A pedestrian infrared image segmentation method and system under complex conditions

Abstract

The application discloses a pedestrian infrared image segmentation method and system under complex conditions, applied to the technical field of infrared image processing, comprising: using the detail information in the visible light image, and adopting a guided image filtering method to carry out denoising processing on the collected pedestrian infrared image, to obtain a processed image; an improved pulse coupled neural network model is constructed, comprising: taking the background score of each pixel in the image as the connection strength, and taking the clustering center of the background area of the image as a dynamic threshold; the improved pulse coupled neural network model is used for image segmentation on the processed image, to obtain a segmented image. In this way, the improved model obtained by the application can automatically stop according to the segmentation result, and over-segmentation phenomenon is avoided, so that the infrared pedestrian images in various complex environments can be effectively segmented, and the infrared images with more heat source interference in the background have a good segmentation effect.

Description

Technical Field

[0001] The present disclosure relates to the field of infrared image processing technology, and in particular to a method and system for segmenting pedestrian infrared images under complex conditions. Background Art

[0002] Infrared imaging is a technology that uses infrared radiation to capture images. Compared to visible light imaging, infrared imaging offers advantages such as nighttime visibility in complex environments, the ability to detect hidden thermal targets, and all-weather operation. Therefore, it is widely used in fields such as pedestrian detection, aerospace, security, military, and assisted driving. Infrared pedestrian segmentation is a key issue in infrared image processing systems. It helps accurately identify human targets in complex, dark environments and obtain more detailed information about them, and has become a research hotspot in the development of machine vision.

[0003] There are many methods for infrared pedestrian segmentation, such as threshold-based segmentation method, clustering-based segmentation method and deep learning-based segmentation method. Among them, the threshold-based segmentation method is simple and easy to implement, but it has poor segmentation effect on infrared images with noise and complex background. The clustering-based method is more sensitive to noise. The deep learning-based segmentation method is too dependent on the completeness of the training set. When the sample labels are few or unlabeled, the segmentation effect is poor. The network is too deep, which increases the application. However, when these methods are applied to complex environment images with mixed target and background grayscale features, over-segmentation or under-segmentation is prone to occur.

[0004] The Pulse Coupled Neural Network (PCNN) is a novel neural network derived from the synchronization phenomenon of gamma-band neurons in the mammalian visual cortex. It simulates the synchronized oscillation phenomenon in the cat visual cortex. PCNN has low computational complexity and high segmentation accuracy, and has been widely used in image segmentation. Existing PCNN models, when directly applied to infrared pedestrian segmentation in complex environments, are prone to uneven brightness distribution and aliasing with the background due to factors such as high noise interference, low contrast, complex backgrounds, interference from external heat sources, and the human body being easily hidden and occluded. This leads to blurred edges of human targets, seriously affecting the effective segmentation of human targets in infrared images under complex environments. Therefore, current models are still limited in their application conditions, and the segmentation effect is poor for low-quality, highly blurred, and complex infrared images. Further research and solutions are still needed.

[0005] Therefore, there is an urgent need for a technical solution to improve the traditional PCNN to achieve accurate segmentation of infrared images. Summary of the Invention

[0006] The present disclosure provides a method and system for pedestrian infrared image segmentation under complex conditions. By improving the traditional PCNN solution, it at least solves the technical problems of poor effect and low efficiency in the infrared image segmentation process of the existing technology.

[0007] According to a first aspect of the present disclosure, a method for segmenting pedestrian infrared images under complex conditions is provided, comprising the following steps:

[0008] The detailed information in the visible light image is used and the guided image filtering method is used to denoise the collected infrared image of the pedestrian to obtain the processed image;

[0009] An improved pulse coupled neural network model is constructed, including: using the background score of each pixel in the image as the connection strength and the cluster center of the background area of ​​the image as the dynamic threshold;

[0010] The improved pulse coupled neural network model is used to perform image segmentation on the processed image to obtain a segmented image.

[0011] According to the above aspects and any possible implementation, a further implementation is provided, wherein the improved pulse coupled neural network model is specifically:

[0012]

[0013]

[0014] E ij [n] = m(n)

[0015] Among them, F ij [n] is the feedback input, L ij [n] is the connection input, U ij [n] is the internal activity item, Y ij [n] is the pulse output, S kl is the external stimulus, corresponding to the pixel intensity in the image, (k,l) is the neighborhood of neuron (i,j), M ij,kl and W ij,kl is the weight matrix, E ij [n] is the dynamic threshold, λ is a coefficient to adjust the dynamic threshold and control the iteration speed of the model. is the connection strength, n is the number of model iterations, and m(n) represents the neuron pulse output.

[0016] According to the above aspects and any possible implementation, an implementation is further provided, wherein the M ij,kl 、W ij,kl and λ are specifically:

[0017]

[0018] λ=0.2

[0019] Among them, (i, j) represents the central neuron and (k, l) represents the area neuron.

[0020] According to the above aspect and any possible implementation, a further implementation is provided, wherein the connection strength is calculated as follows:

[0021] The SLIC superpixel segmentation algorithm is used to segment the infrared image into multiple superpixels, and each superpixel is regarded as a node in the graph;

[0022] The generated superpixels are used as vertices to build an undirected graph, and the brightness similarity is used as the edge weight between superpixels;

[0023] Defining a probability transfer matrix of transfer between vertices on the undirected graph and setting a background score vector of a node;

[0024] The background score vector is iteratively updated, and the updated background score vector is mapped to the original infrared image, and the connection strength is obtained after normalization.

[0025] According to the above aspects and any possible implementation, an implementation is further provided, wherein the process of using the SLIC superpixel segmentation algorithm to segment the infrared image into multiple superpixels is as follows:

[0026] Initialize several superpixel centers to get the initial superpixel center. For an infrared image containing N pixels, the superpixel center C i It can be expressed as (l i ,x i ,y i ), the image is divided into grid steps Divided into k square grids;

[0027] For each initial superpixel center, search the pixels within a 3×3 range around it to find the minimum gradient position;

[0028] The cluster center is changed from the center pixel of the original square pixel group to the minimum gradient point in its neighborhood, and then k initial cluster centers are obtained;

[0029] The characteristic distance D between the pixel points within the 2F×2F range around the initial cluster center and the initial cluster center is calculated, and the pixel points are divided into corresponding superpixel clusters according to the characteristic distance D to obtain multiple superpixels.

[0030] According to the above aspects and any possible implementation, an implementation is further provided, wherein the undirected graph is constructed as follows:

[0031] The random walk method is used to build an undirected graph G with the generated superpixels as vertices.<V,E> , where V represents the set of superpixel vertices and E represents the set of undirected edges connecting two superpixels;

[0032] Connect the two vertices v a and v b The edge is represented as edge e a,b The weight w ab ;

[0033] Gaussian weight is used to represent the weight between two vertices, and brightness similarity is used as the edge weight between superpixels;

[0034] An adjacency matrix of the undirected graph G is constructed based on the edge weights.

[0035] According to the above aspects and any possible implementation, an implementation is further provided, wherein the background score vector is iteratively updated, and the updated background score vector is mapped to the infrared image, and the connection strength is obtained after normalization processing as follows:

[0036] A probability transfer matrix P of transfer between vertices is defined on the undirected graph G;

[0037] For infrared pedestrian images, the edge area is used as the background seed, the center is used as the foreground seed, and an initial background score vector of a node is set based on the background seed and the foreground seed;

[0038] Based on the probability transfer matrix and the initial background score vector, an updated background score vector is obtained in a continuous iterative updating process;

[0039] The updated background score vector is mapped to the infrared image and normalized to obtain the connection strength.

[0040] According to the above aspects and any possible implementation, a further implementation is provided, wherein the cluster center of the background area of ​​the image is used as the dynamic threshold in the following manner:

[0041] Using an iterative and progressive approach, the ignition area is increased from small to large, and the background area with relatively low thermal radiation value in the image is gradually segmented. The expression is as follows:

[0042]

[0043] Among them, I ij is the gray value of the pixel with coordinates (i, j) on the original infrared image, Y ij(n) is the pulse output under the current iterative operation, O is the number of row pixels of the original infrared image, P is the number of column pixels of the original infrared image, and m(n) represents the neuron pulse output Y ij (n) Divide the entire image into background Y ij (n) = 1 and target Y ij Under the specific condition of (n)=0, the cluster center of the background is the dynamic threshold.

[0044] According to the above aspects and any possible implementation, an implementation is further provided, wherein the process of performing image segmentation on the processed image using the improved pulse coupled neural network model to obtain the segmented image is as follows:

[0045] Initializing the improved pulse coupled neural network model and setting model parameters, setting the number of initialization iterations n=1;

[0046] Segmenting the processed image I using the improved pulse coupled neural network model to obtain an output segmented image Y;

[0047] It is determined whether the dynamic threshold value continues to increase. If it increases, the number of iterations is automatically increased by 1, and the image segmentation is continued. Otherwise, the segmented image Y is used as the final segmentation result and output to obtain the final output segmented image.

[0048] According to a second aspect of the present disclosure, there is provided a pedestrian infrared image segmentation system for use under complex conditions, comprising: a pedestrian infrared image acquisition module, a neural network model construction module, and an image segmentation module;

[0049] The pedestrian infrared image acquisition module is used to acquire pedestrian infrared images, and use the detailed information in the visible light image to perform denoising on the infrared image using a guided image filtering method to obtain a processed image;

[0050] The neural network model construction module is used to use the background score of each pixel in the processed image as the connection strength, use the cluster center of the background area of ​​the processed image as the dynamic threshold, and construct an improved pulse coupled neural network model based on the connection strength and the dynamic threshold;

[0051] The image segmentation module is used to collect infrared images of pedestrians in real time and use the improved pulse coupled neural network model to segment the infrared images of pedestrians collected in real time to obtain segmented images.

[0052] Compared with the prior art, the present invention has the following technical effects:

[0053] The present invention takes into account that the proportion of background area in the image is much larger than that of human target area, and the similarity between the background area neighborhoods is relatively large. A new idea of ​​prioritizing the segmentation of background area is proposed, and the infrared pedestrian image is represented as a graph structure. The nodes of the graph represent superpixels in the infrared image, and the weights of the edges reflect the similarity or distance between the nodes. The structure of the graph is analyzed to evaluate the score of each node becoming the background area, and the score is mapped to the original image as the connection strength of the model. Correspondingly, the dynamic threshold is set as the clustering center of the image background, and the background area of ​​the image is preferentially classified and outputted to obtain the final human target segmentation result. Therefore, the parameters in the present invention are determined according to the properties of the image itself. At the same time, the model can automatically stop according to the segmentation result, and will not produce over-segmentation. At the same time, the model can effectively segment infrared pedestrian images of various complex environments, especially infrared images with a lot of heat source interference in the background. It has a better segmentation effect.

[0054] It should be understood that the contents described in the Summary of the Invention section are not intended to limit the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. The accompanying drawings are provided for a better understanding of the present disclosure and do not constitute a limitation of the present disclosure. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, among which:

[0056] Figure 1 A schematic flow chart of a pedestrian infrared image segmentation method under complex conditions according to an embodiment of the present disclosure is shown;

[0057] Figure 2 8 infrared pedestrian images in complex environments collected by a method for segmenting pedestrian infrared images in complex conditions according to an embodiment of the present disclosure are shown;

[0058] Figure 3 8 true value images of infrared pedestrian images in complex environments collected according to a method for segmenting infrared pedestrian images in complex conditions according to an embodiment of the present disclosure are shown;

[0059] Figure 4 A schematic structural diagram of a pedestrian infrared image segmentation system for use under complex conditions according to an embodiment of the present disclosure is shown;

[0060] Figure 5 A schematic diagram of infrared pedestrian image segmentation results using the Ostu algorithm according to an embodiment of a method for infrared image segmentation of pedestrians under complex conditions is shown;

[0061] Figure 6 A schematic diagram of infrared pedestrian image segmentation results using a Canny operator according to an embodiment of a pedestrian infrared image segmentation method under complex conditions is shown;

[0062] Figure 7 A schematic diagram of infrared pedestrian image segmentation results using a Prewitt operator according to an embodiment of a method for infrared pedestrian image segmentation under complex conditions is shown;

[0063] Figure 8 A schematic diagram of infrared pedestrian image segmentation results using a Sobel operator according to an embodiment of a method for infrared pedestrian image segmentation under complex conditions is shown;

[0064] Figure 9 A schematic diagram of infrared pedestrian image segmentation results using a weighted entropy Ostu algorithm according to an embodiment of a method for infrared pedestrian image segmentation under complex conditions is shown;

[0065] Figure 10 A schematic diagram of infrared pedestrian image segmentation results using a traditional PCNN algorithm according to an embodiment of a method for infrared pedestrian image segmentation under complex conditions is shown;

[0066] Figure 11 A schematic diagram of infrared pedestrian image segmentation results based on a spectral residual PCNN algorithm according to an embodiment of a pedestrian infrared image segmentation method for complex conditions is shown;

[0067] Figure 12 A schematic diagram of infrared pedestrian image segmentation results of an improved PCNN algorithm according to an embodiment of a pedestrian infrared image segmentation method for complex conditions according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0068] To make the purpose, technical solutions, and advantages of the embodiments of the present disclosure more clear, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present disclosure.

[0069] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0070] Reference Figure 1 As shown, this embodiment provides a pedestrian infrared image segmentation method for complex conditions, comprising the following steps:

[0071] S101 , using detail information in the visible light image and a guided image filtering method to perform denoising on the collected infrared image of the pedestrian to obtain a processed image.

[0072] like Figure 2 and Figure 3 As shown in this example, an infrared camera is used to capture outdoor pedestrian images. Infrared spectroscopy technology has become a hot topic in pedestrian detection research due to its superior performance at night and in low-visibility conditions. Infrared images can effectively capture the thermal radiation emitted by human body temperature, enabling accurate pedestrian identification in various environments.

[0073] S102. Constructing an improved pulse coupled neural network model, including: using the background score of each pixel in the image as the connection strength, and using the cluster center of the background area of ​​the image as the dynamic threshold.

[0074] In this embodiment, the background score of each pixel in the processed image is used as the connection strength. The cluster center of the background area is used as the dynamic threshold E ij [n], the improved PCNN model is as follows:

[0075]

[0076] E ij [n]=m(n) (5)

[0077] Among them, F ij [n] is the feedback input, L ij [n] is the connection input, U ij [n] is the internal activity item, Y ij [n] is the pulse output, S kl is the external stimulus, usually corresponding to the pixel intensity in the image, (k,l) is the neighborhood of neuron (i,j), M ij,kl and W ij,kl is the weight matrix, E ij [n] is the dynamic threshold, λ is a coefficient to adjust the dynamic threshold and control the iteration speed of the model. For the connection strength.

[0078] Furthermore, in this embodiment, the connection strength The specific calculation process is:

[0079] The SLIC superpixel segmentation algorithm is used to segment the infrared pedestrian image into multiple superpixels. Each superpixel is regarded as a node of the graph, and an undirected graph G is established with the generated superpixels as vertices.<V,E> , using brightness similarity as the edge weight between superpixels, defining a probability transfer matrix P for transfer between vertices on the graph G, setting a background score vector H of a node, and updating H in the iteration, mapping it to the original infrared image, and performing normalization to obtain the improved connection strength

[0080] Furthermore, the process of using the SLIC superpixel segmentation algorithm to segment the infrared pedestrian image into multiple superpixels is as follows:

[0081] First, initialize k super-pixel centers. For an infrared image containing N pixels, the super-pixel center C i It can be expressed as (l i ,x i ,y i ), the image is divided into grid steps Divide into k square grids, each initial superpixel center searches for the position of the minimum gradient by searching the pixels within the 3×3 range around it, and changes the cluster center from the original square pixel center pixel to the point with the minimum gradient in its neighborhood; after obtaining k initial cluster centers, calculate the characteristic distance D between the pixels within the 2F×2F range around the cluster center and the cluster center, and divide the pixel into the corresponding superpixel cluster according to the distance. The distance metric expression is as follows:

[0082]

[0083] Among them, d c Indicates the brightness distance between the pixel and the center pixel, d s Represents the distance in space, and m is a fixed constant.

[0084] Finally, k superpixels L(i,j)∈{1,2,..,k} are obtained, where L(i,j) is the superpixel label to which pixel (i,j) belongs.

[0085] Furthermore, in this embodiment, the construction process of the undirected graph G is:

[0086] Based on the idea of ​​random walk algorithm, the generated superpixels are used as vertices to establish an undirected graph G=<V,E> , V represents the superpixel vertex set, E represents the undirected edge set connecting two superpixels, and the edge connecting two vertices v a and v b The edge is represented as edge e a,b The weight w ab , Gaussian weight is used to represent the weight between two vertices, and brightness similarity is used as the edge weight between superpixels. The expression is as follows:

[0087]

[0088] Among them, C a and C b is the brightness feature of superpixels a and b, σ is the parameter that controls the similarity, so the adjacency matrix A of graph G can be obtained as [w ab ] k .

[0089] Furthermore, in this embodiment, the background score vector H is mapped to the infrared image and normalized to obtain the connection strength The process is:

[0090] Define a probability transfer matrix P for transitions between vertices on the graph G. Since the sum of the transition probabilities starting from a vertex is 1, that is, the sum of the elements in each row of P is 1, the adjacency matrix can be regularized to obtain the transfer matrix:

[0091] P=D -1 A (8)

[0092] Where D is a diagonal matrix consisting of the degree of each vertex, D = diag(d1, d2, ..., d m ), is the degree matrix of graph G.

[0093] For infrared pedestrian images, the edge area is used as the background seed, represented as sn(L(i,j)∈edge)=1, and the center is used as the foreground seed sn(L(i,j)∈center)=2. A background score vector H of a node is set and updated in iterations:

[0094]

[0095] H (t+1) =PH (t) (10)

[0096]

[0097] in, is the initial background score vector of each node, is the background score vector of each node at the t+1th iteration, H (t) is the background score vector of each node at the tth iteration.

[0098] The final background score vector H is mapped to the original infrared image and normalized to obtain the improved connection strength

[0099]

[0100] Among them, normalize(·) is a normalization function that normalizes the value of the connection strength to [0,1].

[0101] Furthermore, in this embodiment, the cluster center of the background area is used as the dynamic threshold E ij [n] The method is:

[0102] When designing the dynamic threshold, an iterative and progressive approach is used to gradually increase the ignition area from small to large, and gradually segment the background area with relatively low thermal radiation value in the image. The expression is as follows:

[0103]

[0104] Where, I ij is the gray value of the pixel with coordinates (i, j) on the original infrared image, Y ij (n) is the pulse output under the current iterative operation, m(n) represents the neuron pulse output Y ij (n) Divide the entire image into background (Y ij (n)=1) and target (Y ij The cluster center of the background under the specific condition of (n)=0), O is the number of row pixels of the original infrared image, P is the number of column pixels of the original infrared image, and n is the number of iterations.

[0105] S103 , collecting infrared images of pedestrians in real time, and segmenting the infrared images of pedestrians collected in real time using an improved pulse coupled neural network model to obtain segmented images.

[0106] In this embodiment, the improved pulse coupled neural network model is first initialized and model parameters are set, and the number of iterations n is initialized to 1.

[0107] Specifically, in this embodiment, the model parameter M that needs to be set is ij,kl 、W ij,kl and λ, the expressions are as follows:

[0108]

[0109] λ=0.2

[0110] Among them, (i, j) represents the central neuron and (k, l) represents the area neuron.

[0111] Then, the improved pulse coupled neural network model is used to segment the real-time pedestrian infrared image I and obtain the output segmented image Y;

[0112] Finally, determine the dynamic threshold E ij[n] Whether it continues to increase, if it increases, the number of iterations will automatically increase by 1, and continue to segment the image, otherwise the final segmentation result of Y will be output to obtain the final output result.

[0113] like Figure 4 As shown, this embodiment also provides a pedestrian infrared image segmentation system for complex conditions, including: a pedestrian infrared image acquisition module 1, a neural network model construction module 2 and an image segmentation module 3;

[0114] The pedestrian infrared image acquisition module 1 is used to acquire pedestrian infrared images, and use the detailed information in the visible light image and the guided image filtering method to perform denoising on the infrared image to obtain a processed image;

[0115] The neural network model construction module 2 is used to use the background score of each pixel in the processed image as the connection strength, use the cluster center of the background area of ​​the processed image as the dynamic threshold, and construct an improved pulse coupled neural network model based on the connection strength and the dynamic threshold;

[0116] The image segmentation module 3 is used to collect infrared images of pedestrians in real time, and uses an improved pulse coupled neural network model to segment the infrared images of pedestrians collected in real time to obtain segmented images.

[0117] Example

[0118] This example is based on the MSRS public dataset, from which 144 infrared images with pedestrian targets in complex environments are selected to form the dataset for verifying the algorithm of the present invention. At the same time, the present invention scheme is compared with Ostu, Canny, Prewitt, weighted entropy Ostu, traditional PCNN, spectral residual PCNN and other methods, and 8 typical infrared pedestrian images in complex environments are selected for demonstration. The segmentation results corresponding to the 8 methods are as follows: Figures 5 to 12 , three evaluation indicators, mean intersection over union (MIoU), pixel accuracy, and F1-Score, are used for evaluation and comparison. The calculation formula is as follows:

[0119]

[0120] Among them, Z is the number of classes, v is the index of this class, True Positive (TP) represents the number of pixels correctly segmented as targets, TN (True Negative) represents the number of pixels correctly segmented as background, FP (False Positive) represents the number of background pixels incorrectly segmented as targets, and FN (False Negative) represents the number of target pixels incorrectly segmented as background.

[0121] The evaluation indicators are calculated according to the above formula. The evaluation indicators of the eight image segmentation algorithms on this dataset are shown in Table 1.

[0122] The MIoU, Pixel Accuracy and F1-score of the method of the present invention are 0.524, 0.985 and 0.665 respectively, achieving the highest index value. The MIoU of the method of the present invention is improved by 19.6% to 44.4% compared with other image segmentation algorithms, the Pixel Accuracy is improved by 0.2% to 7.4% compared with other image segmentation algorithms, and the F1-score is improved by 15.1% to 41.2% compared with other image segmentation algorithms, indicating that the algorithm has a good segmentation effect for human targets in complex environments.

[0123] Table 1 Evaluation indicators of different segmentation algorithms

[0124]

[0125] Taking into account that the background area in the image accounts for a much larger proportion than the human target area, and the similarity between the background area neighborhoods is relatively large, the present invention proposes a new idea for preferentially segmenting the background area, representing the infrared pedestrian image as a graph structure, the nodes of the graph represent superpixels in the infrared image, and the weights of the edges reflect the similarity or distance between the nodes. The structure of the graph is analyzed to evaluate the score of each node becoming the background area, and the score is mapped to the original image as the connection strength of the model. Correspondingly, the dynamic threshold is set as the cluster center of the image background, and the background area of ​​the image is preferentially classified and outputted, thereby obtaining the final human target segmentation result, thereby effectively improving the segmentation effect of infrared images with a lot of heat source interference in the background.

[0126] It should be noted that for the aforementioned method embodiments, for simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present disclosure is not limited by the order of the actions described, because according to the present disclosure, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily required by the present disclosure.

[0127] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not limited herein.

[0128] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A pedestrian infrared image segmentation method for complex conditions, characterized in that: It includes the following steps: Utilize the detailed information in the visible light image, and adopt the guided image filtering method to denoise the collected infrared image of pedestrians, obtaining the processed image; Construct an improved pulse-coupled neural network model, including: taking the background score of each pixel in the image as the connection strength, and taking the clustering center of the background area of the image as the dynamic threshold; Among them, the calculation process of the connection strength is as follows: Adopt the SLIC superpixel segmentation algorithm to segment the infrared image into multiple superpixels, and each superpixel is regarded as a node of the graph; Take the generated superpixels as vertices to establish an undirected graph, and use the brightness similarity as the edge weight between superpixels; Define a probability transition matrix for the transfer between vertices on the undirected graph, and set a background score vector of a node; Iteratively update the background score vector, and map the updated background score vector to the original infrared image. After normalization processing, the connection strength is obtained; Among them, the method of taking the clustering center of the background area of the image as the dynamic threshold is: Adopt an iterative progressive method to gradually increase the ignition area from small to large, and gradually segment the background area with relatively low thermal radiation values in the image. The expression is as follows: Among them, I ij Y is the gray value of the pixel corresponding to the central neuron with coordinates (i, j) on the original infrared image, ij [n] is the pulse output under the current iterative operation, O is the number of row pixels of the original infrared image, P is the number of column pixels of the original infrared image, and m(n) represents the neuron pulse output Y ij [n] Divide the entire image into background Y ij [n] = 1 and target Y ij Under the specific condition of [n]=0, the cluster center of the background is the dynamic threshold; Utilize the improved pulse-coupled neural network model to segment the processed image, obtaining the segmented image.

2. The pedestrian infrared image segmentation method for complex conditions according to claim 1, characterized in that: The improved pulse-coupled neural network model is specifically: E ij [n]=m(n) Among them, F ij [n] is the feedback input, L ij [n] is the connection input, U ij [n] is the internal activity item, Y ij [n] is the pulse output, S kl is the external stimulus, corresponding to the pixel intensity in the image, (k,l) represents the neighborhood of the central neuron with coordinates (i,j), and M ij,kl and W ij,kl is the weight matrix, E ij [n] is the dynamic threshold, λ is a coefficient to adjust the dynamic threshold and control the iteration speed of the model. is the connection strength, n is the number of model iterations, and m(n) represents the neuron pulse output.

3. The pedestrian infrared image segmentation method for complex conditions according to claim 2, characterized in that: The M ij,kl 、W ij,kl and λ are specifically: λ = 0.

2.

4. The pedestrian infrared image segmentation method for complex conditions according to claim 1, characterized in that: The process of adopting the SLIC superpixel segmentation algorithm to segment the infrared image into multiple superpixels is: Initialize several superpixel centers to get the initial superpixel center. For an infrared image containing N pixels, the superpixel center C i It can be expressed as (l i ,x i ,y i ), the image is divided into grid steps Divided into k square grids; For each initial superpixel center, search for the minimum gradient position by searching the pixels within its surrounding 3×3 range; Change the clustering center from the center pixel of the original square pixel cluster to the point with the minimum gradient in its neighborhood, and thus obtain k initial clustering centers; Calculate the feature distance D between the pixel points within the range of 2F×2F around the initial clustering center and the initial clustering center, and divide the pixel points into the corresponding superpixel clusters according to the feature distance D, obtaining multiple superpixels.

5. The pedestrian infrared image segmentation method for complex conditions according to claim 4, characterized in that: The construction process of the undirected graph is: Adopt the random walk method to establish an undirected graph G = <V, E> with the generated superpixels as vertices, where V represents the set of superpixel vertices and E represents the set of undirected edges connecting two superpixels; Connect the two vertices v a and v b The edge is represented as edge e a,b The weight w ab ; Adopt Gaussian weights to represent the weights between two vertices, and use the brightness similarity as the edge weight between superpixels; Based on the edge weights, construct the adjacency matrix of the undirected graph G.

6. The pedestrian infrared image segmentation method for complex conditions according to claim 5, characterized in that: The process of iteratively updating the background score vector, mapping the updated background score vector to the infrared image, and obtaining the connection strength after normalization processing is: Define a probability transition matrix P for the transfer between vertices on the undirected graph G; For the infrared pedestrian image, regard the edge area as the background seed and the center as the foreground seed, and set an initial background score vector of a node based on the background seed and the foreground seed; Based on the probability transition matrix and the initial background score vector, obtain the updated background score vector during the continuous iterative update process; The updated background score vector is mapped to the infrared image and normalized to obtain the connection strength.

7. The pedestrian infrared image segmentation method for complex conditions according to claim 1, characterized in that: The process of performing image segmentation on the processed image using the improved pulse coupled neural network model to obtain the segmented image is as follows: Initializing the improved pulse coupled neural network model and setting model parameters, setting the number of initialization iterations n=1; Segmenting the processed image I using the improved pulse coupled neural network model to obtain an output segmented image Y; It is determined whether the dynamic threshold value continues to increase. If it increases, the number of iterations is automatically increased by 1, and the image segmentation is continued. Otherwise, the segmented image Y is used as the final segmentation result and output to obtain the final output segmented image.

8. A pedestrian infrared image segmentation system for use under complex conditions, for implementing the pedestrian infrared image segmentation method for use under complex conditions according to any one of claims 1 to 7, characterized in that: include: Pedestrian infrared image acquisition module (1), neural network model construction module (2) and image segmentation module (3); The pedestrian infrared image acquisition module (1) is used to acquire pedestrian infrared images, and utilize detail information in visible light images and a guided image filtering method to perform denoising on the infrared images to obtain processed images; The neural network model construction module (2) is used to use the background score of each pixel in the processed image as the connection strength, use the cluster center of the background area of ​​the processed image as the dynamic threshold, and construct an improved pulse coupled neural network model based on the connection strength and the dynamic threshold; The image segmentation module (3) is used to acquire infrared images of pedestrians in real time and segment the infrared images of pedestrians acquired in real time using the improved pulse coupled neural network model to obtain segmented images.

Images