Warehouse fire detection and positioning method based on semantic segmentation

Abstract

The application discloses a warehouse fire detection and positioning method based on semantic segmentation, and belongs to the technical field of image data processing, which comprises the following steps: obtaining a fire data set; constructing an edge feature enhancement module; constructing a fire multi-element segmentation network and training it into a fire segmentation model; obtaining monitoring video in a warehouse and splitting it into single images, which are stored in a time sequence image set A; analyzing every T groups of images in A in turn, wherein one group analysis is as follows: inputting the Tth image in the group into the fire segmentation model, outputting a corresponding semantic segmentation image, if the semantic segmentation image does not simultaneously detect fire and smoke, it is considered safe, otherwise, based on multi-confidence, a fire starting point is identified. T The application can significantly improve the edge segmentation precision of shelves, goods, fire and smoke areas, and based on the law of fire growth and smoke diffusion, multi-confidence fusion calculation is performed, which can effectively distinguish real fire from static interference, and realize accurate positioning of the fire starting point.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image data processing, and in particular to a warehouse fire detection and positioning method based on semantic segmentation. BACKGROUND

[0002] With the increasing demand for safety production and property protection, warehouse fire detection plays a crucial role in ensuring the safety of stored materials, preventing major economic losses, and maintaining the safety of personnel. Especially in modern warehouses that store flammable goods, high-value goods, or automated operations, early fire can not only spread rapidly and cause significant losses, but also disrupt the supply chain and cause serious safety accidents. Therefore, timely and accurate detection of fire and positioning of the fire source is of great significance for intelligent early warning, precise emergency response, and ensuring the continuity of warehouse operations. Currently, the demand for warehouse fire detection and positioning is increasingly urgent, and an efficient and accurate technical solution is needed to improve the intelligent level of safety management.

[0003] However, the current mainstream warehouse fire detection methods have obvious limitations in achieving the core goal of "early warning and accurate positioning":

[0004] (1) Manual inspection and sensor monitoring: Manual inspection requires a large amount of human resources. Due to the limitations of inspection cycle and coverage, it is difficult to detect sudden fires in time. Especially in large warehouses or complex shelving environments, the efficiency and reliability of manual inspection are significantly reduced, resulting in response lag and inability to achieve real-time monitoring in all-weather and non-blind areas. Sensor monitoring based on smoke or temperature is an indirect detection method that requires the fire to develop to a certain stage (such as generating sufficient smoke or heat) to trigger an alarm, making it difficult to detect early fires that are smoldering or have no visible smoke. In addition, sensors are easily disturbed by environmental factors such as warehouse dust and humidity, resulting in a high false alarm rate and inability to provide accurate location information of the fire source.

[0005] (2) Visual detection based on traditional image processing: This method captures visible light images through ordinary monitoring cameras and uses traditional image processing techniques (such as color segmentation and texture analysis) to attempt to identify flames or smoke. Although it can provide visual information, it relies on manually designed features that have poor robustness in complex and variable warehouse environments. When faced with interference such as goods accumulation, light changes, and equipment reflections, traditional algorithms have insufficient generalization ability and detection accuracy for flames and smoke, especially their fuzzy edges and variable shapes, making it impossible to achieve pixel-level accurate positioning. SUMMARY

[0006] The purpose of this invention is to provide a warehouse fire detection and location method based on semantic segmentation that solves the above-mentioned problems, can effectively distinguish between real fires and visual interference such as light reflections and equipment steam, reduces false alarm rate and missed detection rate, and can accurately locate the fire point by utilizing the temporal continuity of video frames.

[0007] To achieve the above objectives, the technical solution adopted by this invention is as follows: a warehouse fire detection and location method based on semantic segmentation, comprising the following steps:

[0008] S1, Obtain the fire dataset, where the target categories of the samples include shelves, goods, flames, and smoke;

[0009] S2, construct an edge feature enhancement module, which is used to input the first feature map C1, and perform fast Fourier transform, 1×1 convolution, activation function activation, 1×1 convolution, and inverse fast Fourier transform in sequence to obtain the second feature map C2. Then, add C1 and C2 element by element to obtain the third feature map C3. Perform coordinate attention on C3 to obtain the fourth feature map C4. Then, add C3 and C4 element by element to obtain the fifth feature map C5.

[0010] S3 constructs a multi-element segmentation network for fire situations, including a ResNet network, a ViT network, an edge feature enhancement module, and a decoder. One sample passes through the ResNet network and the edge feature enhancement module to output C1 and C5 in sequence, and the other passes through the ViT network to output the sixth feature map C6. C5 and C6 are concatenated and decoded back to the sample size by the decoder, and a pixel-level semantic segmentation map is output.

[0011] S4. Use the fire dataset to train the fire multi-element segmentation network to obtain the fire segmentation model;

[0012] S5, acquire the surveillance video inside the warehouse, split it into individual images according to time, and store it in time series atlas A;

[0013] S6, group the images in A into T groups, and analyze each group sequentially. The analysis method for one group is as follows: output the semantic segmentation map Sem from the T-th image of the group using the fire segmentation model. T If Sem T If neither flame nor smoke is detected simultaneously, proceed to step S7; if both flame and smoke are detected simultaneously, proceed to step S8.

[0014] S7, recording Sem T Once it is safe to proceed, end this group's analysis.

[0015] S8, identify the ignition point, including S81~S84;

[0016] S81, sequentially process the T images in this group using the fire segmentation model to obtain the corresponding semantic segmentation maps Sem1~Sem1. T;

[0017] S82, based on Sem1~Sem T The mean and standard deviation of the growth rates in the flame and smoke regions are used to construct the first confidence level IF1, where 0 ≤ IF1 ≤ 1;

[0018] S83, Sem1~Sem T Fit a straight line to the centroid of the flame region and count the number of outliers N. out Calculate the second confidence level IF2=N out / T;

[0019] S83, for t semantic segmentation graphs Sem t For any given condition, 1 ≤ t ≤ T, merge the flame and smoke areas into a single hazardous area. If the shelf intersects with the hazardous area, the hazardous value of the shelf is d1 = 0.5; otherwise, it is 0. If the goods intersect with the hazardous area, the hazardous value of the goods is d2 = 0.5; otherwise, it is 0. Calculate Sem. t The third confidence level ;

[0020] S84, for IF1, IF2, Weighted fusion yields Sem t Danger value d t T hazard values ​​are compared sequentially with a preset threshold. If all are less than the threshold, Sem is recorded. T If the timing is pending, the analysis for this group ends; otherwise, obtain the semantic segmentation map that first exceeds the threshold, take the centroid of the danger zone as the ignition point and issue an early warning, then end the analysis for this group.

[0021] Preferably, the fire dataset is a D2S and a FoggFire Dataset.

[0022] Preferably, the edge feature enhancement module includes a fast Fourier transform layer, a convolutional unit, an inverse fast Fourier transform layer, a first summing layer, a coordinate attention layer, and a second summing layer;

[0023] The Fast Fourier Transform layer is used to transform the C1 two-dimensional Fast Fourier Transform into Fourier features;

[0024] The convolutional unit is used to sequentially perform 1×1 convolution, activation function activation, and 1×1 convolution operations on the Fourier features to obtain convolutional features;

[0025] The inverse fast Fourier transform layer is used to perform inverse fast Fourier transform on the convolutional features to obtain C2;

[0026] The first additive layer is used to obtain C3 according to the formula C3=C1⊕C2, where ⊕ means element-wise addition;

[0027] The coordinate attention layer is used to extract the fourth feature map C4 from C3 by performing coordinate attention.

[0028] The second additive layer is used to obtain C5 according to the formula C5=C3⊕C4.

[0029] Preferably, S82 specifically involves: calculating the mean growth rate μ of the flame region in the first T semantic segmentation images. F and standard deviation σ F The mean growth rate of the smoke area μ S and standard deviation σ S The flame confidence level R is calculated according to the following formula. F And smoke confidence R S Then calculate the first confidence level IF1, IF1=R F +R S ;

[0030] , ,

[0031] In the formula, , These are the standard deviation weights for flame and smoke, respectively.

[0032] As a preferred embodiment, in S82, μ F σ F Calculate according to the following formulas respectively;

[0033] ,

[0034] ,

[0035] ,

[0036] In the formula, For the i-th semantic segmentation map Sem i The growth rate of the flame area, F t F t-i Sem t The ti-th semantic segmentation graph Sem t-i Area of ​​the flame zone.

[0037] As a preferred embodiment, in S82, μ S σ S Calculate according to the following formulas respectively;

[0038] ,

[0039] ,

[0040] ,

[0041] In the formula, For the i-th semantic segmentation map Sem i The growth rate of the smog area, S t S t-i Sem t The ti-th semantic segmentation graph Sem t-i Area of ​​the smoke-filled region.

[0042] As a preferred option, in S84, Sem t Danger value d t Weighted fusion is performed according to the following formula:

[0043] ,

[0044] In the formula, λ1, λ2, and λ3 represent IF1, IF2, IF3, and IF4 respectively. The weights are λ1+λ2+λ3=1.

[0045] Regarding the edge feature enhancement module: The shallow features output by ResNet18 (the first feature map C1 in this invention) are transformed to the frequency domain using Fast Fourier Transform (FFT). High-frequency information of the corresponding edges and textures is enhanced through convolution operations. Then, the module is restored to the spatial domain using Inverse Fast Fourier Transform (IFT), and residual connections are used to preserve the original features and prevent information loss. Subsequently, a coordinate attention mechanism is used to focus on key spatial and channel contexts, suppressing noise and refining feature representation. Through the synergistic effect of frequency domain enhancement and the attention mechanism, the edge segmentation accuracy of target regions such as shelves, goods, flames, and smoke areas can be significantly improved.

[0046] Regarding the multi-element fire segmentation network: It introduces features C1 and C6 from different levels of the output samples of ResNet and ViT networks. C1 is further enhanced by the edge feature enhancement module, which outputs C5 to strengthen the details of the boundaries of shelves, goods, flames, and smoke, maximizing the preservation of detailed information in shallow features and helping to optimize the edge contours of the segmented targets. C6 has a deeper receptive field and global contextual information. The fusion of C5 and C6 constructs a hierarchical and information-complete feature representation, enhancing the robustness of the model and the clarity of the segmentation boundaries.

[0047] Regarding the method for locating the ignition point: Specifically, steps S6 to S8 of this invention are used. This method utilizes the temporal continuity of video frames to dynamically track the growth of flames and the spread of smoke, and introduces three confidence levels IF1 to IF3 to quantify the hazard value.

[0048] Firstly, to save computing power, the monitoring videos in the warehouse will not be semantically segmented frame by frame using the fire segmentation model. Instead, they will be stored in the time series graph set A in chronological order, and monitored by sampling one video per group of T videos. In this invention, the last video of each group (the Tth video) will be sampled, and then the specific situation of the semantic segmentation graph corresponding to the Tth video will be discussed in different cases.

[0049] Scenario 1: If neither flame nor smoke is detected simultaneously in the semantic segmentation map, it indicates that there are no fire-related issues, and the Sem can be recorded. T It ensures safety at all times and jumps to the last image of the next group for semantic segmentation and recognition, thereby saving a lot of computing power and improving work efficiency.

[0050] Scenario 2: If both flame and smoke are detected simultaneously in the semantic segmentation image, it indicates a fire has started. Rapid and accurate location of the ignition point is required. Since flame growth and smoke diffusion are dynamic, and there are interferences such as target occlusion and overlap between the flames and smoke, and between the flames and smoke and shelves / goods, the flame region identified in the T-th semantic segmentation image may not be the accurate ignition point. Therefore, multi-confidence analysis is necessary. This invention introduces a first confidence level (IF1) to a third confidence level (IF3).

[0051] The first confidence level, IF1, comprises two parts. One part is constructed based on the mean and standard deviation of the flame region growth rate; a larger value indicates more and more stable flame region growth, making the model's output image more reliable. Similarly, the second confidence level, IF2, is constructed based on the centroid of the flame region. According to the flame growth pattern, as the flame grows larger, its centroid gradually rises. This invention fits T centroids into a straight line, then identifies outliers. IF2 is calculated based on the proportion of outliers to T. Fewer outliers indicate rapid flame growth, resulting in a large IF2 value. More outliers indicate slow flame development or significant jumps in segmented regions, leading to low reliability. IF1 and IF2 are obtained from T semantic segmentation images, and IF1 and IF2 are equal for the same set of T semantic segmentation images.

[0052] Third confidence level For each semantic segmentation map Sem t Calculated separately, The layout is determined by whether the shelving intersects with the hazardous area and whether the goods intersect with the hazardous area. Finally, IF1, IF2, and IF3 are configured based on experience. Weighted fusion yields each Sem image. t Danger value d t d1~d T The comparisons are then compared with preset thresholds, which may result in two scenarios:

[0053] Case 2.1: d1~dT If all values ​​are less than the threshold, it indicates that the model may have mis-segmented due to environmental and other factors. Therefore, no alarm is issued, and this can be recorded as Sem. T The time is pending, which makes it easy to view the security log, and then jump to the last image of the next group for semantic segmentation and recognition.

[0054] Case 2.2: d1~d T If at least one value is greater than the threshold, find the first dangerous value that is greater than the threshold, then obtain its corresponding semantic segmentation map, take the centroid of the dangerous area (flame area + smoke area) as the ignition point, and then jump to the last image of the next group.

[0055] Compared with the prior art, the advantages of the present invention are as follows:

[0056] (1) An edge feature enhancement module was designed. In the edge feature enhancement module, the shallow features output by the ResNet network are enhanced in the frequency domain by fast Fourier transform. Then, the coordinate attention mechanism is used to focus on the key space and channel context, suppress noise and refine feature expression, effectively optimizing the edge and texture representation of features. Through the synergistic effect of frequency domain enhancement and attention mechanism, the edge segmentation accuracy of shelf, goods, flame and smoke areas can be significantly improved.

[0057] (3) A multi-element fire segmentation network was designed. The shallow features output by the ResNet network were enhanced by the edge feature enhancement module to obtain C5, and then combined with the deep features output by the ViT network for decoding and segmentation, thereby improving the segmentation accuracy of targets related to fire occurrence. Finally, the fire segmentation model trained based on the multi-element fire segmentation network can effectively overcome the problems of insufficient segmentation accuracy in the existing technology due to the diverse categories of targets to be segmented in complex warehouse scenarios, the spatial overlap between goods and fire areas, and the susceptibility of smoke features to environmental dust and light interference. It can effectively improve the segmentation accuracy and boundary integrity of each target, reduce segmentation artifacts caused by smoke or light interference at the edge of the target area, and provide a reliable feature basis for subsequent accurate positioning.

[0058] (3) A method for locating the ignition point is proposed to solve the problem of false alarms and missed alarms in complex warehouse environments caused by goods obstruction and environmental interference. This invention analyzes the temporal area growth trend of flames and smoke, the centroid trajectory of flames, and the overlap between the danger zone and the shelves and goods, and performs multi-confidence fusion calculation to effectively distinguish between real fires and static interference, and can also be based on The method obtains the fire status of each shelf and its goods by calculating the value of the fire and the coordinates of the shelves. Finally, the method obtains a reliable fire segmentation image sequence through multi-dimensional weight calculation, accurately determines the ignition point at the pixel level, realizes the precise location of fire from passive detection to active source tracing, and can provide real-time alarms and output the fire information of each shelf and its goods. Attached Figure Description

[0059] Fig. 1 A schematic diagram of a multi-element fire situation segmentation network;

[0060] Fig. 2 This is a schematic diagram of the edge feature enhancement module;

[0061] Fig. 3 This is a flowchart of the present invention. Detailed Implementation

[0062] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0063] Example 1: See Fig. 1~Fig. 3 A warehouse fire detection and localization method based on semantic segmentation includes the following steps:

[0064] S1, Obtain the fire dataset, where the target categories of the samples include shelves, goods, flames, and smoke;

[0065] S2, construct an edge feature enhancement module, which is used to input the first feature map C1, and perform fast Fourier transform, 1×1 convolution, activation function activation, 1×1 convolution, and inverse fast Fourier transform in sequence to obtain the second feature map C2. Then, add C1 and C2 element by element to obtain the third feature map C3. Perform coordinate attention on C3 to obtain the fourth feature map C4. Then, add C3 and C4 element by element to obtain the fifth feature map C5.

[0066] S3 constructs a multi-element segmentation network for fire situations, including a ResNet network, a ViT network, an edge feature enhancement module, and a decoder. One sample passes through the ResNet network and the edge feature enhancement module to output C1 and C5 in sequence, and the other passes through the ViT network to output the sixth feature map C6. C5 and C6 are concatenated and decoded back to the sample size by the decoder, and a pixel-level semantic segmentation map is output.

[0067] S4. Use the fire dataset to train the fire multi-element segmentation network to obtain the fire segmentation model;

[0068] S5, acquire the surveillance video inside the warehouse, split it into individual images according to time, and store it in time series atlas A;

[0069] S6, group the images in A into T groups, and analyze each group sequentially. The analysis method for one group is as follows: output the semantic segmentation map Sem from the T-th image of the group using the fire segmentation model.T If Sem T If neither flame nor smoke is detected simultaneously, proceed to step S7; if both flame and smoke are detected simultaneously, proceed to step S8.

[0070] S7, recording Sem T Once it is safe to proceed, end this group's analysis.

[0071] S8, identify the ignition point, including S81~S84;

[0072] S81, sequentially process the T images in this group using the fire segmentation model to obtain the corresponding semantic segmentation maps Sem1~Sem1. T ;

[0073] S82, based on Sem1~Sem T The mean and standard deviation of the growth rates in the flame and smoke regions are used to construct the first confidence level IF1, where 0 ≤ IF1 ≤ 1;

[0074] S83, Sem1~Sem T Fit a straight line to the centroid of the flame region and count the number of outliers N. out Calculate the second confidence level IF2=N out / T;

[0075] S83, for t semantic segmentation graphs Sem t For any given condition, 1 ≤ t ≤ T, merge the flame and smoke areas into a single hazardous area. If the shelf intersects with the hazardous area, the hazardous value of the shelf is d1 = 0.5; otherwise, it is 0. If the goods intersect with the hazardous area, the hazardous value of the goods is d2 = 0.5; otherwise, it is 0. Calculate Sem. t The third confidence level ;

[0076] S84, for IF1, IF2, Weighted fusion yields Sem t Danger value d t T hazard values ​​are compared sequentially with a preset threshold. If all are less than the threshold, Sem is recorded. T If the timing is pending, the analysis for this group ends; otherwise, obtain the semantic segmentation map that first exceeds the threshold, take the centroid of the danger zone as the ignition point and issue an early warning, then end the analysis for this group.

[0077] Example 2: See Fig. 1~Fig. 3 More specifically, based on Example 1, the fire dataset is D2S and FoggFire Dataset.

[0078] The edge feature enhancement module includes a Fast Fourier Transform (FFT) layer, a convolutional unit, an Inverse Fast Fourier Transform (IFT) layer, a first addition layer, a coordinate attention layer, and a second addition layer. The FFT layer transforms the two-dimensional FFT of C1 into Fourier features. The convolutional unit sequentially performs 1×1 convolution, activation function activation, and 1×1 convolution operations on the Fourier features to obtain convolutional features. The IFT layer performs an IFT on the convolutional features to obtain C2. The first addition layer obtains C3 according to the formula C3 = C1⊕C2, where ⊕ represents element-wise addition. The coordinate attention layer extracts coordinate attention from C3 to obtain a fourth feature map C4. The second addition layer obtains C5 according to the formula C5 = C3⊕C4.

[0079] S82 specifically involves: calculating the mean growth rate μ of the flame region in the first T semantic segmentation images. F and standard deviation σ F The mean growth rate of the smoke area μ S and standard deviation σ S The flame confidence level R is calculated according to the following formula. F And smoke confidence R S Then calculate the first confidence level IF1, IF1=R F +R S ;

[0080] , ,

[0081] In the formula, , These are the standard deviation weights for flame and smoke, respectively.

[0082] In S82, μ F σ F Calculate according to the following formulas respectively;

[0083] ,

[0084] ,

[0085] ,

[0086] In the formula, For the i-th semantic segmentation map Sem i The growth rate of the flame area, F t F t-i Sem t The ti-th semantic segmentation graph Sem t-i Area of ​​the flame zone.

[0087] In S82, μS σ S Calculate according to the following formulas respectively;

[0088] ,

[0089] ,

[0090] ,

[0091] In the formula, For the i-th semantic segmentation map Sem i The growth rate of the smog area, S t S t-i Sem t The ti-th semantic segmentation graph Sem t-i Area of ​​the smoke-filled region.

[0092] In S84, Sem t Danger value d t Weighted fusion is performed according to the following formula:

[0093] ,

[0094] In the formula, λ1, λ2, and λ3 represent IF1, IF2, IF3, and IF4 respectively. The weights are λ1+λ2+λ3=1. For example, in this embodiment, λ1, λ2, and λ3 are set to 0.4, 0.3, and 0.3 respectively. The actual weight values ​​can be adjusted based on experience.

[0095] Example 3: See Fig. 1~Fig. 3 Based on Example 1, a comparative experiment was conducted using the semantic segmentation models of the present invention and existing technologies. The datasets used in the experiment included D2S and FoggFire Datasets, with a training set to validation set ratio of 8:2. The fire multi-element segmentation network of the present invention and several existing semantic segmentation networks were trained using the training sets, and then validated using the validation sets. The results are shown in Table 1 below:

[0096] Table 1. Comparison of performance indicators of different networks

[0097] Method mAccuracy% mIoU % SegNet 56.35 38.62 DeeplabV3+ 62.29 48.15 HRNet 74.56 63.09 UNet 71.42 58.63 The present invention 78.23 66.78

[0098] In Table 1, mACuracy is the average accuracy and mIoU is the average crossover ratio. As can be seen from Table 1, the two performance indicators of the present invention are significantly better than those of the prior art.

[0099] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A warehouse fire detection and localization method based on semantic segmentation, characterized in that, Includes the following steps: S1, Obtain the fire dataset, where the target categories of the samples include shelves, goods, flames, and smoke; S2, construct an edge feature enhancement module, which is used to input the first feature map C1, and perform fast Fourier transform, 1×1 convolution, activation function activation, 1×1 convolution, and inverse fast Fourier transform in sequence to obtain the second feature map C2. Then, add C1 and C2 element by element to obtain the third feature map C3. Perform coordinate attention on C3 to obtain the fourth feature map C4. Then, add C3 and C4 element by element to obtain the fifth feature map C5. S3. Construct a multi-element segmentation network for fire situation, including a ResNet network, a ViT network, an edge feature enhancement module, and a decoder. One sample goes through the ResNet network and the edge feature enhancement module to output C1 and C5 in sequence, and the other goes through the ViT network to output the sixth feature map C6. C5 and C6 are concatenated and decoded back to the sample size by the decoder, and a pixel-level semantic segmentation map is output. S4. Use the fire dataset to train the fire multi-element segmentation network to obtain the fire segmentation model; S5, acquire the surveillance video inside the warehouse, split it into individual images according to time, and store it in time series atlas A; S6, group the images in A into T groups, and analyze each group sequentially. The analysis method for one group is as follows: output the semantic segmentation map Sem from the T-th image of the group using the fire segmentation model. T If Sem T If neither flame nor smoke is detected simultaneously, proceed to step S7; if both flame and smoke are detected simultaneously, proceed to step S8. S7, recording Sem T Once it is safe to proceed, end this group's analysis. S8, identify the ignition point, including S81~S85; S81, sequentially process the T images in this group using the fire segmentation model to obtain the corresponding semantic segmentation maps Sem1~Sem1. T ; S82, based on Sem1~Sem T The mean and standard deviation of the growth rates in the flame and smoke regions are used to construct the first confidence level IF1, where 0 ≤ IF1 ≤ 1; S83, Sem1~Sem T Fit a straight line to the centroid of the flame region and count the number of outliers N. out Calculate the second confidence level IF2=N out / T; S84, for t semantic segmentation graphs Sem t For any given condition, 1 ≤ t ≤ T, merge the flame and smoke areas into a single hazardous area. If the shelf intersects with the hazardous area, the hazardous value of the shelf is d1 = 0.5; otherwise, it is 0. If the goods intersect with the hazardous area, the hazardous value of the goods is d2 = 0.5; otherwise, it is 0. Calculate Sem. t The third confidence level ; S85, for IF1, IF2, Weighted fusion yields Sem t Danger value d t T hazard values ​​are compared sequentially with a preset threshold. If all are less than the threshold, Sem is recorded. T If the timing is pending, the analysis for this group ends; otherwise, obtain the semantic segmentation map that first exceeds the threshold, take the centroid of the danger zone as the ignition point and issue an early warning, then end the analysis for this group.

2. The warehouse fire detection and location method based on semantic segmentation according to claim 1, characterized in that, The fire dataset consists of D2S and FoggFire Dataset.

3. The warehouse fire detection and location method based on semantic segmentation according to claim 1, characterized in that, The edge feature enhancement module includes a fast Fourier transform layer, a convolutional unit, an inverse fast Fourier transform layer, a first summing layer, a coordinate attention layer, and a second summing layer; The Fast Fourier Transform layer is used to transform the C1 two-dimensional Fast Fourier Transform into Fourier features; The convolutional unit is used to sequentially perform 1×1 convolution, activation function activation, and 1×1 convolution operations on the Fourier features to obtain convolutional features; The inverse fast Fourier transform layer is used to perform inverse fast Fourier transform on the convolutional features to obtain C2; The first additive layer is used to obtain C3 according to the formula C3=C1⊕C2, where ⊕ means element-wise addition; The coordinate attention layer is used to extract the fourth feature map C4 from C3 by performing coordinate attention. The second additive layer is used to obtain C5 according to the formula C5=C3⊕C4.

4. The warehouse fire detection and location method based on semantic segmentation according to claim 1, characterized in that, S82 specifically involves: calculating the mean growth rate μ of the flame region in the first T semantic segmentation images. F and standard deviation σ F The mean growth rate of the smoke area μ S and standard deviation σ S The flame confidence level R is calculated according to the following formula. F And smoke confidence R S Then calculate the first confidence level IF1, IF1=R F +R S ; ， ， In the formula, , These are the standard deviation weights for flame and smoke, respectively.

5. The warehouse fire detection and location method based on semantic segmentation according to claim 4, characterized in that, In S82, μ F σ F Calculate according to the following formulas respectively; ， ， ， In the formula, For the i-th semantic segmentation map Sem i The growth rate of the flame area, F t F t-i Sem t The ti-th semantic segmentation graph Sem t-i Area of ​​the flame zone.

6. The warehouse fire detection and location method based on semantic segmentation according to claim 4, characterized in that, In S82, μ S σ S Calculate according to the following formulas respectively; ， ， ， In the formula, For the i-th semantic segmentation map Sem i The growth rate of the smog area, S t S t-i Sem t The ti-th semantic segmentation graph Sem t-i Area of ​​the smoke-filled region.

7. The warehouse fire detection and location method based on semantic segmentation according to claim 1, characterized in that, In S85, Sem t Danger value d t Weighted fusion is performed according to the following formula: ， In the formula, λ1, λ2, and λ3 represent IF1, IF2, IF3, and IF4 respectively. The weights are λ1+λ2+λ3=1.

Images