Night pedestrian infrared detection method and device and storage medium

Abstract

The invention relates to the technical field of automatic driving pedestrian detection, in particular to a night pedestrian infrared detection method and device and a storage medium. According to the method, on the basis of a lightweight target detection network, shuffled slice down-sampling is designed to replace traditional convolution down-sampling, so that information loss of an original infrared image is reduced, and channel feature information of the infrared image is enhanced; meanwhile, a refined space pooling module is designed, one-dimensional convolution is utilized to establish local feature and channel feature weight mapping under different receptive fields, airspace sparse features extracted by a detection network in an infrared image are enriched, and context-related semantic information is effectively utilized; and finally, designing a self-adaptive feature reinforced connection structure, fusing a learnable weight factor into different scale features, realizing self-adaptive connection among multi-scale features, effectively exciting and inhibiting the multi-scale features of different contributions of pedestrians, and improving the precision of night pedestrian detection.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous driving pedestrian detection, and in particular to a nighttime pedestrian infrared detection method, device and storage medium. Background Art

[0002] Autonomous driving has gradually become an inevitable development direction for the automotive industry, and more and more car companies are investing more money in the research and development of autonomous driving. Under night conditions, infrared imaging systems can achieve unobstructed night vision, but at the same time, infrared images have problems such as lack of color information and insufficient spatial and texture detail information, resulting in a low infrared pedestrian detection rate at night; at the same time, infrared pedestrian detection algorithms are usually deployed on mobile development platforms. Deploying network models with both accuracy and speed on such platforms is a huge challenge to its computing power and memory resources. Therefore, the research on infrared pedestrian detection methods for nighttime assisted driving is of practical significance. Infrared pedestrian detection algorithms are all rooted in existing target detection algorithms, which are composed of two branches. One branch is the traditional machine learning algorithm, which can only detect specific scenes and has problems such as weak generalization ability, slow detection speed and low accuracy; the other branch is deep learning, which has significantly improved the detection effect compared with traditional machine learning algorithms.

[0003] Compared with traditional algorithms, deep learning-based target detection algorithms have the advantages of high accuracy, strong generalization ability, and strong transferability. Currently, the widely used deep learning-based target detection methods can be divided into two categories. One is a two-stage target detection method based on candidate box generation, and the other is a one-stage target detection algorithm. This type of algorithm does not need to generate candidate targets. The single-stage target detection technology unifies all target proposals, feature extraction, classification, and regression steps into a single network, and directly predicts the location and category of the target through the network. It is faster than the two-stage target detection algorithm and is beneficial to the response of the vehicle assisted driving system. For mobile devices, the above target detection algorithm network has high spatiotemporal complexity, large number of parameters, and is heavily dependent on high-performance computers. Therefore, lightweight target detection algorithms such as MobileNet and ShuffleNet have been proposed and applied in the field of pedestrian target detection. With the advantage of low model parameters, lightweight target detection networks are more suitable for mobile development devices such as assisted driving systems.

[0004] However, for mobile devices with limited computing resources, although lightweight target detection algorithms such as ShuffleNet have improved the detection speed, their detection accuracy has also dropped significantly. The accuracy of infrared pedestrian target detection algorithms based on deep learning is not ideal. The main reason is that the network structure is too simple, and the feature extraction of image information is not deep enough. Although the target detection model with high network spatiotemporal complexity and large number of parameters has high detection accuracy, it is heavily dependent on high-performance computers and difficult to deploy to mobile development platforms. At the same time, the difference between infrared images and traditional visible light images mainly lies in the lack of color information, and the lack of rich spatial and texture detail information. The size and position of pedestrian targets in night infrared images vary greatly, and each target has different characteristics. The features extracted by the shallow network are mainly pixel point information such as texture, edge, and corner. As the network level deepens, the image detail information may disappear completely. The infrared image is compressed in the deep network, resulting in the deep features being presented as low-scale, abstract semantic information. Compared with traditional visible light images, infrared images lack color information and are not rich in spatial and texture detail information. Therefore, it is particularly important to reduce the information loss of the original infrared image, enrich the spatial sparse features extracted by the detection network, and effectively utilize context-related semantic information. Since pedestrians at different distances generally have the characteristics of scale scaling and topological deformation, and there is often occlusion between pedestrians or other objects, how to effectively stimulate and suppress the multi-scale features of pedestrians with different contributions is a key issue. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a method, device and storage medium for nighttime pedestrian infrared detection, which achieves the detection results of pedestrian targets by enhancing the expression of pedestrian positions and channel characteristics in infrared image information.

[0006] The present invention discloses a nighttime pedestrian infrared detection method through an embodiment, comprising:

[0007] Acquire an infrared image, and obtain a low-dimensional image of the infrared image by a shuffled slice downsampling method;

[0008] Extract pedestrian feature information from low-dimensional images through lightweight object detection algorithms;

[0009] The pedestrian feature information is refined and pooled to obtain refined fusion information;

[0010] Taking pedestrian feature information and refined fusion information as input, the infrared pedestrian position information and channel feature information are enhanced through the adaptive feature enhancement connection structure to obtain the detection result.

[0011] Furthermore, obtaining a low-dimensional image of the infrared image by the shuffled slice downsampling method includes:

[0012] The infrared image is split into 4 low-resolution infrared images through downsampling operation; the 4 low-resolution infrared images are divided into two groups for channel shuffling operation to obtain a low-dimensional image of the infrared image doubled downsampled. It is worth noting that the convolution downsampling operation adopted by the backbone network of the lightweight target detection algorithm obtains the feature correlation between image pixels, but also causes the loss of features of adjacent pixel points of the infrared image, which is at the expense of the loss of edge information of the infrared image to achieve the reduction of the number of parameters. Therefore, the present invention designs a shuffled slice downsampling module to replace the convolution downsampling operation, thereby reducing the loss of information of the original infrared image. The four low-resolution images are stitched together so that the infrared image information is concentrated in the channel space, and the input channel is expanded by 4 times, resulting in a two-fold downsampled low-resolution image with a channel dimension of 4. Because the downsampling is achieved by slicing only the infrared image pixels, the loss of features of adjacent pixels of the infrared image and the loss of edge information of the infrared image caused by the convolution calculation are avoided, so the downsampled infrared image information obtained by the slicing operation is almost not lost; finally, the 4-channel low-resolution infrared pedestrian image is divided into two groups, and the channel shuffling operation is performed separately to mix the channel order of the original image, so that the feature information of different channels can be connected, and the fusion of information between channels is promoted, and finally a two-fold downsampled feature map with almost no information loss and enhanced image feature expression is obtained, which is the low-dimensional image described in this article. Infrared image Γ∈R H×W×C Split into 4 low-resolution infrared images And the calculation method to obtain the low-dimensional image is as follows:

[0013] ζ(x,y)={(x i ,y i )|i,j=2n-1}

[0014] 0 <i<H,0<j<W,i,j∈Z

[0015] γ=concat[ζ(x,y),ζ(x+1,y),ζ(x,y+1),ζ(x+1,y+1),]

[0016] Among them, x i and i Represents the index of the pixel on the horizontal and vertical axes.

[0017] Furthermore, the pedestrian feature information is refined by spatial pooling to obtain refined fusion information, including: extracting the first channel feature information by Avgpool average pooling calculation of the pedestrian feature information, establishing the channel information connection of the first channel feature information by one-dimensional convolution, normalizing the channel information connection by Sigmoid function, and obtaining the weight information of the first channel feature information; extracting the second channel feature information under different receptive fields by Maxpool maximum pooling calculation of the pedestrian feature information; fusing the weight information of the first channel feature information with the second channel feature information under different receptive fields respectively, and obtaining the fusion information under different receptive fields; adding the fusion information under different receptive fields to obtain refined fusion information. Since the lightweight backbone network has fewer layers and shallower depth than the ordinary backbone network, the feature extraction capability is limited; and the semantic features contained in different receptive fields of the neural network are different, the larger the receptive field value, the larger the original image range it can access, which also means that it may contain more global and higher semantic features; on the contrary, the smaller the value, the more local and detailed the features it contains. Therefore, the present invention designs a refined spatial pooling structure, inserts the refined spatial pooling structure at the tail end of the backbone network, and utilizes the semantic features of different receptive fields to realize the refined fusion of multi-fine-grained local features at the pixel level, enrich the expressive power of the final feature map, enrich the spatial sparse features extracted by the detection network, and effectively utilize context-related semantic information, thereby enhancing the detection effect.

[0018] Furthermore, the number of different receptive fields includes three.

[0019] Furthermore, taking the pedestrian feature information and the refined fusion information as input, the infrared pedestrian position information and channel feature information are enhanced through an adaptive feature enhancement connection structure, and the detection result obtained includes: multiplying the pedestrian feature information by the first self-learning weight parameter to obtain the first pedestrian feature information, upsampling the refined fusion information and multiplying it by the second self-learning weight parameter to obtain the second refined fusion information; adding the first pedestrian feature information and the second refined fusion information and then correcting them through the ReLU function to obtain the corrected information; lightweighting the corrected information through deep separable convolution to obtain the fused feature information; calibrating the fused feature information through the position calibration branch and the channel calibration branch respectively; finally, adding the calibration results of the position calibration branch and the channel calibration branch to obtain the detection result.

[0020] The calculation method of fusion feature information is as follows:

[0021]

[0022] Among them, w1 and w2 are learnable weight parameters, U represents upsampling calculation, is the pedestrian feature information,

[0023] To refine and integrate information.

[0024] Furthermore, the position calibration branch includes: using adaptive pooling in the horizontal and vertical directions to aggregate and fuse feature information to obtain aggregated feature information; dividing the aggregated feature information into highly aggregated feature information and width aggregated feature information along the spatial dimension; activating the highly aggregated feature information and the width aggregated feature information respectively by a Sigmoid function, and then fusing them with the fused feature information to obtain the calibration result of the position calibration branch.

[0025] Furthermore, the channel calibration branch includes: passing the fused feature information through global average pooling and two 1×1 convolution operators in sequence to obtain compressed feature information; activating the compressed feature information through a Sigmoid function, and then fusing it with the fused feature information to obtain a calibration result of the channel calibration branch.

[0026] Furthermore, the lightweight target detection algorithm includes the ShuffleNetv2 algorithm or the Anchor-free target detection algorithm.

[0027] In order to achieve the above-mentioned objectives, the present invention also provides a nighttime pedestrian infrared detection device, comprising an infrared imager for acquiring infrared images; a shuffled slice downsampling module for obtaining a low-dimensional image of the infrared image through a shuffled slice downsampling method; a pedestrian feature information extraction module for extracting pedestrian feature information from the low-dimensional image; a refined spatial pooling module for obtaining refined fusion information through refined spatial pooling of the pedestrian feature information; and an adaptive feature enhancement connection structure for taking the pedestrian feature information and the refined fusion information as input, enhancing the infrared pedestrian position information and the channel feature information, and obtaining a detection result.

[0028] In order to achieve the above-mentioned purpose, the present invention also provides a storage medium storing a plurality of computer instructions for executing the above-mentioned nighttime pedestrian infrared detection method.

[0029] The technical principle of the present invention is as follows: on the basis of a lightweight target detection network, a shuffled slice downsampling is designed to replace the traditional convolution downsampling, and the information loss of the original infrared image is reduced by slicing, and then channel shuffling is performed to strengthen the channel feature information of the infrared image; at the same time, a refined spatial pooling module is designed, and a one-dimensional convolution is used to establish a weight mapping between local features and channel features under different receptive fields, so as to achieve the refined fusion of local features of multiple receptive fields at the pixel level, enrich the spatial sparse features extracted by the detection network in the infrared image, and effectively utilize the context-related semantic information; finally, an adaptive feature enhancement connection structure is designed, and learnable weight factors are integrated into features of different scales to achieve adaptive connection between multi-scale features, and effectively stimulate and inhibit the multi-scale features of pedestrians with different contributions. It improves the detection accuracy under the conditions that pedestrians at a distance generally have the characteristics of scale scaling and topological deformation, and there is often occlusion between pedestrians or other objects.

[0030] Compared with the prior art, the present invention has the following beneficial effects: by back propagation during network training, the parameter values ​​are continuously adjusted, the contribution of different scale features from infrared images to the final prediction results is automatically determined, and the adaptive connection between multi-scale features of infrared pedestrian targets is realized, thereby effectively utilizing multi-scale features. In addition, the input features of different scales are complemented by the channel information and position feature calibration branches, enriching the information features of infrared pedestrian images; enhancing the expression ability of infrared pedestrian feature information, improving detection efficiency, and ensuring the real-time detection of lightweight target detection models on mobile devices. At the same time, on the basis of limited model space complexity and a small number of parameters, the limited color, space and texture information of infrared images are fully utilized, greatly improving the accuracy of pedestrian target detection in night environments, and being more conducive to the response of automobile assisted driving systems. At the same time, it provides an algorithm basis and a practical engineering application strategy for system deployment for the development of automobile assisted driving systems to reduce the incidence of nighttime traffic accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solution of the present invention and facilitate further understanding of the technical effects, technical features and purposes of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings. The accompanying drawings constitute an essential part of the specification and are used together with the embodiments of the present invention to illustrate the technical solution of the present invention, but do not constitute a limitation to the present invention.

[0032] Figure 1 A flow chart of the method of the present invention;

[0033] Figure 2 It is a schematic diagram of the detection process recorded in the embodiment of the method of the present invention, and is also a schematic diagram of the structure of the infrared detection device for pedestrians at night;

[0034] Figure 3Schematic diagram of the refined spatial pooling described in the embodiment of the method of the present invention;

[0035] Figure 4 It is a schematic diagram of the adaptive feature enhancement connection structure recorded in the embodiment of the method of the present invention. DETAILED DESCRIPTION

[0036] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described below are only intended to explain the technical solutions of the present invention, rather than to limit the present invention. In addition, the parts described in the embodiments or the accompanying drawings are only examples of the relevant parts of the present invention, rather than the entirety of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art should naturally belong to the protection scope of the present invention.

[0037] It is worth noting that in the following embodiments, the lightweight target detection algorithm is described using the ShuffleNetv2 backbone network as an example.

[0038] like Figure 1 , 2 As shown, the present invention proposes a nighttime pedestrian infrared detection method through the following embodiments, comprising the following steps: obtaining an infrared image, denoted as Γ∈R H×W×C , the low-dimensional image of the infrared image is obtained by the shuffle slice downsampling method; the pedestrian feature information in the low-dimensional image is extracted by the ShuffleNetv2 backbone network algorithm, which is recorded as The pedestrian feature information is refined through spatial pooling to obtain refined fusion information; the pedestrian feature information and the refined fusion information are used as input, and the infrared pedestrian position information and channel feature information are enhanced through an adaptive feature enhancement connection structure to obtain the detection result.

[0039] In this embodiment, obtaining a low-dimensional image of an infrared image by a shuffled slice downsampling method includes:

[0040] Through the downsampling operation, a value is taken for every other pixel of the high-resolution image, so that the original image is split into 4 low-resolution infrared images with almost no information loss; the 4 low-resolution infrared images are divided into two groups for channel shuffling operation to obtain a low-dimensional image with twice the infrared image downsampled. It is worth noting that the convolution downsampling operation adopted by the backbone network of the lightweight target detection algorithm obtains the feature correlation between image pixels, but also causes the loss of features of adjacent pixels of the infrared image, which is at the expense of the loss of edge information of the infrared image to achieve the reduction of parameters. Therefore, the present invention designs a shuffled slice downsampling module to replace the convolution downsampling operation, thereby reducing the information loss of the original infrared image. Specifically, taking the original input image as Γ∈R H×W×CFor example, the segmentation calculation method for each picture is as follows:

[0041] ζ(x,y)={(x i ,y i )|i,j=2n-1}

[0042] 0 <i<H,0<j<W,i,j∈Z

[0043] Among them, x i and i Represents the index of the pixel on the horizontal and vertical axes.

[0044] Four low-resolution infrared pedestrian images A 4-channel downsampled feature image is obtained by splicing The calculation method of is as follows:

[0045] γ=concat[ζ(x,y),ζ(x+1,y),ζ(x,y+1),ζ(x+1,y+1),]

[0046] Finally, the four low-resolution images are spliced ​​so that the infrared image information is concentrated in the channel space, and the input channel is expanded by 4 times to obtain a two-fold downsampled low-resolution image with a channel dimension of 4. Because the downsampling is achieved by only slicing the infrared image pixels, the loss of adjacent pixel features of the infrared image and the loss of infrared image edge information caused by the convolution calculation are avoided, so the downsampled infrared image information obtained by the slicing operation is almost not lost; Finally, the 4-channel low-resolution infrared pedestrian image is divided into two groups, and the channel shuffling operation is performed separately to mix the original image channel order, so that the feature information of different channels can be connected, and the fusion of information between channels is promoted, and finally a two-fold downsampled feature map with almost no information loss and enhanced image feature expression is obtained. That is the low-dimensional image described in the article.

[0047] In this embodiment, the pedestrian feature information is refined by spatial pooling to obtain refined fusion information, including: extracting first channel feature information through Avgpool average pooling calculation of the pedestrian feature information, using one-dimensional convolution to establish channel information connection of the first channel feature information, and normalizing the channel information connection through Sigmoid function to obtain weight information of the first channel feature information; extracting second channel feature information under different receptive fields through Maxpool maximum pooling calculation of the pedestrian feature information; fusing the weight information of the first channel feature information with the second channel feature information under different receptive fields respectively to obtain fusion information under different receptive fields; adding the fusion information under different receptive fields to obtain refined fusion information, which is recorded as Since the lightweight backbone network has fewer layers and shallower depth than the ordinary backbone network, its feature extraction capability is limited; and the semantic features contained in different receptive fields of the neural network are different. The larger the receptive field value, the larger the original image range it can access, which also means that it may contain more global and higher semantic features; on the contrary, the smaller the value, the more local and detailed the features it contains. Therefore, the present invention designs a refined spatial pooling structure, inserts a refined spatial pooling structure at the end of the backbone network, and uses different receptive field semantic features to achieve refined fusion of multiple fine-grained local features at the pixel level, enriching the expression capability of the final feature map, enriching the spatial sparse features extracted by the detection network, and effectively utilizing context-related semantic information, thereby enhancing the detection effect.

[0048] In this embodiment, the spatial pooling module is composed of three parallel spatial pooling attention modules, each of which is as follows: Figure 3 As shown in the figure, the features passed into the spatial pooling module will pass through the spatial pooling attention modules of different pooling dimensions at the same time, and finally the feature x will be combined with the feature y of the three parallel branches. i Fusion realizes the fusion of multiple receptive field features and enriches the final feature information. Due to the large differences in the size and position of pedestrian targets in night infrared images, each target has different characteristics; the features extracted by the shallow network are mainly pixel information such as texture, edge, and corner. As the network level deepens, the image detail information may disappear completely, and the infrared image is compressed in the deep network, causing the deep features to appear as low-scale, abstract semantic information. When only the feature information extracted by the deep network is used as the input feature of the prediction layer, the detailed features such as the edge and shape of the infrared pedestrian target contained in the shallow features will be missing, which will result in a low detection rate for mid-distance target pedestrians, which is not conducive to the response of the vehicle assisted driving system.

[0049] Therefore, by fully integrating high- and low-level semantic information and different fine-grained features, more contextual information can be obtained during network prediction. The present invention adds a feature fusion line to the ShuffleNetV2 backbone network at three scales: 256×256, 128×128, and 64×64, respectively, and designs an adaptive feature enhancement connection structure to replace the Concat operation, integrating learnable weight factors, and effectively stimulating and suppressing the multi-scale features of pedestrians with different contributions, so as to integrate the feature semantic information of infrared pedestrians at different scales and levels, and enrich the associative expression of semantic information between features of each scale.

[0050] In this embodiment, if Figure 4 As shown, pedestrian feature information and the refined fusion information As input, the infrared pedestrian position information and channel feature information are enhanced through the adaptive feature enhancement connection structure, and the detection results include: Multiply the first self-learning weight parameter w1 to obtain the first pedestrian feature information, upsample the refined fusion information and multiply it by the second self-learning weight parameter w1 to obtain the second refined fusion information; add the first pedestrian feature information and the second refined fusion information and then correct them through the ReLU function to obtain the corrected information; use the deep separable convolution DWConv to lightweight the corrected information to obtain the fused feature information

[0051] The above process can be expressed as:

[0052]

[0053] Among them, w1 and w2 are learnable weight parameters, and U represents upsampling calculation. w1 and w2 automatically adjust parameter values ​​through back propagation during network training, and adaptively determine the contribution value of feature information of different scales to predict sensitive targets.

[0054] Furthermore, the feature information is fused Calibration is performed through the position calibration branch and the channel calibration branch respectively; finally, the calibration results of the position calibration branch and the channel calibration branch are added together to obtain the detection result.

[0055] In this embodiment, the position calibration branch includes: using adaptive pooling in the horizontal and vertical directions to aggregate and fuse feature information to obtain aggregate feature information; dividing the aggregate feature information into high-aggregation feature information and width-aggregation feature information along the spatial dimension; activating the high-aggregation feature information and the width-aggregation feature information by the Sigmoid function respectively, and then fusing them with the fused feature information to obtain the calibration result B1 of the position calibration branch. Specifically, it can be expressed as:

[0056]

[0057]

[0058] Where xavg and yavg represent adaptive average pooling along the X-axis and Y-axis, FC is a fully connected layer; where the aggregated feature g is decomposed into g in the spatial dimension. h and g w , and then by stimulating and expanding q h and q w As the attention weight, the position calibration feature B1 is finally obtained to determine the accurate position of the infrared pedestrian sensitive target.

[0059] In this embodiment, the channel calibration branch includes: fusing feature information Through global average pooling and two 1×1 convolution operators, the compressed feature information is obtained and recorded as The global spatial information is converted into channel descriptors, and then the features of global average pooling are used to capture the connection between nonlinear channels through two 1×1 Conv operators, which can reduce the amount of calculation and thus reduce the complexity of the pedestrian target detection model in infrared images. After being activated by the Sigmoid function, it is fused with the fusion feature information to obtain the calibration result of the channel calibration branch.

[0060] Specifically, the calibration result of the channel calibration branch can be expressed as:

[0061]

[0062] Where avg represents average pooling, The function represents the compression of Conv channel information through two 1×1 operators, specifically by The squeezing and incentives effectively establish interdependence between channels.

[0063] Finally, the image information of the detection result can be expressed as Finally, the pedestrian detection result B is output.

[0064] On the other hand, Figure 2 As shown, in order to achieve the above-mentioned purpose, the present invention also provides a nighttime pedestrian infrared detection device, including an infrared imager for acquiring infrared images; a shuffled slice downsampling module for obtaining a low-dimensional image of the infrared image by a shuffled slice downsampling method; a pedestrian feature information extraction module for extracting pedestrian feature information from the low-dimensional image; a refined spatial pooling module for obtaining refined fusion information by refining spatial pooling of the pedestrian feature information; and an adaptive feature enhancement connection structure for taking the pedestrian feature information and the refined fusion information as input, enhancing the infrared pedestrian position information and the channel feature information, and obtaining the detection result.

[0065] On the other hand, in order to achieve the above-mentioned purpose, the present invention also provides a storage medium storing a plurality of computer instructions for executing the above-mentioned nighttime pedestrian infrared detection method.

[0066] It should be noted that the above embodiments are only for more clearly illustrating the technical solutions of the present invention. Those skilled in the art will understand that the implementation methods of the present invention are not limited to the above contents, and obvious changes, replacements or substitutions based on the above contents do not exceed the scope of the technical solutions of the present invention; other implementation methods will naturally fall within the scope of the present invention without departing from the concept of the present invention.

Claims

1. A nighttime pedestrian infrared detection method, characterized in that: include: Acquire an infrared image, and obtain a low-dimensional image of the infrared image by a shuffled slice downsampling method; Extracting pedestrian feature information from the low-dimensional image using a lightweight target detection algorithm; The pedestrian feature information is refined by spatial pooling to obtain refined fusion information; The pedestrian feature information and the refined fusion information are used as input, the infrared pedestrian position information and channel feature information are enhanced through an adaptive feature enhancement connection structure to obtain a detection result.

2. The nighttime pedestrian infrared detection method according to claim 1, characterized in that: The step of obtaining a low-dimensional image of the infrared image by a shuffled slice downsampling method comprises: Splitting the infrared image into four low-resolution infrared images through a downsampling operation; Dividing the four low-resolution infrared images into two groups and performing a channel shuffling operation to obtain a low-dimensional image of the infrared image that is twice downsampled; The infrared image Split into 4 low-resolution infrared images And the calculation method to obtain the low-dimensional image is expressed as: ζ(x,y)={(x i ,y i )|i,j=2n-1} 0 <i<H,0<j<W,i,j∈Z γ=concat[ζ(x,y),ζ(x+1,y),ζ(x,y+1),ξ(x+1,y+1),] Among them, x i and i Represents the index of the pixel on the horizontal and vertical axes.

3. The nighttime pedestrian infrared detection method according to claim 1, characterized in that: The obtaining refined fusion information by refining spatial pooling of the pedestrian feature information comprises: The pedestrian feature information is subjected to Avgpool average pooling calculation to extract first channel feature information, a channel information connection of the first channel feature information is established using a one-dimensional convolution, and the channel information connection is normalized using a Sigmoid function to obtain weight information of the first channel feature information; The pedestrian feature information is calculated by Maxpool to extract the second channel feature information under different receptive fields; The weight information of the first channel feature information is respectively fused with the second channel feature information under the different receptive fields to obtain fused information under different receptive fields; The fusion information under the different receptive fields is added to obtain refined fusion information.

4. The nighttime pedestrian infrared detection method according to claim 3, characterized in that: The number of different receptive fields includes three.

5. The nighttime pedestrian infrared detection method according to claim 1, characterized in that: The detection result obtained by taking the pedestrian feature information and the refined fusion information as input and strengthening the infrared pedestrian position information and channel feature information through the adaptive feature strengthening connection structure includes: Multiplying the pedestrian feature information by a first self-learning weight parameter to obtain first pedestrian feature information, upsampling the refined fusion information and multiplying it by a second self-learning weight parameter to obtain second refined fusion information; Adding the first pedestrian feature information and the second refined fusion information, and then correcting them through a ReLU function to obtain corrected information; The correction information is lightweighted by depthwise separable convolution to obtain fused feature information; The fused feature information is calibrated respectively through a position calibration branch and a channel calibration branch; Adding the calibration results of the position calibration branch and the channel calibration branch to obtain a detection result; The calculation method of the fusion feature information is as follows: Among them, w1 and w2 are learnable weight parameters, U represents upsampling calculation, is the pedestrian feature information, To refine and integrate information.

6. The nighttime pedestrian infrared detection method according to claim 5, characterized in that: The position calibration branch comprises: Adopting adaptive pooling in horizontal and vertical directions to aggregate the fused feature information to obtain aggregated feature information; Splitting the aggregated feature information along the spatial dimension into height aggregated feature information and width aggregated feature information; The height aggregation feature information and the width aggregation feature information are respectively activated by the Sigmoid function and then fused with the fusion feature information to obtain the calibration result of the position calibration branch.

7. The nighttime pedestrian infrared detection method according to claim 5, characterized in that: The channel calibration branch includes: The fused feature information is sequentially passed through global average pooling and two 1×1 convolution operators to obtain compressed feature information; The compressed feature information is activated by a Sigmoid function and then fused with the fused feature information to obtain a calibration result of a channel calibration branch.

8. The nighttime pedestrian infrared detection method according to claim 1, characterized in that: Lightweight target detection algorithms include ShuffleNetv2 backbone network or Anchor-free target detection algorithm; The normalization function includes the Sigmoid function.

9. A nighttime pedestrian infrared detection device, characterized in that: include: Infrared imager, used to obtain infrared images; A shuffled slice down-sampling module, used for obtaining a low-dimensional image of the infrared image by a shuffled slice down-sampling method; A pedestrian feature information extraction module, used to extract pedestrian feature information from the low-dimensional image; A refined spatial pooling module, used for obtaining refined fusion information by refining spatial pooling of the pedestrian feature information; The adaptive feature enhancement connection structure is used to take the pedestrian feature information and the refined fusion information as input, enhance the infrared pedestrian position information and channel feature information, and obtain the detection result.

10. A storage medium, characterized in that: It stores a number of computer instructions for executing the nighttime pedestrian infrared detection method described in any one of claims 1-8.

Images