Expressway tunnel abnormal event identification method based on multi-scale polarization fusion

Abstract

The invention discloses an expressway tunnel abnormal event recognition method based on multi-scale polarization fusion, relates to the technical field of intelligent recognition of road traffic videos, and is used for solving the problem of robust recognition of trailing collision, fire smoke and abnormal static behaviors in an expressway tunnel scene. The method comprises the following steps: acquiring polarization angle and polarization degree information of each pixel point in a vehicle driving process; then, the image is divided into a plurality of partitions according to longitudinal illumination changes, a time sequence locking window is established in each partition, and the difference between traffic flow reflection characteristics and background interference is extracted; and identifying a smoke area by constructing a disturbance graph and a direction texture graph, calibrating a trailing collision risk, and judging an abnormal static behavior. And finally, an event atlas is generated, full-process identification closed loop and response cooperation are ensured through structure fingerprints and version numbers, and a high-robustness identification means is provided for fire, trailing and abnormal static targets in a tunnel scene.

Description

Technical Field

[0001] This invention relates to the field of intelligent recognition technology for road traffic videos, and more specifically, to a method for identifying abnormal events in highway tunnels based on multi-scale polarization fusion. Background Technology

[0002] Highway tunnels exhibit both an entrance black hole effect and an exit white hole effect. Illuminance exhibits a strong gradient transition along the tunnel's longitudinal direction (s). The superposition of vehicle headlight sweep and internal lighting micro-periods, along with specular reflections introducing highlight bands and ghosting, leads to pulsed pseudo-enhancement of the fused polarization map at the highlight-shadow boundary. Polarization cameras acquire DoP / AoP from multiple angle sequences, with frame intervals between sampling angles. While wide-angle and telephoto channels share the same time base, they still experience micro-delays and field-of-view differences. These factors collectively induce cross-channel alignment drift and fluctuations in DoP / AoP estimation. Existing technologies often employ uniform global exposure and fixed fusion thresholds, processing all-scene data with a single time alignment strategy. They fail to establish independent temporal locking and fusion weights for entrance / exit illuminance zones, and do not explicitly characterize the consistency of illuminance-polarization linkage and boundary drift compensation, easily leading to amplified pseudo-response, early warning jitter, and false triggering of smoke candidates.

[0003] Furthermore, existing methods for addressing rear-end collision risks largely rely on geometric measurements or single-channel motion information, lacking the ability to characterize the enhanced polarization ridge line at the rear of the preceding vehicle using RidgeDoP(s,t). For stationary anomalies, there is a lack of constraints on the consistency of DoP and L joint stable parameters and AoP directions within partitions / sliding windows, making it difficult to maintain stable judgments under continuous disturbances. The lack of cross-channel cross-validation and threshold self-correction / drift synchronization mechanisms results in insufficient reliability of the system under conditions of morning fog, high humidity, strong backlight reflection, and low visibility. Additionally, the lack of structured graphs and fingerprint archiving of event results hinders traceability and coordinated response.

[0004] To address the above problems, this invention proposes a solution. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for identifying abnormal events in highway tunnels based on multi-scale polarization fusion, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: In a preferred embodiment, it includes: A polarization imaging channel is constructed by deploying a four-way rotatable polarizer and a high-sensitivity CMOS, and the polarization degree DoP and polarization angle AoP of each pixel are obtained and a tensor is constructed to generate a DoP feature map containing the reflection background drift band. Construct an illumination observation sequence and divide the tunnel into partitions, obtain the sampling asynchrony and time offset of DoP / AoP, and establish a partition-level temporal locking window; construct a polarization illumination tensor and define a temporal consistency function, calculate the alignment period and boundary drift compensation term, generate fusion weights and label the mask images of four types of structural segments; Generate a multi-scale fused polarization map and a spatiotemporal perturbation tensor, construct event graph nodes and establish logical connections, output a structural summary and record the response state.

[0007] In a preferred embodiment, a polarization imaging channel is constructed by deploying a four-way rotatable polarizer and a high-sensitivity CMOS imaging component in the reverse section of the middle of the tunnel. The polarization degree (DoP) and polarization angle (AoP) information of each pixel are collected with the longitudinal axis of the lane as the polarization angle reference line. A local spatial channel tensor is constructed through the four-way polarization angle, the pixel intensity curve is calculated and sinusoidal fitting is used to obtain the polarization parameter set, and a DoP feature map containing the reflection background drift band is generated.

[0008] In a preferred embodiment, an illumination observation sequence is constructed using the average brightness of each frame in the wide-angle camera image and the vertical coordinates of the corresponding pixels. Based on this, the entire tunnel field of view is divided into four types of regions: entrance transition zone, constant illumination transition zone, internal constant illumination zone, and exit transition zone. A brightness baseline curve set is established for each zone.

[0009] In a preferred embodiment, within each partition, a polarization angle sampling sequence is acquired, and for the DoP and AoP timing information corresponding to the polarization angle sampling sequence, the polarization sampling asynchrony degree within that partition is calculated respectively. The inter-frame mutual information and AoP neighborhood directional spread entropy are jointly evaluated, and the time offset between each polarization angle sampling channel is calculated, thereby establishing a partition-level timing locking window.

[0010] In a preferred embodiment, the illuminance observation sequence, DoP and AoP signals are jointly constructed into a polarization illuminance tensor, and for each pixel position in the tensor, a time consistency function consisting of an orientation consistency term, an intensity steady-state term and an illuminance abrupt change intensity term is defined. The moment when the illuminance transition first reaches a preset threshold in the time domain is calculated based on the temporal consistency function, and the moment when the tensor enters the polarization steady-state region is extracted. The minimum common coverage duration between the two is taken as the alignment period of the partition.

[0011] In a preferred embodiment, the boundary drift compensation term is calculated by combining the temporal consistency function and the variation trend of the partition brightness baseline curve, and the weights are pulled back and the temporal delay is adjusted for the boundary pixels affected by the illuminance gradient interference. Based on the temporal consistency function and boundary drift compensation term, a pixel-level fusion weight vector is constructed. Combined with local temporal features of illumination and polarization characteristics, four types of structural segment labels are annotated, specifically including: Illuminance transition delay period LagL: The rate of change of illuminance per unit time increases significantly, and the temporal consistency score at the same location is low; MisPol polarization sampling misalignment segment: There is a significant alignment error between polarization channels, resulting in asynchronous response between channels; PseudoB (Pseudo-closed segment of transition zone): Boundary pixels are significantly affected by the bright-dark transition zone, which manifests as unstable boundary contrast, abnormally increased local connectivity, and easy formation of pseudo-closed boundaries. GhostR, the peak reflection segment: The dispersion of the polarization angle suddenly increases in a short period of time, accompanied by a short brightness pulse, which is suspected to be a ghosting caused by strong reflection. The above structural segment annotation results are written into the mask map, which only takes effect when the partition locking window completely covers the alignment period of the partition.

[0012] In a preferred embodiment, an abnormal response map is formed by generating a multi-scale fused polarization map and constructing a spatiotemporal perturbation tensor. Then, a hazard score is calculated to trigger collision identification by combining the DoP ridge line, the following speed difference and the distance estimation. Next, color offset and texture consistency are used to mark high-confidence smoke response regions. In the static anomaly regions of the confidence smoke response regions, consistency scores are calculated based on DoP stability and low-perturbation masks, and response perturbation maps are constructed by combining environmental state vectors.

[0013] In a preferred embodiment, an event graph node is constructed that includes structural parameter vectors, spatial locations, temporal windows, risk levels, and image index paths, and the logical connection relationships between events are determined based on temporal continuity, spatial overlap, and consistency of disturbance direction. In the video stream, a red-green difference map and a direction perturbation map are constructed. Then, a cross-channel consistency criterion is established, fire events are marked and written into the main index of the map. Finally, an event structure summary and signature are generated based on the event node number to record the execution status in a closed loop.

[0014] The technical effects and advantages of the present invention regarding the method for identifying abnormal events in highway tunnels based on multi-scale polarization fusion are as follows: This invention addresses the polarization channel alignment mismatch caused by illuminance transitions within tunnels by introducing partitioned illuminance modeling and temporal consistency constraints, effectively suppressing spurious enhanced responses in the transition zone between the entrance and exit. A multi-scale feature fusion mechanism enhances the structural representation of target edges and ridge regions, improving the ability to recognize the response intensity of vehicle rear features. The smoke detection path achieves high-confidence recognition by combining color perturbation trends, directional entropy changes, and texture diffusion structures, while also incorporating temporal continuity judgment to avoid instantaneous misjudgments. The risk scoring for trailing targets integrates speed differences and minimum distance estimation, along with polarization ridge distribution for discrimination. For abnormally stationary areas, a scoring mechanism is constructed based on a triple condition of brightness fluctuation, polarization stability, and directional consistency, effectively distinguishing between non-intrusive dwellings and high-risk targets. The event graph structure enables multi-factor archiving of event recognition results across spatial, temporal, and environmental dimensions, and ensures traceability and subsequent response linkage through structural fingerprinting and version management, significantly improving the stability, accuracy, and engineering adaptability of the intelligent recognition system in tunnel environments. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the implementation of the highway tunnel anomaly event identification method based on multi-scale polarization fusion according to the present invention.

[0016] Figure 2 This is a timing diagram of the highway tunnel anomaly event identification method based on multi-scale polarization fusion according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] This invention discloses a method for identifying abnormal events in highway tunnels based on multi-scale polarization fusion, including: First, to achieve robust recognition of trailing collisions, fire smoke, and abnormal stationary behavior in highway tunnel scenarios, initial image acquisition and feature enhancement are performed based on multi-scale polarization information to address issues such as strong ambient light straying and complex background interference inside the tunnel.

[0019] Specifically, a polarization imaging channel is constructed by deploying a four-way rotatable polarizer and a high-sensitivity CMOS imaging component in a reverse section in the middle of the tunnel. Each imaging node in the channel is bound to a unique imaging task number and imaging direction mark, and the longitudinal axis of the lane is used as the polarization angle reference line. The polarization degree DoP and polarization angle AoP information of each pixel in the image are collected and embedded into the vehicle motion state frame header.

[0020] Among them, coordinates and reference lines: the tunnel longitudinal coordinate s is the main scale, the pixel coordinates are projected onto the s-axis neighborhood through the reference homography Href fixed during the deployment stage, and the AoP uses the lane longitudinal axis reference line φ0 to measure its orientation angle; RGB and polarization channels share the same Href and frame time base; Subsequently, in the image receiving stage, image data from each channel is written to the image frame buffer, triggering the polarization feature extraction module to start. Then, for each image sequence, a local spatial channel tensor is first constructed using the four-way polarization angle. The intensity change curve of each pixel with varying polarization angle is calculated in the tensor, and the corresponding polarization parameter set is obtained through sine fitting, thus completing the separation of high-pass interference reflection background and low-pass moving foreground. The reflection background is represented as an approximate saturation value drift band in the DoP feature map.

[0021] It should be noted that within the region of dramatic illuminance transition between the tunnel entrance black hole effect and the exit white hole effect, the superposition of three factors—illuminance gradient transition, polarization angle sequence sampling rhythm, and traffic sweep reflection—within a short time window can trigger a chain reaction of polarization phase mismatch and cross-channel alignment drift. The specific manifestations of this chain reaction are as follows: The brightness field along the longitudinal axis of the tunnel exhibits monotonic or segmented monotonic changes in the transition zone, and there are moving targets with specular reflection characteristics in some areas. Polarization imaging cameras acquire DoP / AoP in a multi-angle sequence manner, with non-negligible frame intervals between sampling angles; Although the frame time bases of wide-angle and telephoto cameras are synchronized, micro-time delay offset and field-of-view overlap difference exist simultaneously at the moment of illumination transition. The interior artificial lighting exhibits a micro-periodic oscillation in brightness during the transition zone, creating a brief superposition with the sweeping of the headlights.

[0022] At the image representation level, the fused polarization map will produce pulsed pseudo-enhancement at the intersection of the highlight-shadow boundary, the headlight sweep trajectory and the continuous highlight band, and will be amplified in the abnormal response map, causing timing jitter in the detection of smoke candidates and sudden changes in vehicle spacing.

[0023] Existing technologies employ uniform exposure across the entire area and a fixed fusion threshold, and process all scene data using a single time alignment strategy. However, they do not establish independent temporal locking and fusion weight strategies for the entry / exit illuminance zones, resulting in issues such as pseudo-response amplification, fluctuations in DoP / AoP estimation, and jitter in early warnings within the transition zone.

[0024] Therefore, in this embodiment, the average brightness of the wide-angle camera frame and the longitudinal coordinate are first used as the observations to construct the illuminance observation sequence L(s,t). Then, based on the longitudinal illuminance distribution of the tunnel, the field of view is divided into four categories: entrance transition zone, constant illuminance transition zone, internal constant zone, and exit transition zone. A baseline curve set {BZ(s)} is established for the brightness changes in each zone. Here, Z represents the zone label and is only used as a priori reference for illuminance in the method.

[0025] All data sources are from existing image acquisition channels and existing sensor deployments.

[0026] Secondly, the asynchronicity of polarization sampling within the partition is measured and the time offset is locked. Specifically, within each partition Z, for the DoP / AoP timing of the polarization sampling angle sequence {θk}, the partition sampling asynchronicity εZ is calculated. Then, the time offset between sampling angles is jointly measured using inter-frame mutual information and AoP neighborhood directional spread entropy. The time offset ΔτZ is thus obtained within the partition, and a partition timing locking window WZ(t) is established to ensure that pixels corresponding to sampling at different polarization angles complete DoP / AoP estimation and fusion under the same alignment reference.

[0027] Subsequently, the luminance, DoP, and AoP are combined into a tensor within the partition, as follows: ; Define a timing consistency function: ; The three parameters measure directional consistency, steady-state intensity, and abrupt change in illuminance intensity, respectively. The coefficients α, β, and γ are selected based on the prior constraints of the partition, and their specific values ​​are not required.

[0028] Then, using the partitioned temporal locking window WZ(t) as the boundary, the moment when the illuminance transition first crosses the threshold along the time domain is recorded, denoted as tZLt; and the moment when the polarization tensor reaches a steady state within the same sliding window is recorded, denoted as tZΠt.

[0029] Next, we define λZ as the minimum common illuminance-polarization alignment period: ; Next, spatially, using the neighborhood of the partition boundary as a transition zone, a boundary drift compensation term is defined: ; Used to perform weight rollback and timing delay compensation on boundary pixels affected by illumination gradients during fusion.

[0030] Furthermore, within each partition Z, the fusion weight vector is calculated based on CZ(t) and βZ(s,t): ; Furthermore, the pseudo-response within the transition band is classified, suppressed, and labeled into four structural segments, as follows: Illuminance transition delay period LagL: The rate of change of illuminance per unit time increases significantly, and the temporal consistency score at the same location is low. At this time, temporal delay compensation and brightness buffer smoothing are performed first, and the influence weight of this area on the anomaly judgment is temporarily reduced.

[0031] MisPol polarization sampling misalignment segment: Significant alignment errors occur between polarization channels, resulting in asynchronous responses between channels. In this case, cross-channel time alignment and phase correction are performed first; weight rollback is performed on channels with severe misalignment until channel consistency is restored to a safe range.

[0032] PseudoB (Pseudo-closed segment of transition zone): Boundary pixels are significantly affected by the light-dark transition zone, resulting in unstable boundary contrast and abnormally increased local connectivity, which easily leads to pseudo-closed boundaries. In this case, the weight of boundary pixels is reduced, and neighborhood smoothing and structural connectivity re-evaluation are enabled to avoid mistakenly aggregating the transition zone into a complete target contour.

[0033] GhostR (Reflection Peak Ghost Segment): The dispersion of the polarization angle suddenly increases in a short period of time, accompanied by a short brightness pulse, which is suspected to be a ghost caused by strong reflection. At this time, the directional dispersion is suppressed, the diffusion caused by the inconsistency of the directions of adjacent pixels is reduced, the instantaneous peak of polarization is suppressed, and the short pulse period is shielded from participating in subsequent discrimination.

[0034] The four types of segments are labeled with Ltype∈{LagL, MisPol, PseudoB, GhostR} respectively, and the label is written only when the sliding window covers the λZ period within the partition.

[0035] Finally, within each partition, the following are outputs: aligned DoP / AoP estimates, CZ(t) sequences, βZ(s,t) maps, and Ltype marker masks; then, the fusion weight wZ and transition band compensation term βZ are passed as pixel-level weighted preconditions to subsequent multi-scale fusion.

[0036] Baseline initialization and degradation strategy: When entering a new time period for the first time or when there is a lack of recent gaps, the initial estimate of {BZ(s)} is first constructed with a short-term Winit(t) and marked as a cold start; if εZ exceeds the allowed upper bound or λZ=0 causes timing lock instability, wZ is frozen to the previous steady-state value, temporarily regressed to the best single θk channel fusion and writing to Ltype is suspended; when WZ(t) resumes to cover λZ, degradation and freezing are automatically lifted and the marking is resumed.

[0037] Next, multi-scale feature enhancement is applied to the pre-processed image. First, multi-scale high-frequency component maps are constructed with different spatial receptive fields, and contrast stretching and directional gradient extraction are performed at each scale to enhance the response of vehicle contours and boundaries at each scale. Then, the response maps at each scale are weighted and superimposed using a feature pyramid aggregation method to form a fused polarization map. In the fused map, the trailing vehicle and the interfering background show directional polarization differences in local brightness fluctuations, and this difference is presented as a brightness ridge enhancement band on the DoP response map.

[0038] Furthermore, such as Figure 1 As shown, edge confidence calculation and region connectivity thresholding are applied to the fused map to construct a region-stable mask map, and an anomaly response map is generated using the mask map as a reference region. The anomaly response map is constructed by superimposing inter-frame differences to form a response sequence. Each frame in the sequence is processed pixel-by-pixel to form a spatiotemporal perturbation tensor to mark image patches in the tunnel field of view where illumination and object perturbation change drastically at the same time.

[0039] Define the spatiotemporal perturbation tensor ΨZ(s,t):=[ΔLZ(s,t),ΔDoPZ(s,t),∂AoPZ / ∂t(s,t)], where ΔLZ(s,t)=|LZ(t)−LZ(t−1)|; and the anomaly response map RZ(s,t)=Norm(⟨ΨZ>{WZ(t)}), which serves as a prerequisite for subsequent candidate region screening and map entry into the database.

[0040] Where LZ(t) is short for LZ(s,t); it represents the illuminance observation value at the current location s.

[0041] ΔLZ(s,t)=|LZ(s,t)−LZ(s,t−1)|.

[0042] ΔDoPZ(s,t)=|DoPZ(s,t)−DoPZ(s,t−1)|, where DoPZ is the DoP after aggregating the four-way θk channels according to a fixed rule.

[0043] ∂AoPZ / ∂t(s,t): Inter-frame principal angle difference of AoP.

[0044] ⟨ΨZ>{WZ(t)}: Within the partitioned time-locked window WZ(t), perform time statistics on each component of ΨZ point by point.

[0045] Norm(·): A fixed normalization / standardization operator that maps the input component by component to a predefined domain, keeping the output dimension unchanged, and records the version number used.

[0046] After completing the multi-scale fusion of polarization images and the construction of anomaly response maps, the anomaly type-specific identification process is immediately initiated. Firstly, for fire and smoke-related events, smoke candidate regions are extracted from the fused polarization and anomaly response maps. Specifically, a joint mask is constructed using regions with high DoP response and significant perturbations across multiple frames in the anomaly response map. Color shift curve analysis is performed on the regions within this mask, and a two-dimensional color perturbation matrix is ​​constructed using saturation (S) and brightness (V) in the HSV color gamut. Continuous directional anomaly trajectories are identified within this matrix. If this trajectory exhibits a sudden increase in brightness followed by frame-by-frame diffusion over time, with feathered textures at the region boundaries, it is marked as a smoke enhancement candidate region.

[0047] Subsequently, texture consistency detection is applied within the candidate region. A texture stability score map within the region is constructed using the local gray-level co-occurrence matrix, and a discrimination index is constructed based on the magnitude of the texture entropy gradient change at the region boundary. When the score map shows that the candidate region has obvious internal texture inhomogeneity and the texture gradient at the outer edge drops sharply, the region is confirmed as a typical smoke response region that usually exhibits directional disorder and high-frequency oscillation bands in the polarization angle AoP map.

[0048] To further enhance the robustness of smoke event identification, a temporal continuity verification mechanism is initiated after detecting the smoke response area. Specifically, a spatiotemporal convolutional block is constructed for the area over the past N frames, and the changes in its DoP value and brightness value are observed to show a continuous increasing trend. If this trend persists in multiple frame windows and the area diffusion rate meets the stable growth criterion, it is confirmed as a suspected smoke event, and the subsequent spatial localization process is triggered.

[0049] After smoke identification and labeling are completed, the follow-along collision risk identification process is initiated. First, a vehicle boundary gradient map is constructed based on the fused polarization image, and horizontal lane edge lines are detected in the map using Hough transform. Then, longitudinal gradient strip regions are extracted based on the lane centerline.

[0050] Then, for vehicle targets within the strip area, the vehicle position and size are identified using a convolutional neural network model, and the relative offset and width of each target box are labeled.

[0051] Next, the motion trajectory is constructed using the sequence of center points of the target boxes in adjacent frames, and the inter-frame velocity vector is calculated. The velocity difference and minimum time distance estimate of the trailing target are statistically analyzed using a spatiotemporal sliding window structure. When the velocity difference and target spacing of the trailing target in consecutive frames jointly constitute the danger distance criterion, and the rear structure of the vehicle in front of it shows a high DoP enhancement band in the polarization diagram, it is identified as a high-risk target for trailing collision and written into the event response cache pool.

[0052] Define dt as the minimum distance estimate along the lane centerline, and Δvt as the following relative speed; let RidgeDoP(s,t) represent the intensity of the high DoP ridge at the rear of the preceding vehicle; then the dangerous distance score Rtail = u1·g(Δvt,dt) + u2·RidgeDoP(s,t), where g(·) monotonically increases with |Δvt| and monotonically decreases with dt; when Rtail remains above the threshold and the trajectory is continuous within a multi-frame sliding window, a high-risk following collision marker is triggered and written to the event response buffer pool.

[0053] Among them, Rtail: tailing danger distance score, the higher the value, the higher the risk.

[0054] Δvt: Inter-frame estimate of the trailing relative velocity.

[0055] dt: Minimum longitudinal spacing estimate along the lane centerline.

[0056] g(Δvt,dt): Joint scoring function, satisfying ∂g / ∂|Δvt|>0 and ∂g / ∂dt<0 (implementation is fixed, specific form is not given).

[0057] RidgeDoP(s,t): The intensity of the high-response ridge line at the rear of the vehicle on the DoP map (extracted from the multi-scale DoP augmentation map).

[0058] u1, u2: Dimensionless weighting coefficients.

[0059] Furthermore, after the tail-following collision target identification is completed, a preliminary static region screening map is constructed based on the differences in edge structure changes between consecutive frames in the fused polarization image. Specifically, using the pixel structural similarity of three consecutive DoP images as an indicator, pixel-by-pixel similarity integration is performed. If a region maintains high structural stability and does not exhibit boundary perturbations in all frames, it is marked as a potential static region.

[0060] Subsequently, within this static area, the average brightness variation trend is combined with the increase or decrease in the average brightness of the local area to construct an inter-frame difference curve, determine whether there is a continuous situation without significant fluctuations, and refer to whether there is a corresponding low disturbance value band in the abnormal response map. If both meet the threshold requirements, the area is included in the static target candidate set.

[0061] Define the static consistency score Qstatic(s) = ζ1·(1−VarW(DoPZ))+ζ2·(1−VarW(LZ))+ζ3·CAoP(s)+ζ4·MasklowPerturbation(s), where Qstatic(s) is the static consistency score, and the higher the score, the more likely it is to be a static region.

[0062] VarW(·): The time variance operator within a partition / sliding window.

[0063] CAoP(s): AoP directional consistency.

[0064] MasklowPerturbation(s): A mask indicator generated from the low-perturbation band of the anomalous response map RZ(s,t).

[0065] ζ1,ζ2,ζ3,ζ4: Dimensionless weighting coefficients.

[0066] When Qstatic(s) exceeds the adaptive threshold and the region boundary remains structurally closed on the DoP / AoP graph, the static confidence level is confirmed to be improved.

[0067] Then, foreground-background segmentation is performed on the same region in the RGB image to eliminate misclassified regions caused by light occlusion or reflection. Next, consistency analysis is performed on the directional distribution characteristics within the region in the polarization angle AoP image. If the directional angles within the region show a consistent distribution or stable polarization angle characteristics, its static confidence level is further improved.

[0068] The environmental response modeling process then begins. First, tunnel meteorological sensor data and historical event records are retrieved, and an event response sliding window is constructed within a time range T before and after the current frame. For indicators such as light intensity, visibility, and humidity within this window, standardization processing is performed, and an environmental state vector is constructed, labeled E(t).

[0069] Wherein, E(t)=[Illum(t),Vis(t),Hum(t),…] is normalized by z-score or quantile within the same time window; μbase, σbase or quantile curves are obtained by taking the empirical distribution of the event window period of the past 3 months as the baseline; the offset between the current value and the baseline is used to form a response perturbation spectrum with ΔE(t)=Norm(E(t)−Ebase), and the version number and timestamp are written in.

[0070] Illum(t): Illuminance / illuminance.

[0071] Vis(t): Visibility.

[0072] Hum(t): Humidity.

[0073] Ebase: The baseline of the environment state vector.

[0074] ΔE(t)=Norm(E(t)−Ebase): Normalized vector of environment offset; Norm is consistent with the definition above and records the threshold / baseline version and timestamp.

[0075] For each dimension of E(t), a baseline reference line for the mean and variance is established based on historical data from the past three months, and a response perturbation spectrum is established based on the offset of the current value relative to the baseline line.

[0076] In the event graph, if a stationary target region in the current frame is found to be in the spatial location segment corresponding to the abnormal offset of E(t), and its DoP enhancement value is in the top 5% of the high value region of the whole map, combined with the position stability and structural texture consistency, the stationary target is identified as having a high anomaly risk, and the event reporting mechanism is triggered.

[0077] Furthermore, to adapt to recognition errors caused by environmental changes, the system calculates the false alarm rate and false alarm rate based on the most recent N trigger events, and uses a dynamic threshold adjustment factor α(t) to linearly fine-tune all criterion thresholds. For example, when increased humidity causes a decrease in DoP response, the system automatically lowers the high-risk response DoP threshold, ensuring stable recognition capabilities even on slippery surfaces.

[0078] After determining the high risk of static anomaly regions, the event labeling and recording process is immediately initiated. First, the event labeling controller groups all static regions within the current frame and performs structured contour extraction based on the region boundary closure in the DoP image. The regions inside the contours are then binary-masked in the RGB image and marked as a set of candidate event target regions.

[0079] Subsequently, each candidate event target region is assigned a unique event identifier (ID) to construct an event metadata dictionary. This dictionary includes the event ID, trigger frame number, center coordinates of the static region, spatial area pixel value, AoP directional consistency index, DoP enhancement coefficient, structural texture complexity factor, and its corresponding environmental state vector E(t).

[0080] After the metadata recording is completed, based on the frame number timestamp, trace back to the frame in which the stationary target first appeared, and record the initial appearance frame number as the termination condition that the boundary disturbance intensity is always lower than a set threshold within consecutive frames; then trace back to the frame in which the target disappeared or the frame in which the boundary disturbance first occurred, and record it as the disappearance frame number. In this way, a complete event time domain window [Ts,Te] is formed and bound to the event ID.

[0081] Subsequently, the historical event judgment rule set is loaded, and the parameters of the current event are compared item by item with the existing high-risk event template. If all four indicators—DoP enhancement, directional consistency, structural closure, and environmental disturbance state—fall within the 95% confidence interval of historical high-risk events, a high-risk static event label is automatically assigned, with a level of Level-3; if only three match, it is labeled as Level-2; if only two match, it is labeled as Level-1; and the rest are marked as low-risk observation events and reserved as potential early warning targets.

[0082] After label assignment, the main event graph structure is constructed using the event ID as the index node. Each node contains its structural parameter vector, spatial location, temporal window, risk level, and corresponding original image index path. Logical connections are then established between nodes. If two events are temporally consecutive, spatially overlap by more than 30%, and have the same direction of environmental state perturbation, the connection is marked as an evolutionary path; if they do not overlap spatially but have the same direction, they are marked as potential migration channels. The completed graph structure is serialized in JSON-LD format and written to the graph cache. Simultaneously, the graph export process is initiated. First, the corresponding event regions in the DoP graph, AoP graph, and original RGB graph are cropped, retaining the core event region tiles. The image files are renamed according to the event ID and stored uniformly in the event tile archive directory. Exif metadata, including the event ID, time label, risk level, and generation version number, is written to each image file.

[0083] Finally, the event graph is encrypted. A fingerprinting archiving tool is used to integrate and package the event graph structure, image file set, metadata dictionary, and historical comparison records, generating a unique structural signature hash value. A local TLS certificate is used for encryption, generating a two-way authentication key. This key is written to the archive center directory using a version number + event tag naming convention and registered in the regulatory index database. If the system detects consecutively triggered high-risk events, it immediately pushes the latest event summary and graph fingerprint to the early warning center via an HTTPS channel and records the return status.

[0084] After labeling and archiving static anomalies, an RGB channel motion energy map is constructed in the continuous video stream using inter-frame difference. The difference between the red channel of the current frame and the green channel of the previous frame is calculated to form a red-green difference channel map, enhancing the color response characteristics of the flame region. This operation is performed in real-time via a GPU-accelerated frame processing pipeline and written to the image processing cache.

[0085] Simultaneously, edge perturbation enhancement operations are performed on the AoP map. The maximum gradient direction difference of each pixel's direction change is calculated in a 5×5 local window, and the direction entropy fluctuation of continuously changing regions in the whole map is statistically analyzed. If the fluctuation exhibits low-frequency and high-amplitude characteristics in multiple consecutive frames, it is marked as a possible smoke boundary perturbation region.

[0086] Building upon this, a multi-scale DoP feature enhancement process is initiated. For the DoP map, max pooling is performed sequentially with downsampling sizes of 1 / 2, 1 / 4, and 1 / 8 to obtain polarization response maps at different scales. These are then fused with the original DoP map to construct a multi-scale polarization enhancement map, used to extract non-uniform polarization reflection signals from flame edges and smoke disturbances. This fused map is input to the depth discriminator network in a 3-channel format before convolution processing.

[0087] The initial candidate area for a fire incident is triggered when all three types of access routes simultaneously meet the following conditions: It has a high intensity response in the red-green difference map: Let ΔRG(s,t)=|Rt(s)−G{t−1}(s)|, then ⟨ΔRG>{N×N}(s,t)>RG(t); The directional perturbation diagram shows a trend of concentrated rotational perturbation: if RAoP(s,t) = ∂ / ∂t (directional entropy of AoPZ) and its sign forms a single-peaked sign-changing structure in WZ(t), then RAoP(s,t) > dir(t). In multi-scale DoP enhancement maps, this manifests as a locally highlighted discontinuous structure: Let ADoP(s,t) be the local maximum intensity of the multi-scale DoP response, and ρ(s,t) be the connectivity ratio of this maximum. Then ADoP(s,t) > DoP(t) and ρ(s,t) <conn(t); When all three conditions are met simultaneously, it is designated as a high-response zone for fire; RG(t), Θdir(t), DoP(t), and Θconn(t) are adaptively updated by a threshold self-correction and drift synchronization mechanism and their version numbers are recorded in the structure graph.

[0088] Subsequently, region tracking and morphological evolution analysis were performed on the high-response region. The rate of change of the region boundary, the center point drift vector, and the area growth curve were tracked through a three-frame time sliding window. For targets with a significantly higher perturbation rate than the static region, dynamic flame template matching was performed. If the matching score exceeded a set threshold and the height of the region center point moved upward frame by frame, the system marked it as a valid flame leap behavior.

[0089] In parallel, during the smoke detection path, a Laplacian high-pass filter is applied to the DoP image, and the texture entropy gradient change curve of the filtered image is calculated. If a candidate region exhibits a central blurring + edge high-frequency enhancement structure across multiple frames, and the grayscale value in the original RGB image slowly increases, the region is identified as a high-confidence smoke diffusion region.

[0090] When the spatial overlap between the flame rise behavior and the smoke diffusion area exceeds a specified ratio, and the temporal overlap length meets the fire occurrence window standard, the system comprehensively determines that the current state is a fire event triggered. An image buffer locking mechanism is immediately activated, and keyframe images containing the entire event process are sampled at 1-second intervals and stored in the fire event image pool.

[0091] After the image pool is written, the system performs a keyframe event aggregation operation. For each frame, a flame location mask, smoke boundary contour, and DoP mutation distribution map are generated and bound to the event index record, forming the main index node for the fire event. Subsequently, based on the event time-domain window, a fire event timeline structure is constructed, recording the changes in various physical indicators between keyframes. For example, in a tunnel fire test, the DoP change value may show an initial decrease followed by high-frequency fluctuations, with the texture direction entropy increasing to more than twice the baseline value within the maximum amplitude range.

[0092] Furthermore, to enhance system fault tolerance and data integrity, video frames within ±10 seconds before and after the fire are reloaded via a delayed cache channel, and the event feature calculation process is re-executed. If the recalculated result and the initial result have an error in spatial offset and response intensity that is less than the specified redundancy tolerance line, the verification is considered successful; otherwise, the two results are written to the comparison cache, and manual intervention and review prompts are executed.

[0093] Finally, a graph-based structuring process is executed for verified fire events. Referring to the static event processing flow, each event node is written into the fire event graph table, and feature fields such as trigger frame index, response channel category, spatial centroid trajectory, polarization entropy perturbation sequence, and red-green difference dynamic curve are added. Once the graph structure is complete, fingerprinting and archiving are immediately performed, generating a unique event signature and recording it in the secure archive directory, thus completing the closed loop of the entire fire event identification process.

[0094] like Figure 2 As shown, after the fire event identification module completes the image archiving, the environmental response mechanism modeling process for extreme interference scenarios is immediately initiated. First, in the full-time video sequence, the time periods with frequent external interference (such as the day-night transition, backlight at tunnel entrances, and typical interference windows such as dense fog and high humidity in winter) are selected, and interference samples are marked in 3-minute units to construct an environmental interference response feature library.

[0095] Subsequently, in each interference window, DoP, AoP, red-green difference map and RGB image grayscale channel are extracted sequentially, and multi-index feature statistical operations are performed on each frame, including: polarization entropy curve, directional perturbation amplitude, red-green difference noise ratio, grayscale histogram kurtosis, etc.

[0096] During the detection process, if the DoP image in the identification channel shows low amplitude and high frequency fluctuations and the grayscale histogram of the RGB image is flattened at the same time, this phenomenon usually corresponds to areas with abnormal lighting such as fog diffusion or backlight and back reflection. The system will mark the current window as a high reflection interference state, forcibly activate the reflection suppression weight channel in the polarization filter bank, and enable low-pass filtering mode for the red-green difference image to avoid false triggering of high frequency colors.

[0097] In parallel, if a large area of ​​gradually changing structure in the same direction is presented in the AoP direction disturbance map, and the blurred area at the edge of the DoP map continuously exceeds the set frame length, the system judges it to be a vehicle exhaust thermal disturbance or exhaust plume residual area; immediately add a thermal smoke disturbance mask to the area, and shield the flame response calculation process of the area in the abnormal event channel.

[0098] Subsequently, to achieve robust fusion of the system across multiple heterogeneous image channels, a threshold self-correction and drift synchronization mechanism is initiated during the interference detection phase. Specifically, a sliding window time series is constructed for the maximum DoP, average red-green color difference, and grayscale standard deviation of each frame within the current detection window, and compared with the baseline index distribution formed during the past 72-hour event window. If all three deviate from the baseline center value by more than a set amount, the system automatically updates the discrimination threshold and marks the adaptive threshold version number used in the current frame in the structural graph.

[0099] For example, during the morning fog and high humidity phase at the tunnel entrance, the brightness of the area in the red-green contrast image increases significantly, but the DoP volatility and entropy changes remain within the normal range. In this case, the system will determine it as a non-fire behavior based on the DoP steady-state priority strategy to avoid false alarms caused by enhanced illumination. The DoP baseline distribution used in this case was generated by back-calculating images from the past 5 morning periods and then smoothed using Gaussian kernel density to create a reference model.

[0100] During continuous monitoring, the system performs a channel cross-validation mechanism for each event determination result. That is, any event trigger requires independent confirmation from at least two channels, and the difference in their response frame numbers does not exceed the set sliding window width. If an abnormal response occurs only in a single channel, or if there is no temporal consistency between responses from multiple channels, the response record is sent to the low-confidence event buffer pool, awaiting subsequent verification mechanism intervention and confirmation.

[0101] Let the channel set C = {ΔRG, directional perturbation, multi-scale DoP enhancement}; define the same-frame neighborhood consistency window Wcv. If ∑{c∈C}1[c is judged as positive in Wcv]≥2 and the difference between positive frames≤Wcv, then the event is a valid trigger; if only a single channel is positive or there is no temporal consistency between channels, it is downgraded to a low-confidence event and written into the low-confidence event buffer pool; when a conflict occurs between channels, the channel with the largest partition weight CZ(t) is given priority, and the conflict mark and handling path are recorded.

[0102] After completing the above interference correction and adaptive update, the system synchronously updates the interference scene index table in the event graph, recording the current time period, the threshold used in the detection channel, the filtering strategy number, the drift path type, and the model version information.

[0103] Next, using the unique event number of each node in the event graph as an index, a complete identification result set is constructed, which includes: event start and end frame numbers, identification category labels, graph structure, threshold version information used, and the channels used for determination and interference status markers.

[0104] Subsequently, the event graph fingerprinting process is initiated. The aforementioned structure is normalized, hashed, and keyframe feature summaries are extracted to form an event fingerprint digest structure. This structure generates a unique identifier using the SHA-256 algorithm, which is then bound to the event number and written to the graph fingerprint registry. Simultaneously, it is submitted to the platform's local key issuer for signing.

[0105] During this process, each signature record is simultaneously written to both the circular log buffer and the Flash long-term archive area, with a retention time of no less than 72 hours. A fingerprint version snapshot is automatically generated every 4 hours. For example, if an event number is T1245, and it is jointly identified in the DoP channel and the red-green difference channel, its main path in the graph is encoded to form a structural digest {A1 to B3 to C2}. The corresponding graph fingerprint is sha256(A1B3C2), recorded as version V3.4.7, and the timestamp and key number are archived together.

[0106] After fingerprint generation is complete, the system matches the corresponding response script from the preset scheduling strategy template library based on the event type and the response block number in the map. For example, taking fire detection as an example, if the event response area is the channel segment numbered P6-P9, the system automatically matches the "Fire Event · Main Channel · Night Template", which contains the following action chain: Automatically issue start / stop commands for the tunnel ventilation system; Emergency lighting and broadcast announcements should be activated. Send a route closure diagram to the traffic guidance screen; Enable the fire alarm notification interface on the backend platform and push it to the duty terminal.

[0107] Each action chain node has an execution delay setting, an execution confirmation mechanism, and a fault retry channel to ensure that the response process has fault tolerance and a confirmation closed loop.

[0108] Finally, to achieve cross-platform scheduling and response capabilities, all response script execution instructions are sent in a unified JSON encapsulation format, including action type, target device number, execution parameters, expected response status, and timeout threshold. Instructions are sent to each device controller via the scheduling instruction bus, and the device response status and feedback time are collected and written to the execution result record table.

[0109] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0110] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0111] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0112] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0113] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0114] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying abnormal events in highway tunnels based on multi-scale polarization fusion, characterized in that: include: A polarization imaging channel is constructed by deploying a four-way rotatable polarizer and a high-sensitivity CMOS, and the polarization degree DoP and polarization angle AoP of each pixel are obtained and a tensor is constructed to generate a DoP feature map containing the reflection background drift band. Construct an illumination observation sequence and divide the tunnel partitions, obtain the DoP / AoP sampling asynchrony and time offset, and establish a partition-level time-series locking window; Construct a polarization illumination tensor and define a temporal consistency function, calculate the alignment period and boundary drift compensation term, generate fusion weights and label four types of structural segment mask images; Generate a multi-scale fused polarization map and a spatiotemporal perturbation tensor, construct event graph nodes and establish logical connections, output a structural summary and record the response state.

2. The method for identifying abnormal events in highway tunnels based on multi-scale polarization fusion according to claim 1, characterized in that: A polarization imaging channel is constructed by deploying a four-way rotatable polarizer and a high-sensitivity CMOS imaging component in the reverse section of the middle of the tunnel. The polarization degree (DoP) and polarization angle (AoP) information of each pixel are collected with the longitudinal axis of the lane as the polarization angle reference line. The local spatial channel tensor is constructed by the four-way polarization angle, the pixel intensity curve is calculated and sinusoidal fitting is used to obtain the polarization parameter set, and a DoP feature map containing the reflection background drift band is generated.

3. The method for identifying abnormal events in highway tunnels based on multi-scale polarization fusion according to claim 1, characterized in that: An illumination observation sequence was constructed using the average brightness of each frame in the wide-angle camera image and the vertical coordinates of the corresponding pixels. Based on this, the entire tunnel field of view was divided into four types of regions: entrance transition zone, constant illumination transition zone, internal constant illumination zone, and exit transition zone. A brightness baseline curve set was established for each zone.

4. The method for identifying abnormal events in highway tunnels based on multi-scale polarization fusion according to claim 3, characterized in that: Within each partition, polarization angle sampling sequences are acquired, and for the DoP and AoP timing information corresponding to the polarization angle sampling sequences, the polarization sampling asynchrony degree within the partition is calculated respectively. The inter-frame mutual information and AoP neighborhood directional spread entropy are jointly evaluated, and the time offset between each polarization angle sampling channel is calculated, thereby establishing a partition-level timing locking window.

5. The method for identifying abnormal events in highway tunnels based on multi-scale polarization fusion according to claim 4, characterized in that: The illuminance observation sequence, DoP and AoP signals are jointly constructed into a polarization illuminance tensor. For each pixel position in the tensor, a time consistency function consisting of an orientation consistency term, an intensity steady-state term and an illuminance abrupt intensity term is defined. The moment when the illuminance transition first reaches a preset threshold in the time domain is calculated based on the temporal consistency function, and the moment when the tensor enters the polarization steady-state region is extracted. The minimum common coverage duration between the two is taken as the alignment period of the partition.

6. The method for identifying abnormal events in highway tunnels based on multi-scale polarization fusion according to claim 5, characterized in that: By combining the temporal consistency function with the variation trend of the partition brightness baseline curve, the boundary drift compensation term is calculated, and the weight is pulled back and the temporal delay is adjusted for the boundary pixels affected by the illuminance gradient interference. Based on the temporal consistency function and boundary drift compensation term, a pixel-level fusion weight vector is constructed. Combined with local temporal features of illumination and polarization characteristics, four types of structural segment labels are annotated, specifically including: Illuminance transition delay period LagL: The rate of change of illuminance per unit time increases significantly, and the temporal consistency score at the same location is low; MisPol polarization sampling misalignment segment: There is a significant alignment error between polarization channels, resulting in asynchronous response between channels; PseudoB (Pseudo-closed segment of transition zone): Boundary pixels are significantly affected by the bright-dark transition zone, which manifests as unstable boundary contrast, abnormally increased local connectivity, and easy formation of pseudo-closed boundaries. GhostR, the peak reflection segment: The dispersion of the polarization angle suddenly increases in a short period of time, accompanied by a short brightness pulse, which is suspected to be a ghosting caused by strong reflection. The above structural segment annotation results are written into the mask map, which only takes effect when the partition locking window completely covers the alignment period of the partition.

7. The method for identifying abnormal events in highway tunnels based on multi-scale polarization fusion according to claim 1, characterized in that: By generating a multi-scale fused polarization map and constructing a spatiotemporal perturbation tensor, an abnormal response map is formed. Then, by combining the DoP ridge line, the following speed difference and the distance estimation, the hazard score is calculated to trigger collision identification. Next, color offset and texture consistency are used to mark high-confidence smoke response regions. In the static anomaly regions of the confidence smoke response regions, consistency scores are calculated based on DoP stability and low-perturbation masks, and response perturbation maps are constructed by combining environmental state vectors.

8. The method for identifying abnormal events in highway tunnels based on multi-scale polarization fusion according to claim 1, characterized in that: Construct event graph nodes that include structural parameter vectors, spatial locations, temporal windows, risk levels, and image index paths, and determine the logical connections between events based on temporal continuity, spatial overlap, and consistency of disturbance directions. In the video stream, a red-green difference map and a direction perturbation map are constructed. Then, a cross-channel consistency criterion is established, fire events are marked and written into the main index of the map. Finally, an event structure summary and signature are generated based on the event node number to record the execution status in a closed loop.

Images