting spatio-temporal features through local receptive fields, which can suppress noise and capture local patterns to some extent. However, such traditional SNNs lack modeling of long-range dependencies: convolution or local operations are inherently limited by kernel size and stacking depth, making it difficult to effectively capture global dependencies across large temporal steps and large-scale spatial positions, resulting in unstable recognition of small targets, slow-moving targets, or sparse structures.

[0005] To improve non-local modeling capabilities, existing research has preliminarily introduced attention mechanisms into SNNs, such as replacing or enhancing some convolutional operations with self-attention pulse network schemes. Such methods have made some improvements in global interactions in the spatial domain, but overall still have the following common problems: 1) Complexity and stability of spatio-temporal coupled attention: directly extracting spatio-temporal features in the pulse domain has high computational cost and is sensitive to alternative gradient training, making deployment difficult; 2) Insufficient modeling of explicit temporal relationships: Most attention designs focus more on spatial global relationships and lack targeted mechanisms for feature correlations across time steps (such as cross-frame consistency and temporal residual dependence).

[0006] In summary, existing technologies either remain at the level of local convolutional spatiotemporal representation, making it difficult to cover long-range dependencies; or they introduce attention but fail to fully utilize temporal feature associations, resulting in limitations in complexity and training stability. Therefore, there is an urgent need for a network architecture that explicitly decouples temporal and spatial relationships in the pulse domain, balancing low overhead and training stability. This would allow for the improvement of the ability to identify long-range spatiotemporal dependencies and weak targets while maintaining the low-power advantage of SNNs, and enhancing deployability on edge devices. Summary of the Invention

[0007] To address the aforementioned problems in the prior art, this invention provides a target recognition method, electronic device, and storage medium based on a spatiotemporal separation pulse Transformer.

[0008] The technical problem to be solved by this invention is achieved through the following technical solution: In a first aspect, the present invention provides a target recognition method based on a spatiotemporal separation pulse Transformer, comprising: Acquire event stream data from the target neuromorphic vision sensor for target recognition; The event stream data is input into a pre-trained target spiking neural network, which includes an encoding module, a feature extraction module, a spatiotemporal separation spiking Transformer module, and a task output module. The event stream data is pulse-coded by the encoding module, and the pulse-coded features are output. The local spatiotemporal features of the pulse-coded features are extracted by the feature extraction module, and the local spatiotemporal pulse features are output. The global spatiotemporal features of the local spatiotemporal pulse features are extracted by the spatiotemporal separation pulse Transformer module, and the global spatiotemporal enhancement features are output. The target recognition result is output by linearly mapping the global spatiotemporal enhancement features through the task output module.

[0009] In a second aspect, the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a program stored in memory, it implements the method described in the first aspect.

[0010] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0011] This invention provides a target recognition method, electronic device, and storage medium based on a spatiotemporally separated pulse Transformer. By sequentially introducing pulse feature enhancement based on local convolution, spatial awareness feature fusion, and spatiotemporally separated modeling of temporal and spatial self-attention into the target spiking neural network, it achieves efficient characterization of long-range spatiotemporal dependencies in the pulse domain. Compared with existing schemes that simply embed attention operators in SNNs, this invention significantly improves the discrimination stability when dealing with weak, sparse, and slowly moving targets while maintaining event-driven sparse computation and low power consumption characteristics.

[0012] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating a target recognition method based on a spatiotemporal separation pulse Transformer provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a target spiking neural network for a target recognition method based on a spatiotemporal separation spiking Transformer provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0014] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0015] In a first aspect, embodiments of the present invention provide a target recognition method based on a spatiotemporal separation pulse Transformer. See also... Figure 1 and Figure 2 The method includes the following steps: S10. Acquire event stream data output by the target neuromorphic vision sensor for target recognition.

[0016] For example, the target neuromorphic vision sensor can be an event camera, which can be installed on an autonomous vehicle. The event stream data can be dynamic visual signals captured by the event camera, suitable for tasks such as obstacle and pedestrian target recognition and road condition assessment in autonomous driving.

[0017] S20. Input the event stream data into the pre-trained target spiking neural network. The encoding module performs pulse coding on the event stream data and outputs pulse coding features. The feature extraction module extracts local spatiotemporal features from the pulse coding features and outputs local spatiotemporal pulse features. The spatiotemporal separation pulse Transformer module extracts global spatiotemporal features from the local spatiotemporal pulse features and outputs global spatiotemporal enhancement features. The task output module performs linear mapping on the global spatiotemporal enhancement features and outputs the target recognition result.

[0018] The target spiking neural network includes an encoding module, a feature extraction module, a spatiotemporal separation spiking Transformer module, and a task output module.

[0019] Optionally, in step S20, the specific processing procedure of the encoding module includes: S201. Through the encoding module, the event stream data is aggregated according to a fixed time window to obtain an event image sequence with multiple time steps.

[0020] S202. The event image sequence is pulse-coded using an integral-leakage-multipulse quantization firing model, and the pulse coding features are output.

[0021] For example, the encoding module includes an Integral-Leak-Multipulse Quantization (ILIF) neuron. For the input event stream data, it is first aggregated into a multi-time-step event image sequence according to a fixed time window, and then the Integral-Leak-Multipulse Quantization (ILIF) neuron is used to perform membrane potential integration, leakage, and firing to obtain pulse-coded features.

[0022] Unlike the integral-leak neuron, which outputs 0 / 1 pulses only at a single time step, the integral-leak-multipulse quantization firing neuron allows for the output of multiple pulse counts at a single time step. And update the membrane potential with soft reset:

[0023] in, This represents the membrane potential at the current moment. As a leakage factor, For the sudden input current, This represents the quantized transmit pulse count. Indicates a time step. Representing the membrane potential at the previous moment, the discrete quantization function restricts the firing number to a finite set (such as 0 to 4), which can generate a pulse train under strong stimulation, thereby improving the time coding capability and information transmission efficiency.

[0024] Optionally, in step S20, the specific processing procedure of the feature extraction module includes: S203. Through the feature extraction module, perform two-dimensional convolution operation on the pulse code features to obtain local spatial features.

[0025] S204. Perform batch normalization on the local spatial features to obtain normalized features.

[0026] S205. An integral-leakage-multipulse quantization firing model is used to pulse encode the normalized features of multiple time steps, and output local spatiotemporal pulse features.

[0027] For example, to obtain local boundaries / textures and short-range temporal cues, and to suppress widely distributed random thermal noise in the event stream, the feature extraction module consists of four SNN basic modules stacked sequentially, each consisting of a 2D convolutional layer, a normalization layer, and an ILIF spiking neuron. This type of localization operation (convolution + spiking neuron) excels at extracting fine-grained patterns, but its ability to integrate context across larger spatial ranges and time steps is limited. It needs to be coordinated with subsequent decoupled attention to cover long-range dependencies.

[0028] Optionally, the spatiotemporal separation pulse Transformer module includes a spatially aware feature fusion layer, a pulse temporal self-attention layer, a fully connected layer based on a slow-heating strategy, and a pulse spatial self-attention layer.

[0029] In step S20, the specific processing procedure of the spatiotemporal separation pulse Transformer module may include: S206. The spatial perception feature fusion layer performs feature fusion on the local spatiotemporal pulse features and outputs a time token sequence.

[0030] Optionally, the specific processing steps of the spatial perception feature fusion layer may include: The spatially aware feature fusion layer divides the local spatiotemporal pulse features of each time step into non-overlapping windows, and performs max pooling on the local spatiotemporal pulse features within each window to output a time token sequence.

[0031] For example, to address the issues of binary sparsity and sensitivity to subtle displacements in local spatiotemporal impulse features, the local spatiotemporal impulse features at each time step are divided into non-overlapping window blocks. Max pooling is then performed within each window to aggregate the most discriminative responses, generating a time token sequence insensitive to small displacements and discrete noise. This spatial aggregation explicitly robustens the input before temporal modeling, reducing the interference of small displacements and discrete noise on temporal correlation measurements, thereby improving the quality of spatiotemporal modeling in subsequent modules.

[0032] S207. Through the pulse time-series self-attention layer, cross-time step correlation of the time token sequence is obtained in the time dimension, and the time-series enhancement features are output.

[0033] Optionally, the specific processing procedure of the pulsed temporal self-attention layer (STSA) may include: S2071. Through the pulse time series self-attention layer, the time token sequence is transformed to generate time queries, time keys and time value representations for time series correlation modeling.

[0034] S2072. Calculate the relevance weight based on time query and time key in the time dimension.

[0035] For example, based on time query and time key, the initial relevance weight is calculated in the time dimension, and the relevance weight is scaled and batch normalized to obtain the relevance weight, so as to obtain a robust time-series attention allocation.

[0036] S2073. The time value representations are weighted and aggregated according to the correlation weights to obtain the initial time-series enhanced features.

[0037] S2074. Perform pulse activation processing on the initial temporal enhancement features to obtain pulse temporal self-attention results.

[0038] S2075. Perform residual concatenation between the pulse time-series self-attention results and the time token sequence to output the time-series enhancement features.

[0039] The processing steps S2071 to S2075 in this embodiment are used to explicitly characterize the persistence and transition relationships of targets in the event stream, improving the ability to distinguish the temporal sequence of weak, sparse, and slowly moving targets. The order of the above steps can be adjusted according to actual needs and is not limited thereto.

[0040] S208. The temporal enhancement features are linearly transformed and pulsed activated by a fully connected layer based on a slow-heating strategy, and discrete pulse features are output.

[0041] Optionally, the specific processing procedure for the fully connected layer (GH-Linear) based on the slow-heating strategy may include: S2081. Through a fully connected layer based on a slow-heating strategy, the temporal enhancement features are transformed and re-encoded in the channel dimension to obtain re-encoded temporal features.

[0042] S2082. An integral-leakage-multipulse quantization firing model is used to activate the recoded temporal features and output discrete pulse features.

[0043] For example, the fully connected layer based on the slow-heating strategy includes a linear fully connected layer and an integral-leakage-multipulse quantization firing model. The linear fully connected layer uses zero / near-zero initialization of weights and biases to reduce excessive response and gradient fluctuations in the early stages of training.

[0044] A linear fully connected layer performs feature transformation and re-encoding on the temporal enhancement features along the channel dimension to obtain re-encoded temporal features. Then, an integral-leakage-multipulse quantization firing model is used to pulse activate these re-encoded temporal features, outputting sparse discrete pulse features. These discrete pulse features serve as input to subsequent modules, maintaining the sparsity and low-power characteristics of event-driven computation.

[0045] It should be noted that for fully connected layers based on a slow-burn strategy, a gradual strategy of controlled gain or controlled learning rate is adopted for the linear fully connected layer in the first few rounds of model training: initially, low-amplitude updates or partial freezing are maintained, and then the learning rate is gradually increased according to a preset curve (linear, piecewise, or cosine methods are all acceptable), so that the model first consolidates the stable representations obtained from the preceding modules, and then gradually enhances the contribution of the linear layer to feature discrimination. Fixed parameters are used for forward computation during the inference phase.

[0046] The fully connected layer based on the slow-heat strategy in this embodiment can stabilize the alternative gradient propagation in the early stage of training, suppress oscillations, and improve convergence performance.

[0047] S209. Through the pulse space self-attention layer, pulse space self-attention is performed on discrete pulse features in the spatial dimension within the same time step, and global spatiotemporal enhanced features are output.

[0048] Optionally, the specific processing procedure of the pulsed spatial self-attention layer (SSSA) may include: S2091. Through the pulse space self-attention layer, random noise suppression processing is performed on the discrete pulse features to obtain the initial spatial features.

[0049] For example, the discrete pulse features are pulsed preprocessed to suppress random noise while retaining an identity residual branch as the fusion baseline for the subsequent spatial attention output.

[0050] S2092. Perform lightweight feature transformation on the initial spatial features to obtain spatial queries, spatial keys and spatial value representations for spatial correlation modeling.

[0051] For example, lightweight feature transformation can employ 1×1 spatial convolution and normalization, followed by pulsed activation (preferably ILIF) to obtain a pulsed representation. After flattening the initial two-dimensional spatial features into a token sequence, parallel processing is performed in a multi-head manner.

[0052] S2093. Construct a correlation representation based on spatial keys and spatial values ​​in the spatial dimension.

[0053] For example, a correlation representation based on keys and values ​​is constructed in the spatial dimension, and it is recalibrated between heads to improve the collaborative expression of different attention heads.

[0054] S2094. Based on the relevance representation, the spatial query is weighted and aggregated to obtain spatial enhanced features.

[0055] For example, the computation and aggregation of correlation representations preferably employ a sparse / gated approach to reduce computational and storage overhead.

[0056] S2095. Perform linear projection on the spatial enhancement features and perform residual connection with the initial spatial features to output global spatiotemporal enhancement features.

[0057] For example, linear projection can employ 1×1 spatial convolution and normalization. After linear projection of the spatial augmentation features, it is added to the residual branch of step S2091 to obtain the global spatiotemporal augmentation features.

[0058] The processing steps S2091 to S2095 in this embodiment are used to enhance the global spatial dependency and boundary detail representation in the pulse domain, thereby improving the ability to distinguish weak, sparse, and texture-interference scenes. The order of the above steps can be adjusted according to actual needs and is not limited thereto.

[0059] Finally, the task output module can be a classifier, such as a linear classifier head, which takes global spatiotemporal augmentation features as input and outputs the target recognition result.

[0060] In this embodiment of the invention, during the training of the target spiking neural network, the task outputs generated by the same sample at all time steps are uniformly supervised and aggregated according to the time dimension, so that each time step participates in learning. The aggregation process can be averaging, weighted averaging, or adaptive aggregation based on confidence / keyframes; different supervision terms can be selected for different task types (e.g., classification loss is used for classification, and corresponding detection or mask loss is used for detection / segmentation).

[0061] To suppress random fluctuations across time steps and improve temporal stability, a consistency constraint is imposed on the output of the same sample in the time dimension, ensuring stable distribution and smooth trajectory across consecutive time steps. This consistency constraint can be based on: Constrain the deviation of the output at each time step from its time mean (or moving average, teacher time series integration result); or constrain the differences between adjacent time steps; or suppress the variance / jitter in the time dimension. The above methods can be used individually or in combination.

[0062] This invention presents a target recognition method based on a spatiotemporally separated pulse Transformer. This method achieves efficient characterization of long-range spatiotemporal dependencies in the pulse domain by sequentially introducing event-domain ILIF encoding, pulse feature enhancement based on local convolution, spatially perceptual feature fusion, spatiotemporally separated modeling of temporal and spatial self-attention, and a corresponding slow-heating linear layer with temporal consistency supervision into the target spiking neural network. Compared to existing schemes that simply embed attention operators in the SNN, this invention significantly improves the discrimination stability when dealing with weak, sparse, and slowly moving targets while maintaining event-driven sparse computation and low power consumption.

[0063] In autonomous driving applications, this method can process asynchronous event streams in real time, extracting minute and sparse target features, significantly improving the perception and reaction speed of autonomous driving systems in complex and dynamic environments. Furthermore, this invention can process time-series data generated by event cameras in real time, particularly in high-speed driving or low-light environments, significantly improving the robustness of identifying dynamic obstacles, pedestrians, and other slowly moving targets. Compared to traditional vision methods, especially in complex lighting or rapidly changing scenarios, event camera data can quickly capture minute environmental changes, helping autonomous driving systems make more accurate responses in a very short time, thereby effectively improving the safety and reliability of autonomous driving.

[0064] In addition, in experiments on different application scenarios such as defect classification and gesture recognition based on event cameras, the present invention has achieved better results than existing spiking neural network methods, and has good versatility and engineering deployment value.

[0065] Secondly, embodiments of the present invention also provide an electronic device, such as... Figure 3 As shown, it includes a processor 301, a communication interface 302, a memory 303, and a communication bus 304, wherein the processor 301, the communication interface 302, and the memory 303 communicate with each other through the communication bus 304. Memory 303 is used to store computer programs; When the processor 301 executes the program stored in the memory 303, it implements the method steps of the first aspect described above.

[0066] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus.

[0067] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0068] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0069] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0070] Thirdly, this embodiment also provides a computer-readable storage medium. A computer program is stored in this computer-readable storage medium, and when executed by a processor, the computer program implements the method steps described in the first aspect.

[0071] Optionally, the computer-readable storage medium may be non-volatile memory (NVM), such as at least one disk storage device.

[0072] Optionally, the aforementioned computer-readable storage medium may also be at least one storage device located remotely from the aforementioned processor.

[0073] Fourthly, in another embodiment of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to perform the method steps of the first aspect described above.

[0074] It should be noted that the device is basically similar to the method embodiment, so the description is relatively simple. For relevant parts, please refer to the description of the method embodiment.

[0075] It should be noted that the terms "first," "second," etc., are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention.

[0076] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0077] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings and the disclosure in carrying out the claimed invention. In the description of the invention, the word "comprising" does not exclude other components or steps, "a" or "an" does not exclude a plurality, and "a plurality" means two or more, unless otherwise explicitly specified. Furthermore, while different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce good results.

[0078] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A target recognition method based on spatiotemporal separation pulse Transformer, characterized in that, include: Acquire event stream data from the target neuromorphic vision sensor for target recognition; The event stream data is input into a pre-trained target spiking neural network, which includes an encoding module, a feature extraction module, a spatiotemporal separation spiking Transformer module, and a task output module. The event stream data is pulse-coded by the encoding module to output pulse-coded features; the pulse-coded features are extracted locally by the feature extraction module to output local spatiotemporal pulse features; the local spatiotemporal pulse Transformer module is used to extract global spatiotemporal features from the local spatiotemporal pulse features to output global spatiotemporal enhancement features; and the global spatiotemporal enhancement features are linearly mapped by the task output module to output the target recognition result.

2. The target recognition method based on spatiotemporal separation pulse Transformer according to claim 1, characterized in that, The spatiotemporal separation pulse Transformer module includes a spatial awareness feature fusion layer, a pulse temporal self-attention layer, a fully connected layer based on a slow-heating strategy, and a pulse spatial self-attention layer. The step of extracting global spatiotemporal features from the local spatiotemporal pulse features using the spatiotemporal separation pulse Transformer module and outputting global spatiotemporal enhanced features includes: The spatial perception feature fusion layer fuses the local spatiotemporal pulse features to output a time token sequence. Through the pulse-series self-attention layer, cross-time step correlation of the time token sequence is obtained in the time dimension, and time-series enhancement features are output. The temporal enhancement features are linearly transformed and pulsed activated by the fully connected layer based on the slow-heating strategy, and discrete pulse features are output. Through the pulse space self-attention layer, pulse space self-attention is performed on the discrete pulse features in the spatial dimension within the same time step, and global spatiotemporal enhanced features are output.

3. The target recognition method based on spatiotemporal separation pulse Transformer according to claim 2, characterized in that, The step of fusing the local spatiotemporal pulse features through the spatially perceptual feature fusion layer to output a time token sequence includes: The spatial awareness feature fusion layer divides the local spatiotemporal pulse features of each time step into non-overlapping windows, and performs max pooling on the local spatiotemporal pulse features within each window to output a time token sequence.

4. The target recognition method based on spatiotemporal separation pulse Transformer according to claim 2, characterized in that, The step of obtaining cross-time step correlation of the time token sequence through the pulse time-series self-attention layer and outputting time-series enhanced features includes: The pulse time-series self-attention layer performs feature transformation on the time token sequence to generate time queries, time keys, and time value representations for time-series correlation modeling. Calculate the relevance weight based on the time query and the time key in the time dimension; The time value representation is weighted and aggregated according to the correlation weight to obtain the initial time-series enhanced features; The initial temporal enhancement features are subjected to pulse activation processing to obtain pulse temporal self-attention results; The pulse timing self-attention result is residually concatenated with the time token sequence to output timing enhancement features.

5. The target recognition method based on spatiotemporal separation pulse Transformer according to claim 2, characterized in that, The process of performing linear transformation and impulse activation on the temporal enhancement features through the fully connected layer based on the slow-heating strategy to output discrete impulse features includes: The temporal enhancement features are transformed and re-encoded in the channel dimension through the fully connected layer based on the slow-heating strategy to obtain re-encoded temporal features; The recoded timing features are pulse-activated using an integral-leakage-multipulse quantization firing model to output discrete pulse features.

6. The target recognition method based on spatiotemporal separation pulse Transformer according to claim 2, characterized in that, The step of performing pulse space self-attention on the discrete pulse features in the spatial dimension within the same time step through the pulse space self-attention layer, and outputting global spatiotemporal enhanced features, includes: The discrete pulse features are subjected to random noise suppression processing through the pulse space self-attention layer to obtain the initial spatial features; Lightweight feature transformation is performed on the initial spatial features to obtain spatial queries, spatial keys, and spatial value representations for spatial correlation modeling. Based on the spatial keys and spatial value representations in the spatial dimension, a correlation representation is constructed. The spatial query is weighted and aggregated based on the relevance representation to obtain spatially enhanced features; The spatial enhancement features are linearly projected and residually connected with the initial spatial features to output global spatiotemporal enhancement features.

7. The target recognition method based on spatiotemporal separation pulse Transformer according to claim 1, characterized in that, The step of performing pulse coding on the event stream data through the encoding module and outputting pulse coding features includes: The event stream data is aggregated by the encoding module according to a fixed time window to obtain an event image sequence with multiple time steps; The event image sequence is pulse-coded using an integral-leakage-multipulse quantization firing model, and the pulse coding features are output.

8. The target recognition method based on spatiotemporal separation pulse Transformer according to claim 1, characterized in that, The step of extracting local spatiotemporal features from the pulse coding features through the feature extraction module and outputting local spatiotemporal pulse features includes: The feature extraction module performs a two-dimensional convolution operation on the pulse-code features to obtain local spatial features. The local spatial features are batch normalized to obtain normalized features; An integral-leakage-multipulse quantization firing model is used to pulse encode the normalized features of the multiple time steps, and output local spatiotemporal pulse features.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the method as described in any one of claims 1 to 8.

Images