A full-cycle, full-direction, and full-time hazardous chemical storage environment anomaly detection and category analysis method

Abstract

The application discloses a kind of all-cycle all-around all-weather dangerous chemical storage environment anomaly detection and category analysis method. The application first deploys visible light and infrared camera in the key node of storage, synchronously collects double mode video stream, realizes space-time alignment by hardware synchronization, generates multi-modal fusion data;Second, a redundant ring network is established, and a heterogeneous network is constructed in combination with a wireless bridge. The double mode video stream data is compressed using encoding, and the transmission of the double mode video stream and the multi-modal fusion data from the acquisition end to the data center is realized. Then a distributed hierarchical storage architecture is constructed, and a hierarchical storage system is constructed. The RAID disk redundancy and off-site backup strategy form a life cycle management system. Finally, the multi-modal fusion data is processed and analyzed to identify gas leakage and dynamic obstacles. The application realizes all-cycle, all-around, all-weather monitoring of dangerous chemical storage environment, significantly improves detection accuracy and system reliability.

Description

Technical Field

[0001] This invention relates to the field of industrial safety monitoring technology, and in particular to a method for detecting and classifying anomalies in the hazardous chemical storage environment throughout the entire lifecycle, in all directions, and around the clock. Background Technology

[0002] The safety of hazardous chemical storage is of paramount importance in industrial safety management. Currently, this field mainly relies on traditional manual inspection methods, which suffer from problems such as low efficiency, strong subjectivity, high risk, and inability to achieve 24 / 7 monitoring.

[0003] With technological advancements, while some warehouses have begun using thermal imaging cameras for temperature monitoring, vision-based automated inspection technologies face several key bottlenecks: In dynamic obstacle removal and gas leak detection, traditional background modeling algorithms such as Gaussian Mixture Models (GMMs) are poorly adapted to drastic changes in lighting and continuous dynamic disturbances (such as frequent operations). Model updates are slow, and they are prone to misclassifying slowly spreading gas leaks as background and suddenly appearing dynamic targets (such as vehicles or personnel) as foreground, resulting in numerous false alarms, a high false alarm rate, and poor detection robustness. Regarding abnormal object recognition, clustering algorithms such as K-means require pre-setting the number of clusters (K value), making them unable to adaptively identify gas leak areas of unknown shape and quantity. They are also sensitive to noise points and prone to incorrect clustering in complex infrared scenes, leading to ineffective separation of gas from interfering objects. Vision-based automated inspection technologies still face numerous technical bottlenecks and challenges in practical applications. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by providing a method for full-cycle, all-round, and all-weather anomaly detection and category analysis in hazardous chemical storage environments. It aims to solve technical problems such as incomplete data collection, high false alarm rates in anomaly detection, poor system adaptability, and unreliable data transmission and storage in hazardous chemical storage environments. Specifically, it includes the following steps:

[0005] Visible light and infrared cameras are deployed at key nodes in the warehouse to simultaneously collect dual-modal video streams. Spatiotemporal alignment is achieved through hardware synchronization to generate multimodal fusion data.

[0006] A redundant ring network is constructed, and a heterogeneous network is built by combining it with a wireless bridge. The dual-modal video stream data is compressed using encoding to realize the transmission of the dual-modal video stream and the multimodal fused data from the acquisition end to the data center.

[0007] A distributed hierarchical storage architecture is constructed, storing the dual-modal video stream and multi-modal fused data in a time-series database, storing the structured analysis results in a relational database, and constructing a hierarchical storage system, implementing RAID disk redundancy and off-site backup strategies to form a lifecycle management system.

[0008] The multimodal fusion data is processed and analyzed to identify gas leaks and dynamic obstacles.

[0009] The beneficial effects of this invention are:

[0010] First, this invention uses the ViBe background modeling algorithm to replace the traditional GMM algorithm, which greatly improves the background initialization speed, is more robust to sudden changes in illumination and scene noise, effectively solves the problems of slow GMM model updates and susceptibility to dynamic interference, further reduces the false alarm rate of the gas detection process, and improves the real-time performance of the system.

[0011] Secondly, by using the DBSCAN density clustering algorithm instead of the traditional K-means algorithm, the system can automatically identify gas leak areas of arbitrary shapes without pre-setting the number of classifications and effectively filter out noise points generated during feature extraction. This fundamentally solves the problems of K-means relying on a preset K value and being sensitive to noise, thereby improving the accuracy of gas leak detection and its adaptability to complex scenarios.

[0012] Third, this invention fully utilizes the rich texture details of visible light images and the unique temperature sensing advantages of infrared images through a dual-modal collaborative acquisition and synchronization mechanism. It effectively overcomes the monitoring blind spots and performance degradation problems of a single sensor under low visibility or severe weather conditions such as nighttime darkness, rain, snow, fog, and haze, and achieves true all-day monitoring.

[0013] Fourth, this invention, through an industrial-grade reliable transmission and intelligent storage architecture, ensures the security and integrity of monitoring data during transmission, as well as the efficiency and reliability of storage, providing a solid data foundation for subsequent analysis. Attached Figure Description

[0014] Figure 1 An overall architecture diagram provided for embodiments of the present invention;

[0015] Figure 2 This is a flowchart of a data acquisition and transmission network provided in an embodiment of the present invention;

[0016] Figure 3 This is a schematic diagram of the data storage and management architecture provided in an embodiment of the present invention;

[0017] Figure 4 The flowchart of the dynamic obstacle removal and gas leak detection algorithm provided in the embodiment of the present invention is shown. Detailed Implementation

[0018] To make the technical solution, innovative features, and beneficial effects of this invention clearer, the following will, in conjunction with the accompanying drawings and specific embodiments, elaborate on the specific implementation of the method for detecting and classifying anomalies in the hazardous chemical storage environment throughout the entire lifecycle, in all directions, and around the clock provided by this invention. This section aims to explain in detail, through specific operating procedures, parameter configurations, and implementation details, how to apply the aforementioned technical solution to actual hazardous chemical storage monitoring scenarios, thereby providing clear, complete, and highly operable technical guidance for those skilled in the art.

[0019] Example:

[0020] like Figure 1 As shown, this embodiment provides a method for detecting and classifying anomalies in the hazardous chemical storage environment throughout the entire lifecycle, in all directions, and around the clock. It involves a multimodal data acquisition layer, a transmission layer, a data center, and a terminal layer. Specifically, it involves a system and method for simultaneous acquisition of dual-modal data of visible light and infrared thermal imaging, dynamic obstacle removal, gas leak detection, and industrial-grade data transmission and storage.

[0021] Optionally, regarding the explanation of "full cycle, all-round, and all-day": The concept of "full cycle, all-round, and all-day" proposed in this embodiment constitutes the core feature of the system's coverage without blind spots.

[0022] Full lifecycle: refers to the continuous tracking and intervention of a hazardous event throughout its entire lifecycle, from its inception, occurrence, development to its evolution. For example, the system can not only monitor stable gas leak points (occurrence), but also determine the scope and direction of leak spread through feature tracking (development), and provide continuous data support for emergency response until the hazard is eliminated (evolution), forming a complete monitoring closed loop;

[0023] All-round coverage: This refers to achieving comprehensive coverage without blind spots in three-dimensional space. The system deploys a network of visible light and infrared dual-modal cameras to monitor not only planar areas (such as the ground and warehouse floor) but also three-dimensional spaces (such as tanks, pipelines, and overhead lines). For example, one camera monitors for liquid leaks below the tank (bottom), another monitors for gas leaks from the valves on the top of the tank (top), and yet another monitors for intrusions at the perimeter (surroundings), thus achieving comprehensive three-dimensional monitoring of the ground, air, and perimeter.

[0024] 24 / 7 monitoring: This refers to achieving uninterrupted and effective monitoring 24 hours a day. The system utilizes visible light cameras to ensure clear monitoring during the day, while using infrared thermal imaging cameras to overcome the effects of poor lighting conditions such as nighttime, rain, fog, and smoke. For example, even in complete darkness, the system can still effectively detect overheated equipment bearings (high-temperature points) through thermal imaging or identify unauthorized intruders through the thermal radiation characteristics of the human body, truly achieving uninterrupted security protection.

[0025] Based on the above definitions and descriptions, this embodiment specifically includes the following steps:

[0026] S1: Visible light cameras and infrared thermal imaging cameras deployed at key nodes in the warehouse area simultaneously acquire dual-modal video stream data of the monitored scene. The visible light cameras provide a high-definition video stream with a resolution of 1920x1080 pixels and a frame rate of 25fps, accurately capturing texture details; the infrared thermal imaging cameras have a temperature resolution of 0.05℃, effectively detecting temperature anomalies. Hardware synchronization signals achieve spatiotemporal alignment of heterogeneous data, generating multimodal fusion data to realize full-cycle, all-round, and all-day data acquisition.

[0027] Optionally, the acquisition of multimodal fusion data in S1 includes:

[0028] Optionally, a visible light camera is activated to capture clear images, colors, and texture details of the monitored area. This camera features automatic exposure, automatic white balance, and digital noise reduction to ensure a clear video stream with a resolution of at least 1920×1080 pixels and a frame rate of at least 25fps under sunlight conditions. This accurately presents the image, color, and texture details of the monitored area, providing rich visual information for subsequent morphology recognition and target classification.

[0029] Optionally, an infrared thermal imaging camera can be activated simultaneously to generate a temperature distribution image by sensing the thermal radiation from the surface of an object. This thermal imager has a temperature resolution of no less than 0.05℃, which can effectively sense and display the temperature difference distribution in the monitoring scene, thereby accurately identifying abnormal temperature targets such as equipment overheating, gas leaks (small temperature differences between room temperature gas and the background environment), and smoldering flames in the early stages of a fire, making up for the shortcomings of visible light in sensing thermal information.

[0030] Optionally, information acquired by the two heterogeneous sensors can be aligned and fused spatiotemporally using hardware synchronization signals or precise timestamps. Leveraging the rich texture details of visible light images and the unique temperature sensing capabilities of infrared images, pixel-level or feature-level complementary fusion can be performed to generate more comprehensive dual-modal data, providing a more reliable and richer input source for upper-level intelligent analysis algorithms.

[0031] Optionally, this dual-modal collaborative acquisition system effectively overcomes the monitoring blind spots and performance degradation issues inherent in single sensors under low visibility or severe weather conditions such as nighttime darkness, rain, snow, fog, and haze. It achieves reliable, continuous, all-weather, and uninterrupted monitoring of hazardous chemical storage areas under any lighting and weather conditions, fundamentally improving the system's robustness to environmental changes.

[0032] S2: As Figure 2As shown, this embodiment uses a gigabit industrial Ethernet network to build a redundant ring network, ensuring uninterrupted multimodal fusion data transmission. A 5G / Wi-Fi 6 wireless bridge is introduced to build a heterogeneous network, and H.265 encoding is used to compress the dual-modal video stream data. AES-256 encryption ensures transmission security, and the MQTT protocol ensures communication reliability, achieving stable and secure transmission of dual-modal video stream data and multimodal fusion data from the acquisition end to the data center.

[0033] Optionally, to achieve reliable multimodal data transmission in S2, the following methods are included:

[0034] S21: A redundant ring network topology is constructed using gigabit industrial-grade Ethernet switches. When a single point of failure occurs in the network, the system can automatically heal within milliseconds, ensuring uninterrupted monitoring services. This network architecture design guarantees high reliability of data transmission, fully meeting the stringent requirements of 24 / 7 uninterrupted operation of the safety monitoring system.

[0035] S22: Introducing 5G or Wi-Fi 6 wireless bridges as an effective supplement to the wired network, constructing a heterogeneous network architecture that combines wired and wireless technologies. This hybrid networking approach significantly improves the deployment flexibility of the system in complex warehouse environments, and is particularly suitable for monitoring nodes with long-distance transmission or difficult cabling, ensuring comprehensive monitoring network coverage.

[0036] S23: The dual-modal video stream data is efficiently compressed using the advanced H.265 encoding standard, reducing bandwidth usage by approximately 50% while maintaining image quality, significantly reducing network transmission pressure. Simultaneously, all transmitted data packets are encrypted using the AES-256 algorithm to ensure that monitoring data is not stolen or tampered with during transmission.

[0037] S24: An end-to-end encrypted transmission channel is established using secure communication protocols such as MQTT. These protocols feature low bandwidth, low power consumption, and high reliability, making them ideal for industrial IoT environments. Digital certificates and two-way authentication mechanisms ensure the security and integrity of data throughout the entire transmission process from the acquisition point to the data center, providing a reliable data foundation for subsequent analysis.

[0038] S3: As Figure 3As shown, this embodiment constructs a distributed hierarchical storage architecture to store dual-modal video stream data and multi-modal fused data in a time-series database, and stores structured analysis results (dynamic obstacle coordinates or gas leak area) in a relational database. Hot data (the latest dual-modal video stream data and multi-modal fused data) is stored on SSDs to ensure read / write performance, while cold data (historical dual-modal video stream data) is automatically migrated to HDDs for long-term archiving. RAID disk redundancy and off-site backup strategies are implemented to ensure data security and recoverability, forming a complete data lifecycle management system.

[0039] Optionally, the multimodal data storage and management in S3 can be implemented through the following methods:

[0040] Optionally, a distributed storage architecture is employed to store the dual-modal video stream data, multi-modal fusion data, and structured analysis results generated by the system. The dual-modal video stream data and multi-modal fusion data are stored in a time-series database, while the structured analysis results (dynamic obstacle coordinates or gas leak area) are stored in a relational database. The time-series database is optimized for time-series data, greatly improving the read, write, and query efficiency of sensor data that changes over time; the relational database effectively manages structured alarm events and analysis results data, achieving the optimal storage solution for different types of data.

[0041] Optionally, a tiered storage system can be constructed, using high-speed SSDs to store frequently accessed hot data, ensuring the monitoring system's fast read and write performance for real-time data; while cold data with low access frequency can be automatically migrated to large-capacity HDDs or low-cost object storage for long-term archiving, achieving the optimal balance between storage cost and access performance.

[0042] Optionally, RAID disk redundancy technology can be implemented to prevent data loss due to the failure of a single hard drive through data striping and mirroring, ensuring high availability of the storage system. Simultaneously, an off-site backup strategy can be established to back up important data to storage facilities in different physical locations, forming a complete data disaster recovery protection system.

[0043] Optionally, a data lifecycle management mechanism can be established, and automated migration strategies can be formulated based on the importance and access frequency of the data. The system automatically performs integrity checks and regular maintenance on the stored data to ensure that the massive amounts of long-term stored monitoring data remain available, providing reliable assurance for historical data tracing and comprehensive analysis.

[0044] S4: As Figure 4As shown, this embodiment processes and analyzes the stored multimodal fusion data, employs the ViBe algorithm for background modeling and real-time updates, implements adaptive threshold segmentation through an improved OTSU algorithm, and accurately extracts foreground targets using an eight-neighbor connected component algorithm. An improved FAST-1 / FAST-2 feature point operator is used to extract region features, and the DBSCAN density clustering algorithm is combined to automatically identify gas leaks and dynamic obstacles, completing dynamic obstacle removal and gas leak detection.

[0045] Optionally, the multimodal data processing and analysis in S4 includes the following steps:

[0046] S41: The ViBe algorithm is used to model the background and update it in real time for the input multimodal fusion data. The core of this algorithm is that it performs background modeling and real-time updates for each pixel. Maintain a background model containing N samples . It is a pixel. Background model N specific sample values ​​in the current frame. For each pixel in the current frame... value Its foreground determination function is:

[0047] in, It is a distance metric function (such as Euclidean distance). Distance threshold This is the minimum number of matching samples required to classify a pixel as background. It employs a unique random update strategy and spatial propagation mechanism: once a pixel is classified as background, the algorithm updates it with probability... Randomly update its own background model The ViBe algorithm updates a sample value in the background model of a neighboring pixel with equal probability. The output of the ViBe algorithm is a binary mask image with the same size as the original video frame, where the foreground target pixel value is 255 (white) and the background region pixel value is 0 (black).

[0048] S42: By calculating the difference image between the current frame and the background model constructed by the ViBe algorithm in S41, which contains complete grayscale information of the foreground target, the improved OTSU algorithm is used to perform adaptive threshold binarization processing on the difference image, and finally an optimized binary mask image is output. The specific steps include the following:

[0049] S421: Calculate the difference image between the current frame and the reference background frame corresponding to the ViBe background model. This difference image contains complete grayscale information of the foreground target. Given that gas leaks in hazardous chemical storage scenarios typically exhibit blurred edges and low contrast, the OTSU algorithm applies contrast-limited adaptive histogram equalization (CLAHE) to the difference image before calculating the threshold. This step amplifies the grayscale difference in low-contrast areas (such as areas with small temperature differences between the leaked gas and the background) while suppressing overall noise, laying a clear foundation for subsequent threshold segmentation and overcoming the shortcomings of traditional OTSU in directly processing the original difference image, which is insensitive to weak signals.

[0050] S422: Dynamically calculate the optimal threshold and perform posterior validation. An improved OTSU algorithm is used to calculate the inter-class variance of the preprocessed difference image. When the inter-class variance reaches its maximum value, the corresponding candidate gray value is the optimal segmentation threshold automatically calculated by the algorithm. To improve the robustness of threshold calculation, the algorithm adds a posterior verification mechanism: if the calculated threshold... If the value exceeds the preset reasonable range (e.g., minimum grayscale value +10, maximum grayscale value -10), it is determined that the threshold may be invalid due to abnormal image content (e.g., no foreground or full foreground). In this case, the algorithm will use a backup threshold strategy (e.g., the previous valid threshold or a fixed threshold based on global statistics) to effectively prevent meaningless segmentation results in extreme cases and ensure the effectiveness of the segmentation threshold.

[0051] S423: Apply the optimized threshold for binarization segmentation and output the result. The optimal segmentation threshold is obtained. As the criterion, the difference image is judged pixel by pixel: if the pixel gray value is greater than If the image is white, it is considered foreground and assigned a value of 255 (white); otherwise, it is considered background and assigned a value of 0 (black). Finally, the algorithm outputs a binary mask image with the same size as the original frame but significantly improved quality. This improved process, through a closed-loop "enhancement-computation-verification" process, is particularly suitable for detecting low-contrast gas leaks and distinguishing dynamic interference in hazardous chemical storage environments. Compared to the traditional OTSU algorithm, it effectively suppresses missegmentation caused by noise or background fluctuations while preserving the true leak signal, thus improving the accuracy of subsequent target extraction and recognition.

[0052] S43: To reduce computational complexity and adapt to targets of different scales, a Gaussian image pyramid is used to perform multi-scale dimensionality reduction processing on the binary mask image output from S42, finally outputting a multi-scale image pyramid structure. The specific steps include the following:

[0053] S431: Construct the bottom layer of the pyramid by using the binary mask image output by the improved OTSU algorithm as the bottom layer (layer 0) of the pyramid.

[0054] S432: Perform Gaussian smoothing and downsampling on the binary mask image of the current pyramid level (e.g., layer k). Typically, a small convolution kernel (e.g., 3x3 or 5x5, standard deviation σ=1.0) is used for convolution to smooth the binary mask image and suppress noise. Subsequently, the smoothed binary mask image is downsampled, usually using nearest-neighbor interpolation or bilinear interpolation, reducing both the length and width of the binary mask image to half of their original values. The image processed at this stage is the (k+1)th layer of the pyramid.

[0055] S433: Perform iterative construction, taking the binary mask image of the (k+1)th layer as the new input, and repeat the "Gaussian smoothing-downsampling" process of S432 to generate a series of binary mask images with progressively decreasing resolution, forming a pyramid structure.

[0056] S44: Based on the dimension-reduced binary mask image, a series of complete and labeled connected regions are accurately extracted using the eight-neighbor connected component algorithm. The specific steps include the following:

[0057] S441: This algorithm scans and labels a binary mask image with progressively decreasing resolution. Starting from the top left corner of the binary mask image, the algorithm scans row by row, pixel by pixel. For each pixel with a value of 255 (white, representing the foreground), the algorithm checks the labeling of its eight adjacent pixels above and to the left (i.e., eight directions including top, bottom, left, right, top left, top right, bottom left, and bottom right). The current pixel's assignment is determined based on the labels of its neighboring pixels: if all surrounding pixels are background (0) or unlabeled, a new unique label (e.g., 1, 2, 3...) is assigned to the current foreground region; if there are already labeled foreground pixels among the surrounding pixels, the current pixel is labeled with the same label.

[0058] S442: This algorithm performs equivalence parsing and label merging. During the initial labeling process, due to the limitations of the scanning order, the same connected region may be temporarily assigned multiple different labels. After the first scan is completed, the algorithm merges all equivalent labels into a single unique label based on the equivalence table.

[0059] S443: This algorithm calculates and outputs the properties of each connected region. For each ultimately merged independent connected region (i.e., a labeled foreground object), the algorithm calculates a series of geometric properties:

[0060] 1. Circular rectangle: The smallest rectangle that can completely enclose the area. Its properties include the coordinates (x, y) of the top-left vertex, the width and height of the rectangle;

[0061] 2. Region area: the total number of pixels contained within the region;

[0062] 3. Center coordinates: This is the average of the coordinates of all pixels in the region, usually represented as (center_x, center_y).

[0063] 4. Other optional attributes: such as region perimeter, outline, etc.

[0064] S45: For each connected region extracted in S44, the improved FAST-1 and FAST-2 feature point operators are used to extract features from the target, and finally a set of fused feature points is output, which specifically includes the following steps:

[0065] S451: Perform grayscale processing on each independent foreground target region extracted by the eight-neighbor connected domain algorithm, and calculate the local statistical features of each independent foreground target region.

[0066] S452: An improved FAST-1 operator is used to extract features from the target region. Its adaptive threshold mechanism is specifically manifested as follows: for the current target region, the threshold T1 is dynamically adjusted based on the real-time calculated local gray-level variance, i.e. This adaptive mechanism enables the operator to intelligently adjust its sensitivity based on regional characteristics (such as the low contrast characteristics of a gas leak area or the high contrast characteristics of a dynamic obstacle), avoiding the problem of insufficient adaptability of a fixed threshold in different scenarios.

[0067] S453: A modified FAST-2 operator is used to perform supplementary feature extraction on the same target region. This operator employs an expanded neighborhood radius (e.g., radius 4) and a smaller fixed threshold T2, specifically designed to capture subtle grayscale gradient features. The larger neighborhood range allows it to perceive grayscale changes over a wider area, while the lower threshold setting makes it more sensitive to subtle temperature gradient changes, making it particularly suitable for detecting edge diffusion regions of gas leaks.

[0068] S454: The feature points extracted by the FAST-1 and FAST-2 operators are fused to fully utilize the feature extraction capabilities of FAST-1 in the central region and the detection advantages of FAST-2 in the edge gradient region. Through deduplication and optimization, a high-quality set of feature points is formed that can capture both significant features and perceive subtle changes, providing rich and accurate feature input for subsequent DBSCAN density clustering analysis.

[0069] S46: Apply the DBSCAN density clustering algorithm to perform cluster analysis on the feature points extracted in S45. This can be combined with morphological features to determine the target category, including:

[0070] S461: Based on the feature point distribution characteristics of gas leaks and dynamic obstacles, initialize the two core parameters of the DBSCAN algorithm: the neighborhood radius eps and the minimum number of samples (min_samples) required to form core points. Considering that gas leak areas in infrared images typically exhibit a diffuse, uneven density pattern, eps is set to a value suitable for the gas diffusion range (usually 5-15 pixels), and min_samples is set to 3-5 points to distinguish between real clusters and noise. Simultaneously, all feature points previously extracted using the improved FAST-1 / 2 operator are standardized in spatial coordinates to establish a two-dimensional spatial distribution matrix of the feature points, preparing for density calculation.

[0071] S462: For each feature point p, calculate the number of other feature points contained in its eps neighborhood, i.e., N_eps(p). If this number is not less than min_samples, mark point p as a core point and create a new cluster for it or assign it to an existing cluster. Recursively explore all density-reachable points using density accessibility: starting from a core point, add all points in its eps neighborhood to the current cluster, and then continue expanding the neighborhood of these newly added points (if they are also core points) until no new points can be added. This step effectively connects points with similar densities, forming clusters of arbitrary shapes.

[0072] S463: Mark all feature points that cannot be assigned to any cluster as noise points (labeled -1). These points typically correspond to random noise or isolated interfering features. After traversing all points, output the cluster labels (ClusterLabels) for each point, where points in the same cluster have the same label number, and noise points are labeled separately. Simultaneously, the algorithm automatically counts the number of generated clusters and records the core sample index and morphological features of each cluster (such as cluster size, density, spatial distribution range, etc.).

[0073] S464: Based on the generated clustering results, the system distinguishes between gas leak areas and dynamic obstacles by combining the morphological features and spatial distribution characteristics of each cluster. Gas leak areas typically correspond to clusters with a large spatial range, uneven density distribution, and a diffuse pattern; while dynamic obstacles usually form compact clusters with concentrated spatial distribution, clear boundaries, and relatively uniform density. The system automatically distinguishes between gas leak areas and dynamic obstacles by setting appropriate morphological feature thresholds (such as cluster area, perimeter-to-area ratio, density index, etc.) and filtering out noise points. (For example, a cluster area of ​​6200 pixels, a perimeter-to-area ratio of 0.19, and a density standard deviation of 28 are considered gas leaks; a cluster area of ​​1500 pixels, a perimeter-to-area ratio of 0.08, and a density standard deviation of 8 are considered dynamic obstacles.) Ultimately, the system can accurately identify gas leak areas and dynamic obstacles.

[0074] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting and classifying anomalies in the hazardous chemical storage environment throughout its entire lifecycle, in all directions, and around the clock, characterized in that: Includes the following steps: Visible light and infrared cameras are deployed at key nodes in the warehouse to simultaneously collect dual-modal video streams. Spatiotemporal alignment is achieved through hardware synchronization to generate multimodal fusion data. A redundant ring network is constructed, and a heterogeneous network is built by combining it with a wireless bridge. The dual-modal video stream data is compressed using encoding to realize the transmission of the dual-modal video stream and the multimodal fused data from the acquisition end to the data center. A distributed hierarchical storage architecture is constructed, storing the dual-modal video stream and multimodal fused data in a time-series database, storing the structured analysis results in a relational database, and constructing a hierarchical storage system. RAID disk redundancy and off-site backup strategies are implemented to form a lifecycle management system; the structured analysis results include the coordinates of dynamic obstacles or the area of ​​the gas leak region. The multimodal fusion data is processed and analyzed to identify gas leaks and dynamic obstacles. The ViBe algorithm is used to perform background modeling and real-time updates on the input multimodal fusion data; Calculate the difference image between the current frame and the ViBe background model, process the difference image using the improved OTSU algorithm, and output an optimized binary mask image. The binary mask image is subjected to multi-scale dimensionality reduction processing using a Gaussian image pyramid. On the dimension-reduced binary mask image, the eight-neighbor connected component algorithm is used to extract and label complete connected regions; Based on the connected region, features are extracted using the improved FAST-1 and FAST-2 operators, and a fused set of feature points is output. The DBSCAN density clustering algorithm is used to perform cluster analysis on the feature points and determine the target category; The improved OTSU algorithm includes: First, the difference image between the current frame and the ViBe reference background frame is subjected to CLAHE enhancement processing; Then, the optimal threshold for the difference image is calculated and a posterior verification is performed; The improved FAST-1 and FAST-2 operators are used to extract features, outputting a set of feature points. The following steps are then performed: The independent foreground target regions extracted by the eight-neighbor connected component algorithm are grayscaled and their local statistical features are calculated. An improved FAST-1 operator is used to extract features from the target region. For the current target region, the improved FAST-1 operator dynamically adjusts the threshold based on the real-time calculated local gray-level variance. An improved FAST-2 operator is used to extract supplementary features for the same target region. The improved FAST-2 operator uses an extended neighborhood radius and a smaller fixed threshold. The feature points extracted by the improved FAST-1 operator and the improved FAST-2 operator are combined and then deduplicated to form a high-quality feature point set.

2. The method according to claim 1, characterized in that, The hierarchical storage system includes: Hot data is stored using SSDs to ensure real-time read and write performance; Cold data is automatically migrated to HDD or object storage to reduce costs and achieve a balance between performance and cost.

3. The method according to claim 1, characterized in that, The ViBe algorithm maintains a background model containing N samples for each pixel, and combines a random update strategy and a spatial propagation mechanism to generate a binary mask image of the same size as the original video frame.

4. The method according to claim 1, characterized in that, The binary mask image is reduced in dimension using the Gaussian image pyramid by performing the following steps: The bottom layer of the pyramid is constructed by using the binary mask image output by the improved OTSU algorithm. Gaussian smoothing and downsampling are applied to the k-th layer binary mask image to obtain the (k+1)-th layer pyramid image; By iteratively constructing the system, the (k+1)th layer binary mask image is used as input, and the Gaussian smoothing and downsampling process is repeated to generate binary mask images with progressively lower resolution, forming a pyramid structure.

5. The method according to claim 1 or 4, characterized in that, Extracting connected regions using the eight-neighbor connected component algorithm includes the following steps: Scan and label binary mask images with progressively decreasing resolution; The different labels of the same connected region temporarily marked during the scan are merged into a unique label using an equivalence table. Calculate and output the properties of each connected region.

6. The method according to claim 1, characterized in that, The DBSCAN density clustering algorithm is used to perform cluster analysis on feature points, including the following steps: Based on the characteristic point distribution characteristics of gas leaks and dynamic obstacles, the DBSCAN algorithm parameters are initialized, and the feature point coordinates are standardized to construct a two-dimensional distribution matrix in preparation for density calculation. By recursively expanding the neighborhood of the core point, points with similar densities are connected to form clusters of arbitrary shapes; Feature points that cannot be classified into any cluster are marked as noise, the number of clusters is counted, and the core sample index and morphological features are recorded. The discrimination is based on the morphological characteristics and spatial distribution of clusters, and a morphological characteristic threshold is set to distinguish between gas leakage areas and dynamic obstacles.

Images