A warehouse monitoring video enhancement method based on scene state memory

Abstract

This invention discloses a warehouse monitoring video enhancement method based on scene state memory. By filtering low-motion, low-overexposure frames from historical videos of the same camera, lighting topology descriptors are extracted and clustered to form scene state entries, generating anchored brightness maps, confidence maps, and reflection occupancy memory maps. State matching is performed on the current frame, and candidate enhanced brightness maps are generated by combining dynamic occlusion corridor areas and reflection uncertainty areas. The optimized enhanced brightness map is obtained through enhanced confidence back-off fusion and optimization solution, and color channel synchronous mapping is achieved through shared gain. This effectively suppresses false enhancement, reflection interference, dynamic occlusion effects, and comprehensive color shift, improving the stability and adaptability of warehouse monitoring video enhancement results.

Description

Technical Field

[0001] This invention belongs to the fields of image enhancement and video surveillance technology, and specifically relates to a warehouse surveillance video enhancement method based on scene state memory. Background Technology

[0002] With the continuous development of intelligent warehousing and logistics automation systems, warehouse monitoring videos have been widely used in scenarios such as inventory counting, forklift dispatching, personnel operation supervision, aisle safety early warning, and abnormal event tracing. Due to the characteristics of warehouse monitoring images, such as complex lighting conditions, obvious backlighting at entrances, uneven distribution of light strips, strong reflections on shelves and packaging surfaces, and frequent passage of personnel, forklifts, and goods carts, the captured video frames are prone to problems such as local dark areas, local bright areas, uneven brightness distribution, blurred structural boundaries, and overall color shift, which affect the subsequent recognition and monitoring effects.

[0003] Existing video enhancement methods typically employ histogram equalization, Retinex, local contrast enhancement, temporal smoothing, or deep learning-based low-light enhancement strategies. These methods mostly start from single-frame information or short-term temporal information, making it difficult to utilize the stable lighting distribution and static facility information that exist in long-term historical videos from the same fixed camera. As a result, false enhancement, flickering, structural drift, and overall color shift are prone to occur in warehouse entrances, deep passages, loading and unloading areas, and reflective packaging areas.

[0004] In particular, warehouse monitoring scenarios are characterized by fixed cameras, fixed aisle directions, fixed static facilities, and long-term repetitive lighting patterns. At the same time, motion occlusion usually propagates along the aisle direction, and the reflective areas of the mirror surface are highly repetitive in spatial location. Existing enhancement methods usually fail to utilize the above-mentioned scenario patterns simultaneously, making it difficult to improve brightness while taking into account the structural stability of the outline of goods, aisle boundaries, signage, and personnel work areas. Summary of the Invention

[0005] The purpose of this invention is to provide a warehouse monitoring video enhancement method based on scene state memory, so as to solve the problem that existing methods are difficult to simultaneously suppress false enhancement, reflection interference, dynamic occlusion effects and comprehensive color shift in warehouse monitoring scenarios.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a warehouse monitoring video enhancement method based on scene state memory, comprising the following steps.

[0007] S1. Filter low-motion, low-overexposure frames from the historical videos of the same camera, extract the lighting topology descriptor composed of row brightness profile, column brightness profile, light strip peak response, entrance backlight coefficient and static vertical structure density, cluster them to form scene state entries, and generate state descriptors, anchor brightness maps, confidence maps and reflection occupancy memory maps.

[0008] S2. Extract the lighting topology descriptor of the current frame and determine the target scene state entries and matching confidence.

[0009] S3. Determine the motion area based on the temporal motion response, and expand the motion area in a directional manner according to the warehouse passage direction detected in the high-confidence static area to obtain a dynamic shading corridor area.

[0010] S4. Determine the uncertain area of ​​reflection based on the high-brightness, low-color-difference area and the reflection occupancy memory map, and generate a candidate enhanced brightness map by combining the confidence map, the dynamic occlusion corridor area and the uncertain area of ​​reflection.

[0011] S5. Construct an enhanced confidence map based on the matching confidence, confidence map, dynamic occlusion corridor area, reflective uncertainty area, and the degree of deviation between the candidate enhanced brightness map and the anchor brightness map. When the matching confidence is lower than the confidence threshold, backfuse the candidate enhanced brightness map, the current frame brightness map, and the anchor brightness map to obtain the backfuse enhanced brightness map.

[0012] S6. Construct an optimized functional containing log-fidelity terms and gradient regularization terms, and iteratively optimize the backtracking enhanced brightness map to obtain the optimized enhanced brightness map.

[0013] S7. Construct a shared gain based on the ratio of the optimized enhanced luminance map to the current frame luminance map, perform synchronous mapping on each color channel of the current frame, and output the enhanced video frame.

[0014] Preferably, in step S1, a scene state memory is constructed to provide historical prior information corresponding to the fixed camera scene for current frame enhancement. The specific steps are as follows:

[0015] S11. Represent the current color video frame as... Its brightness diagram is shown as follows: ,

[0016] Define the pixel position of the current frame The overall color difference at the location is: ;

[0017] in, Indicates time The current color video frame; , , These represent the red, green, and blue color channels of the current color video frame, respectively. This represents the luminance map corresponding to the current color video frame; , , For the corresponding color channel, the non-negative luminance weighting coefficient is , and satisfies . , ; Indicates time The current frame at pixel position Overall color difference at the location; , , These represent the pixel positions of the current frame. The red, green, and blue color channel values ​​at that location; and These represent the operations of finding the maximum and minimum values, respectively.

[0018] S12, the lighting topology descriptor includes row brightness profile, column brightness profile, LED strip peak response, entrance backlight coefficient, and static vertical structure density; assuming the image pixel domain is... , The total number of pixels in the entire image, for historical frames. Its sports occupancy rate and overexposure They are respectively:

[0019] , ;

[0020] The low-motion, low-overexposure frame is to satisfy and Historical frames;

[0021] in, Represents the pixel domain of an image. Indicates the total number of pixels in the image. Indicates the historical frame index. Representing historical frames The occupancy rate of the exercise, Representing historical frames Overexposure rate and These represent historical frames. and the pixel position of the previous frame The brightness value at that location, The threshold for motion determination. The threshold for determining overexposure, This is an indicator function; it takes the value 1 if the condition inside the parentheses is true, and 0 otherwise. This represents the absolute value operation; The threshold for exercise occupancy rate, The overexposure threshold;

[0022] S13, Set scene state entries The corresponding set of historical frames is The number of historical frames is Scene state entries The state descriptor belongs to Cluster centers of historical frame illumination topology descriptors are determined; corresponding anchored brightness maps are obtained. Confidential graph And reflective memory map They are respectively:

[0023] , , ;

[0024] in, Represents scene state entries The corresponding set of historical frames, This indicates the number of historical frames in the historical frame set. Represents scene state entries At pixel position Anchor brightness value at the location, Represents scene state entries At pixel position Credibility of the location Represents scene state entries At pixel position The reflection at that location occupies memory value. For historical frames The positive weighting coefficients, and , The threshold for highlighting is set. The threshold for judging low color difference is... Representing historical frames At pixel position Overall color difference at the location, This represents the logical AND operation. Indicates the set of historical frames variance calculation on For the variance scaling parameter in the construction of the credibility graph, and , is the positive stability constant in the corresponding formula, and This is used to avoid situations where the denominator is zero, the logarithmic term is undefined, or the numerical value is unstable during calculation. It appears in different formulas. The specific value is preset by the corresponding calculation step. This represents the natural exponential function.

[0025] Preferably, in step S1, low-motion, low-overexposure frames are selected from historical videos of the same camera, and lighting topology descriptors such as row brightness profile, column brightness profile, light strip peak response, entrance backlight coefficient, and static vertical structure density are extracted. These historical frames are then clustered to form multiple scene state entries. Simultaneously, based on the historical frame set corresponding to each scene state entry, state descriptors, anchored brightness maps, confidence maps, and reflectivity occupancy memory maps are generated. By memorizing the long-term stable lighting distribution, static facility brightness distribution, and reflectivity spatial distribution in the warehouse scene, a stable and scene-specific reference can be provided for subsequent enhancement processing, thereby improving the adaptability and stability of the enhancement results.

[0026] Preferably, in step S2, the target scene state entry corresponding to the current frame is determined, and the matching confidence of the current frame calling the historical scene state memory is obtained. The specific steps are as follows:

[0027] The current frame illumination topology descriptor is denoted as Scene state entries The state descriptor is denoted as ;

[0028] Then the current frame and scene state entries Difference Represented as: ;

[0029] in, This represents the current frame's lighting topology descriptor. Represents scene state entries State descriptor, and These represent the current frame and scene state entries, respectively. The line brightness profile, and These represent the current frame and scene state entries, respectively. The column brightness profile, and These represent the current frame and scene state entries, respectively. The peak response of the light strip, and These represent the current frame and scene state entries, respectively. The inlet backlight coefficient, and These represent the current frame and scene state entries, respectively. The static vertical structural density, Indicates the current frame and scene state entries The degree of difference to These are non-negative weight parameters. express Norm; Matching confidence Represented as: ;

[0030] in, Indicates the current frame and scene state entries Match confidence, For difference compression parameters, and The scene state entry with the smallest difference is selected as the target scene state entry. ,when When introducing a low confidence attenuation coefficient It is represented as: ;

[0031] in, This represents the target scene state entry. Indicates the current frame and target scene state entries Match confidence, To match the confidence threshold, This represents the low-confidence attenuation coefficient corresponding to the current frame. This is the lower confidence limit of decay, and ,when At that time, take The scene state memory is updated according to the following conditions: when consecutive The frames satisfy the minimum difference between each frame. All are greater than the preset update threshold At that time, low motion and low overexposure frames are re-selected from the historical candidate frames of the latest time window and the scene state entries, state descriptors, anchor brightness maps, confidence maps and reflection occupancy memory maps are reconstructed.

[0032] in, The threshold number of frames required to continuously meet the update conditions. The frame number, For the first Frame and Scene State Entries The degree of difference For the first The minimum difference between a frame and all scene state entries. Update the threshold for scene state memory. This indicates all scene state entries. Minimum value operation.

[0033] Preferably, in step S2, by extracting the lighting topology descriptor of the current frame and calculating the difference between the current frame lighting topology descriptor and the state descriptors of each scene state entry, the scene state entry with the smallest difference is selected as the target scene state entry, and the corresponding matching confidence is further determined. Through the above matching process, an adaptive association between the current frame and the historical scene state memory can be realized, so that subsequent enhancement processing will prioritize calling historical prior information that is closer to the current lighting conditions. At the same time, the matching confidence can also serve as the basis for subsequent rollback control and scene state memory update judgment, thereby reducing the risk of false enhancement under scene mismatch conditions.

[0034] Preferably, in step S3, motion interference areas are identified and dynamic occlusion corridors are constructed to suppress the influence of moving targets on augmentation decisions. The specific steps are as follows:

[0035] S31. Construct a temporal motion response diagram It is represented as: ;

[0036] The motion region is determined based on the time-series motion response diagram. : Define the high-confidence static region as ;

[0037] in, Indicates time At pixel position The temporal motion response value at the location, The number of preceding frames that participate in the timing comparison. This is the index for the previous frame count. The pixel position of the current frame The local displacement vector at that location. Let be the magnitude of the local displacement vector. , , These are non-negative weight parameters. Indicates time At pixel position The movement area marker at that location, Threshold for determining the motion area, Indicates a high-confidence static region. For high-confidence static region threshold, Indicates set construction, Indicates a "satisfaction" relationship;

[0038] S32, In a high-confidence static region Perform Hough line detection internally, selecting lines whose cumulative length is greater than a preset length threshold. The direction of the main conductor structure serves as the direction of the warehouse passage. When no main conductor structure meeting the conditions is detected, the statistical high-confidence static region is determined. The edge response intensity in different directions is used to select the direction with the highest response intensity as the warehouse passage direction. ;in, The cumulative length threshold of the main conductor structure. This is the direction of the warehouse passageway;

[0039] S33, when the input image width is Height is When, define relative to the reference resolution The scaling factor is And scale the extended length and extended width of the oriented structural element according to the scaling factor: , , utilizing along direction Oriented structural elements For the movement area Perform morphological orientation expansion to obtain the dynamically occluded corridor area. ;

[0040] in, The width of the input image. The height of the input image. This is a scaling factor relative to the reference resolution. and These represent the extended length and extended width of the oriented structuring element, respectively. and These are the reference extension length and reference extension width at the reference resolution, respectively. Indicates along direction , length is Width is Oriented structural elements, Indicates dynamic occlusion of the corridor area. This represents the morphological dilation operation.

[0041] Preferably, in step S3, the motion region in the current frame is determined by constructing a temporal motion response map, and the warehouse passage direction is determined by combining the Hough line detection results or the statistical results of edge response intensity in different directions within the high-confidence static region. Then, morphological orientation expansion of the motion region is performed using directional structural elements set along the warehouse passage direction to obtain a dynamic occlusion corridor area. By expanding the motion region along the warehouse passage direction, the potential occlusion impact range caused by the movement of personnel, forklifts, or cargo carts along the passage can be covered, thereby reducing the interference of dynamic targets and their adjacent impact areas on subsequent brightness enhancement judgments and helping to suppress miscorrection caused by motion occlusion.

[0042] Preferably, in step S4, the reflective uncertainty region is determined and a candidate enhanced brightness map is generated to achieve controlled enhancement of the reliable static region. The specific steps are as follows:

[0043] Extract the highlight and low chromatic aberration regions from the current frame. It is represented as:

[0044] The high-brightness, low-color-difference area Target scene state entries Reflective memory map Combining these elements to determine the area of ​​reflection uncertainty.

[0045] : Construct a static anchor weight graph : ;

[0046] in, Indicates time At pixel position Highlighted, low-color-difference areas are marked. The threshold for highlighting the current frame. The threshold for determining low color difference in the current frame. Indicates time At pixel position The reflective uncertainty area marking at the location, The threshold for determining reflectivity in memory. Indicates time At pixel position The static anchoring weights at the location; the base enhancement gain is determined based on the logarithmic difference between the current frame luminance and the anchor luminance map, where the logarithmic difference... and basic enhancement gain They are respectively: ; ;

[0047] in, By limiting the enhancement gain, the final enhancement gain is obtained. : ;

[0048] in, And satisfy ;

[0049] Candidate Enhanced Brightness Map Represented as: ;

[0050] in, Indicates time At pixel position Logarithmic difference at point, Represents the natural logarithm function. To prevent abnormally positive stability constants in the logarithm and denominator, and ; Indicates pixel position The base enhancement gain, and These are the lower and upper bound truncation parameters for the logarithmic difference, respectively. This represents the interval cutoff function. For the variable to be truncated, and These represent the lower and upper bounds, respectively. Indicates pixel position The final enhancement gain at that point, This represents the clipping function that performs interval clipping on the enhancement gain. and These are the lower limit and upper limit of the enhanced gain, respectively. This is a range clipping function to crop values ​​to a preset brightness range.

[0051] Preferably, in step S4, the high-brightness, low-chromatic-difference region is extracted from the current frame, and the reflectivity occupancy memory map of the target scene state entry is combined to determine the reflectivity uncertainty region. Simultaneously, a static anchoring weight map is constructed based on the confidence map, the dynamic occlusion corridor region, and the reflectivity uncertainty region. A candidate enhanced brightness map is generated based on the logarithmic difference between the current frame brightness and the anchoring brightness map. By jointly constraining the high-brightness, low-chromatic-difference information in the current frame with the historical reflectivity spatial distribution information, unreliable enhancement in the reflectivity region can be effectively limited. Furthermore, by guiding the generation of candidate enhanced brightness through static anchoring weights, more targeted brightness enhancement can be achieved within the confidence static region, thereby balancing the dark area enhancement effect and the false enhancement suppression effect.

[0052] Preferably, in step S5, the candidate enhancement results are evaluated for credibility and then back-down fusion is performed to generate a more robust back-up enhancement brightness map. The specific steps are as follows:

[0053] Constructing an enhanced confidence graph It is represented as: ;

[0054] in, For the Logistic compression function: The result of the static anchor weight map after low-confidence decay is used as the backoff anchor weight. ,Right now ,

[0055] Revert to enhanced brightness map Represented as: ;

[0056] when In this case, the backoff anchoring weight is further expressed as: ;

[0057] in, , Indicates time At pixel position Increase confidence at the location, It is a natural constant. to These are non-negative weight parameters; Indicates the rollback of anchor weights; This indicates the reverted brightness enhancement map at the pixel location. The brightness value at that location; This indicates the lower limit of low-confidence decay.

[0058] Preferably, in step S5, an enhancement confidence map is constructed by comprehensively considering the matching confidence, confidence map, dynamic occlusion corridor area, reflective uncertainty area, and the deviation of the candidate enhanced brightness map from the anchor brightness map. When the matching confidence is lower than a preset confidence threshold, a low-confidence attenuation is applied to the backoff anchor weight, and the candidate enhanced brightness map, the current frame brightness map, and the anchor brightness map are back-fused to obtain a back-fused enhanced brightness map. Through the above enhancement confidence evaluation and low-confidence backoff mechanism, the enhancement effect can be retained when the candidate enhancement result has high confidence, and the retention of the current frame brightness information can be improved under conditions of state mismatch, strong reflection, or dynamic occlusion, thereby effectively suppressing local brightness abrupt changes, flickering, and over-enhancement phenomena.

[0059] Preferably, in step S6, the regressed enhanced brightness map is optimized to balance the brightness enhancement effect and the structure preservation effect. The specific steps are as follows:

[0060] S61. Constructing the task-sensitive area The task-sensitive area consists of the detected cargo outline region. , passage boundary area Identification text area and personnel work area One or more of the components are represented as: ;

[0061] Among them, the cargo outline area Channel boundary regions are extracted using edge detection and morphological closing operations. Text regions are extracted and identified using Hough line detection. Extracted from text region detection, personnel work area Extracted from pedestrian area detection; Indicates time Task-sensitive areas; , , , These respectively represent the cargo outline area, the passage boundary area, the signage text area, and the personnel operation area; This represents the set union operation;

[0062] S62. When the task-sensitive area is empty, in the entire map area... Internal execution optimization; when the task-sensitive area is not empty, in the entire graph region. excluding the pixel domain corresponding to the task-sensitive area Optimize areas outside the region; revert to enhanced brightness map. As a reference brightness diagram And the brightness map to be optimized The initial value is taken as In the pixel domain where optimization is performed Built-in optimized functionals: ;

[0063] in, Indicates the pixel domain for which optimization is performed. , , ;in, Indicates time The optimized functional, Indicates the reference brightness map at the pixel position The brightness value at that location, This indicates the pixel location of the brightness map to be optimized. The brightness value at that location, This represents the pixel domain corresponding to the task-sensitive area. This represents the pixel domain for which optimization is performed, and is true when the task-sensitive region is empty. When the task-sensitive area is not empty, , The gradient regularization term weights are represented by the parameters. This represents the positive stability constant in the logarithmic fidelity term. This represents the positive stability constant in the gradient regularization term. Represents the gradient operator, Indicates pixel position The squared magnitude of the gradient vector at that point. Represents the integral measure over the pixel domain. This represents the set difference operation;

[0064] S63. Update the brightness map to be optimized iteratively as follows: ;

[0065] in, The iteration step size, The iteration number, Indicates the first In the next iteration, the brightness map to be optimized is located at the pixel position. The brightness value at that location; Indicates the first In the next iteration, the brightness map to be optimized is located at the pixel position. The brightness value at that location, Denotes the divergence operator, This indicates a comparison with zero followed by taking the larger value; when the task-sensitive region is not empty, for Pixel preservation Unchanged, for The pixels in the image are updated iteratively as described above, with zero normal derivative boundary conditions applied at the boundaries. After a preset number of iterations, an optimized and enhanced brightness map is obtained. , including the entire map area Except for mission-sensitive areas The area outside is the area where updates will be performed. This indicates the location of the optimized and enhanced brightness map at the pixel position. The brightness value at that location, The preset number of iterations, and It is a positive integer.

[0066] Preferably, in step S6, by constructing a task-sensitive region and establishing an optimized functional containing logarithmic fidelity and gradient regularization terms within the pixel definition domain of the optimization, the backtracking enhanced brightness map is iteratively optimized to obtain an optimized enhanced brightness map. When the task-sensitive region is empty, the pixel definition domain of the optimization is the entire image region; when the task-sensitive region is not empty, the pixel definition domain of the optimization is the region outside the pixel definition domain corresponding to the task-sensitive region within the entire image region. The task-sensitive region includes one or more of the following: cargo outline region, channel boundary region, signage text region, and personnel operation region. Through the above optimization process, while maintaining the overall brightness distribution of the backtracking enhancement result, local discontinuities, noise amplification, and structural distortion can be further suppressed. This also helps maintain the structural stability of the cargo outline, channel boundary, signage text, and personnel operation region, thereby improving the applicability of the enhancement result to subsequent identification, detection, and monitoring analysis tasks.

[0067] Preferably, in step S7, the optimized brightness enhancement result is synchronously mapped to each color channel to output an enhanced video frame with good overall color retention. The specific steps are as follows:

[0068] A shared gain is constructed based on the ratio between the optimized enhanced luminance map and the current frame luminance map. It is represented as: Then, synchronization mapping is performed on each color channel of the original color video frame:

[0069] ; ; ;

[0070] in, Indicates time At pixel position Shared gain at the location, , , These are the red, green, and blue color channel values ​​after synchronous mapping. This is a function for cropping values ​​to a preset color range.

[0071] Preferably, in step S7, a shared gain is constructed based on the ratio between the optimized enhanced brightness map and the current frame brightness map, and the shared gain is synchronously applied to each color channel of the original color video frame to output an enhanced video frame. By performing synchronous mapping on the red, green, and blue color channels, the original proportional relationship between each color channel can be maintained while improving the overall brightness performance, thereby reducing the overall color shift and color distortion, and ensuring that the enhanced video frame maintains good visual naturalness while improving brightness.

[0072] Compared with existing technologies, the present invention has the following beneficial effects: First, by constructing scene state entries under multiple lighting conditions using historical video from the same camera, and providing anchored brightness maps, confidence maps, and reflectivity occupancy memory maps based on the target scene state entries, the scene targeting of current frame brightness enhancement can be improved; Second, by constructing dynamic occlusion corridor areas along the warehouse aisle direction, miscorrection caused by the movement of personnel, forklifts, and cargo carts along the aisle can be reduced; Third, by combining the current frame's high-brightness, low-color-difference areas with historical reflectivity occupancy memory maps to determine reflectivity uncertainty areas, misenhancement of high-brightness reflectivity areas can be suppressed; Fourth, by using enhanced confidence maps, low-confidence attenuation, and back-off fusion mechanisms, local flickering and brightness abrupt changes caused by state mismatch, strong reflectivity, and dynamic occupancy can be reduced; Fifth, by optimizing brightness in task-sensitive areas and using shared gain color synchronization mapping, the structural stability of cargo outlines, aisle boundaries, signage, and personnel work areas can be maintained while enhancing brightness, and overall color shift can be reduced. Attached Figure Description

[0073] Figure 1 This is a flowchart illustrating the overall process of a warehouse monitoring video enhancement method based on scene state memory, according to the present invention.

[0074] Figure 2 This is a flowchart of the scene state memory construction process in this invention.

[0075] Figure 3 This is a flowchart of the current frame state matching and low confidence control process in this invention.

[0076] Figure 4 This is a flowchart of the process of constructing the dynamically obscured corridor area in this invention.

[0077] Figure 5 This is a flowchart of the process for identifying reflective uncertainty areas and generating candidate enhanced brightness maps in this invention.

[0078] Figure 6 This is a flowchart of the enhanced confidence back-off fusion, optimization enhancement, and color synchronization mapping output process in this invention. Detailed Implementation

[0079] The following will combine Figures 1 to 6 The technical solutions in the embodiments of the present invention will be clearly and completely described. 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.

[0080] Please see Figures 1 to 6 This invention provides a warehouse monitoring video enhancement method based on scene state memory. It selects low-motion, low-overexposure frames from historical videos of the same camera and constructs scene state entries to provide historical prior information matching the current lighting state for current frame enhancement. It suppresses the interference of dynamic occlusion on enhancement decisions by constructing a dynamic occlusion corridor area based on temporal motion response and warehouse aisle direction. It improves the stability of enhancement results by constructing a reflection uncertainty area and a static anchoring weight map based on high-brightness, low-color-difference regions, reflection occupancy memory map, and confidence map, and performs confidence back-off fusion in conjunction with candidate enhancement brightness maps. Finally, it performs variational constraint optimization on the back-off enhancement brightness map and synchronously maps each color channel according to the shared gain relationship between the optimized enhancement brightness map and the current frame brightness map, outputting the enhanced video frame.

[0081] Please see Figure 1 As shown in the figure, a warehouse monitoring video enhancement method based on scene state memory in this application embodiment includes the following steps.

[0082] S1. Filter low-motion, low-overexposure frames from the historical videos of the same camera, extract the lighting topology descriptor composed of row brightness profile, column brightness profile, light strip peak response, entrance backlight coefficient and static vertical structure density, cluster them to form scene state entries, and generate state descriptors, anchor brightness maps, confidence maps and reflection occupancy memory maps.

[0083] Furthermore, step S1 specifically includes the following sub-steps.

[0084] S11. Represent the current color video frame as... Its brightness diagram is shown as follows: Defines the pixel position of the current frame. The overall color difference at the location is: ;

[0085] In the formula: Indicates time The current color video frame; , , These represent the red, green, and blue color channels of the current color video frame, respectively. This represents the luminance map corresponding to the current color video frame; , , For the corresponding color channel, the non-negative luminance weighting coefficient is , and satisfies . ; Indicates time The current frame at pixel position In a preferred embodiment, the overall color difference at the location can be taken as follows: , , In an optional implementation, it is possible to take , , In implementation, the pixel values ​​of the current frame and historical frames are preferably normalized to [values ​​to be specified in the original text]. Range; for 8-bit images, you can first use , , For 10-bit images, divide by 1023; for other bit depths, normalize to the full-scale value and then map back to the corresponding output bit depth when display needs to be restored.

[0086] S12, The lighting topology descriptor includes row luminance profile, column luminance profile, peak response of the light strip, entrance backlight coefficient, and static vertical structure density.

[0087] Furthermore, for historical videos from the same warehouse camera, sampling intervals are set... Extract historical frames to obtain a historical brightness map sequence. Sampling interval The settings can be adjusted based on the video frame rate and the speed of scene changes; the optimal range is [insert range here]. To reduce the impact of dynamic targets and extreme overexposure on state memory, historical frames that simultaneously meet the following conditions are preferred as candidate steady-state frames: , In the formula: For the first Motion occupancy rate of each historical frame For the first Overexposure rate of each historical frame, The threshold for exercise occupancy rate, The overexposure rate threshold; where, , In a preferred embodiment, , For any candidate steady-state frame, extract the illumination topology descriptor. Where, if the image height is , width is Then the first The row brightness profile can be represented as: ;No. The column brightness profile can be represented as: ; Entrance backlight coefficient It can be represented as ;

[0088] In the formula: This is the entrance area. For reference area within the library, To prevent the use of tiny constants with a denominator of zero, it is preferable to take a value of [value missing]. In a preferred embodiment, Static vertical structural density It can be represented as In the formula: Image region; For the first At pixel location in a historical frame Vertical edge response at the location; For the vertical edge determination threshold, preferably, Peak response of LED strip This can be obtained through statistics on peak brightness in the top light strip area or a known light strip area, for example. The entrance area Reference area within the library and light strip area It can be manually calibrated during system initialization, or automatically determined based on the structural and brightness statistical characteristics of historical steady-state frames.

[0089] S13. Cluster the candidate steady-state frames according to the lighting topology descriptor to obtain multiple scene state entries, and generate state descriptors, anchor brightness maps, confidence maps and reflection occupancy memory maps.

[0090] Furthermore, the clustering method can employ k-medoids, hierarchical clustering, density clustering, or other methods suitable for state grouping. In a preferred embodiment, k-medoids clustering is used, and the degree of difference between states can be expressed as: ;

[0091] In the formula: to For cluster distance weights, satisfying , ;in, , In a preferred embodiment, it is advisable to , , , , Set scene state entries The corresponding set of historical frames is The number of historical frames is Then anchor brightness map Confidential graph And reflective memory map They are respectively: ; ; ;

[0092] In the formula: For the first The statistical weights of each historical frame, The threshold for highlighting is set. The threshold for judging low color difference is... For the first At pixel location in a historical frame Overall color difference at the location, For credibility normalization parameters, To prevent tiny constants with a denominator of zero;

[0093] in, It can be in the following form: ;in, , ; In a preferred embodiment, , ; Statistical weights A constant of 1 can be used, or a weighted average can be applied based on the quality of historical frames. In the formula: , The weighting coefficients are preferred, with the following range: .

[0094] S2. Extract the lighting topology descriptor of the current frame and determine the target scene state entries and matching confidence.

[0095] Further, step S2 specifically involves: the current frame illumination topology descriptor is denoted as... and the state descriptor of each scene state entry. The comparison is performed; where the current frame illumination topology descriptor is represented as: Scene state entries The state descriptor is represented as: Then the current frame and scene state entries Difference Represented as: ;

[0096] in, This represents the current frame's lighting topology descriptor. Represents scene state entries State descriptor, and These represent the current frame and scene state entries, respectively. The line brightness profile, and These represent the current frame and scene state entries, respectively. The column brightness profile, and These represent the current frame and scene state entries, respectively. The peak response of the light strip, and These represent the current frame and scene state entries, respectively. The inlet backlight coefficient, and These represent the current frame and scene state entries, respectively. The static vertical structural density, Indicates the current frame and scene state entries The degree of difference to These are non-negative weight parameters. express Norm, This represents the absolute value operation, where the weight parameter satisfies... , In a preferred embodiment, based on the different scene state discrimination capabilities of row brightness profile, column brightness profile, light strip peak response, entrance backlight coefficient, and static vertical structure density, higher weights can be assigned to row brightness profile and column brightness profile, while lower weights can be assigned to light strip peak response, entrance backlight coefficient, and static vertical structure density, to match confidence levels. Represented as: ;

[0097] in, Indicates the current frame and scene state entries Match confidence; To prevent positive stability constants with a denominator of zero, where, In a preferred embodiment, take The scene state entry with the smallest difference is selected as the target scene state entry. ,when When introducing a low confidence attenuation coefficient It is represented as: ;

[0098] in, This represents the target scene state entry. Indicates the current frame and target scene state entries Match confidence, To match the confidence threshold, This represents the low-confidence attenuation coefficient corresponding to the current frame. This is the lower confidence limit of decay, and , This indicates the operation of finding the maximum value. At that time, take The scene state memory is updated according to the following conditions: when consecutive The frames satisfy the minimum difference between each frame. All are greater than the preset update threshold At that time, low-motion, low-overexposure frames are re-selected from the historical candidate frames of the latest time window, and scene state entries, state descriptors, anchor brightness maps, confidence maps, and reflection occupancy memory maps are reconstructed; among them, The threshold number of frames required to continuously meet the update conditions; The frame number, For the first Frame and Scene State Entries The degree of difference For the first The minimum difference between a frame and all scene state entries. Update the threshold for scene state memory. This indicates all scene state entries. Minimum value operation.

[0099] S3. Determine the motion area based on the temporal motion response, and expand the motion area in a directional manner according to the warehouse passage direction detected in the high-confidence static area to obtain a dynamic shading corridor area.

[0100] Furthermore, step S3 specifically includes the following sub-steps.

[0101] S31. Construct a temporal motion response diagram It is represented as: The motion region is determined based on the time-series motion response diagram. : The high-confidence static region is defined as follows: ;

[0102] In the formula: The pixel position of the current frame The brightness value at that location, For the first Frame at pixel position The brightness value at that location, The number of preceding frames that participate in the timing comparison. The pixel position of the current frame The local displacement vector at that location. For target scene state entries credible graph, For target scene state entries Anchored brightness map; , , These are weight parameters; The threshold for motion determination. To prevent tiny constants with a denominator of zero;

[0103] in, In a preferred embodiment, The weight parameters satisfy , , , , , , In a preferred embodiment, , , , , Local displacement vector It can be obtained through block matching, adjacent frame optical flow, or feature point tracking. Preferably, it can be calculated using the dense optical flow method. And have .

[0104] S32, In a high-confidence static region Perform Hough line detection internally, selecting lines whose cumulative length is greater than a preset length threshold. The direction of the main conductor structure serves as the direction of the warehouse passage. When no main conductor structure meeting the conditions is detected, the statistical high-confidence static region is determined. The edge response intensity in different directions is used to select the direction with the highest response intensity as the warehouse passage direction. ,in, Indicates a highly reliable static region; The cumulative length threshold of the main conductor structure; when the input image resolution is... hour, Pixel; in a preferred embodiment, Pixel; This refers to the direction of the warehouse passage. In one implementation, this can be achieved in a high-confidence static area. The inner edge map is subjected to Hough line detection, and the length of the detected line segments is accumulated. When the accumulated length in a certain direction exceeds a preset length threshold, the line detection is performed. At that time, this direction is selected as the warehouse passageway direction. When no main conductor structure meeting the conditions is detected, it can be used in the high-confidence static region. The edge response intensity in different directions is statistically analyzed, and the direction with the largest response intensity is selected as the direction of the warehouse passage. .

[0105] S33, when the input image width is Height is When, define relative to the reference resolution The scaling factor is The extended length and extended width of the oriented structural element are scaled according to the scaling factor. , ; Utilize along direction Oriented structural elements For the movement area Perform morphological orientation expansion to obtain the dynamically occluded corridor area. : ;

[0106] in, The width of the input image; The height of the input image; This is the scaling factor relative to the reference resolution; and These are the extended length and extended width of the oriented structural element, respectively; and These are the reference extension length and reference extension width at the reference resolution, respectively; preferably, at the reference resolution... Down, Pixels Pixel; in a preferred embodiment, may be Pixels Pixel; Indicates along direction , length is Width is Oriented structural elements; Indicates the area of ​​motion; Indicates dynamic occlusion of the corridor area; Represents morphological dilation operations; In one embodiment, the reference extension length represents the minimum value operation. and reference extended width The settings can be preset according to the width of the warehouse aisle, the installation height of the camera, and the monitoring angle, so that the dynamic shading corridor area covers the dynamic shading influence range that propagates along the direction of the warehouse aisle.

[0107] S4. Determine the uncertain area of ​​reflection based on the high-brightness, low-color-difference area and the reflection occupancy memory map, and generate a candidate enhanced brightness map by combining the confidence map, the dynamic occlusion corridor area and the uncertain area of ​​reflection.

[0108] Further, step S4 specifically involves: extracting the bright, low-chromatic-difference region from the current frame. It is represented as: ;

[0109] in, The threshold for highlighting the current frame. The threshold for determining low color difference in the current frame. The pixel position of the current frame The overall color difference at the location, of which , In a preferred embodiment, take , The high-brightness, low-color-difference area Target scene state entries Reflective memory map Combining these elements to determine the area of ​​reflection uncertainty. : ;

[0110] in, Indicates time At pixel position The reflective uncertainty area marking at the location, Represents the target scene state entry At pixel position The reflection at that location occupies memory value. The threshold for determining reflectivity occupancy is the memory threshold. A static anchoring weight map can be constructed by pre-setting the degree of repeated occupancy of historical reflective areas in the scene. : ;in, Indicates time At pixel position Static anchor weights at the location, Represents the target scene state entry At pixel position Credibility of the location Indicates time At pixel position Dynamic occlusion corridor markings are used; the base enhancement gain is determined based on the logarithmic difference between the current frame luminance and the anchored luminance map, where the logarithmic difference... and basic enhancement gain They are respectively: ; ;

[0111] in, , To prevent abnormally positive stability constants in the logarithm and denominator, and , and These are the lower and upper bound truncation parameters for the logarithmic difference, respectively. , In a preferred embodiment, take , Limiting the enhancement gain yields the final enhancement gain. ;

[0112] in, and satisfy ;in, and These are the lower limit and upper limit of the enhancement gain, respectively; and Can be preset based on the maximum allowable enhancement amplitude in the dynamic occlusion area and the uncertain reflective area, and candidate enhancement brightness map. Represented as: ;

[0113] in, Indicates the candidate enhanced brightness map at the pixel location. The brightness value at that location; In a preferred embodiment, the cropping function is used to crop pixels to a preset brightness range. This is done when the pixel value is normalized to... When the range is within a certain interval, the range clipping function can be expressed as: ;

[0114] in, Indicates time At pixel position The logarithmic difference at the point; Represents the interval cutoff function; The variable to be truncated; and These represent the lower and upper bounds, respectively. This represents a clipping function that performs interval clipping on the enhancement gain; This represents a clipping function that clips the brightness value range.

[0115] S5. Construct an enhanced confidence map based on the matching confidence, confidence map, dynamic occlusion corridor area, reflective uncertainty area, and the degree of deviation between the candidate enhanced brightness map and the anchor brightness map. When the matching confidence is lower than the confidence threshold, backfuse the candidate enhanced brightness map, the current frame brightness map, and the anchor brightness map to obtain the backfuse enhanced brightness map.

[0116] Furthermore, step S5 specifically involves: constructing an enhanced confidence graph. Its expression is: ;

[0117] In the formula: For target scene state entries Match confidence, For target scene state entries credible graph, To dynamically occlude the corridor area at pixel position The area indicator quantity, For the uncertain area of ​​reflection at the pixel position The area indicator quantity, Candidate brightness enhancement maps, For target scene state entries Anchored brightness map, to These are weight parameters; To prevent tiny constants with a denominator of zero, the function is compressed. The Logistic function is preferred: The weight parameters are preferably satisfied. , ;in, , , , , In a preferred embodiment, , , , , Furthermore, define static anchor weights. ;when When introducing a low confidence attenuation coefficient The backoff anchor weight is represented as: ,when At that time, take Then revert to enhance brightness map It can be represented as: ;

[0118] In the formula: Indicates static anchor weights. Indicates the low-confidence attenuation coefficient. This indicates the lower limit of low-confidence decay. This indicates a rollback to the anchor weight. This represents the current frame's luminance map. This indicates a rollback to enhance brightness.

[0119] S6. Construct an optimized functional containing log-fidelity terms and gradient regularization terms, and iteratively optimize the backtracking enhanced brightness map to obtain the optimized enhanced brightness map.

[0120] Furthermore, step S6 specifically includes the following sub-steps.

[0121] S61. Constructing the task-sensitive area The task-sensitive area From the detected cargo outline area , passage boundary area Identification text area and personnel work area One or more of the components are represented as: ;

[0122] Among them, the cargo outline area Channel boundary regions can be extracted through edge detection and morphological closing operations. Text regions can be extracted and identified using Hough line detection. The personnel work area can be extracted using text detection algorithms. It can be extracted using a pedestrian detection algorithm. If any type of region is not detected, the corresponding region is recorded as an empty set, without affecting the participation of the remaining regions in subsequent optimization.

[0123] S62. Construct an optimized functional. When the task-sensitive region is empty, in the entire graph region... Internal execution optimization; when the task-sensitive area is not empty, in the entire graph region. excluding the pixel domain corresponding to the task-sensitive area Optimize areas outside of the target region to revert to an enhanced brightness map. As a reference brightness diagram And the brightness map to be optimized The initial value is taken as In optimizing the pixel definition domain Built-in optimized functionals: ;

[0124] in, This represents the pixel domain for which optimization is performed, and is true when the task-sensitive region is empty. When the task-sensitive area is not empty, , The weight parameters are for the gradient regularization term. The positive stability constant in the logarithmic fidelity term. This is the positive stability constant in the gradient regularization term. Denotes the gradient operator; where, , , In a preferred embodiment, take , , .

[0125] S63. Iteratively update the brightness map to be optimized and obtain the optimized and enhanced brightness map. Update the brightness map to be optimized iteratively as follows: ,

[0126] in, The iteration step size, Let be the iteration number, where , In a preferred embodiment, take , , Represents the divergence operator, which, when the task-sensitive region is non-empty, for Pixel preservation Unchanged, for The pixels in the image are updated iteratively as described above, with zero normal derivative boundary conditions applied at the boundaries. After a preset number of iterations, an optimized and enhanced brightness map is obtained. , including the entire map area Except for mission-sensitive areas The area outside is the area where updates will be performed. This indicates the location of the optimized and enhanced brightness map at the pixel position. The brightness value at that location, The preset number of iterations, and As a positive integer, the above optimization process can further suppress local discontinuities, noise amplification, and structural distortion while maintaining the overall brightness distribution of the back-up enhancement results. It also helps to maintain the structural stability of cargo outlines, channel boundaries, signage text, and personnel work areas, thereby improving the applicability of the enhancement results to subsequent identification, detection, and monitoring analysis tasks.

[0127] S7. Construct a shared gain based on the ratio of the optimized enhanced luminance map to the current frame luminance map, perform synchronous mapping on each color channel of the current frame, and output the enhanced video frame.

[0128] Further, step S7 specifically involves: constructing a shared gain based on the ratio between the optimized enhanced luminance map and the current frame luminance map. It is represented as: Then, synchronization mapping is performed on each color channel of the original color video frame: , , ;

[0129] in, Indicates time At pixel position Shared gain at the location, This indicates the location of the optimized and enhanced brightness map at the pixel position. The brightness value at that location, Indicates the pixel position of the current frame's luminance map. The brightness value at that location, , , The pixel positions of the current frame are respectively The red, green, and blue color channel values ​​at that location. , , These are the red, green, and blue color channel values ​​after synchronous mapping. The clipping function is for cropping to a preset color range, since the three color channels share the same shared gain. Therefore, it can improve brightness and contrast while maintaining the proportional relationship between the color channels of the original image, reducing the overall color shift and local color distortion problems, thus outputting enhanced video frames.

[0130] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A warehouse monitoring video enhancement method based on scene state memory, characterized in that, Includes the following steps: S1. Filter low motion and low overexposure frames from the historical video of the same camera, extract the lighting topology descriptor composed of row brightness profile, column brightness profile, light strip peak response, entrance backlight coefficient and static vertical structure density, cluster them to form scene state entries, and generate state descriptors, anchor brightness map, confidence map and reflection occupancy memory map. S11. Represent the current color video frame as... ; Its brightness diagram is shown as follows: , Define the pixel position of the current frame The overall color difference at the location is: ; in, Indicates time The current color video frame; , , These represent the red, green, and blue color channels of the current color video frame, respectively. This represents the luminance map corresponding to the current color video frame; , , For the corresponding color channel, the non-negative luminance weighting coefficient is , and satisfies . , ; Indicates time The current frame at pixel position Overall color difference at the location; , , These represent the pixel positions of the current frame. The red, green, and blue color channel values ​​at that location; and These represent the operations of finding the maximum and minimum values, respectively. S12, the lighting topology descriptor includes row brightness profile, column brightness profile, LED strip peak response, entrance backlight coefficient, and static vertical structure density; assuming the image pixel domain is... , The total number of pixels in the entire image, for historical frames. Its sports occupancy rate and overexposure They are respectively: ， ； The low-motion, low-overexposure frame is to satisfy and Historical frames; in, Represents the pixel domain of an image. Indicates the total number of pixels in the image. Indicates the historical frame index. Representing historical frames The occupancy rate of the exercise, Representing historical frames Overexposure rate and These represent historical frames. and the pixel position of the previous frame The brightness value at that location, The threshold for motion determination. The threshold for determining overexposure, This is an indicator function; it takes the value 1 if the condition inside the parentheses is true, and 0 otherwise. This represents the absolute value operation; The threshold for exercise occupancy rate, The overexposure threshold; S13, Set scene state entries The corresponding set of historical frames is The number of historical frames is Scene state entries The state descriptor belongs to Cluster centers of historical frame illumination topology descriptors are determined; corresponding anchored brightness maps are obtained. Confidential graph And reflective memory map They are respectively: ， ， ； in, Represents scene state entries The corresponding set of historical frames, This indicates the number of historical frames in the historical frame set. Represents scene state entries At pixel position Anchor brightness value at the location, Represents scene state entries At pixel position Credibility of the location Represents scene state entries At pixel position The reflection at that location occupies memory value. For historical frames The positive weighting coefficients, and , The threshold for highlighting is set. The threshold for judging low color difference is... Representing historical frames At pixel position Overall color difference at the location, This represents the logical AND operation. Indicates the set of historical frames variance calculation on For the variance scaling parameter in the construction of the credibility graph, and , is the positive stability constant in the corresponding formula, and This is used to avoid situations where the denominator is zero, the logarithmic term is undefined, or the numerical value is unstable during calculation. It appears in different formulas. The specific value is preset by the corresponding calculation step. Represents the natural exponential function; S2. Extract the current frame lighting topology descriptor and determine the target scene state entries and matching confidence. S3. Determine the motion area based on the temporal motion response, and expand the motion area in a directional manner based on the warehouse passage direction detected in the high-confidence static area to obtain the dynamic occlusion corridor area. S4. Determine the uncertain area of ​​reflection based on the high-brightness low-color-difference area and the reflection occupancy memory map, and generate a candidate enhanced brightness map by combining the confidence map, the dynamic occlusion corridor area and the uncertain area of ​​reflection. S5. Construct an enhancement confidence map based on the matching confidence, confidence map, dynamic occlusion corridor area, reflective uncertainty area, and the degree of deviation between the candidate enhanced brightness map and the anchor brightness map. When the matching confidence is lower than the confidence threshold, backfuse the candidate enhanced brightness map, the current frame brightness map, and the anchor brightness map to obtain the backfuse enhanced brightness map. S6. Construct an optimized functional containing log-fidelity terms and gradient regularization terms, and iteratively optimize the back-down enhanced brightness map to obtain the optimized enhanced brightness map; S7. Construct a shared gain based on the ratio of the optimized enhanced luminance map to the current frame luminance map, synchronously map each color channel of the current frame, and output the enhanced video frame.

2. The warehouse monitoring video enhancement method based on scene state memory according to claim 1, characterized in that, In step S2 The current frame illumination topology descriptor is denoted as Scene state entries The state descriptor is denoted as ; Then the current frame and scene state entries Difference Represented as: ; in, This represents the current frame's lighting topology descriptor. Represents scene state entries State descriptor, and These represent the current frame and scene state entries, respectively. The line brightness profile, and These represent the current frame and scene state entries, respectively. The column brightness profile, and These represent the current frame and scene state entries, respectively. The peak response of the light strip, and These represent the current frame and scene state entries, respectively. The inlet backlight coefficient, and These represent the current frame and scene state entries, respectively. The static vertical structural density, Indicates the current frame and scene state entries The degree of difference to These are non-negative weight parameters. express Norm; Matching confidence Represented as: ; in, Indicates the current frame and scene state entries Match confidence, For difference compression parameters, and The scene state entry with the smallest difference is selected as the target scene state entry. ,when When introducing a low confidence attenuation coefficient It is represented as: ; in, This represents the target scene state entry. Indicates the current frame and target scene state entries Match confidence, To match the confidence threshold, This represents the low-confidence attenuation coefficient corresponding to the current frame. This is the lower confidence limit of decay, and ,when At that time, take The scene state memory is updated according to the following conditions: when consecutive The frames satisfy the minimum difference between each frame. All are greater than the preset update threshold At that time, low motion and low overexposure frames are re-selected from the historical candidate frames of the latest time window and the scene state entries, state descriptors, anchor brightness maps, confidence maps and reflection occupancy memory maps are reconstructed. in, The threshold number of frames required to continuously meet the update conditions. The frame number, For the first Frame and Scene State Entries The degree of difference For the first The minimum difference between a frame and all scene state entries. Update the threshold for scene state memory. This indicates all scene state entries. Minimum value operation.

3. The warehouse monitoring video enhancement method based on scene state memory according to claim 2, characterized in that, Step S3 includes: S31. Construct a temporal motion response diagram It is represented as: ; The motion region is determined based on the time-series motion response diagram. : Define the high-confidence static region as ; in, Indicates time At pixel position The temporal motion response value at the location, The number of preceding frames that participate in the timing comparison. This is the index for the previous frame count. The pixel position of the current frame The local displacement vector at that location. Let be the magnitude of the local displacement vector. , , These are non-negative weight parameters. Indicates time At pixel position The movement area marker at that location, Threshold for determining the motion area, Indicates a high-confidence static region. For high-confidence static region threshold, Indicates set construction, Indicates a "satisfaction" relationship; S32, In a high-confidence static region Perform Hough line detection internally, selecting lines whose cumulative length is greater than a preset length threshold. The direction of the main conductor structure serves as the direction of the warehouse passage. When no main conductor structure meeting the conditions is detected, the statistical high-confidence static region is determined. The edge response intensity in different directions is used to select the direction with the highest response intensity as the warehouse passage direction. ;in, The cumulative length threshold of the main conductor structure. This is the direction of the warehouse passageway; S33, when the input image width is Height is When, define relative to the reference resolution The scaling factor is And scale the extended length and extended width of the oriented structural element according to the scaling factor: , , utilizing along direction Oriented structural elements For the movement area Perform morphological orientation expansion to obtain the dynamically occluded corridor area. ; in, The width of the input image. The height of the input image. This is a scaling factor relative to the reference resolution. and These represent the extended length and extended width of the oriented structuring element, respectively. and These are the reference extension length and reference extension width at the reference resolution, respectively. Indicates along direction , length is Width is Oriented structural elements, Indicates dynamic occlusion of the corridor area. This represents the morphological dilation operation.

4. The warehouse monitoring video enhancement method based on scene state memory according to claim 3, characterized in that, In step S4, the bright and low chromatic aberration regions are extracted from the current frame. It is represented as: ； The high-brightness, low-color-difference area Target scene state entries Reflective memory map Combining these elements to determine the area of ​​reflection uncertainty. : Construct a static anchor weight graph : ; in, Indicates time At pixel position Highlighted, low-color-difference areas are marked. The threshold for highlighting the current frame. The threshold for determining low color difference in the current frame. Indicates time At pixel position The reflective uncertainty area marking at the location, The threshold for determining reflectivity in memory. Indicates time At pixel position The static anchoring weights at the location; the base enhancement gain is determined based on the logarithmic difference between the current frame luminance and the anchor luminance map, where the logarithmic difference... and basic enhancement gain They are respectively: ; ; in, By limiting the enhancement gain, the final enhancement gain is obtained. : ; in, And satisfy ; Candidate Enhanced Brightness Map Represented as: ; in, Indicates time At pixel position Logarithmic difference at point, Represents the natural logarithm function; Indicates pixel position The base enhancement gain, and These are the lower and upper bound truncation parameters for the logarithmic difference, respectively. This represents the interval cutoff function. For the variable to be truncated, and These represent the lower and upper bounds, respectively. Indicates pixel position The final enhancement gain at that point, This represents the clipping function that performs interval clipping on the enhancement gain. and These are the lower limit and upper limit of the enhanced gain, respectively. This is a range clipping function to crop values ​​to a preset brightness range.

5. The warehouse monitoring video enhancement method based on scene state memory according to claim 4, characterized in that, In step S5, an enhanced confidence graph is constructed. It is represented as: ; in, For the Logistic compression function: The result of the static anchor weight map after low-confidence decay is used as the backoff anchor weight. ,Right now , Revert to enhanced brightness map Represented as: ; when In this case, the backoff anchoring weight is further expressed as: ; in, , Indicates time At pixel position Increase confidence at the location, It is a natural constant. to These are non-negative weight parameters; Indicates the rollback of anchor weights; This indicates the reverted brightness enhancement map at the pixel location. The brightness value at that location; This indicates the lower limit of low-confidence decay.

6. The warehouse monitoring video enhancement method based on scene state memory according to claim 5, characterized in that, Step S6 includes: S61. Constructing the task-sensitive area The task-sensitive area consists of the detected cargo outline region. , passage boundary area Identification text area and personnel work area One or more of the components are represented as: ; Among them, the cargo outline area Channel boundary regions are extracted using edge detection and morphological closing operations. Text regions are extracted and identified using Hough line detection. Extracted from text region detection, personnel work area Extracted from pedestrian area detection; Indicates time Task-sensitive areas; , , , These respectively represent the cargo outline area, the passage boundary area, the signage text area, and the personnel operation area; This represents the set union operation; S62. When the task-sensitive area is empty, in the entire map area... Internal execution optimization; when the task-sensitive area is not empty, in the entire graph region. excluding the pixel domain corresponding to the task-sensitive area Optimize areas outside the region; revert to enhanced brightness map. As a reference brightness diagram And the brightness map to be optimized The initial value is taken as In the pixel domain where optimization is performed Built-in optimized functionals: ; in, Indicates the pixel domain for which optimization is performed. , , ;in, Indicates time The optimized functional, Indicates the reference brightness map at the pixel position The brightness value at that location, This indicates the pixel location of the brightness map to be optimized. The brightness value at that location, This represents the pixel domain corresponding to the task-sensitive area. This represents the pixel domain for which optimization is performed, and is true when the task-sensitive region is empty. When the task-sensitive area is not empty, , The gradient regularization term weights are represented by the parameters. This represents the positive stability constant in the logarithmic fidelity term. This represents the positive stability constant in the gradient regularization term. Represents the gradient operator, Indicates pixel position The squared magnitude of the gradient vector at that point. Represents the integral measure over the pixel domain. This represents the set difference operation; S63. Update the brightness map to be optimized iteratively as follows: ; in, The iteration step size, The iteration number, Indicates the first In the next iteration, the brightness map to be optimized is located at the pixel position. The brightness value at that location; Indicates the first In the next iteration, the brightness map to be optimized is located at the pixel position. The brightness value at that location, Denotes the divergence operator, This indicates a comparison with zero followed by taking the larger value; when the task-sensitive region is not empty, for Pixel preservation Unchanged, for The pixels in the image are updated iteratively as described above, with zero normal derivative boundary conditions applied at the boundaries. After a preset number of iterations, an optimized and enhanced brightness map is obtained. , including the entire map area Except for mission-sensitive areas The area outside is the area where updates will be performed. This indicates the location of the optimized and enhanced brightness map at the pixel position. The brightness value at that location, The preset number of iterations, and It is a positive integer.

7. The warehouse monitoring video enhancement method based on scene state memory according to claim 6, characterized in that, In step S7, a shared gain is constructed based on the ratio between the optimized enhanced luminance map and the current frame luminance map. It is represented as: Then, synchronization mapping is performed on each color channel of the original color video frame: ； ； ； in, Indicates time At pixel position Shared gain at the location, , , These are the red, green, and blue color channel values ​​after synchronous mapping. This is a function for cropping values ​​to a preset color range.

Images