Multi-agent distributed perception fusion system for underground rapid target detection

Abstract

The present invention provides a multi-agent distributed perception fusion system for underground rapid target detection, belonging to the field of multi-agent distributed perception fusion. The system includes: a compound eye-like data acquisition module for rapidly perceiving underground spatial data using multiple agents equipped with multiple types of sensors and mimicking the visual system of insects; wherein the multiple types of sensors include: inertial sensors, visual sensors with compound eye lenses, and infrared sensors with compound eye lenses; the acquired underground spatial data includes: the status of the multiple agents and images of underground rapid targets; and a distributed perception data fusion module for fusing the underground spatial data perceived by the compound eye-like data acquisition module in spatial and temporal dimensions using a distributed perception data fusion strategy to obtain a global information map for underground rapid target detection. The present invention can greatly improve the efficiency of underground rapid target detection.

Description

Technical Field

[0001] The present invention relates to the field of multi-agent distributed perception fusion, and in particular to a multi-agent distributed perception fusion system for underground rapid target detection. Background Art

[0002] In recent years, multi-agent distributed perception technology has garnered significant attention. It aims to leverage modern technologies such as reinforcement learning algorithms and communication technologies to achieve collaborative tracking of multiple targets. By leveraging the autonomy and mobility of multi-agents and their adaptability to complex environments, it aims to comprehensively address the numerous challenges currently facing multi-agent distributed perception. The ever-expanding application areas of multi-agents are placing increasing demands on the efficiency of distributed perception algorithms. Only by accurately and in real time acquiring data can multi-agents be able to rapidly and autonomously adapt to diverse application areas.

[0003] Current multi-agent distributed perception technology involves agents understanding their own state and the state of their targets. However, due to limitations in data acquisition speed, traditional multi-agent distributed perception strategies often neglect the detection of fast-moving targets (referred to as "fast targets") and are no longer sufficient for underground fast-moving target detection. Therefore, distributed perception of fast targets is being widely researched.

[0004] Detecting fast-moving underground targets requires faster image acquisition and more agile multi-agent decision-making compared to conventional target detection. Multi-agents must rapidly acquire real-time state space based on the low-light underground environment (including hidden inanimate and animate objects) and develop globally optimal strategies to avoid multi-agent collisions and attacks from fast-moving targets. Compared to conventional mobile target detection, detecting fast-moving underground targets requires a larger state space. Therefore, designing a rational multi-agent distributed perception fusion strategy requires critical considerations, including sensor data acquisition efficiency and efficient fusion strategies for perception data.

[0005] However, most current sensor data collection efficiency is low, the collected images are not clear enough, and there are problems such as data sparsity, slow collection efficiency, and redundant data structure. These deficiencies restrict the accuracy of multi-agent distributed perception fusion strategies, and thus reduce the efficiency of multi-agents in completing rapid target detection. Summary of the Invention

[0006] The embodiments of the present invention provide a multi-agent distributed perception fusion system for underground fast target detection, which can greatly improve the efficiency of underground fast target detection.

[0007] The multi-agent distributed perception fusion system for underground rapid target detection provided by an embodiment of the present invention includes:

[0008] A compound eye-like data acquisition module is used to rapidly perceive underground space data using a multi-agent system equipped with multiple sensors and mimicking the visual system of insects. The multiple sensors include inertial sensors, visual sensors with compound eye lenses, and infrared sensors with compound eye lenses. The acquired underground space data includes the status of the multi-agent system and images of underground fast-moving targets, which include visual images and infrared images.

[0009] The distributed perception data fusion module is used to fuse the underground space data perceived by the compound eye-like data acquisition module in spatial and temporal dimensions using a distributed perception data fusion strategy to obtain a global information map for underground rapid target detection.

[0010] Furthermore, the inertial sensor is used to obtain the state of the multi-agent to form an inertial data set; wherein the state of the multi-agent includes: the acceleration and angular velocity of the multi-agent;

[0011] The visual sensor is used to imitate the insect visual system to sense low-intensity light underground, obtain compound eye images of underground fast targets, and form a compound eye image set;

[0012] The infrared sensor is used to divide the underground space into a bright area and a dark area, obtain infrared images of underground fast targets, and form an infrared image set.

[0013] Furthermore, the distributed perception data fusion module is used to preprocess the inertial data set, compound eye image set and infrared image set respectively, input the preprocessed inertial data set, compound eye image set and infrared image set into the spatial fusion network to fuse spatial features, and then fuse the vectors output by the spatial fusion network at different times in the time dimension to obtain a multi-agent distributed perception map Map for underground rapid target detection.

[0014] Furthermore, the distributed perception data fusion module is specifically used to preprocess the inertial data set using a multidimensional Kalman algorithm based on an autoregressive sliding average model.

[0015] Furthermore, the distributed perception data fusion module is specifically used to convert the underground RGB image acquired by the visual sensor into an HSV image, so that the low-light underground RGB image is converted into a low-brightness channel image through HSV conversion; the feature information of the low-brightness channel image is extracted by connecting two 3×3 convolutional layers in series to obtain feature L1, L1 is input into the brightening residual structure to obtain the residual image feature L2, and L1 and L2 are fused to obtain feature L3; wherein the size of feature L1 is X L ×Y L ×Z L , X LIndicates the width of the feature, Y L Indicates the height of the feature, Z L The number of channels representing the features, L1 and L2 have different brightness information; feature L3 is fused with L1 and L2 respectively, and the fusion results are input into the brightening residual structure to obtain features L4 and L5 respectively; the five features L1, L2, L3, L4 and L5 containing different structural information and brightness information are element-wise added, and then processed by the convolution layer and added to the original low-brightness channel image to obtain the final brightened image.

[0016] Furthermore, the distributed perception data fusion module is specifically used to perform pixel value mapping according to the grayscale of the image to achieve preprocessing of the infrared image set.

[0017] Furthermore, the spatial fusion network includes: a first convolutional layer, a first attention module, a second convolutional layer, a second attention module, a third convolutional layer and a third attention module;

[0018] The distributed perception data fusion module is specifically used to process the preprocessed inertial data set using the first convolutional layer and the first attention module to obtain a multi-agent motion state set MA1 with different weights:

[0019] MA1=X1(E,δ1)

[0020] Where X1 represents the mapping parameter between the preprocessed inertial data set and MA1, E represents the preprocessed inertial data set, and δ1 represents the training weight parameter of the mapping parameter X1;

[0021] The preprocessed compound eye image set is processed using the second convolutional layer and the second attention module to obtain a high-resolution scene set MA2 with different weights:

[0022] MA2=X2(F,δ2)

[0023] Where X2 represents the mapping parameter between the preprocessed compound eye image set and MA2, F represents the preprocessed compound eye image set, and δ2 represents the training weight parameter of the mapping parameter X2;

[0024] The preprocessed infrared image set is processed using the third convolutional layer and the third attention module to obtain the high-definition infrared imaging set MA3 with different weights:

[0025] MA3=X3(G,δ3)

[0026] Among them, X3 represents the mapping parameter between the preprocessed infrared image set and MA3, G represents the preprocessed infrared data set, and δ3 represents the training weight parameter of the mapping parameter X3;

[0027] The spatial fusion feature vector S is obtained by fusing the spatial features of the multi-agent motion state set, high-resolution scene set and high-definition infrared imaging set with different weights. f .

[0028] Furthermore, the spatial fusion feature vector S f Expressed as:

[0029] S f =ω1·MA1+ω2·MA2+ω3·MA3

[0030] Among them, ω1 is the spatial fusion feature vector S of MA1 f The weight coefficient of MA2 is the spatial fusion feature vector S f The weight coefficient of MA3 is the spatial fusion feature vector S f The weight coefficient of .

[0031] Furthermore, the distributed perception data fusion module is specifically used to combine the spatial fusion feature vectors S output by the spatial fusion network at different times f Through fusion, we can obtain a multi-agent distributed perception map for underground rapid target detection:

[0032]

[0033] Among them, S f1 ,S f2 ,...S fm is the spatial fusion feature vector of m moments, corresponding to time t1, t2, ..., t m , t1<t2<…<t m ;z i Represents the weight of the i-th moment, and the weight of each moment is proportional to the exponential function of the current time, that is, a is a constant used to control the rate of weight decrease.

[0034] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0035] 1. The compound eye-like data acquisition scheme proposed in this embodiment has the advantages of being fast and having a wide viewing angle, enabling distributed multi-agents to sensitively and quickly detect fast-moving targets, improving the ability of multi-agents to autonomously adapt to complex and low-light underground environments, and is more suitable for low-light underground fast-moving target detection scenarios.

[0036] 2. The distributed sensing data fusion strategy proposed in this embodiment can process low-resolution images quickly captured by sensors, extract effective features to obtain high-resolution images, and solve problems such as low color saturation and loss of detail caused by weak underground lighting. This strategy greatly helps in obtaining rich, high-definition underground spatial information, and has the advantages of high accuracy, rapid information acquisition, and wide applicability.

[0037] 3. The distributed perception data fusion strategy proposed in this embodiment fuses the multi-agent motion state set, high-resolution scene set and high-definition infrared imaging set with different weights in spatial and temporal dimensions, effectively improving the detection efficiency of underground fast targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0039] Figure 1 A schematic diagram of the structure of a multi-agent distributed perception fusion system for rapid underground target detection provided by an embodiment of the present invention;

[0040] Figure 2 A schematic diagram of a residual network framework for brightening low-light images provided by an embodiment of the present invention;

[0041] Figure 3 A schematic diagram of the workflow of the distributed perception data fusion module provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0043] The purpose of this embodiment is to provide an accurate and complete multi-agent distributed perception fusion system for underground space detection with low light and fast-moving targets, laying the foundation for rapid target detection in underground spaces.

[0044] The multi-agent distributed perception fusion system for underground fast target detection described in this embodiment must not only consider the accuracy of perception fusion, but also the characteristics of weak light, complex environment, and fast target movement underground. In addition, the data acquisition technology should have the advantages of fast speed and wide viewing angle, and be able to capture fast targets well. Therefore, the multi-agent distributed perception fusion system T for underground fast target detection proposed in this embodiment consists of two parts: a compound eye-like data acquisition module C and a distributed perception data fusion module D. Figure 1 As shown, that is:

[0045] T=(C,D) (1)

[0046] The compound eye-like data acquisition module C is used to rapidly perceive underground space data using a multi-agent system equipped with multiple sensors and mimicking the insect vision system. The multiple sensors include inertial sensors, visual sensors with compound eye lenses, and infrared sensors with compound eye lenses. The acquired underground space data includes the states of the multi-agent system and images of underground fast-moving targets, which include visual images and infrared images.

[0047] The distributed perception data fusion module D is used to fuse the underground space data perceived by the compound eye-like data acquisition module in spatial and temporal dimensions using a distributed perception data fusion strategy to obtain a global information map for underground rapid target detection.

[0048] In this embodiment, the inertial sensor is equipped with a high-precision 3-axis accelerometer and a 3-axis gyroscope, which are used to obtain the status of multiple intelligent agents and form an inertial data set; among them, the 3-axis accelerometer measures the acceleration of multiple intelligent agents; the 3-axis gyroscope measures the angular velocity of multiple intelligent agents.

[0049] In this embodiment, the visual sensor is used to imitate the insect visual system to sense low-intensity light underground, obtain compound eye images of underground fast targets, and form a compound eye image set; the visual sensor includes: a compound eye lens, a visual processing tool and a communication component; among them, the working principle of the compound eye lens is similar to that of an insect compound eye. This component has a strong focusing function, can sense low-intensity light underground, and can accurately capture fast targets, but has a low resolution. It is mainly responsible for collecting underground images and sending them to the visual processing component; the visual processing tool is responsible for the preliminary preprocessing of the collected images, mainly including operations such as denoising, edge detection, and target recognition; the communication component sends data via a serial connection.

[0050] In this embodiment, the infrared sensor is used to divide the underground space into bright and dark areas, obtain infrared images of underground fast-moving targets, and form an infrared image set; the infrared sensor includes: a compound eye lens, a detection element and a conversion circuit; among them, the main function of the compound eye lens is to divide the underground space into bright and dark areas, and can quickly transmit fast-moving targets (people, objects) in the underground space to the detection element in the form of temperature changes. The detection element receives the signal transmitted by the compound eye lens, amplifies and encodes it, and finally the conversion circuit converts the result of the detection element into an electrical signal for output.

[0051] From the above scheme, we can see that, in response to the characteristics of weak light and fast target movement underground, inspired by the working principle of insect compound eyes, a compound eye-like data acquisition scheme was designed. It has the advantages of speed and wide viewing angle. Multi-agents integrate multiple types of sensors and imitate the visual system of insects. By reducing the resolution, it can quickly capture fast targets underground and form multi-angle imaging of the underground. However, the resolution of the acquired images is low, and the subsequent distributed perception fusion module will process these low-resolution images.

[0052] In this embodiment, the workflow of the distributed perception data fusion module includes:

[0053] Step A1: inputting the inertial data set, compound eye image set, and infrared image set collected by the inertial sensor, visual sensor, and infrared sensor, and preprocessing the input data;

[0054] In this embodiment, a multidimensional Kalman algorithm based on an autoregressive moving average model is used to preprocess the input inertial data set, the purpose of which is to reduce the noise of the gyroscope.

[0055] In this embodiment, the preprocessing process of the compound eye image set collected by the visual sensor is relatively complicated. This is because the characteristics of weak light in underground space result in the collection of weak light and low resolution images. The purpose of this preprocessing is to deal with weak light images with deficiencies such as loss of detail information, uneven brightening, and blurred edges in the process of underground space data collection. The enhancement function of weak light images is mainly achieved based on the brightening residual network. The framework diagram of the brightening residual network for weak light images is shown in the figure below. Figure 2 As shown, the specific process is as follows:

[0056] a. Convert the RGB image acquired by the visual sensor into an HSV image. This allows the low-light underground RGB image to be converted into a low-brightness channel image. This is to resolve the color deviation problem easily caused by RGB images. HSV images are more in line with the intuitive vision of the human eye.

[0057] b. Extract the feature information of the low-brightness channel image through two 3×3 convolutional layers in series to obtain feature L1, whose feature size is X L ×Y L ×Z L , X L Indicates the width of the feature, Z L Indicates the height of the feature, Z L Indicates the number of channels of the feature. L1 is input into the brightened residual structure to obtain the residual image feature L2. At this time, the brightness information of L1 and L2 is different. L1 and L2 are fused to obtain feature L3;

[0058] L2=Brn(L1) (2)

[0059]

[0060] Among them, Brn(·) represents the extraction operation of the brightness feature of the residual image in the incremental residual structure, Represents feature fusion operation;

[0061] c. Fuse L3 with L1 and L2 respectively, and input the fusion results into the brightening residual structure to obtain features L4 and L5 respectively; continuously neutralizing different brightness features can prevent overexposure of low-light images due to excessive brightness correction;

[0062]

[0063]

[0064] d. Add the five features L1, L2, L3, L4, and L5 containing different structural information and brightness information element-wise, and then add them to the original low-brightness channel image after processing through the convolution layer to obtain the final brightened image.

[0065] In this embodiment, the infrared image set collected by the infrared sensor is preprocessed, and pixel value mapping is mainly performed based on the grayscale of the image. The purpose is to solve the problem of low resolution of the infrared imaging image formed due to the small temperature difference between the underground object and the underground space.

[0066] Step A2: input the pre-processed inertial data set, compound eye image set and infrared image set into the spatial fusion network to fuse spatial features;

[0067] In this embodiment, since the images collected in the underground space have complex feature information, in order to obtain the main feature information of the underground fast target image, an attention mechanism (implemented by the attention module) is added to the spatial fusion network. This can effectively improve the accuracy of distributed perception fusion and reduce the amount of computation. The spatial fusion network includes: a first convolutional layer, a first attention module, a second convolutional layer, a second attention module, a third convolutional layer, and a third attention module; wherein the first convolutional layer, the second convolutional layer, and the third convolutional layer are two 3×3 convolutional layers in series. This is because the images collected by the compound eye lens for fast targets contain more information and have a lower resolution. If a single convolutional layer is used, the complete features cannot be extracted, resulting in deviations in the fusion results. The spatial fusion network uses two 3×3 convolutional layers in series. This design can effectively extract richer feature information.

[0068] In this embodiment, Figure 3 As shown, the step A2 may specifically include the following steps:

[0069] The first convolutional layer and the first attention module are used to process the preprocessed inertial data set. The first convolutional layer is used to perform a convolution operation on the inertial data set to extract feature information from the data. The second convolutional layer is then used to perform further convolution operations on the feature information to extract more advanced features. Finally, the first attention module is used to enhance the adaptability of the network and output a multi-agent motion state set MA1 containing different weights:

[0070] MA1=X1(E,δ1) (6)

[0071] Where X1 represents the mapping parameter between the preprocessed inertial data set and MA1, E represents the preprocessed inertial data set, and δ1 represents the training weight parameter of the mapping parameter X1;

[0072] The preprocessed compound eye image set is processed using the second convolutional layer and the second attention module. First, the preprocessed compound eye image set is input into the first convolutional layer to extract some basic features; then, the output of the first convolutional layer is sent to the second convolutional layer to further extract high-level features of the image; then, the output of the second convolutional layer is sent to the second attention module. During the training process, the attention mechanism adds corresponding weights to each input feature map and outputs a high-resolution scene set MA2 containing different weights:

[0073] MA2=X2(F,δ2) (7)

[0074] Where X2 represents the mapping parameter between the preprocessed compound eye image set and MA2, F represents the preprocessed compound eye image set, and δ2 represents the training weight parameter of the mapping parameter X2;

[0075] The preprocessed infrared image set is processed using the third convolutional layer and the third attention module. First, the preprocessed infrared image set is input into the first convolutional layer to extract basic features. Then, the output of the first convolutional layer is input into the second convolutional layer to further extract higher-level features. Finally, the output of the second convolutional layer is passed to the third attention module to adaptively adjust the weights and output a high-definition infrared imaging set MA3 containing different weights:

[0076] MA3=X3(G,δ3) (8)

[0077] Among them, X3 represents the mapping parameter between the preprocessed infrared image set and MA3, G represents the preprocessed infrared data set, and δ3 represents the training weight parameter of the mapping parameter X3;

[0078] The spatial fusion feature vector S is obtained by fusing the spatial features of the multi-agent motion state set, high-resolution scene set and high-definition infrared imaging set with different weights. f :

[0079] S f =ω1·MA1+ω2·MA2+ω3·MA3 (9)

[0080] Among them, ω1 is the spatial fusion feature vector S of MA1 f The weight coefficient of MA2 is the spatial fusion feature vector S f The weight coefficient of MA3 is the spatial fusion feature vector S f The weight coefficient of .

[0081] Step A3: The vectors output by the spatial fusion network at different times are fused in the time dimension to obtain a multi-agent distributed perception map for underground rapid target detection:

[0082]

[0083] Among them, S f1 ,S f2 ,...S fm is the spatial fusion feature vector of m moments, corresponding to time t1, t2, ..., t m , t1<t2<…<t m ;z i Represents the weight of the i-th moment, and the weight of each moment is proportional to the exponential function of the current time, that is, a is a constant used to control the rate of weight decrease.

[0084] In summary, the multi-agent distributed perception fusion system for underground rapid target detection described in the embodiments of the present invention has at least the following beneficial effects:

[0085] 1. The compound eye-like data acquisition scheme proposed in this embodiment has the advantages of being fast and having a wide viewing angle, enabling distributed multi-agents to sensitively and quickly detect fast-moving targets, improving the ability of multi-agents to autonomously adapt to complex and low-light underground environments, and is more suitable for low-light underground fast-moving target detection scenarios.

[0086] 2. The distributed sensing data fusion strategy proposed in this embodiment can process low-resolution images quickly captured by sensors, extract effective features to obtain high-resolution images, and solve problems such as low color saturation and loss of detail caused by weak underground lighting. This strategy greatly helps in obtaining rich, high-definition underground spatial information, and has the advantages of high accuracy, rapid information acquisition, and wide applicability.

[0087] 3. The distributed perception data fusion strategy proposed in this embodiment fuses the multi-agent motion state set, high-resolution scene set and high-definition infrared imaging set with different weights in spatial and temporal dimensions, effectively improving the detection efficiency of underground fast targets.

[0088] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-agent distributed perception fusion system for underground rapid target detection, characterized by: include: A compound eye-like data acquisition module is used to rapidly perceive underground space data using a multi-agent system equipped with multiple sensors and mimicking the visual system of insects. The multiple sensors include inertial sensors, visual sensors with compound eye lenses, and infrared sensors with compound eye lenses. The acquired underground space data includes the status of the multi-agent system and images of underground fast-moving targets, which include visual images and infrared images. The distributed perception data fusion module is used to fuse the underground spatial data perceived by the compound eye-like data acquisition module in spatial and temporal dimensions using a distributed perception data fusion strategy to obtain a global information map for underground rapid target detection. The distributed perception data fusion module is specifically used to convert the underground RGB image acquired by the visual sensor into an HSV image, so that the low-light underground RGB image is converted into a low-brightness channel image through HSV conversion; the feature information of the low-brightness channel image is extracted by connecting two 3×3 convolutional layers in series to obtain feature L1, and L1 is input into the brightening residual structure to obtain the residual image feature L2, and L1 and L2 are fused to obtain feature L3; wherein the size of feature L1 is X L ×Y L ×Z L , X L Indicates the width of the feature, Y L Indicates the height of the feature, Z L The number of channels representing the features, L1 and L2 have different brightness information; feature L3 is fused with L1 and L2 respectively, and the fusion results are input into the brightening residual structure to obtain features L4 and L5 respectively; the five features L1, L2, L3, L4 and L5 containing different structural information and brightness information are element-wise added, and then processed by the convolution layer and added to the original low-brightness channel image to obtain the final brightened image.

2. The multi-agent distributed perception fusion system for underground rapid target detection according to claim 1 is characterized in that: The inertial sensor is used to obtain the state of the multi-agent to form an inertial data set; wherein the state of the multi-agent includes: the acceleration and angular velocity of the multi-agent; The visual sensor is used to imitate the insect visual system to sense low-intensity light underground, obtain compound eye images of underground fast targets, and form a compound eye image set; The infrared sensor is used to divide the underground space into a bright area and a dark area, obtain infrared images of underground fast targets, and form an infrared image set.

3. The multi-agent distributed perception fusion system for underground rapid target detection according to claim 2 is characterized in that: The distributed perception data fusion module is used to preprocess the inertial data set, compound eye image set and infrared image set respectively, input the preprocessed inertial data set, compound eye image set and infrared image set into the spatial fusion network to fuse spatial features, and then fuse the vectors output by the spatial fusion network at different times in the time dimension to obtain a multi-agent distributed perception map Map for underground rapid target detection.

4. The multi-agent distributed perception fusion system for underground rapid target detection according to claim 1 is characterized in that: The distributed perception data fusion module is specifically used to preprocess the inertial data set using a multidimensional Kalman algorithm based on an autoregressive sliding average model.

5. The multi-agent distributed perception fusion system for underground rapid target detection according to claim 1 is characterized in that: The distributed perception data fusion module is specifically used to perform pixel value mapping according to the grayscale of the image to achieve preprocessing operations of the infrared image set.

6. The multi-agent distributed perception fusion system for underground rapid target detection according to claim 3 is characterized in that: The spatial fusion network includes: a first convolutional layer, a first attention module, a second convolutional layer, a second attention module, a third convolutional layer and a third attention module; The distributed perception data fusion module is specifically used to process the preprocessed inertial data set using the first convolutional layer and the first attention module to obtain a multi-agent motion state set MA1 with different weights: MA1=X1(E,δ1) Where X1 represents the mapping parameter between the preprocessed inertial data set and MA1, E represents the preprocessed inertial data set, and δ1 represents the training weight parameter of the mapping parameter X1; The preprocessed compound eye image set is processed using the second convolutional layer and the second attention module to obtain a high-resolution scene set MA2 with different weights: MA2=X2(F,δ2) Where X2 represents the mapping parameter between the preprocessed compound eye image set and MA2, F represents the preprocessed compound eye image set, and δ2 represents the training weight parameter of the mapping parameter X2; The preprocessed infrared image set is processed using the third convolutional layer and the third attention module to obtain the high-definition infrared imaging set MA3 with different weights: MA3=X3(G,δ3) Among them, X3 represents the mapping parameter between the preprocessed infrared image set and MA3, G represents the preprocessed infrared data set, and δ3 represents the training weight parameter of the mapping parameter X3; The spatial fusion feature vector S is obtained by fusing the spatial features of the multi-agent motion state set, high-resolution scene set and high-definition infrared imaging set with different weights. f .

7. The multi-agent distributed perception fusion system for underground rapid target detection according to claim 6 is characterized in that: Spatial fusion feature vector S f Expressed as: <h2 style=";text-align:left;direction:ltr">S<h2 style=";text-align:left;direction:ltr"> f <h2 style=";text-align:left;direction:ltr"> (ω1·MA1+ω2·MA2+ω3·MA3) Among them, ω1 is the spatial fusion feature vector S of MA1 f The weight coefficient of MA2 is the spatial fusion feature vector S f The weight coefficient of MA3 is the spatial fusion feature vector S f The weight coefficient of .

8. The multi-agent distributed perception fusion system for underground rapid target detection according to claim 7 is characterized in that: The distributed perception data fusion module is specifically used to combine the spatial fusion feature vectors S output by the spatial fusion network at different times f Through fusion, we can obtain a multi-agent distributed perception map for underground rapid target detection: Among them, S f1 ,S f2 ,...S fm is the spatial fusion feature vector of m moments, corresponding to time t1, t2, ..., t m , t1<t2<…<t m ;z i Represents the weight of the i-th moment, and the weight of each moment is proportional to the exponential function of the current time, that is, a is a constant used to control the rate of weight decrease.

Images