A warehouse smoke and fire detection method and system based on image analysis

Abstract

This application relates to the field of smoke and fire detection technology, specifically disclosing a warehouse smoke and fire detection method and system based on image analysis. The method includes: dividing the monitoring area into sub-regions with different fire risk levels, and acquiring image data and environmental sensor data of the sub-regions; assessing the interference level of the current environment based on the image data and environmental sensor data; dynamically adjusting the image analysis parameters applied to the corresponding sub-regions according to the fire risk level and interference level; detecting and capturing moving targets in the corresponding sub-regions based on the adjusted image analysis parameters; performing visual feature analysis on the moving targets to obtain a fire situation determination result; when the fire situation determination result indicates the presence of a fire, marking the location of the sub-region with the fire and issuing an alarm message. This application improves the performance bottleneck of traditional fixed parameters in terms of "false alarms, missed alarms, and inaccurate positioning" by introducing an adaptive parameter adjustment mechanism.

Description

Technical Field

[0001] This application relates to the field of smoke and fire detection technology, and in particular to a warehouse smoke and fire detection method and system based on image analysis. Background Technology

[0002] In modern smart warehousing environments, image-based smoke and fire detection systems are widely deployed to improve safety and operational efficiency. These systems utilize high-definition cameras for real-time monitoring and employ image processing algorithms to identify the visual characteristics of smoke and flames. However, the harsh and dynamic operating environment of warehouses introduces a series of complex and interconnected challenges, leading to frequent false alarms, missed alarms, and inaccurate fire source location, severely impacting the system's reliability.

[0003] Specifically, dust, fibers, and other suspended particulate matter generated during daily warehouse operations continuously adhere to the surface of camera lenses, forming an uneven coating that reduces light transmittance and causes light scattering, resulting in decreased image clarity and contrast. This optical degradation interferes with the denoising logic in image preprocessing, potentially leading to over-smoothing of images or misjudging noise, thereby weakening the basis for recognizing subtle features such as smoke.

[0004] At night or in low-light conditions, the system activates infrared (IR) supplemental lighting. When IR light passes through a contaminated lens, it produces a severe "IR glare" phenomenon, forming halos or "dirt spots" in the image, making the overall image "foggy," drastically reducing contrast, and further obscuring the already faint initial smoke and fire characteristics.

[0005] Furthermore, the densely stacked goods in the warehouse will cast blurry, moving, and irregular shadows under IR glare due to layout adjustments. The visual characteristics of these shadows (such as slow diffusion and blurred edges) are highly similar to the diffusion pattern of the initial smoke, which seriously interferes with the system's preset smoke pattern recognition rules.

[0006] More importantly, to avoid false alarms triggered by routine activities such as forklift dust, traditional systems typically employ aggressive background modeling update strategies and strict filtering mechanisms for small, slowly moving targets. However, when image quality is already severely degraded due to the aforementioned issues, the actual initial smoke diffusion also presents as a weak and slow movement. This makes it highly susceptible to misinterpreting real initial smoke signals as background changes and incorporating them into the background model, or having them directly rejected by the filtering mechanism, resulting in significant missed detections. Ultimately, traditional systems face a detection dilemma due to the interplay of multiple real-world factors, making accurate identification and location difficult in the early stages of a fire. Summary of the Invention

[0007] This application proposes a warehouse smoke and fire detection method and system based on image analysis, aiming to solve the technical problems of false alarms, missed alarms and inaccurate fire source location faced by existing image analysis systems in warehouse smoke and fire detection, especially the detection difficulties caused by factors such as lens contamination, infrared glare, blurred shadows caused by high-density stacked goods and overly aggressive background modeling strategies in complex and ever-changing warehouse environments.

[0008] In a first aspect, this application provides a warehouse smoke and fire detection method based on image analysis, used to detect fires in a monitored area of ​​a warehouse. The method includes the following steps:

[0009] The monitoring area is divided into sub-areas with different fire risk levels, and image data and environmental sensor data of the sub-areas are acquired.

[0010] The degree of interference in the current environment is assessed based on the high-frequency noise characteristics in the image data and the particulate matter concentration indicated by the environmental sensor data.

[0011] Based on the fire risk level and interference level, the image analysis parameters applied to the corresponding sub-region are dynamically adjusted; wherein, the image analysis parameters include: the update rate of the background modeling algorithm, and the sensitivity threshold for moving target detection;

[0012] The adjusted update rate for each sub-region is applied to the background modeling algorithm to control the absorption rate of the background model for potential targets that conform to the preset physical characteristics of smoke diffusion; and based on the sensitivity threshold adjusted for each sub-region, moving targets are detected and captured in the corresponding sub-region; wherein the captured moving targets include the potential targets separated from the background;

[0013] Visual feature analysis is performed on the moving target to obtain the fire situation determination result;

[0014] When the fire situation determination result indicates that a fire exists, the location of the sub-area where the fire exists is marked and an alarm message is issued.

[0015] As some embodiments of this application, the step of dividing the monitoring area into sub-areas with different fire risk levels and acquiring image data and environmental sensor data of the sub-areas includes:

[0016] Call the data interface of the warehouse management system to read the cargo attribute information associated with a specific storage location;

[0017] Read the shelf height sensor data associated with a specific storage location or the stacking status information entered manually, as the shelf stacking density information;

[0018] Based on the cargo attribute information and the stacking density information, the monitoring area is divided into sub-areas with different fire risk levels;

[0019] Based on the cameras deployed in the sub-region, real-time video streams of the sub-region are acquired as image data.

[0020] Based on the laser scattering particulate matter sensor deployed in the sub-region, the concentration data of particulate matter in the air of the sub-region is acquired as environmental sensor data.

[0021] As some embodiments of this application, the step of assessing the level of interference in the current environment based on the high-frequency noise features in the image data and the particulate matter concentration indicated by the environmental sensor data includes:

[0022] Calculate the average energy of high-frequency noise in the image data within a preset frequency band;

[0023] Obtain the real-time particulate matter concentration value indicated by the environmental sensor data;

[0024] If the average energy is lower than a preset energy threshold and the real-time particulate matter concentration is lower than a preset concentration threshold, then the interference level of the current environment is determined to be low interference level.

[0025] Otherwise, the interference level of the current environment is determined to be high interference level.

[0026] As some embodiments of this application, the step of assessing the level of interference in the current environment based on the high-frequency noise features in the image data and the particulate matter concentration indicated by the environmental sensor data further includes:

[0027] The image data is converted to a preset color space, and the overall average saturation and average brightness of the low-brightness areas of the image data are calculated; wherein, the low-brightness areas are the regions composed of pixels in the preset color space whose brightness components are lower than a preset brightness threshold.

[0028] Within a preset time period, when the average brightness of the low-brightness area shows an upward trend and the overall average saturation of the image shows a downward trend, and the real-time particulate matter concentration value is continuously lower than the preset cleaning threshold, the humidity data of the corresponding sub-area within the preset historical time period is obtained.

[0029] When the fluctuation range of the humidity data during the historical time period exceeds the preset fluctuation threshold, the current environmental interference level is determined to be latent environmental interference level.

[0030] As some embodiments of this application, the step of dynamically adjusting the image analysis parameters applied to the corresponding sub-region based on the fire risk level and the degree of interference includes:

[0031] Based on the fire risk level, a background modeling algorithm update rate is set for the corresponding sub-area, and the update rate is negatively correlated with the fire risk level.

[0032] Based on the level of interference, a sensitivity threshold for moving target detection is set for the corresponding sub-region, and the sensitivity threshold is negatively correlated with the level of interference.

[0033] When the interference level is the latent environmental interference level, the update rate of the corresponding sub-region is adjusted to a lower preset value than that of the high interference level, and the sensitivity threshold of the corresponding sub-region is adjusted to a lower preset value than that of the high interference level.

[0034] As some embodiments of this application, the step of detecting and capturing a moving target in a corresponding sub-region based on the sensitivity threshold adjusted for each sub-region includes:

[0035] Within each sub-region, pixel change regions in the image data that conform to the adjusted sensitivity threshold for capturing potential targets that meet the preset physical characteristics of smoke diffusion are identified, and spatially connected pixel change regions are merged into motion connected regions; wherein, determining whether a pixel change region conforms to the adjusted sensitivity threshold includes: setting a pixel area threshold and a duration frame number threshold for the sub-region based on the adjusted sensitivity threshold; when the pixel area of ​​the pixel change region is not less than the pixel area threshold, and / or the number of times the pixel change region appears continuously in consecutive image frames is not less than the duration frame number threshold, it is determined to conform;

[0036] To establish tracking identifiers for the motion-connected regions that persist across multiple consecutive image frames, a target trajectory is formed;

[0037] The motion connectivity region, the target trajectory, and the pixel coordinates of the target trajectory in each image frame are used as the motion target of the corresponding sub-region.

[0038] As some embodiments of this application, the step of performing visual feature analysis on the moving target to obtain the fire situation determination result includes:

[0039] Extract at least one visual feature of the moving target and generate a corresponding feature vector;

[0040] Calculate the similarity between the feature vector and the preset smoke feature template vector;

[0041] When the similarity exceeds a preset matching threshold, a determination result indicating the presence of a fire is obtained.

[0042] As some embodiments of this application, the preset smoke feature template vector includes a seepage smoke feature template vector; the seepage smoke feature template vector is pre-trained or set based on seepage smoke data with a downward or horizontal diffusion trend, a reference morphological feature, and a non-periodic trajectory.

[0043] The step of extracting at least one visual feature of the moving target and generating a corresponding feature vector includes:

[0044] Calculate the motion vector of the moving target across consecutive image frames, and calculate the average motion direction vector based on the motion vector as a motion direction feature;

[0045] The shape irregularity, edge ambiguity, and diffusion velocity of the moving target are calculated as morphological features;

[0046] The center-of-gravity position of the moving target in consecutive image frames is recorded to form a trajectory sequence, and the trajectory sequence is subjected to time-frequency analysis to obtain non-periodic features;

[0047] The motion direction features, morphological features, and non-periodic features are used as the feature vectors corresponding to the moving target.

[0048] As some embodiments of this application, the step of dividing the monitoring area into sub-areas with different fire risk levels and acquiring image data and environmental sensor data of the sub-areas further includes:

[0049] The camera is geometrically calibrated to establish a positional mapping relationship between the camera's image pixel coordinate system and the warehouse's actual spatial coordinate system;

[0050] When the fire assessment result indicates the presence of a fire, the step of marking the location of the sub-area with the fire and issuing an alarm message includes:

[0051] When the fire situation determination result indicates that a fire exists, the location of the sub-region where the fire exists is marked;

[0052] Based on the location mapping relationship, the pixel coordinates of the moving target corresponding to the fire determination result in the image data are converted into coordinates in the actual space of the warehouse to mark the actual location of the fire source.

[0053] An alarm message is issued that includes the location of the sub-area where the fire is located and the actual location of the fire source.

[0054] Secondly, this application also provides a warehouse smoke and fire detection system based on image analysis for detecting fires in the monitored area of ​​a warehouse. The system includes:

[0055] The data acquisition module is used to divide the monitoring area into sub-areas with different fire risk levels, and acquire image data and environmental sensor data of the sub-areas.

[0056] An interference assessment module is used to assess the degree of interference in the current environment based on the high-frequency noise characteristics in the image data and the particulate matter concentration indicated by the environmental sensor data.

[0057] The parameter adjustment module is used to dynamically adjust the image analysis parameters applied to the corresponding sub-region according to the fire risk level and the degree of interference; wherein, the image analysis parameters include: the update rate of the background modeling algorithm and the sensitivity threshold of moving target detection;

[0058] The target capture module is used to apply the adjusted update rate of each sub-region to the background modeling algorithm to control the absorption speed of the background model for potential targets that conform to the preset physical characteristics of smoke diffusion; and to detect and capture moving targets in the corresponding sub-regions based on the sensitivity threshold adjusted for each sub-region; wherein the captured moving targets include the potential targets separated from the background.

[0059] The fire assessment module is used to perform visual feature analysis on the moving target and obtain the fire assessment result.

[0060] The alarm triggering module is used to mark the location of the sub-area where the fire exists and issue an alarm message when the fire determination result indicates that a fire exists.

[0061] The technical solution according to the embodiments of this application has at least the following beneficial effects:

[0062] This application overcomes the problems of missed detection caused by lens contamination, IR glare leading to image quality degradation, and aggressive background strategies designed to avoid false alarms from daily dust by dynamically adjusting the background update rate, ensuring that initial faint smoke is not incorrectly absorbed. By dynamically adjusting motion detection sensitivity based on the degree of interference, it enhances the ability to capture initial smoke signals in low-contrast, blurred images, reducing missed detections. Simultaneously, by combining risk area delineation and subsequent visual feature analysis, it can better distinguish between real smoke and interference such as blurred moving shadows caused by high-density goods, thereby reducing false alarms. Therefore, by introducing an adaptive parameter adjustment mechanism, this application can flexibly cope with multiple complex interferences in warehouses, such as lens contamination, lighting changes, and goods shadows, improving the performance bottleneck of traditional fixed parameters in terms of "false alarms, missed detections, and inaccurate positioning."

[0063] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0064] The accompanying drawings are used to provide a further understanding of the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application and do not constitute a limitation on the technical solutions of this application.

[0065] Figure 1 This is a flowchart illustrating a warehouse smoke and fire detection method based on image analysis, provided in an embodiment of this application.

[0066] Figure 2 This is a flowchart of S110.

[0067] Figure 3 This is a flowchart of S120.

[0068] Figure 4 This is a flowchart of S140.

[0069] Figure 5 This is a schematic diagram of the architecture of a warehouse smoke detection system based on image analysis, provided in an embodiment of this application. Detailed Implementation

[0070] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0071] like Figure 1 As shown in the figure, a specific embodiment of this application also discloses a warehouse smoke and fire detection method based on image analysis, used to detect fires in the monitored area of ​​a warehouse. The method includes the following steps:

[0072] S110, the monitoring area is divided into sub-areas with different fire risk levels, and image data and environmental sensor data of the sub-areas are acquired;

[0073] S120, assess the level of interference in the current environment based on the high-frequency noise characteristics in the image data and the particulate matter concentration indicated by the environmental sensor data;

[0074] S130, dynamically adjust the image analysis parameters applied to the corresponding sub-region according to the fire risk level and interference level; wherein, the image analysis parameters include: the update rate of the background modeling algorithm and the sensitivity threshold for moving target detection;

[0075] S140, the adjusted update rate of each sub-region is applied to the background modeling algorithm to control the absorption speed of the background model for potential targets that conform to the preset physical characteristics of smoke diffusion; and based on the sensitivity threshold adjusted for each sub-region, moving targets are detected and captured in the corresponding sub-region; wherein, the captured moving targets include the potential targets separated from the background;

[0076] S150, Perform visual feature analysis on the moving target to obtain the fire situation determination result;

[0077] S160, when the fire situation determination result indicates that a fire exists, the location of the sub-area where the fire exists is marked and an alarm message is issued.

[0078] The monitored area is divided into sub-zones with different fire risk levels. This can be done based on the warehouse's physical layout, the types of goods stored (e.g., flammable materials, general goods), and historical fire data. Pre-defined rules can be used to divide the entire monitored area into several sub-zones, each assigned a fixed fire risk level. For example, an area storing flammable chemicals might be designated as high-risk, while an area storing general office supplies might be designated as low-risk. Image data can be acquired using general-purpose network cameras deployed throughout the warehouse. These cameras can periodically capture still images or record fixed-length video clips. Environmental sensor data can be obtained through general-purpose air quality sensors installed in each sub-zone. These sensors periodically measure the particulate matter content in the air and upload the data to a central processing unit.

[0079] The high-frequency noise feature is used to quantify high-frequency random fluctuations in image data caused by environmental interference. Its acquisition method includes: performing high-pass or band-pass filtering on the image data to extract a preset high-frequency component; subsequently, calculating the statistical characteristics of this high-frequency component as a quantification index, such as subband variance, mean absolute deviation, information entropy, or residual energy. An increase in the value of the quantification index indicates an enhancement of high-frequency random disturbances caused by particulate scattering, sensor noise, and fine contamination, i.e., a greater degree of visual interference. Simultaneously, real-time particulate concentration values ​​are read from a laser scattering particulate sensor. The evaluation logic is as follows: the quantification index is compared with a preset index threshold, and the particulate concentration value is compared with a preset concentration threshold. If the quantification index is lower than the preset index threshold, and / or the particulate concentration value is higher than the preset concentration threshold, the current environmental interference level is determined to be high interference; otherwise, it is determined to be low interference.

[0080] Based on this, the step of dynamically adjusting the image analysis parameters applied to the corresponding sub-region according to the fire risk level and interference level can be performed as follows: A parameter adjustment table can be preset, which contains fixed values ​​for the background modeling algorithm update rate and moving target detection sensitivity threshold under different combinations of fire risk levels and interference levels. Once the fire risk level and current interference level of a sub-region are identified, the corresponding parameter value is directly retrieved from the parameter adjustment table and applied. It should be noted that the meaning of "dynamic adjustment" above is that the values ​​of the image analysis parameters (update rate, sensitivity threshold) are not fixed, but can be changed in real time according to the changes in the two input conditions of the real-time acquired fire risk level and interference level, thereby adapting to the environmental conditions of different sub-regions and at different times. The aforementioned method of retrieving and applying corresponding fixed parameter values ​​through the parameter adjustment table is a typical and efficient implementation method for achieving dynamic adjustment.

[0081] In practical applications, background modeling algorithms can employ methods such as inter-frame differencing or Gaussian mixture models. The update rate determines how quickly the background model adapts to environmental changes. Adjusting the update rate alters the absorption rate of newly emerging pixel changes that match smoke diffusion characteristics. For example, a lower update rate allows the background model to retain a longer "memory" of slowly changing potential smoke targets, making them less likely to be misclassified as background. Simultaneously, the sensitivity threshold for moving target detection can be simply set as the minimum pixel change amplitude; any area with a pixel change amplitude exceeding this threshold is initially identified as a moving target. By adjusting this threshold, the system's sensitivity to subtle movements can be controlled.

[0082] Then, visual features such as color, texture, and shape can be extracted from the captured moving target. For example, the average color value and edge complexity of the moving target can be calculated. These extracted features are then compared with preset smoke feature samples, such as by calculating the Euclidean distance. If the calculated distance is less than a preset matching threshold, the moving target is determined to be smoke, thus indicating the presence of a fire.

[0083] Finally, the step of marking the location of the sub-area with the fire and issuing an alarm when the fire assessment result indicates the existence of a fire can be implemented as follows: Once the system determines that a fire exists in a certain sub-area, it will immediately highlight the sub-area's number or name on the monitoring interface, and may trigger an audible and visual alarm. The alarm message can simply include the identifier of the sub-area where the fire occurred and be sent to the pre-set fire safety supervisor via SMS or email.

[0084] The method provided in this application firstly divides the warehouse monitoring area into sub-areas with different fire risk levels based on cargo attributes and stacking density, enabling focused monitoring. Secondly, it analyzes the high-frequency noise characteristics of image data and combines them with particulate matter sensor data to quantitatively assess the degree of environmental interference in real time. Based on this, the most crucial step is to dynamically and differentially adjust two core image processing parameters according to the risk level and real-time interference level of each sub-area: the update rate of the background modeling algorithm and the sensitivity threshold for moving target detection. In high-risk or high-interference areas, the system employs a lower background update rate (to protect weak, slowly changing initial smoke from being absorbed too quickly by the background model) and a lower motion detection sensitivity threshold (to improve the ability to capture weak motion signals). Subsequently, based on the adjusted parameters, moving targets are captured, and visual feature analysis is performed to determine the fire situation, ultimately achieving alarm and location.

[0085] In summary, this application overcomes the problems of missed detection caused by lens contamination, IR glare leading to image quality degradation, and aggressive background strategies designed to avoid false alarms from daily dust by dynamically adjusting the background update rate, ensuring that initial faint smoke is not incorrectly absorbed. By dynamically adjusting motion detection sensitivity based on the degree of interference, the ability to capture initial smoke signals in low-contrast, blurred images is enhanced, reducing missed detections. Simultaneously, by combining risk area delineation and subsequent visual feature analysis, it can better distinguish between real smoke and interference such as blurred moving shadows caused by high-density goods, thereby reducing false alarms. Therefore, by introducing an adaptive parameter adjustment mechanism, this application can flexibly cope with multiple complex interferences in warehouses, such as lens contamination, lighting changes, and goods shadows, improving the performance bottleneck of traditional fixed parameters in terms of "false alarms, missed detections, and inaccurate positioning."

[0086] It should be noted that, as Figure 2 As shown, the step of dividing the monitoring area into sub-areas with different fire risk levels and acquiring image data and environmental sensor data of the sub-areas preferably includes:

[0087] S111, call the data interface of the warehouse management system to read the cargo attribute information associated with a specific storage location;

[0088] S112, Read the shelf height sensor data associated with a specific storage location or the stacking status information entered manually, as the shelf stacking density information;

[0089] S113, Based on the cargo attribute information and the stacking density information, the monitoring area is divided into sub-areas with different fire risk levels;

[0090] S114, Based on the cameras deployed in the sub-region, acquire the real-time video stream of the sub-region as image data;

[0091] S115, Based on the laser scattering particulate matter sensor deployed in the sub-region, acquire the concentration data of particulate matter in the air of the sub-region as environmental sensor data.

[0092] The system calls the data interface of the warehouse management system to obtain cargo attribute information directly associated with each specific storage location in the warehouse. This cargo attribute information may include, but is not limited to, the cargo's flammability rating, storage quantity, value, and chemical properties. This information forms the basis for assessing the fire risk of that storage location.

[0093] The stacking density information reflects the density of goods on the shelf. Generally speaking, the higher the stacking density, the greater the speed and intensity of fire spread, and therefore it is an important factor to consider when assessing fire risk.

[0094] Based on the acquired cargo attribute and stacking density information, the monitored area can be finely divided into sub-areas with different fire risk levels. For example, an area with high-density stacking of flammable materials may be classified as a high-risk sub-area, while an area with sparsely stored low-flammable materials may be classified as a low-risk sub-area. This classification allows subsequent detection strategies to be adjusted accordingly.

[0095] The image data is the primary input for visual feature analysis of fireworks. The particulate matter concentration data reflects the presence of interfering substances such as smoke or dust in the current environment, providing an objective basis for subsequent assessment of the degree of environmental interference.

[0096] This application's solution, by acquiring detailed information on the attributes and stacking density of goods within the warehouse, enables a quantitative assessment of the fire risk at different storage locations, thereby achieving a refined division of the monitoring area. This division method makes the setting of fire risk levels more targeted and accurate. Simultaneously, the deployment of cameras and laser-scattered particulate sensors ensures real-time and accurate acquisition of image data and environmental sensor data, providing a reliable data foundation for subsequent dynamic adjustment of image analysis parameters and fire detection.

[0097] In some embodiments of this application, the step of assessing the level of interference in the current environment based on high-frequency noise characteristics in image data and particulate matter concentration indicated by environmental sensor data preferably includes:

[0098] S121, Calculate the average energy of high-frequency noise in the image data within a preset frequency band;

[0099] S122, acquire the real-time particulate matter concentration value indicated by environmental sensor data;

[0100] S123, if the average energy is lower than a preset energy threshold and the real-time particulate matter concentration is lower than a preset concentration threshold, then the interference level of the current environment is determined to be low interference level; otherwise, the interference level of the current environment is determined to be high interference level.

[0101] Image data typically contains various frequency components, with high-frequency components often related to image details, texture, and noise. In warehouse environments, high-frequency noise can be caused by factors such as the characteristics of the camera sensor itself, changes in lighting, and small particulate matter in the air (such as dust). By calculating the average energy of high-frequency noise within a specific preset frequency band, the intensity of random or unstructured interference present in the image can be quantified. This preset frequency band can be empirically set based on the actual application scenario and noise characteristics, or determined by analyzing historical data, aiming to effectively capture high-frequency noise related to environmental interference.

[0102] Environmental sensors, such as laser scattering particulate matter sensors, can monitor the concentration of particulate matter in warehouse air in real time. Particulate matter concentration is an important indicator for measuring air quality and potential disturbances (such as dust and smoke precursors). Real-time particulate matter concentration values ​​provide direct physical evidence for assessing environmental disturbances.

[0103] If the average energy is lower than a preset energy threshold and the real-time particulate matter concentration is lower than a preset concentration threshold, this means that the high-frequency noise level in the image is low and the particulate matter content in the air is within an acceptable range, indicating a relatively clear and stable environment. Otherwise, if any of the above conditions are not met, it indicates that the image quality may be significantly affected, or that there is a high concentration of particulate matter in the air, which may affect the accuracy of smoke detection. In this case, the interference level of the current environment is determined to be high interference.

[0104] The above technical solution enables a refined assessment of the degree of interference in the warehouse environment. This dual-judgment mechanism, which simultaneously considers high-frequency noise in the image and the concentration of particulate matter in the environment, allows the system to more accurately distinguish between normal environmental fluctuations and potential interference factors when facing complex and ever-changing warehouse environments. This provides a more reliable basis for the dynamic adjustment of subsequent image analysis parameters, avoiding false alarms or missed alarms caused by inaccurate environmental interference assessments, thereby improving the accuracy and reliability of the entire smoke and fire detection system.

[0105] In a further embodiment of this application, such as Figure 3 The step of assessing the level of interference in the current environment based on the high-frequency noise features in the image data and the particulate matter concentration indicated by the environmental sensor data preferably further includes:

[0106] S124, convert the image data to a preset color space, and calculate the overall average saturation and average brightness of the low-brightness areas of the image data; wherein, the low-brightness areas are the regions composed of pixels in the preset color space whose brightness components are lower than a preset brightness threshold;

[0107] S125, within a preset time period, when the average brightness of the low-brightness area shows an upward trend and the overall average saturation of the image shows a downward trend, and the real-time particulate matter concentration value is continuously lower than the preset cleaning threshold, the humidity data of the corresponding sub-area within the preset historical time period is obtained.

[0108] S126, when the fluctuation range of the humidity data during the historical time period exceeds the preset fluctuation threshold, the current environmental interference level is determined to be a latent environmental interference level.

[0109] Converting image data to a preset color space, such as HSV, HSL, or YUV, aims to separate the image's luminance and color information, facilitating independent analysis of specific visual attributes. The overall average saturation of the image refers to the average saturation component of all pixels in the converted color space, characterizing the vividness or purity of the colors. The average luminance of low-brightness areas refers to the average luminance value of the area comprised of pixels whose luminance components are below a preset luminance threshold in the converted color space. This area typically corresponds to the darker parts of the image and is more sensitive to subtle changes in the environment. The preset luminance threshold can be set according to the actual application scenario and image characteristics; for example, it can be set to 20% or 30% of the maximum luminance component value.

[0110] Within a preset time period, when the average brightness of low-brightness areas shows an upward trend while the overall average saturation of the image shows a downward trend, this typically indicates the presence of some diffuse medium in the environment, such as slight fog or water vapor, causing the overall image to become blurry and color saturation to decrease, while the brightness of low-brightness areas is relatively increased due to scattering effects. Simultaneously, if the real-time particulate matter concentration value remains below a preset cleaning threshold, interference caused by conventional smoke or particulate matter is ruled out, thus focusing attention on latent interference of non-particulate nature. Under this condition, the system will acquire humidity data for the corresponding sub-region within a preset historical time period to further verify whether latent environmental interference exists. The preset time period and historical time period can be configured according to actual needs; for example, the preset time period can be several minutes, and the historical time period can be the past hour.

[0111] When the fluctuation range of humidity data exceeds a preset fluctuation threshold within a historical time period, the current environmental interference level is determined to be latent environmental interference. The fluctuation range of humidity data can reflect the stability of ambient humidity. Large fluctuations may indicate changes in the water vapor content in the air, such as condensation or evaporation. Combined with the aforementioned image feature changes (increased average brightness in low-brightness areas, decreased overall average saturation), latent environmental interference caused by humidity changes can be identified more accurately. The fluctuation threshold can be empirically set based on the normal humidity variation range of the warehouse environment.

[0112] This application's solution, by introducing a comprehensive analysis of image color space, saturation, brightness trends in low-brightness areas, and humidity data fluctuations, can identify hidden environmental interferences that are difficult to detect using traditional methods. Specifically, when slight fog or moisture in the environment causes a decrease in overall image saturation and an increase in brightness in low-brightness areas, and the concentration of common particulate matter is not high, the system will further examine the fluctuations in humidity data. If the humidity fluctuations are drastic, it indicates the presence of hidden interference that is not smoke but may affect visual detection, thus avoiding misjudging such interference as a fire.

[0113] The following is a specific example to illustrate this.

[0114] Suppose that over a period of time, the average brightness of low-brightness areas in a certain sub-area of ​​a warehouse gradually increases in the monitoring images, while the overall average saturation of the image shows a decreasing trend. Simultaneously, data from a laser scattering particulate sensor deployed in this sub-area shows that the real-time particulate matter concentration remains below a preset cleaning threshold. In this situation, relying solely on high-frequency noise and particulate matter concentration for judgment might fail to identify potential interference. However, according to the solution in this application, the system further acquires the humidity data of this sub-area over the past hour. If the analysis reveals that the fluctuation range of this humidity data exceeds a preset fluctuation threshold (e.g., a humidity change exceeding 5% RH), the system determines that there is a latent environmental interference level. This latent interference might be due to slight fog or water vapor condensation caused by localized humidity changes within the warehouse. These phenomena may visually resemble early smoke, but are not a true fire. By identifying this latent interference, the system can avoid misjudging it as a fire, thereby preventing unnecessary alarm triggering and improving the accuracy and reliability of detection.

[0115] In a further embodiment of this application, the step of dynamically adjusting the image analysis parameters applied to the corresponding sub-region based on the fire risk level and the degree of interference includes:

[0116] Based on the fire risk level, a background modeling algorithm update rate is set for the corresponding sub-area, and the update rate is negatively correlated with the fire risk level.

[0117] Based on the level of interference, a sensitivity threshold for moving target detection is set for the corresponding sub-region, and the sensitivity threshold is negatively correlated with the level of interference.

[0118] When the interference level is the latent environmental interference level, the update rate of the corresponding sub-region is adjusted to a lower preset value than that of the high interference level, and the sensitivity threshold of the corresponding sub-region is adjusted to a lower preset value than that of the high interference level.

[0119] The update rate of a background modeling algorithm refers to the speed at which the background model learns and adapts to changes in the environment. When the fire risk level is high, it means that the sub-area is more likely to experience a fire. In this case, it is necessary to ensure that even weak smoke signals are captured promptly. Therefore, the background model should be updated at a slower rate to avoid absorbing potential smoke targets into the background too quickly. Thus, the update rate is negatively correlated with the fire risk level; that is, the higher the fire risk level, the lower the update rate.

[0120] The sensitivity threshold for moving target detection controls how well the system responds to changes in pixels within an image. When environmental interference is high, such as the presence of significant non-smoke-induced motion or lighting changes, sensitivity needs to be reduced to minimize false negatives. Therefore, the sensitivity threshold is negatively correlated with the level of interference; that is, the higher the level of interference, the lower the sensitivity threshold.

[0121] When the interference level is classified as latent environmental interference, it indicates the presence of subtle but potentially impactful environmental factors, such as humidity fluctuations. In this case, a more conservative strategy is needed to further improve detection accuracy and anti-interference capabilities. Specifically, the update rate of the corresponding sub-region is adjusted to a lower preset value than that used at high interference levels. This means the background model will adapt to environmental changes more slowly, thus preserving potential smoke features to the maximum extent. Simultaneously, the sensitivity threshold of the corresponding sub-region is adjusted to a lower preset value than that used at high interference levels, making the moving target detection more cautious in responding to minute changes.

[0122] Through the above technical solution, this application can achieve more intelligent and refined dynamic adjustment of image analysis parameters, especially in the presence of hidden environmental interference, which can significantly improve the accuracy and robustness of smoke and fire detection. This solution effectively avoids false alarms or missed alarms caused by complex types of environmental interference, enabling the system to maintain efficient and stable operation in various complex warehouse environments, thereby improving the reliability of warehouse fire early warning.

[0123] The following is a specific example to illustrate this.

[0124] Suppose a sub-area of ​​a warehouse is classified as having a high fire risk level, for example, if the area stores flammable chemicals. Furthermore, through analysis of image data and environmental sensor data, this sub-area is further classified as having a level of latent environmental disturbance. For instance, although high-frequency noise and particulate matter concentrations are at low levels, humidity data fluctuates beyond a preset fluctuation threshold over a historical period, indicating the presence of subtle airflow or temperature changes.

[0125] In this context, the detection method of this application first sets a low update rate for the background modeling algorithm of the sub-region based on a high fire risk level. This ensures that even slowly spreading smoke can be effectively captured without being absorbed too quickly by the background model. Next, due to the presence of latent environmental interference, the system further adjusts the sensitivity threshold for moving target detection to a preset value lower than that used for high interference levels. This dual adjustment strategy makes the system more cautious in detecting moving targets, effectively filtering out minute, smoke-free pixel changes caused by latent environmental interference, thus avoiding false alarms. Simultaneously, the extremely low update rate ensures that even under latent interference, the background model absorbs real smoke very slowly, thereby improving the ability to identify potential smoke targets. Therefore, even in sub-regions with complex environmental conditions and concealed interference, the scheme of this application can ensure the accuracy and reliability of smoke detection.

[0126] In embodiments of this application, the step of detecting and capturing moving targets in the corresponding sub-regions based on the sensitivity thresholds adjusted for each sub-region preferably includes:

[0127] Within each sub-region, the pixel change regions in the image data that conform to the adjusted sensitivity threshold for capturing potential targets that meet the preset physical characteristics of smoke diffusion are identified, and spatially connected pixel change regions are merged into a motion connected region; wherein, determining whether a pixel change region conforms to the adjusted sensitivity threshold includes: setting a pixel area threshold and a duration frame number threshold for the sub-region based on the adjusted sensitivity threshold; when the pixel area of ​​the pixel change region is not less than the pixel area threshold, and / or the number of times the pixel change region appears continuously in consecutive image frames is not less than the duration frame number threshold, it is determined to conform;

[0128] To establish tracking labels for motion-connected components that persist across multiple consecutive image frames, thereby forming the target trajectory;

[0129] The motion connectivity, target trajectory, and pixel coordinates of the target trajectory in each image frame are used as the motion target of the corresponding sub-region.

[0130] When analyzing image data from each sub-region, the first step is to identify pixels or pixel regions that differ from the background, as these differences may be caused by potential smoke targets. To avoid false alarms and improve detection accuracy, these pixel change regions need to be screened. The screening is based on a pre-adjusted sensitivity threshold for the sub-region. This sensitivity threshold is not a single value but can include multiple criteria, such as a pixel area threshold and a duration frame count threshold. When the pixel area of ​​a pixel change region is not less than the preset pixel area threshold, it indicates that the change region is large enough and may be a real moving target. Simultaneously, when the number of times the pixel change region appears continuously in consecutive image frames is not less than the preset duration frame count threshold, its stability is further confirmed, eliminating transient noise or interference. Pixel change regions that meet these conditions are considered valid regions that conform to the adjusted sensitivity threshold.

[0131] To integrate these discrete pixel variation regions into a meaningful moving target, spatially connected pixel variation regions are merged to form a complete motion connected region. This motion connected region represents an independent moving entity in the image with a certain shape and size.

[0132] Building upon this, to track these moving entities and analyze their dynamic behavior, a unique tracking identifier is assigned to each motion-connected component that persists across multiple consecutive image frames. This tracking identifier allows association of the same motion-connected component across different frames, thus forming a continuous target trajectory. This trajectory records the temporal positional changes of the motion-connected component.

[0133] Ultimately, these identified, merged, and tracked motion-connected components, the target trajectories they form, and the specific pixel coordinates of these target trajectories in each image frame are collectively considered as the motion targets of this sub-region. This information provides a comprehensive data foundation for subsequent visual feature analysis.

[0134] The proposed solution, by dynamically adjusting the sensitivity threshold and combining multiple judgment criteria (such as pixel area and continuous frame count), enables the system to more effectively suppress environmental noise and non-fire interference, reducing false alarms. Simultaneously, by merging pixel variation regions into motion connected domains and establishing target trajectories, the system can continuously and stably track potential smoke targets, providing more reliable and comprehensive motion feature data for subsequent fire assessment, thereby improving the overall reliability and real-time performance of the detection method.

[0135] It should be noted that, as Figure 4 As shown, the preferred steps for performing visual feature analysis on the moving target to obtain the fire situation determination result include:

[0136] S141, extract at least one visual feature of the moving target and generate a corresponding feature vector;

[0137] S142, Calculate the similarity between the feature vector and the preset smoke feature template vector;

[0138] S143, when the similarity exceeds the preset matching threshold, a determination result indicating the presence of a fire is obtained.

[0139] The visual features may include, but are not limited to, the shape features of the moving target (such as irregularity and aspect ratio), texture features (such as gray-level co-occurrence matrix and local binary pattern), color features (such as saturation and brightness distribution), motion features (such as diffusion speed and direction of motion), and temporal evolution features. The feature vector is a numerical representation of these visual features after quantization and encoding, and its generation can be achieved through various feature extraction algorithms, such as statistical methods, transform domain methods, or deep learning methods.

[0140] The smoke feature template vector is obtained in advance through analysis, training, or manual setting of a large number of real smoke samples, and is used to characterize the visual characteristics of typical smoke. Similarity calculation can employ various metrics, such as cosine similarity, Euclidean distance, Mahalanobis distance, or correlation coefficient, to quantify the degree of matching between the moving target features and the smoke feature template.

[0141] The matching threshold is a critical value pre-set through experiments, statistical analysis, or expert experience based on the actual application scenario and the requirements for detection accuracy, used to distinguish between smoke and non-smoke targets.

[0142] The proposed solution extracts the specific visual features of moving targets and performs similarity matching with a preset smoke feature template, enabling more precise differentiation between smoke and other similar moving targets, thereby effectively reducing the false alarm rate. Furthermore, by setting a matching threshold, the detection sensitivity can be adjusted according to actual needs, further optimizing system performance and ensuring timely and accurate detection of fires in complex warehouse environments.

[0143] It is worth mentioning that the preset smoke feature template vector preferably includes a seepage smoke feature template vector; the seepage smoke feature template vector is pre-trained or set based on seepage smoke data with a downward or horizontal diffusion trend, a reference morphological feature, and a non-periodic trajectory.

[0144] The step of extracting at least one visual feature of the moving target and generating a corresponding feature vector preferably includes:

[0145] The shape irregularity, edge ambiguity, and diffusion velocity of the moving target are calculated as morphological features;

[0146] The center-of-gravity position of the moving target in consecutive image frames is recorded to form a trajectory sequence, and the trajectory sequence is subjected to time-frequency analysis to obtain non-periodic features;

[0147] The motion direction features, morphological features, and non-periodic features are used as the feature vectors corresponding to the moving target.

[0148] The seepage smoke feature template vector refers to the feature model established for seepage smoke. Seepage smoke typically exhibits diffusion characteristics different from traditional smoke; for example, its movement direction may show a downward or horizontal trend, rather than the typical upward diffusion. To accurately identify this type of smoke, the template vector is designed to capture these unique physical characteristics. Specifically, the template vector is constructed based on a large amount of seepage smoke data, pre-trained or set through machine learning or expert experience, and its core lies in its ability to characterize the unique patterns of seepage smoke in terms of movement direction, morphology, and trajectory.

[0149] The extraction of motion direction features aims to quantify the movement trend of a moving target in space. Specifically, by analyzing the positional changes of a moving target between consecutive image frames, its instantaneous motion vector can be calculated. Subsequently, these instantaneous motion vectors are averaged to obtain an average motion direction vector representing the overall movement trend of the moving target. This feature is crucial for distinguishing different types of smoke (e.g., upward-spreading fire smoke versus downward or horizontally spreading seepage smoke).

[0150] Morphological features are used to describe the geometry and dynamic changes of moving targets. Shape irregularity can reflect the irregular boundaries of a smoke plume; edge ambiguity can indicate the transition area between smoke and background, and smoke edges are usually relatively blurred; diffusion rate quantifies the rate at which the smoke plume expands or contracts over time. These features together depict the physical morphology of smoke and its evolution, helping to identify the unique visual characteristics of smoke.

[0151] Extracting aperiodic features aims to capture the randomness and irregularity of smoke trajectories. Specifically, a trajectory sequence can be constructed by tracking the centroid position of a moving target in consecutive image frames. Subsequently, time-frequency analysis, such as through Fourier transform or wavelet transform, is performed on this trajectory sequence to identify the presence of periodic patterns. Since smoke diffusion is typically aperiodic, the lack of obvious periodicity is a crucial distinguishing feature.

[0152] The feature vector comprehensively describes the dynamic and static visual characteristics of the moving target, providing rich and discriminative information for subsequent similarity calculation with the feature template vector of the seepage smoke.

[0153] Through the above technical solution, this application can significantly improve the accuracy and reliability of detecting special smoke types, especially seepage smoke, in warehouse environments. Compared with methods that rely solely on general smoke feature templates, this solution introduces seepage smoke feature template vectors and combines them with the extraction of refined visual features such as motion direction, morphology, and aperiodicity. This enables the system to more effectively capture the unique physical characteristics and visual appearance of seepage smoke. Therefore, it can effectively reduce the false negative rate for this type of smoke, ensuring that fires can be detected promptly and accurately, thus providing a more robust guarantee for warehouse safety management.

[0154] The following is a concrete example. Suppose that in a sub-area of ​​a warehouse, a surveillance camera captures a moving target resembling smoke. The system first calculates the motion vector of this target across consecutive image frames and derives its average motion direction vector. For example, if the average motion direction vector shows a clear downward or horizontal diffusion trend, it is initially judged to be likely seepage smoke. Simultaneously, the system analyzes the irregularity of the target's shape, the blurriness of its edges, and its diffusion speed. For instance, if its shape is irregular and its edges are blurry, and its diffusion speed is moderate, it further supports the judgment that it is smoke. Furthermore, the system records the centroid position of the target in consecutive image frames, forming a trajectory sequence, and performs time-frequency analysis on it. If the analysis results show that the trajectory sequence lacks obvious periodicity, it further strengthens the judgment that it is smoke. Finally, these extracted motion direction features, morphological features, and non-periodic features are combined into a feature vector, and its similarity is calculated with a pre-trained or preset seepage smoke feature template vector. If the similarity exceeds a preset matching threshold, the system determines that a fire exists and triggers the corresponding alarm mechanism. In this way, even seepage smoke with special diffusion patterns can be accurately identified and detected by the system.

[0155] In embodiments of this application, the step of dividing the monitoring area into sub-areas with different fire risk levels and acquiring image data and environmental sensor data of the sub-areas preferably further includes:

[0156] Geometric calibration of the camera is performed to establish a positional mapping relationship between the camera's image pixel coordinate system and the warehouse's actual spatial coordinate system.

[0157] When the fire assessment indicates the presence of a fire, the preferred steps for identifying the location of the affected sub-area and issuing an alarm message include:

[0158] When the fire situation assessment result indicates that a fire exists, the location of the sub-area where the fire exists is marked;

[0159] Based on the location mapping relationship, the pixel coordinates of the moving target corresponding to the fire situation determination result in the image data are converted into coordinates in the actual space of the warehouse to pinpoint the actual location of the fire source.

[0160] It issues an alarm message that includes the location of the sub-area where the fire is located and the actual location of the fire source.

[0161] Camera geometric calibration refers to determining the camera's internal parameters (such as focal length, principal point, and distortion coefficient) and external parameters (such as the camera's position and orientation within the warehouse space) through a series of pre-defined calibration procedures and tools. For example, a checkerboard or other known geometrically shaped calibration board can be used to capture multiple images from different positions and angles, and then computer vision algorithms (such as the Zhang Zhengyou calibration method) can be used to calculate these parameters. These parameters establish a precise correspondence between the pixels in the images captured by the camera and the physical points in the actual three-dimensional space of the warehouse—a positional mapping relationship. The purpose is to provide fundamental data for subsequent precise location calibration of fire sources.

[0162] The moving target corresponding to the fire assessment result refers to the visual object identified by the system as smoke or flame. In image data, this moving target occupies a certain pixel area and has corresponding pixel coordinates. By utilizing the previously established position mapping relationship, these pixel coordinates can be accurately converted into three-dimensional coordinates in the actual warehouse space (e.g., shelf number, shelf height, specific storage location coordinates, etc.). Thus, the actual location of the fire source can be obtained. In practical applications, the alarm information not only includes the name or number of the sub-area where the fire is located, but also further includes the actual location information of the fire source obtained through coordinate transformation, such as "5th floor of shelf 3 in area A, coordinates X:10m, Y:5m, Z:3m". The purpose is to provide emergency responders with accurate fire source location so that they can take rapid action.

[0163] This application's solution introduces camera geometric calibration, enabling a precise correspondence between visual information in image data and the actual physical space of the warehouse. When the system detects a potential moving fire target, its pixel position in the image can be accurately converted into three-dimensional coordinates in the actual warehouse space through a pre-established position mapping relationship. It is precisely this ability to convert pixel coordinates to actual spatial coordinates that allows the system to move beyond simply identifying the sub-region where the fire is located to accurately pinpointing the actual location of the fire source. This precise positioning capability effectively compensates for the shortcomings of relying solely on sub-region division for fire location, providing crucial, high-precision information support for subsequent emergency response.

[0164] The following is a specific example to illustrate this.

[0165] Suppose a warehouse's monitoring area is divided into several sub-areas, one of which is "Area A". Within Area A, a camera is deployed. This camera has been geometrically calibrated to establish a precise mapping between its image pixel coordinates and the warehouse's actual spatial coordinates (e.g., a three-dimensional Cartesian coordinate system with the warehouse entrance as the origin). When the system detects a moving target in Area A that exhibits smoke characteristics—for example, identifying a pixel area in an image frame showing smoke diffusion—the system first determines that a fire exists in Area A. Further, the system extracts the center pixel coordinates of this moving smoke target in the image, for example, (500, 300). Using the pre-established positional mapping, these pixel coordinates are converted to coordinates in the actual warehouse space, for example, (X=15.2m, Y=8.7m, Z=4.1m), which may correspond to a specific location on a shelf within Area A. Finally, the system will issue an alarm message. This message will not only indicate that "there is a fire in Zone A," but will also specify the "actual location of the fire: Zone A, shelf number: H3, shelf height: L4, specific coordinates: X=15.2m, Y=8.7m, Z=4.1m." Upon receiving this alarm, firefighters can quickly locate the fire source based on the precise coordinates and take appropriate firefighting measures.

[0166] like Figure 5 As shown, this application also discloses a warehouse smoke and fire detection system based on image analysis, used for fire detection in the monitored area of ​​a warehouse. The system includes:

[0167] The data acquisition module 210 is used to divide the monitoring area into sub-areas with different fire risk levels, and acquire image data and environmental sensor data of the sub-areas.

[0168] Interference assessment module 220 is used to assess the degree of interference in the current environment based on the high-frequency noise characteristics in the image data and the particulate matter concentration indicated by the environmental sensor data.

[0169] The parameter adjustment module 230 is used to dynamically adjust the image analysis parameters applied to the corresponding sub-region according to the fire risk level and the degree of interference; wherein, the image analysis parameters include: the update rate of the background modeling algorithm and the sensitivity threshold of moving target detection;

[0170] The target capture module 240 is used to apply the updated rate of each sub-region to the background modeling algorithm to control the absorption speed of the background model for potential targets that conform to the preset physical characteristics of smoke diffusion; and to detect and capture moving targets in the corresponding sub-regions based on the sensitivity thresholds adjusted for each sub-region; wherein the captured moving targets include the potential targets separated from the background.

[0171] The fire assessment module 250 is used to perform visual feature analysis on the moving target and obtain the fire assessment result.

[0172] The alarm triggering module 260 is used to mark the location of the sub-area where the fire exists and issue an alarm message when the fire determination result indicates that a fire exists.

[0173] The data acquisition module 210 may include a configuration unit for manually inputting or pre-setting the boundaries and corresponding fire risk levels of each sub-area. Simultaneously, the data acquisition module 210 may integrate a standard video capture interface to receive real-time video streams captured by cameras deployed in each sub-area, as image data. Furthermore, the data acquisition module 210 can also connect to and read data output from environmental sensors (e.g., sensors for measuring particulate matter concentration) deployed in each sub-area via a universal sensor interface, as environmental sensor data.

[0174] The interference assessment module 220 may include an image processing unit for performing frequency domain analysis on image data and calculating the high-frequency noise energy in a specific frequency band. Simultaneously, the interference assessment module 220 may include a data analysis unit for processing environmental sensor data and obtaining particulate matter concentration values. Based on these calculation results, the interference assessment module 220 can classify the interference level into different grades through preset logical rules or simple threshold comparisons.

[0175] The parameter adjustment module 230 may include a strategy engine that, based on the received fire risk level and interference level information, determines the update rate of the background modeling algorithm and the sensitivity threshold for moving target detection by looking up parameters in a preset parameter table or through a simple linear interpolation method. These adjusted parameters are then passed to the target acquisition module.

[0176] The target capture module 240 may include a background modeler that maintains and updates the image background model according to an adjusted update rate to adapt to environmental changes. Simultaneously, the target capture module 240 may include a motion detector that identifies pixel regions in the image that differ from the background based on an adjusted sensitivity threshold, capturing them as potential moving targets.

[0177] The fire detection module 250 may include a feature extractor for extracting visual features such as color, texture, shape, and trajectory from the image region of a moving target. Subsequently, the fire detection module 250 may include a classifier that compares the extracted features with a preset smoke or flame feature model, for example, by calculating similarity or matching degree, to determine whether a fire exists.

[0178] The alarm triggering module 260 may include a location calibration unit for calibrating the area where the fire occurred based on the sub-area information associated with the fire determination result. Simultaneously, the alarm triggering module 260 may include a communication interface for sending alarm information (e.g., including the location information of the fire sub-area) via a network to the warehouse management system or fire control center, and can also drive local audible and visual alarms to sound.

[0179] It should be noted that the information interaction and execution process between the above modules are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0180] The preferred embodiments of this application have been described in detail above, but this application is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application.

Claims

1. A warehouse smoke and fire detection method based on image analysis, used for fire detection in the monitored area of ​​a warehouse, characterized in that, The method includes the following steps: The monitoring area is divided into sub-areas with different fire risk levels, and image data and environmental sensor data of the sub-areas are acquired. The degree of interference in the current environment is assessed based on the high-frequency noise characteristics in the image data and the particulate matter concentration indicated by the environmental sensor data. Based on the fire risk level and interference level, the image analysis parameters applied to the corresponding sub-region are dynamically adjusted; wherein, the image analysis parameters include: the update rate of the background modeling algorithm, and the sensitivity threshold for moving target detection; The adjusted update rate for each sub-region is applied to the background modeling algorithm to control the absorption rate of the background model for potential targets that conform to the preset physical characteristics of smoke diffusion; and based on the sensitivity threshold adjusted for each sub-region, moving targets are detected and captured in the corresponding sub-region; wherein the captured moving targets include the potential targets separated from the background; Visual feature analysis is performed on the moving target to obtain the fire situation determination result; When the fire situation determination result indicates that a fire exists, the location of the sub-area where the fire exists is marked and an alarm message is issued.

2. The warehouse smoke and fire detection method based on image analysis according to claim 1, characterized in that, The steps of dividing the monitored area into sub-areas with different fire risk levels and acquiring image data and environmental sensor data of the sub-areas include: Call the data interface of the warehouse management system to read the cargo attribute information associated with a specific storage location; Read the shelf height sensor data associated with a specific storage location or the stacking status information entered manually, as the shelf stacking density information; Based on the cargo attribute information and the stacking density information, the monitoring area is divided into sub-areas with different fire risk levels; Based on the cameras deployed in the sub-region, real-time video streams of the sub-region are acquired as image data. Based on the laser scattering particulate matter sensor deployed in the sub-region, the concentration data of particulate matter in the air of the sub-region is acquired as environmental sensor data.

3. The warehouse smoke and fire detection method based on image analysis according to claim 1, characterized in that, The step of assessing the level of interference in the current environment based on the high-frequency noise features in the image data and the particulate matter concentration indicated by the environmental sensor data includes: Calculate the average energy of high-frequency noise in the image data within a preset frequency band; Obtain the real-time particulate matter concentration value indicated by the environmental sensor data; If the average energy is lower than a preset energy threshold and the real-time particulate matter concentration is lower than a preset concentration threshold, then the interference level of the current environment is determined to be low interference level. Otherwise, the interference level of the current environment is determined to be high interference level.

4. The warehouse smoke and fire detection method based on image analysis according to claim 3, characterized in that, The step of assessing the level of interference in the current environment based on the high-frequency noise features in the image data and the particulate matter concentration indicated by the environmental sensor data further includes: The image data is converted to a preset color space, and the overall average saturation and average brightness of the low-brightness areas of the image data are calculated; wherein, the low-brightness areas are the regions composed of pixels in the preset color space whose brightness components are lower than a preset brightness threshold. Within a preset time period, when the average brightness of the low-brightness area shows an upward trend and the overall average saturation of the image shows a downward trend, and the real-time particulate matter concentration value is continuously lower than the preset cleaning threshold, the humidity data of the corresponding sub-area within the preset historical time period is obtained. When the fluctuation range of the humidity data during the historical time period exceeds the preset fluctuation threshold, the current environmental interference level is determined to be latent environmental interference level.

5. The warehouse smoke and fire detection method based on image analysis according to claim 4, characterized in that, The step of dynamically adjusting the image analysis parameters applied to the corresponding sub-region based on the fire risk level and interference level includes: Based on the fire risk level, a background modeling algorithm update rate is set for the corresponding sub-area, and the update rate is negatively correlated with the fire risk level. Based on the level of interference, a sensitivity threshold for moving target detection is set for the corresponding sub-region, and the sensitivity threshold is negatively correlated with the level of interference. When the interference level is the latent environmental interference level, the update rate of the corresponding sub-region is adjusted to a lower preset value than that of the high interference level, and the sensitivity threshold of the corresponding sub-region is adjusted to a lower preset value than that of the high interference level.

6. The warehouse smoke and fire detection method based on image analysis according to claim 1, characterized in that, Based on the sensitivity thresholds adjusted for each sub-region, the step of detecting and capturing moving targets in the corresponding sub-region includes: Within each sub-region, pixel change regions in the image data that conform to the adjusted sensitivity threshold for capturing potential targets that meet the preset physical characteristics of smoke diffusion are identified, and spatially connected pixel change regions are merged into motion connected regions; wherein, determining whether a pixel change region conforms to the adjusted sensitivity threshold includes: setting a pixel area threshold and a duration frame number threshold for the sub-region based on the adjusted sensitivity threshold; when the pixel area of ​​the pixel change region is not less than the pixel area threshold, and / or the number of times the pixel change region appears continuously in consecutive image frames is not less than the duration frame number threshold, it is determined to conform; To establish tracking identifiers for the motion-connected regions that persist across multiple consecutive image frames, a target trajectory is formed; The motion connectivity region, the target trajectory, and the pixel coordinates of the target trajectory in each image frame are used as the motion target of the corresponding sub-region.

7. The warehouse smoke and fire detection method based on image analysis according to claim 1, characterized in that, The step of performing visual feature analysis on the moving target to obtain the fire situation determination result includes: Extract at least one visual feature of the moving target and generate a corresponding feature vector; Calculate the similarity between the feature vector and the preset smoke feature template vector; When the similarity exceeds a preset matching threshold, a determination result indicating the presence of a fire is obtained.

8. The warehouse smoke and fire detection method based on image analysis according to claim 7, characterized in that, The preset smoke feature template vector includes a seepage smoke feature template vector; the seepage smoke feature template vector is pre-trained or set based on seepage smoke data with a downward or horizontal diffusion trend, a reference morphological feature, and a non-periodic trajectory. The step of extracting at least one visual feature of the moving target and generating a corresponding feature vector includes: Calculate the motion vector of the moving target across consecutive image frames, and calculate the average motion direction vector based on the motion vector as a motion direction feature; The shape irregularity, edge ambiguity, and diffusion velocity of the moving target are calculated as morphological features; The center-of-gravity position of the moving target in consecutive image frames is recorded to form a trajectory sequence, and the trajectory sequence is subjected to time-frequency analysis to obtain non-periodic features; The motion direction features, morphological features, and non-periodic features are used as the feature vectors corresponding to the moving target.

9. The warehouse smoke and fire detection method based on image analysis according to claim 2, characterized in that, The step of dividing the monitoring area into sub-areas with different fire risk levels and acquiring image data and environmental sensor data of the sub-areas further includes: The camera is geometrically calibrated to establish a positional mapping relationship between the camera's image pixel coordinate system and the warehouse's actual spatial coordinate system; When the fire assessment result indicates the presence of a fire, the step of marking the location of the sub-area with the fire and issuing an alarm message includes: When the fire situation determination result indicates that a fire exists, the location of the sub-region where the fire exists is marked; Based on the location mapping relationship, the pixel coordinates of the moving target corresponding to the fire determination result in the image data are converted into coordinates in the actual space of the warehouse to mark the actual location of the fire source. An alarm message is issued that includes the location of the sub-area where the fire is located and the actual location of the fire source.

10. A warehouse smoke and fire detection system based on image analysis, used for fire detection in the monitored area of ​​a warehouse, characterized in that, The system includes: The data acquisition module is used to divide the monitoring area into sub-areas with different fire risk levels, and acquire image data and environmental sensor data of the sub-areas. An interference assessment module is used to assess the degree of interference in the current environment based on the high-frequency noise characteristics in the image data and the particulate matter concentration indicated by the environmental sensor data. The parameter adjustment module is used to dynamically adjust the image analysis parameters applied to the corresponding sub-region according to the fire risk level and the degree of interference; wherein, the image analysis parameters include: the update rate of the background modeling algorithm and the sensitivity threshold of moving target detection; The target capture module is used to apply the adjusted update rate of each sub-region to the background modeling algorithm to control the absorption speed of the background model for potential targets that conform to the preset physical characteristics of smoke diffusion; and to detect and capture moving targets in the corresponding sub-regions based on the sensitivity threshold adjusted for each sub-region; wherein the captured moving targets include the potential targets separated from the background. The fire assessment module is used to perform visual feature analysis on the moving target and obtain the fire assessment result. The alarm triggering module is used to mark the location of the sub-area where the fire exists and issue an alarm message when the fire determination result indicates that a fire exists.

Images