An early flame feature recognition method and system based on AI video analysis

Abstract

The application discloses an early flame feature recognition method and system based on AI video analysis, relates to the technical field of flame feature recognition, and comprises the following steps: acquiring a continuous video frame sequence collected by a monitoring camera device, and acquiring a suspected flame candidate region set sequence; performing linkage flame region prediction; performing semantic uncertainty analysis to obtain a semantic uncertainty coefficient set; performing spatiotemporal consistency constraint enhancement on the combined suspected flame candidate region set to obtain an enhanced video frame sequence; and performing early flame feature recognition to obtain a target recognition result. The application solves the technical problems that early weak flame features are easily ignored, flame and light reflection scenes are easily confused, and flame recognition accuracy is low and stability is poor in a complex environment in the prior art, and achieves the technical effects of realizing accurate recognition of early weak flames and easily confused scenes and improving the accuracy and stability of early flame feature recognition in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of flame feature recognition technology, and specifically to an early flame feature recognition method and system based on AI video analysis. Background Technology

[0002] Early flame recognition played a crucial role in security monitoring and industrial production. Traditional point-sampling sensors relied on indirect information such as smoke particles and temperature for detection, which resulted in high response delays, susceptibility to environmental interference, limited applicability, and high costs. Furthermore, they could not record the spatiotemporal process of flame development. While existing video flame detection technologies have mitigated some of these shortcomings through multi-feature fusion and moving target tracking, they still face significant bottlenecks: early, weak flame signals are easily masked by noise and have low distinguishability from light reflections and sunlight, leading to insufficient recognition accuracy. Simultaneously, changes in lighting and scene interference in complex environments can easily trigger false alarms, and the algorithms' insufficient utilization of spatiotemporal correlations makes it difficult to meet practical requirements in terms of recognition stability and anti-interference capabilities.

[0003] Existing technologies suffer from problems such as the easy neglect of early weak flame features, the easy confusion between flame and light reflection, and low accuracy and poor stability of flame recognition in complex environments. Summary of the Invention

[0004] This application provides an early flame feature recognition method and system based on AI video analysis, which addresses the technical problems in the prior art where early weak flame features are easily ignored, flames and light reflections are easily confused, and flame recognition accuracy and stability are low in complex environments.

[0005] In view of the above problems, this application provides an early flame feature recognition method and system based on AI video analysis.

[0006] The first aspect of this application provides an early flame feature recognition method based on AI video analysis, the method comprising:

[0007] A continuous sequence of video frames captured by a surveillance camera is acquired. An AI video analyzer is used to analyze each video frame from three dimensions: color features, brightness variation features, and spatial connectivity, to obtain a sequence of suspected flame candidate regions. The sequence of suspected flame candidate regions is then traversed to perform linked flame region prediction, resulting in a joint suspected flame candidate region set and a semantic prediction probability set. Semantic uncertainty analysis is performed on the joint suspected flame candidate region set based on the semantic prediction probability set, yielding a semantic uncertainty coefficient set. Based on the semantic uncertainty coefficient set, spatiotemporal consistency constraints are enhanced on the joint suspected flame candidate region set in the video frame sequence to obtain an enhanced video frame sequence. Finally, a flame feature recognizer is used to perform early flame feature recognition on the enhanced video frame sequence to obtain the target recognition result.

[0008] A second aspect of this application provides an early flame feature recognition system based on AI video analysis, the system comprising:

[0009] The video frame analysis module acquires a continuous sequence of video frames captured by a surveillance camera. It uses an AI video analyzer to analyze each video frame from three dimensions: color features, brightness variation features, and spatial connectivity, to obtain a set of suspected flame candidate regions. The linked flame region prediction module iterates through the set of suspected flame candidate regions to predict linked flame regions, obtaining a joint set of suspected flame candidate regions and a semantic-side prediction probability set. The semantic uncertainty analysis module performs semantic uncertainty analysis on the joint set of suspected flame candidate regions based on the semantic-side prediction probability set, obtaining a set of semantic uncertainty coefficients. The constraint enhancement module enhances the spatiotemporal consistency of the joint set of suspected flame candidate regions in the video frame sequence based on the set of semantic uncertainty coefficients, obtaining an enhanced video frame sequence. The target recognition result acquisition module uses a flame feature recognizer to perform early flame feature recognition on the enhanced video frame sequence, obtaining the target recognition result.

[0010] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0011] A continuous video frame sequence captured by a surveillance camera is acquired to obtain a set of suspected flame candidate regions. The sequence of suspected flame candidate regions is then traversed to perform linked flame region prediction, resulting in a joint set of suspected flame candidate regions and a set of semantic prediction probabilities. Semantic uncertainty analysis is performed to obtain a set of semantic uncertainty coefficients. Spatiotemporal consistency constraints are applied to the joint set of suspected flame candidate regions in the video frame sequence to enhance its consistency, resulting in an enhanced video frame sequence. An early flame feature recognition function is then used to identify the early flame features in the enhanced video frame sequence, yielding the target recognition result. This method achieves accurate identification of early, weak flames and easily confused scenes, improving the accuracy and stability of early flame feature recognition in complex environments. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A schematic diagram of an early flame feature recognition method based on AI video analysis provided for an embodiment of this application;

[0014] Figure 2 This is a schematic diagram of an early flame feature recognition system based on AI video analysis, provided as an embodiment of this application.

[0015] Figure labeling: Video frame analysis module 10, linked flame area prediction module 20, semantic uncertainty analysis module 30, constraint enhancement module 40, target recognition result acquisition module 50. Detailed Implementation

[0016] This application provides an early flame feature recognition method and system based on AI video analysis, which addresses the technical problems in the prior art where early weak flame features are easily ignored, flames and light reflections are easily confused, and flame recognition accuracy and stability are low in complex environments.

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0018] Example 1, as Figure 1As shown, this application provides an early flame feature recognition method based on AI video analysis, the method comprising:

[0019] Step S100: Obtain a continuous video frame sequence captured by the monitoring camera device, and use an AI video analyzer to analyze each video frame from three dimensions: color features, brightness change features, and spatial connectivity, to obtain a set sequence of suspected flame candidate regions.

[0020] Specifically, the process first acquires a continuous sequence of video frames captured by surveillance cameras and performs tensor transformation on them to generate a network input tensor sequence containing a time dimension. This tensor sequence is then fed into the multi-layer coding structure of the AI ​​video analyzer. Color information is encoded sequentially through a low-level feature coding layer to learn the flame color distribution characteristics and output a low-level feature map sequence. This low-level feature map sequence is then passed to a middle-level temporal feature coding layer, where convolutional operations along the time dimension are used to jointly model the features of adjacent video frames, mining the dynamic features of brightness changes over time and obtaining a middle-level feature map sequence. Finally, the middle-level feature map sequence... The data is fed into a high-level spatial structure encoding layer, where a convolutional structure with progressively expanding receptive fields is used to construct the spatial distribution of candidate regions, outputting a high-level feature map sequence. Then, the low, medium, and high-level feature map sequences are downsampled to achieve spatial scale alignment. The aligned three types of feature map sequences are then weighted and fused according to preset weights to generate a comprehensive flame response feature map sequence. This sequence is then nonlinearly mapped to obtain a flame response probability map sequence. Binarization is performed based on a preset threshold to obtain a candidate region mask map sequence. Finally, through region extraction, a set of suspected flame candidate regions containing suspected flame regions is obtained.

[0021] Step S200: Traverse the sequence of suspected flame candidate regions to perform linked flame region prediction, and obtain a joint suspected flame candidate region set and a semantic side prediction probability set.

[0022] Specifically, based on the acquired sequence of suspected flame candidate regions, a joint temporal adjacency analysis is conducted to predict linked flame regions. First, the first set of suspected flame candidate regions is extracted from this sequence. Then, a first suspected flame candidate region is randomly selected from this set without replacement. This first suspected flame candidate region is then combined with the second set of suspected flame candidate regions using two dimensions: spatial intersection-union ratio (IUU) and feature similarity. If the maximum joint similarity score between this first suspected flame candidate region and the second set is greater than or equal to a preset threshold, the corresponding suspected flame candidate region is included in the neighborhood set of the first joint suspected flame candidate region. Simultaneously, the first suspected flame candidate region is added to the joint suspected flame candidate region set, and this process continues to be performed with suspected flame candidate regions in subsequent video frames. The above joint similarity analysis and threshold judgment are repeated for the selected region set until the maximum joint similarity is less than the preset threshold. If the initial maximum joint similarity does not reach the threshold, the next suspected flame candidate region is randomly extracted from the first suspected flame candidate region set without replacement, and the above process is repeated. After all regions in the first set have been analyzed, the process jumps to the second suspected flame candidate region set, and so on until the end of the sequence, finally obtaining the joint suspected flame candidate region set and the corresponding joint suspected flame candidate region neighborhood set. Then, the mean joint similarity between each region in the joint suspected flame candidate region set and the regions in its corresponding neighborhood set is calculated, thereby forming a semantic side prediction probability set that can characterize the confidence of the flame attribute of the region.

[0023] Step S300: Perform semantic uncertainty analysis on the joint suspected flame candidate region set based on the semantic side prediction probability set to obtain the semantic uncertainty coefficient set.

[0024] Specifically, for the obtained semantic-side predicted probability set and joint suspected flame candidate region set, the sum of all predicted probabilities in the semantic-side predicted probability set is first calculated to obtain the semantic-side predicted probability sum. Then, based on this semantic-side predicted probability sum, each predicted probability in the semantic-side predicted probability set is normalized to eliminate the dimensional differences between different probability values ​​and ensure that the processed data is on a uniform scale. Finally, the difference between each normalized predicted probability and 1 is used as the semantic uncertainty coefficient of the corresponding joint suspected flame candidate region. This quantifies the degree of hesitation in classifying flames with similar targets such as light reflections and welding sparks, ultimately forming a semantic uncertainty coefficient set that can accurately reflect the semantic ambiguity of each joint suspected flame candidate region.

[0025] Step S400: Based on the set of semantic uncertainty coefficients, perform spatiotemporal consistency constraint enhancement on the set of joint suspected flame candidate regions in the video frame sequence to obtain an enhanced video frame sequence.

[0026] Specifically, based on the obtained set of semantic uncertainty coefficients, a preset maximum semantic uncertainty coefficient threshold is first obtained. This threshold is then used as a screening criterion to verify each semantic uncertainty coefficient in the set, selecting qualified semantic uncertainty coefficient sets that meet the threshold requirements. Subsequently, through the mapping relationship between coefficients and regions, a set of qualified joint suspected flame candidate regions is selected from the obtained set of joint suspected flame candidate regions, eliminating regions with semantic ambiguity and high risk of misjudgment. Next, the difference between 1 and each coefficient in the qualified semantic uncertainty coefficient set is used as an enhancement weight to form an enhancement weight set, where regions with lower semantic uncertainty correspond to larger enhancement weights. Finally, combining this enhancement weight set, the temporal correlation and spatial continuity of adjacent frames in the video frame sequence are used to enhance the spatiotemporal consistency of the selected joint suspected flame candidate region set, compensating for the impact of geometric uncertainties such as flame boundary jitter and temporal feature drift, strengthening the feature expression of effective regions, and ultimately generating an enhanced video frame sequence with clearer features and stronger stability.

[0027] Step S500: Use a flame feature recognizer to perform early flame feature recognition on the enhanced video frame sequence to obtain target recognition results.

[0028] Specifically, the enhanced video frame sequence is input into a flame feature recognizer. This recognizer targets the enhanced features of suspected flame areas in the enhanced video frames, and combines the core attributes of early flames in terms of color distribution, dynamic changes in brightness, and spatial connectivity to accurately distinguish flames from easily confused targets such as light reflections and welding sparks. This effectively solves the problem of category judgment hesitation caused by semantic uncertainty, and finally outputs target recognition results that can accurately locate the early flames and clarify the flame attributes, providing a reliable basis for fire early warning.

[0029] In one possible implementation, step S100 further includes:

[0030] Step S110: Obtain a continuous video frame sequence captured by the monitoring camera device, and perform tensor transformation to obtain a network input tensor sequence containing the time dimension.

[0031] Step S120: The network input tensor sequence is fed into the low-level feature encoding layer of the AI ​​video analyzer to encode the color information, learn the flame color distribution features, and obtain a low-level feature map sequence.

[0032] Step S130: The low-level feature map sequence is fed into the middle temporal feature coding layer of the AI ​​video analyzer. The features of adjacent video frames are jointly modeled by convolution operations unfolded along the time dimension to learn the dynamic features of brightness changes over time and obtain the middle-level feature map sequence.

[0033] Step S140: The middle-layer feature map sequence is fed into the high-level spatial structure coding layer of the AI ​​video analyzer. The spatial distribution of candidate regions is constructed through a convolutional structure with progressively expanding receptive fields to obtain the high-level feature map sequence.

[0034] Step S150: Map and fuse the low-level feature map sequence, the middle-level feature map sequence, and the high-level feature map sequence respectively to obtain a set sequence of suspected flame candidate regions.

[0035] Specifically, the system first uses cameras deployed in the monitoring scene to capture continuous sequences of raw video frames in real time. This sequence is then standardized to unify frame resolution and color channels; color videos retain three color channels, while grayscale videos retain one. Next, a tensor transformation operation is performed to convert each frame of two-dimensional image data into a tensor format suitable for neural network input, while incorporating temporal information. This results in a 5D network input tensor sequence with dimensions Batch_Size, Time_Step, Height, Width, and Channel. Batch_Size represents the number of frames processed in a batch, Time_Step represents the time step of consecutive video frames, Height and Width correspond to the height and width of the video frame, and Channel represents the number of color channels. This tensor sequence fully preserves the spatiotemporal characteristics of the video.

[0036] The obtained 5D network input tensor sequence containing the time dimension is fed into the low-level feature encoding layer of the AI ​​video analyzer. This encoding layer adopts a pre-trained and optimized lightweight convolutional neural network structure. First, the color information in the tensor sequence is converted from RGB to HSV and YCrCb dual color spaces to accurately lock the color spectrum range unique to flames. Then, the converted color features are encoded in multiple channels through 3×3 small-size convolutional kernels. Batch normalization and ReLU activation functions are combined to suppress noise interference and enhance effective feature responses, while preserving local details of the flame color distribution. Finally, after convolutional sliding with a stride of 1 and feature aggregation, the output is a low-level feature map sequence with the same dimension as the input tensor time step and focusing on the flame color attributes.

[0037] The output low-level feature map sequence focusing on the color attributes of the flame is fed into the mid-level temporal feature encoding layer of the AI ​​video analyzer. This encoding layer adopts a temporal convolutional network (TCN) architecture, which integrates causal convolution and dilated convolution modules at its core. Causal convolution ensures that feature modeling relies only on historical and current frame information, avoiding future information leakage. At the same time, dilated convolution exponentially expands the receptive field to accurately capture long-term brightness change patterns. First, the brightness-related features in the low-level feature map sequence are separated and enhanced. Then, a 3×3×3 spatiotemporal convolution kernel unfolded along the time dimension is used to jointly model the brightness features of adjacent video frames. Batch normalization and ReLU activation functions are combined to suppress noise and enhance effective dynamic feature responses. The focus is on learning the dynamic features of brightness changes over time, such as irregular flickering and gradual enhancement, which are unique to early flames. Finally, after feature aggregation and dimension normalization, a mid-level feature map sequence with the same time step as the input sequence and focusing on the temporal dynamic attributes of the flame is output.

[0038] The output sequence of mid-level feature maps focusing on the temporal dynamic attributes of flames is fed into the high-level spatial structure coding layer of the AI ​​video analyzer. This coding layer employs a multi-scale convolutional structure with progressively expanding receptive fields. By setting convolutional kernels of different sizes, such as 3×3, 5×5, 7×7, and progressive convolutional levels, it gradually captures the spatial correlation information of candidate regions from local to global. First, the spatial features in the mid-level feature map sequence are enhanced, focusing on mining the spatial connectivity, contour morphology, and pixel distribution patterns of the flame region. Then, hierarchical convolutional operations are used to gradually expand the receptive field, accurately covering suspected flame regions of different sizes. At the same time, pooling operations are combined to retain key spatial features and compress redundant information. Finally, through feature integration and channel fusion, a high-level feature map sequence with the same time step as the input sequence and focusing on the spatial structural attributes of the flame is output.

[0039] For the obtained low-level feature map sequences focusing on flame color attributes, mid-level feature map sequences focusing on temporal dynamic attributes, and high-level feature map sequences focusing on spatial structure attributes, a downsampling operation is first used to spatially align the three types of feature map sequences to eliminate the resolution differences between different levels of feature maps, resulting in aligned low-level, aligned mid-level, and aligned high-level feature map sequences with uniform dimensions. Then, according to preset feature importance weights, such as color features, temporal features, and spatial features, weights are respectively assigned to the corresponding adaptation weights, and weighted fusion calculations are performed on the three types of aligned feature map sequences to aggregate multi-dimensional effective information and generate a comprehensive flame response feature map sequence. Subsequently, nonlinear mapping processing is performed on the comprehensive flame response feature map sequence to convert feature values ​​into flame existence probability values, resulting in a flame response probability map sequence. Finally, the flame response probability map sequence is binarized according to a preset probability threshold, and regions with flame response probabilities higher than the threshold are selected to form a candidate region mask map sequence. Suspected flame regions are separated from the mask map using a region extraction algorithm, ultimately obtaining a set sequence of suspected flame candidate regions containing suspected flame regions from each video frame.

[0040] In one possible implementation, step S150 further includes:

[0041] Step S151: Perform spatial scale alignment on the low-level feature map sequence, the middle-level feature map sequence, and the high-level feature map sequence by downsampling, respectively, to obtain aligned low-level feature map sequence, aligned middle-level feature map sequence, and aligned high-level feature map sequence.

[0042] Step S152: The aligned low-level feature map sequence, aligned middle-level feature map sequence, and aligned high-level feature map sequence are weighted according to preset weights to obtain a comprehensive flame response feature map sequence.

[0043] Step S153: Perform nonlinear mapping on the obtained comprehensive flame response feature map sequence to generate a flame response probability map sequence, and perform binarization processing on the flame response probability map sequence according to a preset threshold to obtain a candidate region mask map sequence, perform region extraction, and obtain the suspected flame candidate region set sequence.

[0044] Specifically, for the obtained low-level feature map sequences focusing on flame color attributes, the obtained mid-level feature map sequences focusing on temporal dynamic attributes, and the obtained high-level feature map sequences focusing on spatial structure attributes, the three types of feature map sequences have inconsistent spatial resolutions due to differences in convolution operations, receptive field expansion, and feature extraction logic at different coding levels. Therefore, max pooling or average pooling downsampling algorithms are used to unify them. By setting the same target spatial scale parameters, i.e., uniform width and height dimensions, downsampling operations are performed on each feature map in the low-level, mid-level, and high-level feature map sequences respectively. While preserving the core feature information, the spatial resolution of the three types of feature map sequences is adjusted to be completely consistent, and finally, aligned low-level feature map sequences, aligned mid-level feature map sequences, and aligned high-level feature map sequences are obtained that are spatially scale aligned and can be directly fused into subsequent operations.

[0045] After completing spatial scale alignment, for the aligned low-level feature map sequence, mid-level feature map sequence, and high-level feature map sequence, differentiated preset weights are pre-configured based on the core needs and feature importance of early flame recognition. For example, the weight ratio of color features, temporal dynamic features, and spatial structure features is set according to scene requirements to ensure that key features play a leading role in fusion. Subsequently, a pixel-by-pixel weighted summation fusion strategy is adopted to multiply the pixel feature values ​​at corresponding positions in the three types of aligned feature map sequences by their preset weights and then perform cumulative calculations to deeply aggregate effective information in multiple dimensions such as color, temporal sequence, and space, and finally generate a comprehensive flame response feature map sequence that can comprehensively and accurately reflect the intensity of flame feature response.

[0046] For the obtained comprehensive flame response feature map sequence, a Sigmoid nonlinear activation function is used for mapping processing, transforming the feature response value of each pixel in the feature map into a flame presence probability value in the 0-1 range, generating a flame response probability map sequence that can intuitively reflect the degree of flame suspicion in each region. Subsequently, based on the accuracy requirements of early flame recognition, a preset probability threshold is set, and each probability map in the flame response probability map sequence is binarized. Pixels with probability values ​​higher than the preset threshold are marked as foreground, i.e., suspected flame regions, and their pixel values ​​are set to 1. Pixels with probability values ​​lower than or equal to the preset threshold are marked as background, i.e., non-flame regions, and their pixel values ​​are set to 0, resulting in a candidate region mask map sequence that retains only the outline of suspected flame regions. Finally, a connected component extraction algorithm is used to filter and calibrate the boundaries of the foreground regions in the mask map sequence, eliminating noise regions with too small an area, and extracting complete and independent suspected flame regions. The position and boundary range of each suspected flame region in the original video frame are determined by coordinate transformation, and finally, a suspected flame candidate region set sequence containing all effective suspected flame regions in each video frame is obtained.

[0047] In one possible implementation, step S200 further includes:

[0048] Step S210: Perform regional temporal adjacency joint analysis based on the suspected flame candidate region set sequence to obtain the joint suspected flame candidate region set and the joint suspected flame candidate region neighborhood set.

[0049] Step S220: Calculate the mean joint similarity between each joint suspected flame candidate region in the joint suspected flame candidate region set and the corresponding joint suspected flame candidate region neighborhood in the joint suspected flame candidate region neighborhood set, and obtain the semantic side prediction probability set.

[0050] Specifically, based on the acquired sequence of suspected flame candidate regions, a joint analysis process of temporal adjacency is initiated. First, the first set of suspected flame candidate regions is extracted from this sequence, and a first suspected flame candidate region is randomly selected from it without replacement. Then, this first suspected flame candidate region is matched with a second set of suspected flame candidate regions, and the spatial intersection-union ratio (IUU) is calculated to characterize the degree of overlap and feature similarity of the regions, reflecting the matching degree of color, temporal, and spatial features. It is also determined whether the maximum joint similarity score between the region and the second set reaches a preset threshold. If the threshold is reached, the corresponding suspected flame candidate region is included in the neighborhood set of the first joint suspected flame candidate region, and the first suspected flame candidate region is... The flame candidate region is added to the joint suspected flame candidate region set, and the joint similarity analysis and threshold judgment are repeated with the third and subsequent suspected flame candidate region sets until the maximum joint similarity value is lower than the preset threshold. If the initial joint similarity does not reach the threshold, the next suspected flame candidate region is extracted from the first set without replacement and the analysis process is re-executed. After all regions in the first set have been analyzed, the process jumps to the second suspected flame candidate region set as the starting set, and the entire sequence is traversed in the same way. Finally, the joint suspected flame candidate region set with continuous temporal correlation and the joint suspected flame candidate region neighborhood set composed of the neighborhood regions corresponding to each joint region are selected.

[0051] For the obtained set of joint suspected flame candidate regions and the corresponding set of joint suspected flame candidate region neighborhoods, a one-to-one correspondence between each joint suspected flame candidate region and its neighborhood set is first established to ensure that each joint region can be matched with its associated neighborhood region in the temporally adjacent video frames. Then, for each joint suspected flame candidate region, the spatial intersection-union ratio and feature similarity with each joint suspected flame candidate region in its corresponding neighborhood set are extracted one by one, and the joint similarity of a single neighborhood is obtained by weighted calculation. Then, the joint similarity values ​​of all corresponding neighborhoods of the joint region are counted, and their arithmetic mean is calculated. This mean is used as the semantic prediction probability of the joint suspected flame candidate region to quantify its confidence that it belongs to a real flame region. After the similarity mean of all regions in the set of joint suspected flame candidate regions is calculated in sequence, a set of semantic prediction probabilities corresponding one-to-one with the joint regions is finally formed.

[0052] In one possible implementation, step S210 further includes:

[0053] Step S211: Extract the first suspected flame candidate region set from the sequence of suspected flame candidate regions, and randomly extract the first suspected flame candidate region from the first suspected flame candidate region set without replacement.

[0054] Step S212: Perform joint similarity analysis on the first suspected flame candidate region and the second suspected flame candidate region set from two dimensions: spatial intersection-union ratio and feature similarity. Determine whether the maximum joint similarity value between the first and second suspected flame candidate regions is greater than or equal to a preset threshold. If not, randomly extract the second suspected flame candidate region from the first suspected flame candidate region set without replacement and repeat the joint similarity analysis and threshold judgment until all suspected flame candidate regions in the first suspected flame candidate region set have been analyzed. Then, jump to the second suspected flame candidate region set in the suspected flame candidate region set sequence, and so on, until the end of the sequence is reached, to obtain the joint suspected flame candidate region set and the joint suspected flame candidate region neighborhood set.

[0055] Specifically, the temporal order of the output sequence of suspected flame candidate regions is first determined, and the first suspected flame candidate region set at the beginning of the sequence is accurately extracted. This set covers all the preliminary suspected flame regions obtained after multi-dimensional feature fusion and filtering in the current corresponding video frame. Each region contains location coordinates, boundary range and core feature information. Then, a random extraction method without replacement is used to randomly select a suspected flame region from this first suspected flame candidate region set as the first suspected flame candidate region. This ensures that all suspected flame regions in the set can be traversed one by one in the subsequent analysis process, avoiding repeated analysis or omissions, and laying the foundation for the joint similarity analysis with candidate regions in adjacent frames.

[0056] Taking the extracted first suspected flame candidate region as the analysis object, a joint similarity analysis is performed between it and the second suspected flame candidate region set in the suspected flame candidate region set sequence. The spatial intersection-union ratio (IUU) is calculated simultaneously to quantify the degree of positional overlap between the first suspected flame candidate region and each region in the second set, reflecting spatial correlation and feature similarity. The core feature matching degree of the two in color distribution, temporal dynamics, and spatial structure is compared to reflect attribute consistency. The joint similarity of each compared region is obtained through weighted fusion. Then, it is determined whether the maximum joint similarity between all regions in the second set and the first suspected flame candidate region is greater than or equal to a preset threshold. If the threshold is not reached, the analysis continues from the first set. From a set of suspected flame candidate regions, the next region is randomly extracted without replacement as the second suspected flame candidate region. The joint analysis of spatial intersection-union ratio and feature similarity and the threshold judgment process are repeated. This operation continues until all suspected flame candidate regions in the first set have been analyzed. Then, the process jumps to the second set of suspected flame candidate regions in the sequence as the new starting set. All subsequent candidate region sets are traversed in the same way until the entire sequence of suspected flame candidate regions is covered. Finally, a joint set of suspected flame candidate regions with continuous temporal correlation is selected, as well as a joint set of suspected flame candidate region neighborhoods composed of temporally adjacent regions corresponding to each joint region.

[0057] In one possible implementation, step S212 further includes:

[0058] If the maximum joint similarity between the second suspected flame candidate region set and the first suspected flame candidate region is greater than or equal to a preset threshold, then the suspected flame candidate region in the second suspected flame candidate region set corresponding to the maximum joint similarity is added to the neighborhood set of the first joint suspected flame candidate region, and the first suspected flame candidate region is added to the joint suspected flame candidate region set. The joint similarity analysis and threshold judgment are repeated between the first suspected flame candidate region and the third suspected flame candidate region set, and so on, until the maximum joint similarity is less than the preset threshold, thus obtaining the first joint suspected flame candidate region and the neighborhood of the first joint suspected flame candidate region.

[0059] Add the first joint suspected flame candidate region and the neighborhood of the first joint suspected flame candidate region to the joint suspected flame candidate region set and the joint suspected flame candidate region neighborhood set.

[0060] Specifically, after performing joint similarity analysis on the first and second sets of suspected flame candidate regions, simultaneously calculating the spatial intersection-union ratio and feature similarity, and fusing them to obtain the joint similarity, if the maximum joint similarity value between the first suspected flame candidate region and the second set satisfies the condition of being greater than or equal to a preset threshold, then the specific suspected flame candidate region in the second set corresponding to the maximum value is first added to the neighborhood set of the first joint suspected flame candidate region, and the first suspected flame candidate region is simultaneously included in the joint suspected flame candidate region set; then, keeping the first suspected flame candidate region unchanged, the same logic of jointing it with the third set of suspected flame candidate regions is repeated. The similarity analysis and threshold judgment process involves recalculating the spatial intersection-union ratio and feature similarity, fusing them into a joint similarity, and determining whether the maximum joint similarity value in the set meets the threshold. If it does, the corresponding region is added to the neighborhood set of the first joint suspected flame candidate region. This process is repeated, iterating through subsequent sets in the suspected flame candidate region set sequence, and continuously performing joint similarity analysis and threshold judgment operations until the maximum joint similarity value of a subsequent set with the first suspected flame candidate region is less than a preset threshold. At this point, the traversal process stops, and the first joint suspected flame candidate region is finally determined, which is the initial first suspected flame candidate region and the neighborhood of the first joint suspected flame candidate region containing all neighborhood regions that meet the threshold during the traversal process.

[0061] After the first joint suspected flame candidate region and its corresponding neighborhood are selected and determined, a set integration operation is performed: the first joint suspected flame candidate region confirmed by multiple rounds of temporal adjacency joint analysis, that is, the core suspected flame region that meets the temporal association condition in the initial traversal, is directly added to the global joint suspected flame candidate region set to ensure that the core region is included in the scope of subsequent semantic uncertainty analysis and constraint enhancement; at the same time, the neighborhood of the first joint suspected flame candidate region, which consists of all the neighborhood regions that meet the joint similarity threshold requirement with the first joint suspected flame candidate region during the traversal, is completely added to the corresponding joint suspected flame candidate region neighborhood set, so that each core joint region can be associated with its temporally adjacent matching neighborhood region, realizing a one-to-one correspondence between the core region and the neighborhood region.

[0062] In one possible implementation, step S300 further includes:

[0063] Step S310: Calculate the sum of the semantic side prediction probability set to obtain the semantic side prediction probability sum.

[0064] Step S320: Normalize the set of semantic side prediction probabilities based on the sum of the semantic side prediction probabilities, and use the difference between the normalization result and 1 as the semantic uncertainty coefficient to obtain the set of semantic uncertainty coefficients.

[0065] Specifically, for the acquired semantic prediction probability set, each element in the set corresponds to the semantic prediction probability of a region in the joint suspected flame candidate region set. The higher the probability value, the higher the confidence that the region is a real flame. By traversing all semantic prediction probability values ​​in the set and performing an accumulation and summation operation, the cumulative result of all probability values ​​is taken as the sum of semantic prediction probabilities. This provides key total reference data for the subsequent normalization processing of the semantic prediction probability set, ensuring the accuracy and rationality of the subsequent semantic uncertainty coefficient calculation.

[0066] Based on the sum of the calculated semantic prediction probabilities, normalization is performed on each probability value in the semantic prediction probability set. This involves dividing each individual semantic prediction probability value by the sum, mapping all processed probability values ​​to the 0-1 range, and ensuring that the sum of all normalized probability values ​​equals 1. This forms a normalized probability set that intuitively reflects the relative confidence of each joint suspected flame candidate region. Following the core logic that higher probabilities indicate a higher likelihood of a flame and lower uncertainty, the difference between each normalization result in the normalized probability set and 1 is calculated. This difference is the semantic uncertainty coefficient of the corresponding joint suspected flame candidate region. A higher normalized probability results in a smaller difference from 1, indicating a lower semantic uncertainty coefficient and a stronger certainty that the region is a real flame; conversely, a smaller difference indicates higher uncertainty. After calculating the coefficients for all joint suspected flame candidate regions, a semantic uncertainty coefficient set corresponding one-to-one with the set of joint suspected flame candidate regions is finally generated.

[0067] In one possible implementation, step S400 further includes:

[0068] Step S410: Obtain the preset maximum semantic uncertainty coefficient threshold, filter the set of semantic uncertainty coefficients, and determine the qualified set of semantic uncertainty coefficients.

[0069] Step S420: Map and filter the set of joint suspected flame candidate regions according to the set of qualified semantic uncertainty coefficients to obtain a filtered set of joint suspected flame candidate regions.

[0070] Step S430: Combining the qualified semantic uncertainty coefficient set, the spatiotemporal consistency constraint of the selected joint suspected flame candidate region set is enhanced according to the adjacent frames of the video frame sequence to obtain the enhanced video frame sequence.

[0071] Specifically, the process first retrieves a pre-defined maximum semantic uncertainty coefficient threshold, which is based on the accuracy requirements of early flame recognition, application scenario characteristics, and actual business needs. This threshold serves as the core criterion for distinguishing the certainty of flame attributes in joint suspected flame candidate regions. Regions with semantic uncertainty coefficients lower than or equal to this threshold have a high degree of confidence in being real flames, while regions with coefficients higher than this threshold have insufficient certainty. Subsequently, based on this pre-defined threshold, the obtained set of semantic uncertainty coefficients is compared and filtered one by one. Each semantic uncertainty coefficient in the set is traversed, and all coefficients with values ​​less than or equal to the pre-defined maximum semantic uncertainty coefficient threshold are retained, while coefficients exceeding the threshold are removed. Finally, a qualified set of semantic uncertainty coefficients containing only high-confidence coefficients is formed.

[0072] Based on the one-to-one correspondence between the set of qualified semantic uncertainty coefficients and the set of joint suspected flame candidate regions (i.e., each semantic uncertainty coefficient uniquely corresponds to a joint suspected flame candidate region), a precise mapping and screening operation is performed. All regions in the set of joint suspected flame candidate regions are traversed, and through the mapping between coefficients and regions, all joint suspected flame candidate regions corresponding to coefficients in the set of qualified semantic uncertainty coefficients are retained, while regions whose corresponding semantic uncertainty coefficients exceed a preset maximum threshold and whose flame attribute certainty is low are removed. Finally, a set of joint suspected flame candidate regions focused on high-confidence regions is formed, ensuring that subsequent spatiotemporal consistency constraint enhancement operations can be specifically applied to core regions with strong flame attribute certainty, improving the efficiency and accuracy of subsequent processing.

[0073] After obtaining the qualified semantic uncertainty coefficient set and the screening set of joint suspected flame candidate regions, the difference between 1 and the qualified semantic uncertainty coefficient is first used as the enhancement weight of each corresponding joint suspected flame candidate region. The lower the coefficient, the higher the certainty that the region is a flame, and the greater the enhancement weight. It is necessary to focus on strengthening the spatiotemporal feature stability of the region and construct an enhancement weight set. Then, combined with the temporal correlation characteristics of adjacent frames in the video frame sequence, spatiotemporal consistency constraint enhancement is carried out on the screening set of joint suspected flame candidate regions. From the time dimension, the continuity of brightness change, flicker frequency stability and position movement trajectory of the same candidate region in adjacent frames are checked, and the temporal feature drift caused by jitter and noise is corrected. From the spatial dimension, the smoothness of the region boundary contour is optimized, the boundary jitter is suppressed, and the spatial connectivity of the region is strengthened. At the same time, the enhancement intensity is adjusted according to the enhancement weight set, and stronger constraint optimization is given to high weight regions. Through the synergistic enhancement of the spatiotemporal dimensions, an enhanced video frame sequence with clearer regional features, more stable spatiotemporal correlation and stronger anti-interference ability is finally generated.

[0074] In one possible implementation, step S430 further includes:

[0075] Step S431: Take the difference between 1 and the qualified semantic uncertainty coefficient in the qualified semantic uncertainty coefficient set as the enhancement weight to obtain the enhancement weight set.

[0076] Step S432: Perform spatiotemporal consistency constraint enhancement on the set of suspected flame candidate regions selected by the screening based on the enhanced weight set and the adjacent frames of the video frame sequence to obtain the enhanced video frame sequence.

[0077] Specifically, for the set of qualified semantic uncertainty coefficients that have been selected, each qualified semantic uncertainty coefficient in the set is traversed, and the calculation is performed one by one according to the calculation rule of enhancement weight = 1 - qualified semantic uncertainty coefficient. The lower the qualified semantic uncertainty coefficient, the higher the certainty that the corresponding joint suspected flame candidate region is a real flame. The larger the difference with 1, the higher the generated enhancement weight, which means that the region needs to obtain a stronger feature enhancement in the subsequent spatiotemporal consistency constraint enhancement. After completing the above difference calculation for all qualified semantic uncertainty coefficients in the set, the enhancement weight set is finally formed that corresponds one-to-one with the set of qualified semantic uncertainty coefficients and the set of selected joint suspected flame candidate regions.

[0078] By combining an enhanced weight set (higher weights correspond to a stronger certainty that the selected joint suspected flame candidate region is a real flame) with the temporal correlation characteristics of adjacent frames in the video frame sequence, spatiotemporal consistency constraints are enhanced for weak flame regions that may exist in the selected joint suspected flame candidate region set—regions with indistinct features and easily overlooked. In the temporal dimension, flame features from corresponding regions in adjacent video frames are fused differently based on the magnitude of the enhanced weights. For regions with higher weights, the fusion strength of temporal features such as the continuity of brightness changes and the stability of flicker frequency in adjacent frames is strengthened, and clear flame features from adjacent frames are superimposed onto the corresponding region of the current video frame. For regions with lower weights, features from adjacent frames are moderately fused to correct temporal drift. In the spatial dimension, based on the positional correlation of regions in adjacent frames, the boundary contour smoothness and spatial connectivity of the candidate region in the current frame are optimized to suppress boundary jitter and noise interference. Through this weight-guided adjacent frame feature fusion and spatiotemporal consistency verification method, the feature representation of weak flames is effectively enhanced, compensating for the deficiencies of single-frame features, reducing the difficulty of subsequent flame feature recognition, and ultimately generating an enhanced video frame sequence with clear features, spatiotemporal stability, and high recognizability.

[0079] Example 2 is based on the same inventive concept as the early flame feature recognition method based on AI video analysis in the foregoing examples, such as... Figure 2 As shown, this application provides an early flame feature recognition system based on AI video analysis. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0080] The video frame analysis module 10 is used to acquire a continuous sequence of video frames collected by the monitoring camera. It uses an AI video analyzer to analyze each video frame from three dimensions: color features, brightness change features, and spatial connectivity, to obtain a set of suspected flame candidate regions.

[0081] The linked flame region prediction module 20 is used to traverse the sequence of suspected flame candidate regions to perform linked flame region prediction, and obtain a joint suspected flame candidate region set and a semantic side prediction probability set.

[0082] The semantic uncertainty analysis module 30 is used to perform semantic uncertainty analysis on the joint suspected flame candidate region set based on the semantic side prediction probability set to obtain a set of semantic uncertainty coefficients.

[0083] The constraint enhancement module 40 is used to perform spatiotemporal consistency constraint enhancement on the set of joint suspected flame candidate regions in the video frame sequence based on the set of semantic uncertainty coefficients, so as to obtain an enhanced video frame sequence.

[0084] The target recognition result acquisition module 50 is used to perform early flame feature recognition on the enhanced video frame sequence using a flame feature recognizer to obtain target recognition results.

[0085] Furthermore, the system is also used to implement the following functions:

[0086] A continuous sequence of video frames captured by a surveillance camera is acquired and tensor transformation is performed to obtain a network input tensor sequence containing a time dimension. This network input tensor sequence is then fed into the low-level feature encoding layer of the AI ​​video analyzer to encode color information, learn the flame color distribution characteristics, and obtain a low-level feature map sequence. The low-level feature map sequence is then fed into the mid-level temporal feature encoding layer of the AI ​​video analyzer, where convolutional operations along the time dimension are used to jointly model the features of adjacent video frames, learning the dynamic characteristics of brightness changes over time, and obtaining a mid-level feature map sequence. The mid-level feature map sequence is then fed into the high-level spatial structure encoding layer of the AI ​​video analyzer, where a convolutional structure with progressively expanding receptive fields is used to construct the spatial distribution of candidate regions, and obtain a high-level feature map sequence. Finally, the low-level, mid-level, and high-level feature map sequences are mapped and fused to obtain a set of potential flame candidate regions.

[0087] Furthermore, the system is also used to implement the following functions:

[0088] The low-level feature map sequence, mid-level feature map sequence, and high-level feature map sequence are spatially scale aligned by downsampling to obtain aligned low-level feature map sequence, aligned mid-level feature map sequence, and aligned high-level feature map sequence, respectively. These aligned low-level feature map sequence, aligned mid-level feature map sequence, and aligned high-level feature map sequence are then weighted according to preset weights to obtain a comprehensive flame response feature map sequence. The obtained comprehensive flame response feature map sequence is then nonlinearly mapped to generate a flame response probability map sequence, and binarized according to a preset threshold to obtain a candidate region mask map sequence. Region extraction is then performed to obtain the set sequence of suspected flame candidate regions.

[0089] Furthermore, the system is also used to implement the following functions:

[0090] Based on the sequence of suspected flame candidate regions, a joint temporal adjacency analysis is performed to obtain a joint suspected flame candidate region set and a joint suspected flame candidate region neighborhood set. The mean joint similarity between each joint suspected flame candidate region in the joint suspected flame candidate region set and the corresponding joint suspected flame candidate region neighborhood in the joint suspected flame candidate region neighborhood set is calculated to obtain a semantic prediction probability set.

[0091] Furthermore, the system is also used to implement the following functions:

[0092] The first set of suspected flame candidate regions is extracted from the sequence of suspected flame candidate regions, and a first suspected flame candidate region is randomly extracted from the first set of suspected flame candidate regions without replacement. The first set of suspected flame candidate regions and the second set of suspected flame candidate regions are subjected to joint similarity analysis from two dimensions: spatial intersection-union ratio and feature similarity. It is determined whether the maximum value of the joint similarity between the first set of suspected flame candidate regions and the second set of suspected flame candidate regions is greater than or equal to a preset threshold. If not, the second set of suspected flame candidate regions is randomly extracted from the first set of suspected flame candidate regions without replacement, and the joint similarity analysis and threshold determination are repeated until all the suspected flame candidate regions in the first set of suspected flame candidate regions have been analyzed. Then, the process jumps to the second set of suspected flame candidate regions in the sequence of suspected flame candidate regions, and so on, until the end of the sequence is reached, thus obtaining a joint set of suspected flame candidate regions and a joint set of suspected flame candidate region neighborhoods.

[0093] Furthermore, the system is also used to implement the following functions:

[0094] If the maximum joint similarity between the second suspected flame candidate region set and the first suspected flame candidate region is greater than or equal to a preset threshold, then the suspected flame candidate region in the second suspected flame candidate region set corresponding to the maximum joint similarity is added to the first joint suspected flame candidate region neighborhood set, and the first suspected flame candidate region is added to the joint suspected flame candidate region set. The joint similarity analysis and threshold judgment are repeated between the first suspected flame candidate region and the third suspected flame candidate region set, and so on, until the maximum joint similarity is less than the preset threshold, thus obtaining the first joint suspected flame candidate region and the first joint suspected flame candidate region neighborhood. The first joint suspected flame candidate region and the first joint suspected flame candidate region neighborhood are then added to the joint suspected flame candidate region set and the joint suspected flame candidate region neighborhood set.

[0095] Furthermore, the system is also used to implement the following functions:

[0096] Calculate the sum of the semantic side prediction probability set to obtain the semantic side prediction probability sum; normalize the semantic side prediction probability set based on the semantic side prediction probability sum, and use the difference between the normalization result and 1 as the semantic uncertainty coefficient to obtain the semantic uncertainty coefficient set.

[0097] Furthermore, the system is also used to implement the following functions:

[0098] A preset maximum semantic uncertainty coefficient threshold is obtained, and the set of semantic uncertainty coefficients is filtered to determine a qualified set of semantic uncertainty coefficients. The set of joint suspected flame candidate regions is mapped and filtered according to the qualified set of semantic uncertainty coefficients to obtain a filtered set of joint suspected flame candidate regions. Combined with the qualified set of semantic uncertainty coefficients, the set of filtered joint suspected flame candidate regions is subjected to spatiotemporal consistency constraint enhancement according to the adjacent frames of the video frame sequence to obtain an enhanced video frame sequence.

[0099] Furthermore, the system is also used to implement the following functions:

[0100] The difference between 1 and the qualified semantic uncertainty coefficient in the qualified semantic uncertainty coefficient set is used as the enhancement weight to obtain the enhancement weight set; the spatiotemporal consistency constraint enhancement of the screening joint suspected flame candidate region set is performed based on the enhancement weight set and the adjacent frames of the video frame sequence to obtain the enhanced video frame sequence.

[0101] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0102] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0103] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A method for early flame feature recognition based on AI video analysis, characterized in that, The method includes: The system acquires a continuous sequence of video frames captured by surveillance cameras. It then uses an AI video analyzer to analyze each video frame from three dimensions: color features, brightness change features, and spatial connectivity, to obtain a set of suspected flame candidate regions. Traverse the sequence of suspected flame candidate regions to perform linked flame region prediction, and obtain a joint suspected flame candidate region set and a semantic side prediction probability set; Semantic uncertainty analysis is performed on the joint set of suspected flame candidate regions based on the semantic prediction probability set to obtain a set of semantic uncertainty coefficients, including: Calculate the sum of the semantic side prediction probability set to obtain the semantic side prediction probability sum; The set of semantic side prediction probabilities is normalized based on the sum of the semantic side prediction probabilities, and the difference between the normalization result and 1 is used as the semantic uncertainty coefficient to obtain the set of semantic uncertainty coefficients. Based on the set of semantic uncertainty coefficients, the spatiotemporal consistency constraint enhancement is performed on the set of joint suspected flame candidate regions in the video frame sequence to obtain an enhanced video frame sequence. The enhanced video frame sequence is subjected to early flame feature recognition using a flame feature recognizer to obtain target recognition results. The sequence of suspected flame candidate regions is traversed to perform linked flame region prediction, resulting in a joint suspected flame candidate region set and a semantic-side prediction probability set, including: Based on the sequence of suspected flame candidate regions, regional temporal adjacency joint analysis is performed to obtain the joint suspected flame candidate region set and the joint suspected flame candidate region neighborhood set. Calculate the mean joint similarity between each joint suspected flame candidate region in the joint suspected flame candidate region set and the corresponding joint suspected flame candidate region neighborhood in the joint suspected flame candidate region neighborhood set to obtain the semantic side prediction probability set; Based on the set of semantic uncertainty coefficients, spatiotemporal consistency constraint enhancement is applied to the set of joint suspected flame candidate regions in the video frame sequence to obtain an enhanced video frame sequence, including: Obtain a preset maximum semantic uncertainty coefficient threshold, filter the set of semantic uncertainty coefficients, and determine a qualified set of semantic uncertainty coefficients; The set of joint suspected flame candidate regions is mapped and filtered according to the set of qualified semantic uncertainty coefficients to obtain a filtered set of joint suspected flame candidate regions. By combining the qualified semantic uncertainty coefficient set, the spatiotemporal consistency constraint of the selected joint suspected flame candidate region set is enhanced based on the adjacent frames of the video frame sequence to obtain an enhanced video frame sequence.

2. The early flame feature recognition method based on AI video analysis as described in claim 1, characterized in that, A continuous sequence of video frames captured by surveillance cameras is acquired. An AI video analyzer is used to analyze each video frame from three dimensions: color features, brightness variation features, and spatial connectivity, to obtain a set of potential flame candidate regions, including: Acquire a continuous sequence of video frames captured by a surveillance camera and perform tensor transformation to obtain a network input tensor sequence containing the time dimension; The network input tensor sequence is fed into the low-level feature encoding layer of the AI ​​video analyzer to encode the color information, learn the flame color distribution features, and obtain a low-level feature map sequence. The low-level feature map sequence is fed into the mid-level temporal feature coding layer of the AI ​​video analyzer. The features of adjacent video frames are jointly modeled by convolution operations unfolded along the time dimension to learn the dynamic features of brightness changes over time and obtain the mid-level feature map sequence. The middle-layer feature map sequence is fed into the high-level spatial structure coding layer of the AI ​​video analyzer. The spatial distribution of candidate regions is constructed through a convolutional structure with progressively expanding receptive fields, thereby obtaining the high-level feature map sequence. The low-level feature map sequence, the middle-level feature map sequence, and the high-level feature map sequence are mapped and fused respectively to obtain a set of suspected flame candidate regions.

3. The method for early flame feature recognition based on AI video analysis as described in claim 2, characterized in that, The low-level feature map sequence, the mid-level feature map sequence, and the high-level feature map sequence are mapped and fused respectively to obtain a set of potential flame candidate regions, including: Spatial scale alignment is performed on the low-level feature map sequence, the middle-level feature map sequence, and the high-level feature map sequence by downsampling, respectively, to obtain aligned low-level feature map sequence, aligned middle-level feature map sequence, and aligned high-level feature map sequence. The aligned low-level feature map sequence, the aligned middle-level feature map sequence, and the aligned high-level feature map sequence are weighted according to preset weights to obtain a comprehensive flame response feature map sequence. The obtained comprehensive flame response feature map sequence is nonlinearly mapped to generate a flame response probability map sequence. The flame response probability map sequence is then binarized according to a preset threshold to obtain a candidate region mask map sequence. Region extraction is then performed to obtain the set sequence of suspected flame candidate regions.

4. The early flame feature recognition method based on AI video analysis as described in claim 1, characterized in that, Based on the sequence of suspected flame candidate regions, a joint temporal adjacency analysis is performed to obtain a joint set of suspected flame candidate regions and a joint set of suspected flame candidate region neighborhoods, including: Extract the first set of suspected flame candidate regions from the sequence of suspected flame candidate regions, and randomly extract the first suspected flame candidate region from the first set of suspected flame candidate regions without replacement. The first set of suspected flame candidate regions and the second set of suspected flame candidate regions are subjected to joint similarity analysis from two dimensions: spatial intersection-union ratio and feature similarity. It is determined whether the maximum value of the joint similarity between the first set of suspected flame candidate regions and the second set of suspected flame candidate regions is greater than or equal to a preset threshold. If not, the second set of suspected flame candidate regions is randomly extracted from the first set of suspected flame candidate regions without replacement, and the joint similarity analysis and threshold judgment are repeated until all the suspected flame candidate regions in the first set of suspected flame candidate regions have been analyzed. Then, the process jumps to the second set of suspected flame candidate regions in the suspected flame candidate region set sequence, and so on, until the end of the sequence is reached, thus obtaining the joint set of suspected flame candidate regions and the joint set of suspected flame candidate region neighborhoods.

5. The early flame feature recognition method based on AI video analysis as described in claim 4, characterized in that, If the maximum joint similarity between the second suspected flame candidate region set and the first suspected flame candidate region is greater than or equal to a preset threshold, then the suspected flame candidate region in the second suspected flame candidate region set corresponding to the maximum joint similarity is added to the neighborhood set of the first joint suspected flame candidate region, and the first suspected flame candidate region is added to the joint suspected flame candidate region set. The joint similarity analysis and threshold judgment are repeated between the first suspected flame candidate region and the third suspected flame candidate region set until the maximum joint similarity is less than the preset threshold, thus obtaining the first joint suspected flame candidate region and the neighborhood of the first joint suspected flame candidate region. Add the first joint suspected flame candidate region and the neighborhood of the first joint suspected flame candidate region to the joint suspected flame candidate region set and the joint suspected flame candidate region neighborhood set.

6. The early flame feature recognition method based on AI video analysis as described in claim 1, characterized in that, By combining the qualified semantic uncertainty coefficient set, and enhancing the spatiotemporal consistency of the selected joint suspected flame candidate region set based on adjacent frames of the video frame sequence, an enhanced video frame sequence is obtained, including: The difference between 1 and the qualified semantic uncertainty coefficient in the qualified semantic uncertainty coefficient set is used as the enhancement weight to obtain the enhancement weight set; The spatiotemporal consistency constraint enhancement of the selected joint suspected flame candidate region set is performed based on the enhanced weight set and the adjacent frames of the video frame sequence to obtain the enhanced video frame sequence.

7. An early flame feature recognition system based on AI video analysis, characterized in that, The system is used to implement the early flame feature recognition method based on AI video analysis as described in any one of claims 1-6, the system comprising: The video frame analysis module is used to acquire a continuous sequence of video frames captured by the surveillance camera. It uses an AI video analyzer to analyze each video frame from three dimensions: color features, brightness change features, and spatial connectivity, to obtain a set of suspected flame candidate regions. The linked flame region prediction module is used to traverse the sequence of suspected flame candidate regions to perform linked flame region prediction, and obtain a joint suspected flame candidate region set and a semantic side prediction probability set. The semantic uncertainty analysis module is used to perform semantic uncertainty analysis on the joint set of suspected flame candidate regions based on the set of semantic prediction probabilities, and obtain a set of semantic uncertainty coefficients. The constraint enhancement module is used to perform spatiotemporal consistency constraint enhancement on the set of joint suspected flame candidate regions in the video frame sequence based on the set of semantic uncertainty coefficients, so as to obtain an enhanced video frame sequence. The target recognition result acquisition module is used to perform early flame feature recognition on the enhanced video frame sequence using a flame feature recognizer to obtain target recognition results.

Images