Active flame recognition and localization method

Abstract

This invention provides an active flame recognition and localization method, belonging to the fields of fire early warning and safety monitoring. The technical solution is as follows: A binocular thermal imaging device simultaneously acquires left and right thermal images of the monitored scene and performs preprocessing; based on historical thermal frames, the left and right background thermal models are constructed and updated respectively, generating left and right anomalous thermal difference maps; anomalous heating areas are extracted to form a set of left and right candidate thermal areas; edge ambiguity features are extracted and flame authenticity is determined, resulting in a set of left and right target flame areas; binocular matching is performed and the spatial position of the flame is calculated, outputting the active recognition result. The beneficial effects of this invention are: it can effectively suppress complex thermal background interference, reduce the false alarm rate, and improve the accuracy and stability of flame recognition and localization.

Description

Technical Field

[0001] This invention relates to fields such as fire early warning and safety monitoring, and in particular to an active flame identification and location method. Background Technology

[0002] In industrial workshops, warehouses, equipment rooms, power distribution areas, hazardous materials storage areas, and forest areas, timely detection and accurate location of fires are crucial for minimizing losses. Existing fire detection methods typically employ smoke detectors, temperature sensors, or video monitoring for identification. However, smoke detectors and temperature sensors have limited response locations and cannot directly pinpoint the location of the fire source; while ordinary video monitoring methods are easily affected by changes in lighting, smoke, and environmental background.

[0003] Thermal imaging technology can identify high-temperature targets by detecting the distribution of thermal radiation, offering the advantage of being independent of ambient light. However, existing thermal imaging flame detection methods still have shortcomings in complex thermal environments. For example, in industrial sites and equipment rooms, stable heat sources such as waste heat from equipment, hot pipes, heating components, hot air vents, and heat reflection areas can easily interfere with flame identification. If detection is based solely on absolute temperature thresholds or ordinary heat zone extraction methods, non-flame thermal targets can easily be misidentified as flames. Furthermore, monocular thermal imaging struggles to directly obtain spatial depth information of the target; even with binocular thermal imaging, if the front-end cannot effectively distinguish between real flames and non-flame heat sources, subsequent binocular matching and localization results are easily affected. Summary of the Invention

[0004] The purpose of this invention is to provide an active flame identification and localization method that can suppress stable thermal background interference, accurately identify the real flame in complex thermal environments, and achieve reliable localization on this basis.

[0005] This invention is achieved through the following measures: An active flame identification and localization method, characterized by the following steps: S1, synchronously acquiring left and right thermal images of the monitored scene using a binocular thermal imaging device, and preprocessing the left and right thermal images to obtain preprocessed left current thermal frames and right current thermal frames; S2, establishing and updating left and right background thermal models based on historical thermal frames, comparing the left current thermal frame with the left background thermal model to obtain a left abnormal thermal difference map, and comparing the right current thermal frame with the right background thermal model to obtain a right abnormal thermal difference map; S3, extracting abnormal heating regions based on the left and right abnormal thermal difference maps to obtain a left candidate hot zone set. S4. Extract edge ambiguity features from each candidate hot zone in the left and right candidate hot zone sets respectively, and perform flame authenticity discrimination based on the edge ambiguity features. Confirm the candidate hot zones that pass the discrimination as target flame zones to obtain left target flame zone sets and right target flame zone sets; S5. Perform binocular matching on the left and right target flame zone sets to obtain corresponding flame zone pairs, and calculate the flame spatial position based on the corresponding flame zone pairs; S6. Output active recognition results containing flame confirmation information and three-dimensional positioning information based on the flame spatial position and target flame zone status.

[0006] The invention also has the following specific features: In step S1, the preprocessing of the left and right thermal images includes one or more of the following: bad pixel correction, thermal noise suppression, non-uniformity correction, temperature normalization, local contrast enhancement, and time synchronization verification.

[0007] In step S2, the historical thermal frames are the left and right thermal frame sequences obtained by the binocular thermal imaging device before the current time and after preprocessing. A left background thermal model and a right background thermal model are established based on the left and right thermal frame sequences, respectively. The left and right background thermal models are continuously updated as a new frame of thermal image data arrives. The update prioritizes the absorption of thermal changes with temporal continuity and spatial stability, avoiding the direct incorporation of sudden, locally rapidly enhanced thermal anomalies into the left and right background thermal models.

[0008] In step S2, the left current hot frame is compared position by position with the left background thermal model to obtain the thermal anomaly distribution of the left current hot frame relative to the left background thermal model, and a left anomaly thermal difference map is generated; the right current hot frame is compared position by position with the right background thermal model to obtain the thermal anomaly distribution of the right current hot frame relative to the right background thermal model, and a right anomaly thermal difference map is generated; wherein, for stable heat source regions, since their heat distribution has been absorbed by the left or right background thermal model, the response in the left or right anomaly thermal difference map is weakened.

[0009] In step S3, when extracting abnormal heating regions, threshold filtering, connected region extraction, and region validity screening are performed on the left and right abnormal thermal difference maps respectively. The region validity screening includes one or more of the following: region area screening, time persistence screening, region thermal difference concentration screening, and region morphological integrity screening. The abnormal heating regions after screening form the left candidate thermal region set and the right candidate thermal region set respectively.

[0010] In step S4, the edge blurring features include two or more of the following: edge temperature gradient width, edge temperature gradient magnitude, edge sharpness, edge drift, and contour irregularity. Specifically, the edge temperature gradient width characterizes the spatial change width of the candidate hot area as it transitions from an inner high-temperature region to an outer background region; the edge temperature gradient magnitude characterizes the steepness of the temperature change at the boundary of the candidate hot area; the edge sharpness characterizes whether the boundary of the candidate hot area exhibits a clear and sharp contour; the edge drift characterizes the degree of positional change of the candidate hot area edge in consecutive hot frames; and the contour irregularity characterizes the randomness and complexity of the candidate hot area's shape boundary.

[0011] In step S4, when determining the authenticity of a flame based on the edge ambiguity features, a multi-dimensional feature fusion discrimination logic is constructed based on the edge ambiguity features. Specifically, the edge temperature gradient width, edge drift, and contour irregularity are treated as positively correlated features, while the edge temperature gradient amplitude and edge sharpness are treated as negatively correlated features. All edge ambiguity features are then uniformly scaled and comprehensively evaluated to form a flame confidence score. When the flame confidence score reaches or exceeds a preset authenticity confirmation threshold, the corresponding candidate hot zone is confirmed as the target flame zone.

[0012] In step S5, when performing binocular matching on the left and right target flame area sets, the target flame areas in the left and right target flame area sets are matched at the regional level by comprehensively utilizing regional positional relationships, regional range relationships, thermal distribution similarity relationships, and binocular imaging constraints. Specifically, for each left target flame area in the left target flame area set, one or more candidate corresponding areas are determined in the right target flame area set based on the binocular imaging constraints. Then, the candidate corresponding areas are optimized by combining regional positional relationships, regional range relationships, and thermal distribution similarity relationships to determine the right target flame area that matches the left target flame area, and the two are constructed as a corresponding flame area pair. After obtaining the corresponding flame area pair, the representative position of the region used for positioning is determined from the corresponding flame area pair, and the spatial depth information and lateral and longitudinal position information of the target flame are determined based on the calibration parameters of the binocular thermal imaging device and the positional differences between the left and right target flame areas, thereby obtaining the spatial position of the flame.

[0013] In step S6, the active identification result includes at least flame confirmation information and three-dimensional positioning information. The target flame zone state includes at least one of the following: the existence state of the target flame zone, the area change state, the boundary change state, the location change state, and the continuous existence state. Based on the flame spatial location and the target flame zone state, the target flame is comprehensively confirmed, and the active identification result is output.

[0014] A system related to active flame recognition and localization methods includes a binocular thermal imaging device, a processing unit, a storage unit, and an output unit. The binocular thermal imaging device is used to simultaneously acquire left and right thermal images of a monitored scene. The storage unit is used to store historical thermal frames, a left background thermal model, a right background thermal model, and intermediate processing results. The processing unit is used to preprocess the left and right thermal images to obtain preprocessed left and right current thermal frames. Based on the historical thermal frames, the left and right background thermal models are constructed and updated respectively. The left current thermal frame is compared with the left background thermal model to obtain a left abnormal thermal difference map, and the right current thermal frame is compared with the right background thermal model to obtain a right abnormal thermal difference map. Based on the left abnormal thermal difference map... Abnormal heating regions are extracted from the normal thermal difference map and the right abnormal thermal difference map, respectively, to obtain a left candidate hot zone set and a right candidate hot zone set. Edge ambiguity features are extracted from each candidate hot zone in the left and right candidate hot zone sets, and flame authenticity is judged based on the edge ambiguity features. The candidate hot zones that pass the judgment are confirmed as target flame zones, to obtain a left target flame zone set and a right target flame zone set. Binocular matching is performed on the left and right target flame zone sets to obtain corresponding flame zone pairs, and the flame spatial position is calculated based on the corresponding flame zone pairs. The output unit is used to output an active recognition result containing flame confirmation information and three-dimensional positioning information based on the flame spatial position and the target flame zone state.

[0015] The beneficial effects of this invention are as follows: This invention simultaneously acquires left and right thermal images using a binocular thermal imaging device, and generates left and right anomalous thermal difference maps by combining the left and right background thermal models. This effectively reduces the interference of stable thermal backgrounds such as equipment waste heat, hot pipes, and heat reflection areas on flame identification. Based on this, by extracting edge ambiguity features from the left and right candidate hot zone sets, the authenticity of the candidate hot zones is determined, further distinguishing real flames from non-flame heat sources and reducing the false alarm rate. Subsequently, binocular matching is performed based on the left and right target flame zone sets to construct corresponding flame zone pairs and calculate the flame's spatial position, improving the accuracy and stability of flame localization. Therefore, this invention can reliably identify and accurately locate real flame targets in complex thermal environments, with advantages such as low false alarm rate, high identification accuracy, good localization effect, and strong continuous monitoring stability. Attached Figure Description

[0016] Figure 1 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0018] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0019] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0020] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to explain the solution.

[0021] Example 1 See Figure 1 This embodiment provides an active flame recognition and localization method. This method is applicable to complex thermal environment monitoring scenarios such as industrial workshops, storage areas, equipment rooms, power distribution areas, hazardous materials storage areas, and other environments with stable heat sources, heat reflection areas, and fluctuating ambient temperatures. In these scenarios, high-temperature equipment surfaces, hot pipes, heating elements, hot air vents, heat reflection areas, and slowly heating backgrounds can easily interfere with traditional flame recognition methods, leading to a high false alarm rate and further affecting the accuracy of flame localization. Therefore, this embodiment combines environmental thermal background self-learning with flame edge ambiguity discrimination. First, it extracts abnormally heating targets from the thermal background, then performs flame authenticity discrimination on the abnormally heating targets, and based on this, performs binocular matching and flame spatial position calculation, thereby achieving active recognition and localization of real flame targets in complex thermal environments.

[0022] In this embodiment, the system for executing the active flame recognition and localization method includes at least a binocular thermal imaging device, a processing unit, a storage unit, and an output unit. The binocular thermal imaging device is used to simultaneously acquire left and right thermal images of the monitored scene; the processing unit is used to perform image preprocessing, background thermal model establishment and updating, abnormal temperature rise area extraction, flame authenticity determination, binocular matching, and flame spatial position calculation; the storage unit is used to store historical thermal frames, the background thermal model, and intermediate processing results; and the output unit is used to output the active recognition result containing flame confirmation information and three-dimensional positioning information. The binocular thermal imaging device is preferably installed at a fixed position above, diagonally above, or slightly above the side of the area to be monitored to ensure a good thermal image acquisition field of view for the monitored area and reduce the impact of obstruction and blind spots on the recognition results.

[0023] This invention achieves active flame identification and localization in complex thermal environments through binocular thermal imaging image recognition, background thermal model construction and updating, flame authenticity determination, and binocular localization.

[0024] In this embodiment, the active flame identification and localization method includes the following steps.

[0025] Step S1: Simultaneously acquire the left and right thermal images of the monitored scene using a binocular thermal imaging device, and preprocess the left and right thermal images to obtain the preprocessed left current thermal frame and right current thermal frame.

[0026] This step is used to acquire binocular thermal image data that reflects the current thermal distribution state of the monitoring scene, and to improve the stability, consistency and analyzability of the thermal image data through preprocessing, so as to provide reliable input for subsequent background thermal model establishment and updating, abnormal thermal difference map generation, candidate hot zone extraction and flame authenticity determination.

[0027] In practical implementation, the binocular thermal imaging device includes a left thermal imaging acquisition unit and a right thermal imaging acquisition unit. The left thermal imaging acquisition unit acquires thermal images from a left-view perspective to obtain a left thermal image; the right thermal imaging acquisition unit acquires thermal images from a right-view perspective to obtain a right thermal image. The left and right thermal imaging acquisition units acquire images synchronously at the same time to ensure that the left and right thermal images correspond to the same thermal state in the monitored scene. The processing unit extracts the left and right thermal images corresponding to the current moment from the thermal imaging data stream continuously output by the binocular thermal imaging device and uses them as the raw inputs for the left current thermal frame and the right current thermal frame, respectively. The left current thermal frame and the right current thermal frame refer to the two thermal imaging images acquired by the binocular thermal imaging device at the same acquisition moment, used to jointly characterize the thermal distribution state of the monitored scene at that moment.

[0028] Considering that the original thermal images may be affected by factors such as bad pixels, thermal noise, imaging non-uniformity, and insufficient local thermal contrast during actual acquisition, in this embodiment, the processing unit performs preprocessing on the left and right current thermal frames after acquisition. The preprocessing preferably includes one or more of the following: bad pixel correction, thermal noise suppression, non-uniformity correction, temperature normalization, local contrast enhancement, and time synchronization verification. Specifically, bad pixel correction corrects abnormal isolated pixels in the thermal image; thermal noise suppression reduces random fluctuations in the thermal imaging data; non-uniformity correction improves the overall consistency of the thermal image; temperature normalization ensures a uniform scale between the left and right thermal images and thermal images from different times; local contrast enhancement enhances the difference between high-temperature areas and the surrounding background; and time synchronization verification ensures that the left and right thermal images belong to the same or approximately the same time.

[0029] The above preprocessing steps constitute image enhancement processing for the left and right thermal images to improve the accuracy of subsequent candidate thermal area extraction, flame authenticity determination, and binocular matching. After preprocessing, the processing unit outputs the preprocessed left current thermal frame and the preprocessed right current thermal frame as direct input to step S2.

[0030] Step S2: Establish and update the left background thermal model and the right background thermal model according to the historical hot frames. Compare the left current hot frame with the left background thermal model to obtain the left abnormal thermal difference map. Compare the right current hot frame with the right background thermal model to obtain the right abnormal thermal difference map.

[0031] This step is used to separate the long-term stable thermal background information within the monitoring scene from continuous thermal imaging data, and based on this, identify the abnormal temperature rise region relative to the thermal background at the current moment, thus providing a basis for subsequent candidate thermal area extraction. This step does not directly rely on the absolute high temperature regions in the left and right current thermal frames for identification, but instead uses the left and right background thermal models to perform background stripping on the current heat distribution, shifting the focus of identification from absolute high temperature to abnormal temperature rise relative to the background.

[0032] In this embodiment, the preprocessed left current hot frame and the preprocessed right current hot frame output in step S1 are used as the current input data for this step, and the historical hot frames stored in the storage unit serve as the data basis for establishing and updating the background thermal model. The historical hot frames refer to the left and right hot frame sequences obtained by the binocular thermal imaging device before the current moment and after preprocessing. To ensure consistent processing scale, the historical hot frames preferably use the same preprocessing method as the left and right current hot frames.

[0033] The processing unit establishes independent left and right background thermal models for the left and right thermal imaging channels, respectively. The left background thermal model describes the stable thermal distribution state of the monitoring scene under the left thermal imaging channel, and the right background thermal model describes the stable thermal distribution state of the monitoring scene under the right thermal imaging channel. The left and right background thermal models are used to characterize the relatively stable thermal information in the monitoring scene over a long period of time. The thermal information may include waste heat from equipment, heat pipes, heating components, heated areas of walls, thermal reflection background, and other background thermal factors that do not suddenly disappear or increase in a short period of time.

[0034] During the background thermal model establishment phase, the processing unit reads historical thermal frames from the storage unit at multiple consecutive time points, forming a left historical thermal frame sequence and a right historical thermal frame sequence, respectively. It then generates a left background thermal model based on the left historical thermal frame sequence and a right background thermal model based on the right historical thermal frame sequence. During the background thermal model update phase, the processing unit continuously updates the left and right background thermal models as new thermal image data arrives, enabling the background thermal models to adaptively adjust to changes in ambient temperature, equipment operating status, or sunlight conditions.

[0035] Specifically, the background thermal model update formula is: in, Indicates thermal imaging channel At any moment Background thermal model values, Indicates thermal imaging channel The background thermal model value at the previous moment, Indicates thermal imaging channel At any moment The current hot frame pixel value, Indicates the background update coefficient. and Represents pixel position coordinates, Indicates the current moment. Indicates the thermal imaging channel identifier, where Indicates the left thermal imaging channel. This indicates the right thermal imaging channel.

[0036] During the update process, the processing unit prioritizes absorbing thermal changes that have temporal continuity and spatial stability, avoiding the direct incorporation of sudden, locally rapidly amplified thermal anomalies into the background thermal model, thereby ensuring that the background thermal model remains sensitive to sudden flames.

[0037] After obtaining the left and right background thermal models, the processing unit compares the current left hot frame with the left background thermal model position by position to obtain the thermal anomaly distribution of the current left hot frame relative to the left background thermal model. Simultaneously, it compares the current right hot frame with the right background thermal model position by position to obtain the thermal anomaly distribution of the current right hot frame relative to the right background thermal model. Based on this, the processing unit generates a left anomaly thermal difference map and a right anomaly thermal difference map. The left anomaly thermal difference map characterizes the degree of thermal anomaly at each position in the current left hot frame relative to the left background thermal model, and the right anomaly thermal difference map characterizes the degree of thermal anomaly at each position in the current right hot frame relative to the right background thermal model.

[0038] Specifically, the formula for generating an abnormal heat map is:

[0039] in, Indicates thermal imaging channel At any moment Abnormal thermal difference value, Indicates thermal imaging channel At any moment The current hot frame pixel value. When When, a left-side anomalous thermal difference map can be generated; when At this time, a right anomalous thermal difference map can be generated. For stable heat source regions, their heat distribution has been absorbed by the background thermal model, so the response in the anomalous thermal difference map is weakened; for newly emerging, locally enhanced heat targets, they appear as more obvious anomalous regions in the anomalous thermal difference map. Therefore, the output of step S2 includes the left anomalous thermal difference map, the right anomalous thermal difference map, and the updated left and right background thermal models, where the left and right anomalous thermal difference maps serve as direct inputs to step S3.

[0040] Step S3: Extract the abnormal heating regions based on the left and right abnormal thermal difference maps to obtain the left candidate thermal region set and the right candidate thermal region set.

[0041] This step is used to filter out regions with abnormal temperature rise characteristics from the left and right anomalous thermal difference maps, and organize these regions into the objects to be judged in the subsequent flame authenticity determination, namely the left candidate hot zone set and the right candidate hot zone set. This step does not directly confirm the flame, but first extracts the anomalous thermal targets that may belong to the flame from the anomalous thermal difference maps, so that the subsequent step S4 can perform edge ambiguity analysis and flame authenticity determination around these anomalous thermal targets.

[0042] In this embodiment, the processing unit performs region analysis on the left and right anomalous thermal difference maps respectively to identify continuous thermal difference regions where the degree of thermal anomaly is significantly higher than that of the surrounding background. The anomalous heating region refers to a local area in the anomalous thermal difference map where the thermal difference value is relatively concentrated, and this local area has a certain degree of spatial connectivity and temporal persistence. Therefore, the anomalous heating region is not a single isolated pixel, nor is it a random interference point that appears instantaneously and disappears immediately, but rather a local thermal anomaly target that can form a certain range and has a certain degree of significant thermal difference.

[0043] To extract anomalous heating regions, the processing unit preferably performs threshold filtering, connected component extraction, and region validity screening on the left and right anomalous thermal difference maps, respectively. Threshold filtering identifies locations in the anomalous thermal difference maps where the thermal difference response exceeds a preset level and marks these locations as potential anomalies. Connected component extraction groups spatially adjacent or continuously distributed potential anomalies into locally continuous regions. Region validity screening determines whether locally continuous regions meet the basic conditions for being candidate hot zones. In this embodiment, region validity screening preferably includes one or more of the following: region area screening, time persistence screening, regional thermal difference concentration screening, and region morphological integrity screening. Region area screening eliminates invalid regions with excessively small areas; time persistence screening excludes transient thermal disturbances that occur only at a single moment or in a very short period; regional thermal difference concentration screening identifies whether the internal thermal difference distribution has a certain degree of concentration; and region morphological integrity screening ensures that candidate hot zones have a basically analyzable regional morphology.

[0044] After the above screening, the processing unit designates the abnormally heated areas that passed the screening as left and right candidate hot zones, respectively, and aggregates all candidate hot zones in the same thermal imaging channel to form left and right candidate hot zone sets. Each candidate hot zone in the left and right candidate hot zone sets represents a regional target that exhibits a significant abnormal temperature rise relative to the background thermal model and meets the basic regional validity requirements; however, whether it is a real flame is not yet determined in this step. Therefore, the output of step S3 is the left and right candidate hot zone sets, which serve as the direct input for step S4.

[0045] Step S4: Extract edge ambiguity features from each candidate hot zone in the left candidate hot zone set and the right candidate hot zone set respectively, and perform flame authenticity discrimination based on the edge ambiguity features. Confirm the candidate hot zones that pass the discrimination as target flame zones to obtain the left target flame zone set and the right target flame zone set. The flame authenticity discrimination is a pattern recognition process based on edge ambiguity features, which is used to distinguish between real flames and non-flame heat sources.

[0046] This step further verifies the authenticity of the left and right candidate hot zone sets obtained in step S3, distinguishing those candidate hot zones that truly possess the essential attributes of a flame, and confirming the candidate hot zones that pass the verification as the target flame zone. This step does not simply rely on the temperature or size of the candidate hot zones, but analyzes the edge heat distribution characteristics and edge dynamic change characteristics of the candidate hot zones to identify the essential differences between real flames and non-flame heat sources.

[0047] In this embodiment, the processing unit extracts edge ambiguity features from each candidate hot zone in the left candidate hot zone set and each candidate hot zone in the right candidate hot zone set. The edge ambiguity features refer to a set of features that reflect the boundary transition state, boundary stability, and boundary morphological characteristics of the candidate hot zone. Real flames are not rigid entities; their boundaries are affected by the combustion process, the movement of the heat plume, and local turbulent disturbances, typically exhibiting wide boundary transitions, irregular outlines, and edges that jitter and are not clear over time. In contrast, non-flame heat sources, such as metal heating elements, heating components, or heat reflection areas, often have sharper, more stable, and more regular boundaries. Therefore, by analyzing the edge ambiguity features of the candidate hot zones, real flames and static heat sources can be better distinguished.

[0048] In this embodiment, the processing unit preferably extracts at least two or more of the following edge blurring features for each candidate hot region, and preferably extracts all the main features. First, edge temperature gradient width, used to characterize the spatial change width of the candidate hot region as it transitions from an inner high-temperature region to an outer background region; second, edge temperature gradient magnitude, used to characterize the steepness of the temperature change at the boundary of the candidate hot region; third, edge sharpness, used to characterize whether the boundary of the candidate hot region exhibits a clear and sharp outline; fourth, edge drift, used to characterize the degree of positional change of the edge of the candidate hot region in consecutive hot frames; fifth, outline irregularity, used to characterize the randomness and complexity of the outline boundary of the candidate hot region.

[0049] After extracting edge ambiguity features, the processing unit constructs a multi-dimensional feature fusion discrimination logic based on the extracted features to comprehensively determine candidate hot zones. In this embodiment, the multi-dimensional feature fusion discrimination logic preferably follows the following discrimination principles: for features such as edge temperature gradient width, edge drift, and contour irregularity, the processing unit treats them as positively correlated features; that is, when these features are more significant, the discrimination tendency of the candidate hot zone to belong to a real flame is increased. For features such as edge temperature gradient amplitude and edge sharpness, the processing unit treats them as negatively correlated features; that is, when these features are more significant, the discrimination tendency of the candidate hot zone to belong to a real flame is reduced. In specific implementation, the processing unit first performs uniform scaling on each edge ambiguity feature, then assigns enhancement weights to positively correlated features and suppression weights to negatively correlated features, and comprehensively evaluates them based on the importance of each feature to form a flame confidence score that characterizes the strength of the flame attributes of the candidate hot zone.

[0050] Specifically, the formula for flame confidence scoring is:

[0051] in, This indicates the confidence score for the flame. Indicates the width of the edge temperature gradient. Indicates the amount of edge drift. Indicates the irregularity of the outline. Indicates the magnitude of the edge temperature gradient. Indicates edge sharpness, This represents the enhancement weights corresponding to positively correlated features. This represents the suppression weights corresponding to negatively correlated features. Among them, edge temperature gradient width, edge drift, and contour irregularity are positively correlated features; as their values ​​increase, the flame confidence score increases. Edge temperature gradient magnitude and edge sharpness are negatively correlated features; as their values ​​increase, the flame confidence score decreases.

[0052] After obtaining the flame confidence score, the processing unit compares it with a preset authenticity confirmation threshold.

[0053] Specifically, the formula for determining the authenticity confirmation threshold is as follows:

[0054] in, This indicates the confidence score for the flame. This represents the authenticity confirmation threshold. When the flame confidence score reaches or exceeds the authenticity confirmation threshold, it indicates that the candidate hot zone possesses both significant boundary feathering and dynamic blurring characteristics, conforming to the essential boundary attributes of a real flame. Therefore, the candidate hot zone is confirmed as the target flame zone. When the flame confidence score is below the authenticity confirmation threshold, it indicates that the candidate hot zone is more likely to be a static heat source, a heat reflection zone, or other non-flame anomalous heat targets. Even if the area has a high temperature or a large area, it is not confirmed as a target flame zone and is eliminated in this step. After the above processing, the processing unit obtains the left target flame zone set and the right target flame zone set, which serve as the direct input for step S5.

[0055] Step S5: Perform binocular matching on the left target flame zone set and the right target flame zone set to obtain corresponding flame zone pairs, and calculate the flame spatial position based on the corresponding flame zone pairs.

[0056] This step utilizes the left and right target flame area sets obtained in step S4 to establish a correspondence between the same flame target in the left and right thermal imaging channels, and determines the spatial location of the flame in the monitored scene based on this. This step does not perform indiscriminate matching of the entire left and right thermal images; instead, it uses the target flame areas already identified in the previous step as the matching objects. This avoids background areas and non-flame thermal targets from entering the matching process, improving matching accuracy and subsequent positioning reliability.

[0057] In practical implementation, the processing unit comprehensively utilizes regional positional relationships, regional extent relationships, thermal distribution similarity relationships, and binocular imaging constraints to perform region-level matching of target flame areas in the left and right target flame area sets. The binocular imaging constraints are determined by the installation and calibration results of the binocular thermal imaging device. Regarding regional positional relationships, the processing unit performs preliminary screening based on the location range of the left and right target flame areas in their respective thermal images. Regarding regional extent relationships, the processing unit compares the regional area, boundary range, and overall scale of the left and right target flame areas to eliminate obviously unreasonable matching pairs. Regarding thermal distribution similarity relationships, the processing unit analyzes the heat concentration locations, thermal difference distribution, and edge transition states within the left and right target flame areas to enhance matching reliability. Regarding binocular imaging constraints, the processing unit limits the reasonable search range of the right target flame area based on the installation and calibration results of the binocular thermal imaging device, thereby reducing blind matching.

[0058] The processing unit first identifies one or more candidate corresponding regions in the set of right target flame regions for each left target flame region in the set of left target flame regions, based on the binocular imaging constraints. Then, it optimizes the candidate corresponding regions by considering regional positional relationships, regional extent relationships, and thermal distribution similarity, ultimately determining the right target flame region that matches the left target flame region. When a left target flame region and a right target flame region meet preset matching conditions, the processing unit treats them as corresponding representations of the same real flame target in the left and right thermal images, constructing them as a pair of corresponding flame regions. When multiple target flames exist in the monitored scene, the processing unit can establish multiple pairs of corresponding flame regions.

[0059] After obtaining the corresponding flame zone pairs, the processing unit further calculates the flame spatial position based on the corresponding flame zone pairs. The flame spatial position refers to the spatial position of the flame target relative to the binocular thermal imaging device or the preset monitoring scene coordinate system, used to characterize the three-dimensional positioning information of the flame in the monitoring scene. In specific implementation, the processing unit determines the representative position of the region for positioning from the corresponding flame zone pairs, such as the region center position, the heat concentration area position, or the main contour center position, to improve positioning stability. Subsequently, the processing unit recovers the spatial depth information and lateral and longitudinal position information of the target flame based on the calibration parameters of the binocular thermal imaging device and the positional differences between the left and right target flame zones, thereby forming a complete flame spatial position result. Therefore, the output of step S5 includes the corresponding flame zone pairs and the flame spatial position, where the flame spatial position serves as the direct input to step S6.

[0060] Step S6: Based on the spatial location of the flame and the state of the target flame zone, output an active identification result containing flame confirmation information and three-dimensional positioning information.

[0061] This step integrates the flame spatial location obtained in step S5 with the target flame zone information confirmed in step S4 to form an active identification result that can be used for monitoring, display, recording, linkage, or alarm. This step does not simply display the aforementioned intermediate processing results; instead, it unifies the flame authenticity confirmation result with the flame spatial location result, ensuring that the system ultimately outputs an active identification result with clear target attributes and spatial location information.

[0062] In this embodiment, the flame spatial location output in step S5 serves as one of the core inputs of this step, while the target flame area information obtained in step S4 serves as another input basis for this step. Specifically, the flame spatial location characterizes the spatial position of the target flame within the monitoring scene, and the target flame area state characterizes the regional changes of the target flame area obtained after authenticity verification at the current and subsequent times. The target flame area state may include at least one of the following: the existence state of the target flame area, the regional area change state, the regional boundary change state, the regional position change state, and the persistent existence state.

[0063] In practical implementation, the processing unit first reads the flame spatial location obtained in step S5 and uses it as the basis for three-dimensional positioning of the target flame in the monitoring scene. If only a single flame spatial location exists at the current moment, the processing unit directly forms the corresponding active identification result around that flame spatial location; if multiple flame spatial locations exist at the current moment, the processing unit establishes corresponding active identification results for each flame spatial location. Subsequently, the processing unit combines the state information of the corresponding target flame area to comprehensively confirm the current state of the target flame. For example, when a target flame area exists continuously in consecutive moments, and its regional boundary and internal heat distribution continuously maintain flame characteristics, the processing unit increases the credibility of the active identification result corresponding to the target flame; when a target flame area has passed the flame authenticity judgment, but its area shrinks rapidly, its duration is short, or its boundary state weakens rapidly, the processing unit still retains the confirmation conclusion that it is a target flame in the active identification result, while reflecting its current state change characteristics.

[0064] In this embodiment, the active identification result includes at least flame confirmation information and three-dimensional positioning information. Flame confirmation information indicates that the current output object has been identified as the target flame by the system through flame authenticity judgment; three-dimensional positioning information indicates the spatial location of the target flame in the monitoring scene. The output unit, according to the system configuration, outputs the flame confirmation information and three-dimensional positioning information to a display terminal, monitoring platform, alarm system, or linkage control system in the form of display information, recording information, communication information, or control information. If necessary, the active identification result may also include a description of the target flame zone's state to enhance the ability to express the current changes in the flame.

[0065] The active identification result in this embodiment is not a passive alarm based on a single temperature threshold, but a comprehensive identification result based on synchronous thermal image acquisition, background thermal model self-learning, abnormal temperature rise area extraction, flame authenticity determination, binocular matching, and flame spatial location calculation. Therefore, the active identification result has high reliability and clear target orientation. Thus, steps S1 to S6 sequentially complete the complete processing flow from binocular thermal image acquisition, background thermal model establishment and updating, abnormal temperature rise area extraction, flame authenticity determination, binocular matching and spatial positioning to active identification result output, enabling the present invention to reliably identify and locate real flame targets in complex thermal environments.

[0066] Example 2 See Figure 1 This embodiment provides a method for verifying the effectiveness of an active flame recognition and localization method. This embodiment is also applicable to complex thermal environment monitoring scenarios such as industrial workshops, storage areas, equipment rooms, power distribution areas, hazardous materials storage areas, and other environments with stable heat sources, heat reflection areas, and fluctuating ambient temperatures. Unlike Embodiment 1, this embodiment focuses on verifying the technical effectiveness of the invention in identifying real flames and determining their spatial location under complex thermal backgrounds. In particular, it verifies the improvement in false alarm rate, false negative rate, flame recognition accuracy, localization error, and continuous monitoring stability achieved by combining environmental thermal background self-learning with flame edge ambiguity discrimination. As described in Embodiment 1, the present invention utilizes a binocular thermal imaging device, processing unit, storage unit, and output unit working collaboratively to complete the entire process from binocular thermal image acquisition, background thermal model establishment and updating, abnormal temperature rise area extraction, flame authenticity discrimination, binocular matching and flame spatial location calculation to active recognition result output, sequentially following steps S1 to S6. This embodiment further introduces exemplary test conditions and comparison schemes within the same technical framework to quantitatively illustrate the technical effectiveness of the invention.

[0067] In this embodiment, the test scenario is a device room with a typical complex thermal background. This device room contains power distribution cabinets, ventilation openings, metal supports, heat pipes, and local reflective surfaces, thus creating a test environment where a stable heat source, localized thermal reflection, and a slowly rising temperature background coexist. A binocular thermal imaging device is fixedly installed on a support above the area to be monitored, providing a continuous and stable thermal image acquisition field of view. The processing unit continuously acquires the left and right thermal images from the data stream output by the binocular thermal imaging device, and extracts the left and right current thermal frames at the same acquisition time. Subsequently, the same preprocessing procedure as in Embodiment 1 is performed on the left and right current thermal frames, including bad pixel correction, thermal noise suppression, non-uniformity correction, temperature normalization, local contrast enhancement, and time synchronization verification, to ensure data consistency and comparability when establishing subsequent left and right background thermal models. Before the test, historical thermal frames were continuously collected under flameless operating conditions to establish left and right background thermal models respectively. During the test, the processing unit continued to update the left and right background thermal models based on newly arrived thermal frames, enabling them to adapt to changes in equipment residual heat, environmental thermal drift, and slow temperature rise background, while avoiding direct absorption of sudden flames into the background. Subsequently, the current left thermal frame was compared with the left background thermal model to generate a left anomalous thermal difference map; the current right thermal frame was compared with the right background thermal model to generate a right anomalous thermal difference map. Based on the left and right anomalous thermal difference maps, anomalous temperature rise areas were extracted, and left and right candidate thermal area sets were formed through area screening, time persistence screening, area thermal difference concentration screening, and area morphological integrity screening. For each candidate hot zone, the edge temperature gradient width, edge temperature gradient amplitude, edge sharpness, edge drift, and contour irregularity are further extracted. Based on multi-dimensional feature fusion discrimination logic, flame authenticity is determined. Candidate hot zones that pass the discrimination are confirmed as the left and right target flame zones, thus forming a set of left and right target flame zones. For the confirmed sets of left and right target flame zones, region-level matching is performed based on regional positional relationships, regional extent relationships, thermal distribution similarity relationships, and binocular imaging constraints to construct corresponding flame zone pairs. The spatial position of the flame is determined based on the calibration parameters of the binocular thermal imaging device and the positional differences between the left and right target flame zones. Finally, the active recognition result is output based on the flame spatial position and target flame zone status. To verify the technical advantages of the present invention, the method of the present invention and five comparison schemes were set up under the same test scenario, the same sample conditions and the same collection time. Among them, comparison scheme A is the absolute temperature threshold method, comparison scheme B is the background thermal model method only, comparison scheme C is the edge ambiguity discrimination method only, comparison scheme D is the conventional binocular thermal imaging method, and comparison scheme E is the monocular thermal imaging threshold method.The test sample consisted of real flame samples and non-flame thermal interference samples. Real flame samples were used to examine flame recognition accuracy, false alarm rate, average positioning error, and response time. Non-flame thermal interference samples were used to examine false alarm rate, effective alarm rate under complex thermal backgrounds, and continuous monitoring stability. The test results for each scheme under the same conditions are as follows:

[0068] Table 1. Verification Data of Active Flame Recognition and Localization Effect in Example 2 In this embodiment, as shown in Table 1, the method of the present invention outperforms the comparative schemes in key indicators such as flame recognition accuracy, false alarm rate, false alarm rate, average positioning error, effective alarm rate under complex thermal backgrounds, and continuous monitoring stability. This indicates that the present invention does not only perform local optimization in a single processing step, but forms a continuous, synergistic, and mutually supportive technical chain at three levels: environmental thermal background self-learning, flame authenticity judgment, and binocular positioning. First, in terms of flame recognition accuracy, the method of the present invention achieves 95.7%, which is significantly higher than the 84.2% of the absolute temperature threshold method, the 82.8% of the conventional binocular thermal imaging method, and the 78.6% of the monocular thermal imaging threshold method. It is also higher than the 89.4% of the method using only the background thermal model and the 87.9% of the method using only the edge ambiguity judgment. This result shows that although relying solely on thermal background processing can reduce some interference caused by stable heat sources, it is still insufficient to accurately identify real flames; although relying solely on edge ambiguity judgment can identify some flame boundary features, it is still significantly affected in the presence of complex thermal backgrounds. Only by combining the left and right anomalous thermal difference maps generated by the left and right background thermal models with the edge ambiguity feature discrimination of the left and right candidate hot zone sets can the system truly focus on regional targets that "have anomalous temperature rise relative to the background and possess the essential attributes of flame boundaries." Secondly, regarding the false alarm rate, the method of this invention has a false alarm rate of 4.8%, significantly lower than the 17.6% of the absolute temperature threshold method, the 14.9% of the conventional binocular thermal imaging method, and the 21.4% of the monocular thermal imaging threshold method. It is also significantly lower than the 9.7% of the method using only the background thermal model and the 11.3% of the method using only edge ambiguity discrimination. This result indicates that the solution of this invention has significant advantages in suppressing non-flame interference such as residual heat from distribution cabinets, heat flow from ventilation openings, heat reflection areas, and heated areas of metal supports. More importantly, this advantage is not due to a single temperature threshold adjustment, but rather to first weaken the response of stable heat sources through environmental thermal background self-learning, and then further eliminate sharp and stable non-flame hot zones through flame authenticity discrimination, thus exhibiting stronger targeting and stability. Furthermore, regarding the average positioning error, the method of this invention achieves 0.36 meters, which is superior to 0.58 meters using only the background thermal model method, 0.67 meters using only the edge ambiguity discrimination method, and 0.63 meters using the conventional binocular thermal imaging method. It is also significantly superior to 1.47 meters using the monocular thermal imaging threshold method. This indicates that the present invention not only improves the accuracy of flame recognition but also improves the quality of the input objects entering the binocular matching stage. Since the matching and positioning are based on the set of left and right target flame regions confirmed in step S4, rather than all unfiltered abnormal thermal targets, the construction of corresponding flame region pairs is more reliable, thereby reducing the calculation error of the flame's spatial position.Furthermore, the effective alarm rate under complex thermal backgrounds reached 93.8%, and the continuous monitoring stability reached 96.1%, both higher than the comparative schemes. This indicates that the present invention can output active identification results more stably under long-term continuous monitoring conditions, rather than only performing well at individual moments. Although the average identification response time of the present invention is 1.42 seconds, slightly higher than the absolute temperature threshold method and the monocular thermal imaging threshold method, the increase is small, and it results in a significant reduction in false alarm rate, false negative rate, and positioning error. This shows that the present invention achieves higher identification accuracy and positioning reliability with limited processing cost while maintaining engineering feasibility. In summary, the data in Table 1 demonstrates that the present invention exhibits beneficial effects compared to the prior art in at least the following aspects: First, it significantly reduces the false alarm rate under complex thermal backgrounds; second, it improves the ability to distinguish between real flames and non-flame heat sources; third, it improves the purity of the binocular matching input object, thereby reducing the error in calculating the spatial position of the flame; and fourth, it improves the effective alarm rate and stability in continuous monitoring scenarios. Therefore, this embodiment can fully demonstrate the innovation, rationality, and technical advantages of the present invention in active flame identification and positioning under complex thermal environments.

[0069] The technical features of this invention not described can be implemented by or using existing technology, and will not be repeated here. Of course, the above description is not a limitation of this invention, and this invention is not limited to the examples above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of this invention should also be within the protection scope of this invention.

Claims

1. An active flame identification and localization method, characterized in that, The process includes the following steps: S1. Simultaneously acquire the left and right thermal images of the monitored scene using a binocular thermal imaging device, and preprocess the left and right thermal images to obtain the preprocessed left current thermal frame and right current thermal frame. S2. Based on historical hot frames, establish and update the left background hot model and the right background hot model respectively. Compare the left current hot frame with the left background hot model to obtain the left abnormal hot difference map. Compare the right current hot frame with the right background hot model to obtain the right abnormal hot difference map. S3. Extract the abnormal heating regions based on the left and right abnormal thermal difference maps to obtain a left candidate hot zone set and a right candidate hot zone set; S4. Extract edge ambiguity features from each candidate hot zone in the left and right candidate hot zone sets, and perform flame authenticity discrimination based on the edge ambiguity features. Confirm the candidate hot zones that pass the discrimination as target flame zones to obtain a left target flame zone set and a right target flame zone set; S5. Perform binocular matching on the left and right target flame zone sets to obtain corresponding flame zone pairs, and calculate the flame spatial position based on the corresponding flame zone pairs; S6. Output an active recognition result containing flame confirmation information and three-dimensional positioning information based on the flame spatial position and the target flame zone status.

2. The active flame recognition and localization method according to claim 1, characterized in that, In step S1, the preprocessing of the left and right thermal images includes one or more of the following: bad pixel correction, thermal noise suppression, non-uniformity correction, temperature normalization, local contrast enhancement, and time synchronization verification.

3. The active flame identification and localization method according to claim 1, characterized in that, In step S2, the historical thermal frames are the left thermal frame sequence and the right thermal frame sequence obtained by the binocular thermal imaging device before the current time through continuous acquisition and preprocessing. The left background thermal model and the right background thermal model are established based on the left thermal frame sequence and the right thermal frame sequence, respectively, and the left background thermal model and the right background thermal model are continuously updated as a new frame of thermal image data arrives.

4. The active flame identification and localization method according to claim 3, characterized in that, In step S2, the left current hot frame is compared with the left background thermal model position by position to obtain the thermal anomaly distribution of the left current hot frame relative to the left background thermal model, and a left anomaly thermal difference map is generated. The current right hot frame is compared position by position with the right background thermal model to obtain the thermal anomaly distribution of the current right hot frame relative to the right background thermal model, and a right anomaly thermal difference map is generated.

5. The active flame identification and localization method according to claim 1, characterized in that, In step S3, when extracting abnormal heating regions, threshold filtering, connected region extraction, and region validity screening are performed on the left and right abnormal thermal difference maps respectively. The region validity screening includes one or more of the following: region area screening, time persistence screening, region thermal difference concentration screening, and region morphological integrity screening. The abnormal heating regions after screening form the left candidate thermal region set and the right candidate thermal region set respectively.

6. The active flame identification and localization method according to claim 1, characterized in that, In step S4, the edge blurring features include two or more of the following: edge temperature gradient width, edge temperature gradient magnitude, edge sharpness, edge drift, and contour irregularity. Specifically, the edge temperature gradient width characterizes the spatial change width of the candidate hot area as it transitions from an inner high-temperature region to an outer background region; the edge temperature gradient magnitude characterizes the steepness of the temperature change at the boundary of the candidate hot area; the edge sharpness characterizes whether the boundary of the candidate hot area exhibits a clear and sharp contour; the edge drift characterizes the degree of positional change of the candidate hot area edge in consecutive hot frames; and the contour irregularity characterizes the randomness and complexity of the candidate hot area's shape boundary.

7. The active flame identification and localization method according to claim 6, characterized in that, In step S4, when determining the authenticity of a flame based on the edge ambiguity features, a multi-dimensional feature fusion discrimination logic is constructed based on the edge ambiguity features. Specifically, the edge temperature gradient width, edge drift, and contour irregularity are treated as positively correlated features, while the edge temperature gradient amplitude and edge sharpness are treated as negatively correlated features. All edge ambiguity features are then uniformly scaled and comprehensively evaluated to form a flame confidence score. When the flame confidence score reaches or exceeds a preset authenticity confirmation threshold, the corresponding candidate hot zone is confirmed as the target flame zone.

8. The active flame identification and localization method according to claim 1, characterized in that, In step S5, when performing binocular matching on the left and right target flame area sets, the target flame areas in the left and right target flame area sets are matched at the regional level by comprehensively utilizing regional positional relationships, regional range relationships, thermal distribution similarity relationships, and binocular imaging constraints. Specifically, for each left target flame area in the left target flame area set, one or more candidate corresponding areas are determined in the right target flame area set based on the binocular imaging constraints. Then, the candidate corresponding areas are optimized by combining regional positional relationships, regional range relationships, and thermal distribution similarity relationships to determine the right target flame area that matches the left target flame area, and the two are constructed as a corresponding flame area pair. After obtaining the corresponding flame area pair, the representative position of the region used for positioning is determined from the corresponding flame area pair, and the spatial depth information and lateral and longitudinal position information of the target flame are determined based on the calibration parameters of the binocular thermal imaging device and the positional differences between the left and right target flame areas, thereby obtaining the spatial position of the flame.

9. The active flame identification and localization method according to claim 1, characterized in that, In step S6, the active identification result includes at least flame confirmation information and three-dimensional positioning information. The target flame zone state includes at least one of the following: the existence state of the target flame zone, the area change state, the boundary change state, the location change state, and the continuous existence state. Based on the flame spatial location and the target flame zone state, the target flame is comprehensively confirmed, and the active identification result is output.

Images