An Infrared-Visible Light Joint Target Detection Method

Abstract

This invention discloses an infrared-visible light joint target detection method. It extracts multi-scale features through a backbone network and employs a cross-modal feature fusion module to fuse spatial and semantic information from infrared and visible light data. Simultaneously, an improved same-scale feature interaction module and a convolution-based cross-scale feature fusion module are designed to further enhance feature representation capabilities. The fused features are dynamically selected for high-quality initial object queries via a Query Selection module, and then the decoder outputs the target category and location information. This network model possesses powerful multi-modal data fusion capabilities and efficient target detection performance, making it suitable for various application scenarios such as security monitoring and autonomous driving, and exhibiting particularly excellent performance in complex environments.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and artificial intelligence, and in particular to an infrared-visible light joint target detection method. Background Technology

[0002] In the fields of computer vision and image processing, object detection is a key technology widely used in security monitoring, autonomous driving, medical diagnosis, industrial inspection, and many other areas. Traditional object detection methods mainly rely on visible light images, which perform well in well-lit and good weather conditions. However, in complex or adverse environmental conditions, such as at night, in fog, haze, or smoke, the quality of visible light images often deteriorates significantly, leading to a substantial reduction in detection performance.

[0003] Infrared imaging technology generates images by capturing the infrared radiation (thermal radiation) emitted by objects, unlike visible light imaging. Infrared images have the ability to penetrate smoke and fog and detect targets in dark environments, thus offering unique advantages in complex environments. However, infrared images often lack the detailed information of visible light images, leading to certain limitations in some applications. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an infrared-visible light joint target detection method.

[0005] The objective of this invention is achieved through the following technical solution: an infrared-visible light joint target detection method, comprising:

[0006] S1. Input the infrared light image and the visible light image into the backbone network respectively to obtain feature maps of three different spatial resolution levels for infrared light and visible light.

[0007] S2. Perform cross-modal feature fusion on the feature maps of the first two lower spatial resolution levels of the infrared light image and the feature maps of the first two lower spatial resolution levels of the visible light image to obtain a feature image after fusion of two different spatial resolution levels.

[0008] S3. Perform same-scale feature interaction on the high-level feature maps of infrared and visible light to enhance the global context representation and obtain the high-level feature map after feature interaction.

[0009] S4. Upsample the high-level feature map after feature interaction and fuse it with the mid-level fusion feature map to obtain the first fusion feature map; upsample the first fusion feature map and fuse it with the low-level fusion feature map to obtain the second fusion feature map; downsample the second fusion feature map and fuse it with the first fusion feature map again to obtain the third fusion feature map; downsample the third fusion feature map and fuse it with the high-level feature map after feature interaction to obtain the fourth fusion feature map; convolve the second, third and fourth fusion feature maps, unfold them and concatenate them to obtain the final feature map.

[0010] S5. Predict the initial class confidence for the final feature map, take the top k indices with the highest class confidence, and select the corresponding k class confidence and the corresponding k anchor boxes;

[0011] S6. Input the final feature map, k class confidence scores and the corresponding k anchor boxes into the decoder to predict the class confidence scores and predicted boxes.

[0012] Furthermore, the backbone network is PREsNet, and its output includes: feature maps with high spatial resolution, feature maps with medium spatial resolution but rich semantic information, and feature maps with low spatial resolution but rich semantic information.

[0013] Furthermore, the cross-modal feature fusion specifically involves: the input being two feature maps of the same shape, and the output being a feature map of the same shape as the input. For example, if the inputs are A1 and A2, then...

[0014] C1(x) = SiLU(BN(Conv(x)))

[0015] C2(x)=SiLu(BN(Conv(SiLU(BN(Conv(x)))))

[0016]

[0017] Here, `concat` means concatenating along the feature dimension, and `Add` means adding two feature maps together. This represents repeating the C2 operation N times.

[0018] Furthermore, the same-scale feature interaction includes: flattening the high-level feature maps of infrared and visible light to form infrared feature X1 and visible light feature X2.

[0019] Through a learnable weight matrix W Q W K and W V (Calculate query, key and):

[0020] Q = X1W Q K = X²W KV = X²W V

[0021] Utilizing a self-attention mechanism to calculate attention weights and capture cross-modal feature correlations

[0022]

[0023] The obtained attention output is added to the original input features using residual addition, followed by layer normalization.

[0024] X′=LayerNorm(X+Attention(Q,K,V))

[0025] Nonlinear transformation is performed using a two-layer fully connected network.

[0026] FFN(X′)=W2·ReLU(W1·X′)

[0027] Where W1 and W2 are learnable activation functions, and ReLU is the activation function;

[0028] Residual addition and layer normalization are performed again on the feedforward network output.

[0029] Output=LayerNorm(X′+FFN(X′))

[0030] The output retains the size of the input feature map as H×W×C, but has an enhanced global context representation.

[0031] Furthermore, the preliminary class confidence score prediction for the final feature map, specifically selecting the k indices with the highest class confidence scores, is as follows:

[0032] The final feature map passes through a linear layer to predict preliminary class confidence. The top k indices with the highest class confidence are denoted as ind.

[0033] class = linear(F1)

[0034] Suppose that `class` is a tensor representing a two-dimensional array, where each row is a set of classification scores.

[0035] Use max(A) ij ) represents the maximum value in the last dimension, therefore

[0036] V i =max(class[i,c])

[0037] Here V i It is the maximum value of all columns in the i-th row;

[0038] ind=topk(max(class[i,c]),k)),i=1,2,…M

[0039] Where k is the required first k indices, M is the number of rows in the matrix, and topk returns the indices of the k largest values ​​in each row of the tensor.

[0040] Furthermore, the selection of the corresponding k category confidence scores and the corresponding k anchor boxes is as follows:

[0041] Based on the index ind, select the corresponding k category confidence scores; the final feature map passes through an MLP layer to predict the preliminary anchor box coordinates in the channel dimension, and select the corresponding k anchor boxes according to the calculated index ind.

[0042] Furthermore, the decoder includes:

[0043] It includes multi-head self-attention, multi-scale deformable attention, and feedforward neural network modules, specifically:

[0044] Multi-head self-attention module:

[0045] N1 = Add(MLP(anchor),target)

[0046] Q′=W′ q N1,K′=W′ k N1,V′=W′ v target

[0047] D1=AddNorm(Attention(Q′,K′,V′),target)

[0048] Multi-scale deformation attention module:

[0049] N2 = Add(MLp(anchor), D1)

[0050] Q″=Wq″N2,K″=W″ k ·anchor,V″=Wv″F1

[0051] D2=AddNorm(MDAttention(Q″,K″,V″),D1)

[0052] Feedforward neural network module:

[0053] D3 = AddNorm(FFN(D2),D2)

[0054] Prediction module:

[0055] class = Linear(D3)

[0056] bbox = Add(MLP(D3), anchor)

[0057] Where target represents the confidence scores of the k input classes, anchor represents the k input anchor boxes, class represents the confidence scores of the output classes, and bbox represents the predicted bounding boxes of the output.

[0058] Furthermore, after the decoder outputs the class confidence score and the predicted bounding box (bbox), it serves as the target and anchor input to the next decoder module. The next decoder has the same structure as the above decoder. After passing through a total of 6 decoder modules, the final prediction result is output, which is the target information and location information in the image.

[0059] On the other hand, this invention provides an infrared-visible light joint target detection device, including a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it implements the infrared-visible light joint target detection method.

[0060] On the other hand, this specification provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the infrared-visible light joint target detection method.

[0061] The beneficial effects of this invention are:

[0062] Traditional infrared and visible light target detection methods typically employ independent detection strategies. This makes it difficult for models to fully leverage the complementary advantages of infrared and visible light images when dealing with problems such as low light, strong reflection, and complex backgrounds. This invention employs an attention mechanism for cross-modal feature fusion and global context modeling, overcoming the limitations of traditional methods that rely solely on a single modality. By effectively fusing different levels of low, medium, and high resolution, and capturing cross-modal correlations between infrared and visible light images through a self-attention mechanism, this invention can dynamically adjust the information weights of different modalities, achieving complementary advantages between infrared and visible light images and significantly improving target detection performance under complex conditions.

[0063] In practical applications, targets may appear in regions of different scales, especially in infrared images. Due to variations in target distance, the scale span of targets is significant, making it difficult for traditional methods to simultaneously handle multi-scale targets, particularly when targets are widely distributed. This invention utilizes a multi-scale deformable attention mechanism (MDAttention) to effectively model targets of different scales, avoiding the limitations of traditional methods in multi-scale scenes. By combining a multi-head self-attention module with multi-scale deformable attention, it can flexibly handle targets of different scales and enhance feature interactions between targets of various scales. This significantly improves the detection accuracy of multi-scale targets, especially in complex backgrounds, enabling the simultaneous and effective detection of small targets at a distance and large targets at close range.

[0064] Traditional object detection methods often suffer from accuracy issues in class prediction and bounding box regression, especially in multimodal image scenarios. Accurately predicting the class and location of each object is crucial for improving detection accuracy. This invention performs class confidence prediction on the final feature map and iteratively improves the accuracy of both class prediction and location regression through multiple decoder module operations. Furthermore, by selecting the top k class confidence scores and their corresponding anchor boxes, the model's prediction process is optimized. This significantly improves the accuracy of object class prediction and bounding box regression, especially in complex scenes and multi-object scenarios, enabling efficient and accurate prediction of object class and location. Attached Figure Description

[0065] Figure 1 This is an overall network structure diagram provided for an embodiment of the present invention;

[0066] Figure 2 This is a diagram of the backbone network structure for extracting image features provided in an embodiment of the present invention.

[0067] Figure 3 A flowchart of the Fusion module provided in an embodiment of the present invention;

[0068] Figure 4 A flowchart of the decoder module provided in an embodiment of the present invention;

[0069] Figure 5 This is a schematic diagram of an infrared-visible light combined target detection device provided in an embodiment of the present invention. Detailed Implementation

[0070] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0071] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0072] The objective of this invention is achieved through the following technical solution: an infrared-visible light joint target detection method. This method simultaneously inputs infrared and visible light images into a model, fuses the features of the two images, and outputs the type and location of the target in the image. The network structure diagram is shown below. Figure 1 As shown;

[0073] The input to the model is a set of corresponding infrared and visible light images;

[0074] The network comprises three main modules: a backbone network, an efficient hybrid encoder, and a Transformer decoder with an auxiliary prediction head.

[0075] The backbone network is responsible for extracting image features, typically based on existing convolutional neural networks (CNNs) or other efficient backbone network models, such as classic feature extraction models like PREsNet or Swing Transformer. Taking PREsNet as an example, its structure is based on residual blocks, using skip connections to solve the vanishing gradient problem. Features are output from three different levels of the backbone network, designated S3, S4, and S5, with the network structure as follows: Figure 2 As shown. The feature map at each level contains spatial information at different scales:

[0076] S3: Stage 3 output of PResNet, with higher spatial resolution.

[0077] S4: Stage 4 output of PResNet, with medium spatial resolution and rich semantic information.

[0078] S5: Stage 5 output of PResNet, with lower spatial resolution but rich semantic information.

[0079] Cross-modal feature fusion module

[0080] After the infrared and visible light images are processed by the backbone network to extract S3, S4, and S5 feature maps respectively, the S3 and S4 feature maps of the infrared image and the corresponding S3 and S4 feature maps of the visible light image are input into the Fusion module for cross-modal feature fusion to obtain the fused S3 and S4.

[0081] The structure of the Fusion module is as follows: Figure 3As shown, its input consists of two feature maps of the same shape, and the output is a feature map of the same shape as the input. If the input is A1 and A2, then...

[0082] C1(x) = SiLU(BN(Conv(x)))

[0083] C2(x)=SiLu(BN(Conv(SiLU(BN(Conv(x)))))

[0084]

[0085] Here, `concat` means concatenating along the feature dimension, and `Add` means adding two feature maps together. This represents repeating operation C2 N times;

[0086] Improved Attention-Based Same-Scale Feature Interaction Module (AIFI):

[0087] (1) This invention proposes an improved attention-based same-scale feature interaction (AIFI) module for effectively fusing high-level features of infrared and visible light images (S5):

[0088] (2) Flatten the S5 feature maps of infrared and visible light to form infrared feature X1 and visible light feature X2.

[0089] Through a learnable weight matrix W Q W K and W V Calculate the query (Q), key (K), and value (V):

[0090] Q = X1W Q K = X²W K V = X²W V

[0091] (3) Calculate attention weights using the self-attention mechanism to capture cross-modal feature correlations.

[0092]

[0093] (4) Add the attention output to the original input features using residual addition, and then perform layer normalization.

[0094] X′=LayerNorm(X+Attention(Q,K,V))

[0095] (5) Nonlinear transformation is performed through a two-layer fully connected network.

[0096] FFN(X′)=W2·ReLU(W1·X′)

[0097] Where W1 and W2 are learnable activation functions, and ReLU is the activation function;

[0098] (6) Perform residual addition and layer normalization on the feedforward network output again.

[0099] Output=LayerNorm(X′+FFN(X′))

[0100] The output retains the size of the input feature map as H×W×C, but has an enhanced global context representation, denoted as F5;

[0101] Convolution-based cross-scale feature fusion module

[0102] This invention designs a convolution-based cross-scale feature fusion module for processing and fusing multi-scale features extracted from the backbone network:

[0103] The feature map F5 obtained in the previous stage is upsampled to make its spatial dimensions consistent with the intermediate-level feature map S4. Bilinear interpolation or transposed convolution is used to maintain the smoothness and spatial consistency of the features. The upsampled F5 and S4 are then input into the Fusion module, which integrates the semantic information of the high-level features and the detailed information of the mid-level features.

[0104] F fused1 =Fusion(upsample(Conv(F5)),S4)

[0105] F fused2 =Fusion(upsample(Conv(F fused1 )),S3)

[0106] F fused3 =Fusion(downsample(Conv(F fused2 )),Conv(F fused1 ))

[0107] F fused4 =Fusion(downsample(Conv(F fused3 )),Conv(F5))

[0108] The obtained F fused2 F fused3 F fused4 After convolution and unfolding, the data is concatenated to obtain the feature map F required for subsequent decoding by the transformer. 1, Right now

[0109] C3(x) = flatten(Conv(x))

[0110]

[0111] Query Selection module

[0112] The feature map F1 output by the CCFM module is input into the Query Selection module to dynamically select high-quality initial object queries.

[0113] The input F1 score is processed through a linear layer to predict initial class confidence. The top k indices with the highest class confidence are denoted as ind. Then we have...

[0114] class = linear(F1)

[0115] Suppose that `class` is a tensor representing a two-dimensional array, where each row is a set of classification scores.

[0116] Use max(A) ij ) represents the maximum value in the last dimension, therefore

[0117] V i =max(class[i,c])

[0118] Here V i It is the maximum value of all columns in the i-th row;

[0119] ind=topk(max(class[i,c]),k)),i=1,2,…M

[0120] Where k is the first k indices required, M is the number of rows in the matrix, and topk is used to return the indices of the k largest values ​​in each row of the tensor;

[0121] Based on the index ind, select the corresponding k category confidence scores, denoted as target;

[0122] The input F1 score is processed through an MLP layer, which predicts the initial anchor box coordinates in the channel dimension, and selects the corresponding k anchor boxes according to the index ind calculated above, which are denoted as anchors.

[0123] Decoder:

[0124] The decoder receives the feature map F1 from the encoder, the target and anchor output from the Query Selection module, and outputs the predicted target category and bounding box. The decoder includes multi-head self-attention, multi-scale deformable attention, and feedforward neural network (FFN) modules, as shown in the specific structure below. Figure 4 As shown:

[0125] Multi-head self-attention module:

[0126] N1 = Add(MLP(anchor),target)

[0127] Q′=W′ q N1,K′=W′ k N1,V′=W′ v target

[0128] D1=AddNorm(Attention(Q′,K′,V′),target)

[0129] Multi-scale deformation attention module:

[0130] N2 = Add(MLp(anchor), D1)

[0131] Q″=Wq″N2,K″=W″ k ·anchor,V″=Wv″F1

[0132] D2=AddNorm(MDAttention(Q″,K″,V″),D1)

[0133] Feedforward neural network module:

[0134] D3 = AddNorm(FFN(D2),D2)

[0135] Prediction module:

[0136] class = Linear(D3)

[0137] bbox = Add(MLP(D3), anchor)

[0138] After the prediction module outputs the class confidence score and the predicted bounding box (bbox), these are used as the target and anchor input to the next decoder module. The next decoder has the same structure as the previous decoder. After passing through a total of 6 such decoder modules, the final prediction result is output, which is the target information and location information in the image.

[0139] Corresponding to the aforementioned embodiment of the infrared-visible light joint target detection method based on the TRANSFORMER architecture, the present invention also provides an embodiment of an infrared-visible light joint target detection device based on the TRANSFORMER architecture.

[0140] See Figure 5 The present invention provides an infrared-visible light joint target detection device based on the TRANSFORMER architecture, comprising a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement an infrared-visible light joint target detection method based on the TRANSFORMER architecture in the above embodiment.

[0141] The present invention provides an embodiment of an infrared-visible light joint target detection device based on the TRANSFORMER architecture, which can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 5 The diagram shown is a hardware structure diagram of any device with data processing capabilities, including the infrared-visible light joint target detection device based on the TRANSFORMER architecture provided by this invention. (Except for...) Figure 5 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0142] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0143] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0144] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements an infrared-visible light joint target detection method based on the TRANSFORMER architecture described in the above embodiments.

[0145] The computer-readable storage medium can be an internal storage unit of any data processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0146] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the infrared-visible light joint target detection method based on the TRANSFORMER architecture.

[0147] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0148] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this application. This application is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. An infrared-visible light joint target detection method, characterized in that, The method includes: S1. Input the infrared light image and the visible light image into the backbone network respectively to obtain feature maps of three different spatial resolution levels for infrared light and visible light. S2. Perform cross-modal feature fusion on the feature maps of the first two lower spatial resolution levels of the infrared image and the first two lower spatial resolution levels of the visible light image to obtain a feature image fused from two different spatial resolution levels. Specifically, the cross-modal feature fusion involves inputting two feature maps of the same shape and outputting a feature map of the same shape as the input. If the inputs are A1 and A2, then... ； ； ； Here, `concat` means concatenating along the feature dimension, and `Add` means adding two feature maps together. This represents repeating operation C2 N times; S3. Perform same-scale feature interaction on the high-level feature maps of infrared and visible light to enhance the global context representation, resulting in a high-level feature map after feature interaction; the same-scale feature interaction includes: flattening the high-level feature maps of infrared and visible light to form infrared feature X1 and visible light feature X2. Through a learnable weight matrix W Q W K and W V Calculate the query, key, and value: ； Utilizing a self-attention mechanism to calculate attention weights and capture cross-modal feature correlations ； The obtained attention output is added to the original input features using residual addition, followed by layer normalization. ； Nonlinear transformation is performed using a two-layer fully connected network. ； Where W1 and W2 are learnable activation functions, and ReLU is the activation function; Residual addition and layer normalization are performed again on the feedforward network output. ； The output retains the size of the input feature map as H×W×C, but has an enhanced global context representation; S4. Upsample the high-level feature map after feature interaction and fuse it with the mid-level fusion feature map to obtain the first fusion feature map; upsample the first fusion feature map and fuse it with the low-level fusion feature map to obtain the second fusion feature map; downsample the second fusion feature map and fuse it with the first fusion feature map again to obtain the third fusion feature map; downsample the third fusion feature map and fuse it with the high-level feature map after feature interaction to obtain the fourth fusion feature map; convolve the second, third and fourth fusion feature maps, unfold them and concatenate them to obtain the final feature map. S5. Predict the initial class confidence for the final feature map, take the top k indices with the highest class confidence, and select the corresponding k class confidence and the corresponding k anchor boxes; S6. Input the final feature map, k class confidence scores and the corresponding k anchor boxes into the decoder to predict the class confidence scores and predicted boxes.

2. The infrared-visible light joint target detection method according to claim 1, characterized in that, The backbone network is PREsNet, and its outputs are: feature maps with high spatial resolution, feature maps with medium spatial resolution but rich semantic information, and feature maps with low spatial resolution but rich semantic information.

3. The infrared-visible light joint target detection method according to claim 1, characterized in that, The preliminary class confidence score for predicting the final feature map is determined by selecting the top k indices with the highest class confidence scores, specifically as follows: The final feature map passes through a linear layer to predict preliminary class confidence. The top k indices with the highest class confidence are denoted as ind. ； Suppose that `class` is a tensor representing a two-dimensional array, where each row is a set of classification scores. use This represents the maximum value in the last dimension, therefore ； here It is the maximum value of all columns in the i-th row; ； Where k is the first k indices required, M is the number of rows in the matrix, and topk is used to return the indices of the k largest values ​​in each row of the tensor.

4. The infrared-visible light joint target detection method according to claim 3, characterized in that, The specific steps for selecting the corresponding k category confidence scores and the corresponding k anchor boxes are as follows: Based on the index ind, select the corresponding k category confidence scores; the final feature map passes through an MLP layer, which predicts the preliminary anchor box coordinates in the channel dimension, and selects the corresponding k anchor boxes according to the calculated index ind.

5. The infrared-visible light joint target detection method according to claim 1, characterized in that, The decoder includes: It includes multi-head self-attention, multi-scale deformable attention, and feedforward neural network modules, specifically: Multi-head self-attention module: ； Multi-scale deformation attention module: ； ； Feedforward neural network module: ； Prediction module: ； in For the confidence scores of the input k categories, The input consists of k anchor boxes, class represents the output class confidence score, and bbox represents the output predicted bounding box.

6. The infrared-visible light joint target detection method according to claim 5, characterized in that, After the decoder outputs the class confidence score and the predicted bounding box (bbox), it serves as the target and anchor input to the next decoder module. The next decoder has the same structure as the above decoder. After passing through a total of 6 decoder modules, the final prediction result is output, which is the target information and location information in the image.

7. An infrared-visible light joint target detection device, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code, it implements an infrared-visible light joint target detection method as described in any one of claims 1-6.

8. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements an infrared-visible light joint target detection method as described in any one of claims 1-6.

Images