Adaptive scene-aware security protection method and system

Abstract

The present application relates to the technical field of life scene space safety protection, in particular to a safety protection method and system with adaptive scene sensing; the method comprises: establishing a far field mapping entry set and a far field static texture database, and calculating a far field similarity label corresponding to a candidate object and a crosstalk reduction weight coefficient; extracting an edge gradient response and determining a bimodal structure to generate a first reflection label; when an image displacement ratio falls within a preset double-layer reflection displacement interval, a second reflection label is generated; an occlusion relationship graph is established, consistency verification is performed according to an occlusion direction and a double-layer reflection parameter constraint, and a third reflection label is generated; an event judgment vector is generated, and the event judgment vector is matched with a preset gate trigger condition to determine whether the candidate object enters a safety protection alarm process. The present application identifies and filters visual interference targets caused by double-layer glass reflection, and adaptively protects safety from the actual problem of false image collision caused by double-layer glass reflection.

Description

Technical Field

[0001] This invention relates to the field of spatial safety protection technology in living scenarios, specifically to a safety protection method and system with adaptive scene perception. Background Technology

[0002] With the widespread application of embedded video surveillance terminals, intelligent recognition algorithms, and embedded behavior modeling technology, the deployment density of intelligent security systems in spatial scenarios such as urban traffic, building entrances and exits, and industrial parks is gradually increasing. These systems mostly rely on image perception and multi-target tracking mechanisms to achieve spatial perception and recognition of moving targets such as people, vehicles, and objects in the scene, and then execute security control logic. However, in complex visual reflection environments such as glass curtain walls, reflective display cabinets, and screen covering panels, conventional image processing and target recognition algorithms have difficulty effectively distinguishing reflective virtual images, which can easily lead to logical judgment errors and thus interfere with the accurate triggering of high-priority protection processes.

[0003] In typical urban public space perception and security scenarios, the following two types of technical problems often exist: First, due to the reflection caused by double-glazed windows, the virtual image in the image has clear edges, a complete structure, and a motion trajectory that approximates the real target. It is easy for the system to misjudge it as the actual foreground object, leading to misidentification, false alarms, or false intervention. Second, in dynamic scenes with occlusion, movement, or multiple objects coexisting, the reflected virtual image is very likely to occlude with the real target. If the system cannot correctly identify its spatial hierarchy and physical imaging consistency, the reflected virtual image may mistakenly enter the foreground path and trigger misleading security strategies, which is especially serious in high-reflectivity spatial scenes. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive scene-aware security protection method and system to solve the two types of technical problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a security protection method for adaptive scene perception, comprising:

[0006] S1. Collect background images of the area outside the double-layer glass, establish a set of far-field mapping entries and a far-field static texture database, and combine the candidate objects and their neighboring background blocks based on the target detection output as the candidate object region of interest. Perform far-distance scene feature mapping matching, and calculate the far-field similarity label and crosstalk weighting coefficient corresponding to the candidate object.

[0007] S2. In the region of interest of the candidate object, extract the edge gradient response along the glass normal and determine the bimodal structure to generate the first reflection label; calculate the image displacement ratio of the candidate object in adjacent frames, and generate the second reflection label when the image displacement ratio falls into the preset double-layer reflection displacement interval; based on the first reflection label and the second reflection label, use the dual label joint rule to determine whether the candidate object conforms to the reflection candidate set.

[0008] S3. Establish an occlusion relationship graph between candidate objects and other objects. For all candidate objects, perform consistency verification based on the occlusion direction and the double-layer reflection parameter constraints. When the occlusion direction of a candidate object in the occlusion relationship graph is inconsistent with the double-layer reflection parameter constraints, generate a third reflection label.

[0009] S4. Generate an event determination vector based on far-field similarity markers, crosstalk weighting coefficients, and three reflection tags, and match the event determination vector with preset gating trigger conditions to determine whether the candidate object enters the security protection alarm process.

[0010] Preferably, in S1, long-distance scene feature mapping matching refers to: matching the image texture features of the region of interest of the candidate object and the multi-scale texture vector of the neighboring background block with the far-field mapping entries stored in the far-field static texture database; the long-distance scene feature mapping matching adopts a dual constraint matching strategy of histogram of directional gradients and image brightness consistency.

[0011] The far-field similarity marker is a texture similarity index between the candidate object's region of interest and the matched far-field entry; the crosstalk weighting coefficient is a dynamic adjustment factor calculated based on the gradient perturbation ratio between the candidate object's central region and edge region.

[0012] Preferably, in step S2, extracting the edge gradient response along the glass normal and determining the bimodal structure includes the following steps:

[0013] Within the region of interest of the candidate object, local edge image regions are extracted according to the preset glass normal direction, and the edge gradient response function along the normal direction in the local edge image region is calculated. In the edge gradient response function, the distribution of local maxima in the normal direction is detected. If there are two main peak responses within a preset spatial distance range, and the amplitudes of the two main peak responses both exceed the background response threshold, it is determined that a bimodal structure exists. Based on the geometric symmetry of the bimodal structure and the response amplitude ratio, the first reflection label of the candidate object is generated.

[0014] The first reflection label is an identifier data object generated based on the bimodal structure of the edge gradient response extracted along the glass normal, used to identify candidate objects in the image that have the edge bimodal structure response features caused by double-layer glass reflection.

[0015] Preferably, in S2, the image displacement ratio of the candidate object is the ratio of the pixel displacement of the candidate object in adjacent frame images to the expected disparity displacement calculated by the camera pose estimation; the preset double-layer reflection displacement range is a range of pixel displacement ratio values ​​calculated by physical modeling based on the camera imaging parameters, the distance between the front and rear surfaces of the double-layer glass, the glass normal direction, and the camera shooting angle.

[0016] The second reflection tag is a logical marker generated based on the matching result of the displacement ratio of the candidate object image in adjacent frame images and the preset double-layer reflection displacement interval. It is used to identify whether the candidate object conforms to the parallax motion characteristics of double-layer glass reflection.

[0017] Preferably, in S2, the dual-label joint rule is a joint judgment mechanism triggered when the candidate object simultaneously meets the generation conditions of the first reflection label and the second reflection label, which is used to improve the accuracy of double-sided glass reflection recognition and avoid misjudgment by a single label.

[0018] The dual-label joint rule is used to determine whether a candidate object fits the reflection candidate set, as follows:

[0019] When a candidate object has both a first reflection label and a second reflection label, the candidate object is directly included in the reflection candidate set and marked as a highly suspected virtual image object.

[0020] The reflection candidate set is a collection of objects classified as highly likely to be reflection virtual images, serving as the basis for consistency verification.

[0021] Preferably, in step S3, the occlusion relationship graph is a directed graph structure constructed based on the spatial position, edge overlap area, and boundary direction of the candidate object and other objects in the image frame. The occlusion relationship graph is constructed as follows: calculate the edge overlap area, occlusion boundary direction, and depth ranking index between the candidate object and other objects, and establish a directed graph structure representing the occlusion relationship between targets in the two-dimensional image plane. The nodes in the graph are the candidate object and other objects, the edge direction represents the front and rear occlusion relationship, and the edge weight represents the occlusion confidence level.

[0022] Preferably, in S3, the occlusion direction refers to the spatial relative hierarchical direction of the candidate object in the occlusion relationship diagram; the double-layer reflection parameter constraint means that the reflecting object corresponding to the virtual image reflected by the double-layer glass can only appear in space at the relatively far end position of a specific reflection layer.

[0023] The consistency verification based on the occlusion direction and the double-layer reflection parameter constraints specifically includes: for candidate objects in the reflection candidate set, comparing whether the occlusion direction of the candidate object in the occlusion relationship diagram is consistent with the double-layer reflection parameter constraints; when inconsistency occurs, the candidate object is marked as physically inconsistent with reflection, and a third reflection label is generated.

[0024] Preferably, in step S4, the preset gating trigger condition is used as a gating determination mechanism to determine whether a candidate object triggers the security protection alarm process;

[0025] The preset gating trigger conditions include the following three types:

[0026] Far-field matching tag joint triggering condition: Identify candidate objects with high far-field similarity tags and at least one reflection tag, and determine them as candidate objects with potential reflection false alarm risk; Dual tag joint triggering condition: Identify candidate objects in the reflection candidate set, directly regard them as highly suspected double-layer glass reflection virtual image targets, and trigger entry into the reflection confirmation path and blocking path; Tag occlusion collaborative triggering condition: Identify candidate objects that simultaneously have at least one reflection tag and whose occlusion consistency verification fails, as the criterion for reflection artifacts entering the foreground, and trigger the high-priority security protection alarm process.

[0027] Preferably, if the preset gate triggering condition is not triggered, it is determined that the candidate object does not pose a risk of interference from double-layer glass reflection, and then the candidate object is allowed to pass.

[0028] On the other hand, the present invention provides an adaptive scene-aware security protection system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned adaptive scene-aware security protection method.

[0029] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects:

[0030] 1. In this invention, by identifying the edge structure and motion features of suspected reflections in double-layered glass images, virtual image targets formed by reflections from double-layered glass in the images can be effectively identified, thereby improving the detection accuracy of false alarm targets and reducing the false judgment rate of conventional visual recognition systems in scenarios with reflective interference.

[0031] 2. In this invention, consistency judgment is made by combining occlusion relationship and imaging physical parameters, and different levels of security response are triggered according to preset conditions. This enables graded response and dynamic scheduling of security protection process for highly suspected reflective targets, thereby ensuring that the system can still stably execute high-priority protection strategy in multi-target dynamic interaction scenarios, and improving the overall scene adaptability and security response accuracy. Attached Figure Description

[0032] Figure 1 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0033] Example 1, as Figure 1As shown, the specific implementation steps of the adaptive scene-aware security protection method proposed in this invention are as follows:

[0034] S1. Collect background images of the area outside the double-layer glass, establish a set of far-field mapping entries and a far-field static texture database, and combine the candidate objects and their neighboring background blocks based on the target detection output as the candidate object region of interest. Perform far-distance scene feature mapping matching, and calculate the far-field similarity label and crosstalk weighting coefficient corresponding to the candidate object.

[0035] S2. In the region of interest of the candidate object, extract the edge gradient response along the glass normal and determine the bimodal structure to generate the first reflection label; calculate the image displacement ratio of the candidate object in adjacent frames, and generate the second reflection label when the image displacement ratio falls into the preset double-layer reflection displacement interval; based on the first reflection label and the second reflection label, use the dual label joint rule to determine whether the candidate object conforms to the reflection candidate set.

[0036] S3. Establish an occlusion relationship graph between candidate objects and other objects. For all candidate objects, perform consistency verification based on the occlusion direction and the double-layer reflection parameter constraints. When the occlusion direction of a candidate object in the occlusion relationship graph is inconsistent with the double-layer reflection parameter constraints, generate a third reflection label.

[0037] S4. Generate an event determination vector based on far-field similarity markers, crosstalk weighting coefficients, and three reflection tags, and match the event determination vector with preset gating trigger conditions to determine whether the candidate object enters the security protection alarm process.

[0038] In this embodiment S1, long-distance scene feature mapping matching refers to: matching the image texture features of the region of interest of the candidate object and the multi-scale texture vector of the neighboring background block with the far-field mapping entries stored in the far-field static texture database; the long-distance scene feature mapping matching adopts a dual constraint matching strategy of histogram of directional gradients and image brightness consistency.

[0039] The far-field similarity marker is a texture similarity index between the candidate object's region of interest and the matched far-field entry. Its value is generated based on the normalized cross-correlation coefficient and local gray-level difference measure, used to identify whether the candidate object has features of overlapping with the background at a distance. The crosstalk weighting coefficient is a dynamic adjustment factor calculated based on the gradient perturbation ratio between the candidate object's central and edge regions. It reflects the degree of background crosstalk in the candidate object's region and is used to adjust the weight of the candidate object in the event determination process, suppressing false alarm interference caused by far-field reflection.

[0040] The calculation of the far-field similarity marker and crosstalk weighting coefficient corresponding to the candidate object is as follows:

[0041] The similarity is calculated between the current frame texture vector of the region of interest of the candidate object and the historical background texture vector of the best matching entry. The similarity index includes the normalized cross-correlation coefficient and the local mean error. When the similarity exceeds a set threshold, the candidate object is marked as having far-field correspondence, and a far-field similarity label is generated. Based on the establishment of far-field matching, the gradient perturbation ratio between the central region and the edge region of the candidate object is calculated to reflect the degree of crosstalk. The perturbation ratio is proportional to the degree of crosstalk, and a crosstalk weighting coefficient is generated based on the ratio as the basis for subsequent event judgment weight adjustment.

[0042] In this embodiment S1, to achieve long-distance scene feature mapping matching of candidate objects, the system first captures images of the monitoring area obstructed by double-layer glass using a camera, and extracts the background region located outside the double-layer glass from multiple clear image frames using image segmentation technology or static foreground modeling methods to construct a far-field background sample set. Based on this background sample set, a far-field static texture database is generated using K-means clustering and background stability judgment rules. This database is based on image regions with high frequency of occurrence and strong texture consistency within a time window, and records multiple far-field texture blocks and their corresponding viewpoint, illumination, and boundary feature information. Furthermore, according to different viewpoints and shooting depths, a far-field mapping entry set is constructed, and each entry includes: the orientation gradient histogram of the far-field texture fragment, the image brightness histogram, the matching confidence score, and the spatial scale annotation.

[0043] In this embodiment S1, the candidate object is extracted using YOLOv7 or other deep learning object detection algorithms with bounding box output. The output includes the center position, class confidence, and bounding box size. Based on the candidate object, a region of interest for the candidate object is constructed. This region includes the target bounding box and the surrounding background block of a fixed scale.

[0044] In long-distance scene feature mapping matching, texture features of the region of interest are extracted and matched against a far-field static texture database. A dual constraint strategy of histogram of oriented gradients (HAR) and image brightness consistency is employed during the matching process. The HAR constraint ensures the consistency of edge structure directionality, while the image brightness consistency constraint suppresses mismatches caused by illumination variations. The matching results are used to generate far-field similarity markers and crosstalk weighting coefficients. The far-field similarity markers are calculated jointly based on the normalized cross-correlation coefficient between candidate regions and the best matching entry, and the root mean square value of gray-level difference. A quantifiable similarity index is used to determine whether the region overlaps with the far-field background for reflection. The crosstalk weighting coefficient is calculated based on the gradient perturbation ratio between the central and edge regions of the candidate object ROI. This perturbation ratio is used to characterize whether the texture undergoes unstable diffusion in the boundary region, thereby reflecting whether the candidate object is affected by far-field reflection crosstalk. In the specific calculation, the mean and variance of the gradients of the central and edge regions are extracted respectively to construct the perturbation ratio. After normalization using the Sigmoid function, it is used as a dynamic weighting factor to adjust the influence weight of the candidate object on the security status recognition result, avoiding misjudgment caused by far-field virtual images.

[0045] In this embodiment S2, extracting the edge gradient response along the glass normal and determining the bimodal structure includes the following steps:

[0046] Within the region of interest of the candidate object, local edge image regions are extracted according to the preset glass normal direction, and the edge gradient response function along the normal direction in the local edge image region is calculated. In the edge gradient response function, the distribution of local maxima in the normal direction is detected. If there are two main peak responses within a preset spatial distance range, and the amplitudes of the two main peak responses both exceed the background response threshold, it is determined that a bimodal structure exists. Based on the geometric symmetry of the bimodal structure and the response amplitude ratio, the first reflection label of the candidate object is generated.

[0047] The first reflection label is an identifier data object generated based on the bimodal structure of the edge gradient response extracted along the glass normal, used to identify candidate objects in the image that have the edge bimodal structure response features caused by double-layer glass reflection.

[0048] In this embodiment S2, to determine whether a candidate object exhibits the edge ghosting phenomenon characteristic of double-layer glass reflection, an edge gradient response analysis along the glass normal is performed on the region of interest of the candidate object. The glass normal direction is set by the camera mounting parameters and the geometric layout of the monitored scene, and is usually a direction vector perpendicular to the glass surface and towards the lens. In a static scene, this direction can be constructed by fitting a straight line to the glass edge in the calibration image or by combining it with a laser rangefinder. In a dynamic scene, the direction needs to be determined by introducing the scene geometry preset or feedback data from the inertial measurement unit.

[0049] In this embodiment S2, the system extracts the local edge image region in the candidate ROI according to the glass normal direction. This region is composed of multiple orthogonal sampling lines used to scan the pixel intensity gradient in the glass normal direction. The gradient value on each sampling line is calculated by operators such as Sobel and Prewitt to form an edge gradient response function, which represents the edge intensity distribution of the image along the normal direction at each pixel. Next, the distribution of local maxima is detected in the edge gradient response function. First, the first derivative zero-crossing method is used to locate the maxima, and then the position of the main peak is detected based on the dynamic neighborhood window. If there are two main peak responses within a preset spatial distance range, and the response amplitudes of the two main peaks are both higher than the set background response threshold, then the current region is considered to have a bimodal structure. The set background response threshold is calculated by adding twice the standard deviation to the mean edge intensity of the background region.

[0050] After confirming the existence of a bimodal structure, the system further models the structural features and generates the first reflection label based on the following two dimensions: the first is the geometric symmetry index, which calculates the symmetry offset distance and angle of the two main peaks relative to the center line of the candidate object or the glass normal center axis, sets a symmetry threshold range, and if the offset is less than the threshold, it is considered to meet the symmetry; the second is the response amplitude ratio index, which constructs the ratio of the response amplitudes of the two main peaks and normalizes it to a symmetry score between 0 and 1; the first reflection label is generated based on the fusion of the geometric symmetry index and the response amplitude ratio index.

[0051] In this embodiment S2, the image displacement ratio of the candidate object is the ratio of the pixel displacement of the candidate object in adjacent frame images to the expected disparity displacement calculated by the camera pose estimation; the preset double-layer reflection displacement range is a range of pixel displacement ratio values ​​calculated by physical modeling based on the camera imaging parameters, the distance between the front and rear surfaces of the double-layer glass, the glass normal direction, and the camera shooting angle.

[0052] The second reflection tag is a logical marker generated based on the matching result of the displacement ratio of the candidate object image in adjacent frame images and the preset double-layer reflection displacement interval. It is used to identify whether the candidate object conforms to the parallax motion characteristics of double-layer glass reflection.

[0053] In this embodiment S2, an image displacement ratio is introduced to determine whether a candidate object has a parallax shift feature caused by double-layer glass reflection. The image displacement ratio of the candidate object is a quantitative index obtained by calculating the ratio of the actual pixel displacement of the candidate object in adjacent frame images to the expected parallax displacement obtained by camera pose estimation. Specifically, the actual pixel displacement is calculated from the difference in center point coordinates of the candidate object in two consecutive frame images. Considering scale consistency, distortion correction and registration preprocessing are required for the image. The expected displacement is calculated by inputting the camera pose parameters (including displacement vector and rotation matrix) of the current frame and the pixel position of the candidate object in the first frame image to deduce the position that the object should appear in the next frame under no reflection conditions, thereby forming the expected parallax displacement. The ratio of the above two is calculated as the image displacement ratio.

[0054] By analyzing the parallax shift characteristics of virtual images in images reflected by double-layer glass, a preset double-layer reflection displacement range is established. The construction of the preset double-layer reflection displacement range is based on camera imaging parameters, the thickness distance between the front and rear surfaces of the double-layer glass, the glass normal direction, and the fixed installation angle of the camera. The displacement deviation range of the reflected virtual image in the image coordinates is derived through physical modeling. In this process, the imaging shift path after the light undergoes two refractions on the front and rear glass surfaces is considered. Finally, a set of pixel displacement ratio numerical ranges adapted to different scenarios is constructed through simulation data training and empirical correction. This is the preset double-layer reflection displacement range. This range is used to measure whether there is abnormal behavior in the image displacement ratio that conforms to the reflection motion trajectory.

[0055] In this embodiment S2, the second reflection tag serves as a logical determination identifier to identify whether the candidate object has a parallax offset mode consistent with the reflection characteristics of the double-layer glass. When the image displacement ratio falls into the preset double-layer reflection displacement range, the corresponding second reflection tag is generated.

[0056] In this embodiment, S2, the dual-label joint rule is a joint judgment mechanism triggered when the candidate object simultaneously meets the generation conditions of the first reflection label and the second reflection label. It is used to improve the accuracy of double-sided glass reflection recognition and avoid misjudgment by a single label.

[0057] The dual-label joint rule is used to determine whether a candidate object fits the reflection candidate set, as follows:

[0058] When a candidate object has both a first reflection label and a second reflection label, the candidate object is directly included in the reflection candidate set and marked as a highly suspected virtual image object.

[0059] The reflection candidate set is a collection of objects classified as highly likely to be reflection virtual images, serving as the basis for consistency verification.

[0060] In this embodiment S2, the dual-label joint rule is used to improve the recognition accuracy of virtual image candidate objects in double-glass reflection scenes and reduce the risk of misjudgment caused by single feature judgment, and is the core triggering logic of the reflection candidate set screening mechanism. Specifically, when a candidate object simultaneously meets the conditions of the first reflection label and the second reflection label in the current analysis frame, that is, its image region has significant edge bimodal structure features and exhibits reflection parallax shift behavior that conforms to the physical model derivation in adjacent frames, the system triggers the dual-label joint rule and directly classifies the candidate object into the reflection candidate set.

[0061] The dual-label joint rule avoids misjudgment caused by the overlap of a single reflective feature with a non-reflective target. This is especially true in complex real-world scenarios, such as when static objects have light and shadow interference at their edges, resulting in a bimodal response, or when dynamic objects are affected by camera blur and exhibit abnormal displacement. In such cases, relying solely on the first or second label may lead to a decrease in the accuracy of reflective recognition.

[0062] In this embodiment S3, the occlusion relationship graph is a directed graph structure constructed based on the spatial position, edge overlap area and boundary direction of the candidate object and other objects in the image frame. The occlusion relationship graph is constructed as follows: calculate the edge overlap area, occlusion boundary direction and depth ranking index between the candidate object and other objects, and establish a directed graph structure representing the occlusion relationship between targets in the two-dimensional image plane. The nodes in the graph are the candidate object and other objects, the edge direction represents the front and back occlusion relationship, and the edge weight represents the occlusion confidence level.

[0063] In this embodiment S3, the occlusion relationship graph is used for reflection consistency verification and false positive elimination based on occlusion relationships, and to express the before-and-after occlusion relationships and their confidence levels between candidate objects and other targets in the scene. The occlusion relationship graph is constructed using the following method:

[0064] First, for all target objects in each frame of the image, including candidate objects and other objects, edge contour extraction and spatial location analysis are performed to obtain the bounding box coordinates, edge pixel set, and target region center point position of each object. Taking the candidate object as the center, the area of ​​the edge overlap region between it and other objects is calculated one by one to measure whether the two targets overlap and occlude. If the overlap area exceeds the occlusion judgment threshold, the occlusion boundary direction is further calculated. This direction is the main direction gradient vector of the overlapping area of ​​the two object boundaries. All objects and their occlusion relationships are organized into a complete directed graph structure, where each node represents a candidate or non-candidate object, the directed edges in the graph represent the occlusion direction between two objects, and the edge weights characterize the occlusion confidence level.

[0065] In this embodiment S3, the occlusion direction refers to the spatial relative hierarchical direction of the candidate object in the occlusion relationship diagram. It is calculated based on the image perspective geometry by the boundary position, size ratio, and edge blurring change of the overlapping area between the candidate object and its occluded object. It is used to indicate whether the candidate object is located in the foreground layer or the background layer from the current viewpoint. The double-layer reflection parameter constraint means that the reflecting object corresponding to the virtual image reflected by the double-layer glass can only appear in the space at the relatively far end position of a specific reflection layer. The double-layer reflection parameter constraint is determined by the virtual image imaging range calculated by the camera intrinsic parameters, the glass layer spacing, the incident angle, and the reflection angle.

[0066] The consistency verification based on the occlusion direction and the double-layer reflection parameter constraints specifically includes: for candidate objects in the reflection candidate set, comparing whether the occlusion direction of the candidate object in the occlusion relationship diagram is consistent with the double-layer reflection parameter constraints; when inconsistency occurs, the candidate object is marked as physically inconsistent with reflection, and a third reflection label is generated.

[0067] In this embodiment S3, the process of obtaining the occlusion direction is as follows: First, extract the binary mask of the overlapping area of ​​the candidate object and the target with overlapping edges, calculate the main direction vector and boundary normal direction of the overlapping area, and count the area of ​​the overlapping area and the boundary length as the overlapping strength index. At the same time, calculate the sharpness gradient difference and contrast difference on both sides of the overlapping boundary as the front and back layer criteria. Then, combine the perspective prior (pixel size of the same type of target, image lower edge position, vanishing point direction) to generate foreground / background binary layer labels and confidence. Finally, the directed edge direction, edge weight and layer label in the occlusion relationship graph are used as the output of the occlusion direction. The process of generating the double-layer reflection parameter constraint is as follows: Based on the camera intrinsic parameters, extrinsic parameters and double-layer glass geometric parameters (layer spacing, glass normal, incident / reflection angle), establish the virtual image imaging range template. The template includes the allowable displacement range along the glass normal, the allowable visible layer and the relative depth relationship with the real target, and provide it to the consistency verification module for calling in the form of a lookup table.

[0068] The consistency check is performed based on the occlusion direction and dual-layer reflection parameter constraints, and is executed on an object-by-object basis. The hierarchical label of the occlusion direction is mapped and matched with the allowed hierarchical level in the virtual image imaging range template. If a contradiction occurs, such as "the candidate object is judged as a foreground occluding the real object" or "the candidate object is in the prohibited hierarchical level of the template", a reflection physical inconsistency mark is output and a third reflection label is generated. If there are multiple occlusion pairs, an edge weighted voting strategy is used to determine the final consistency result, and the results of consecutive frames are subjected to majority voting and hysteresis processing within a time window to suppress instantaneous jitter. When the object being checked belongs to the reflection candidate set, the side with the higher edge weight is given priority for a first judgment. In case of conflict, the decision is made according to the hierarchical priority rules preset in the template. All thresholds and weight parameters are managed by the strategy mapping table in an environmental gradation manner to ensure consistency under different lighting and viewing angle conditions.

[0069] The preset hierarchical priority rule is used to arbitrate conflicting situations when there is a conflict between the occlusion direction and the reflection imaging parameters, ensuring the robustness and physical consistency of the judgment. The preset hierarchical priority rule sets the following decision order based on the imaging characteristics of glass double-layer reflection: the priority of background reflection far away from the glass surface is greater than the priority of foreground virtual image close to the glass surface, and the priority of foreground virtual image close to the glass surface is greater than the priority of real object occluding the foreground. That is, when occlusion occurs, if the candidate object is identified as a background virtual image and meets the physical imaging position range of the virtual image, it is given priority to be identified as a reflection object; if the candidate object occludes the real object but is located in a position that does not meet the physical reflection area, the reflection identity label of the object is given priority to be eliminated.

[0070] The strategy mapping table is a parameter configuration system used to support consistency judgment under different scenarios, lighting, shooting angles, and glass parameter combinations. It includes recommended parameter sets at multiple levels, including occlusion judgment weight factors (edge ​​sharpness difference weight, occlusion area weight, etc.), edge weight voting threshold, front and back level comparison threshold, reflection imaging offset tolerance, dual-label fusion weight ratio, etc. In addition, it also includes imaging position template selection items and imaging level judgment accuracy levels used for physical consistency verification. The levels of the strategy mapping table are automatically matched and selected based on image brightness histogram, reflection spectrum distribution, viewing angle direction, and motion characteristics, and support online learning updates or manual adjustment by administrators to adapt to specific monitoring scenarios.

[0071] In this embodiment S4, the preset gating trigger condition is used to determine whether the candidate object triggers the security protection alarm process.

[0072] The preset gating trigger conditions include the following three types:

[0073] Far-field matching tag joint triggering condition: Identify candidate objects with high far-field similarity tags and at least one reflection tag, and determine them as candidate objects with potential reflection false alarm risk; Dual tag joint triggering condition: Identify candidate objects in the reflection candidate set, directly regard them as highly suspected double-layer glass reflection virtual image targets, and trigger entry into the reflection confirmation path and blocking path; Tag occlusion collaborative triggering condition: Identify candidate objects that simultaneously have at least one reflection tag and whose occlusion consistency verification fails, as the criterion for reflection artifacts entering the foreground, and trigger the high-priority security protection alarm process.

[0074] In this embodiment S4, the preset gating trigger condition serves as the logical gating criterion for entering the security protection alarm process, used to filter out target candidate objects with the risk of reflection interference or the possibility of misjudging actual occlusion; the three gating trigger conditions included in the preset gating trigger condition are constructed independently, or they can be judged collaboratively, as follows:

[0075] Far-field matching tag joint triggering condition: By performing far-field semantic matching analysis on candidate objects between the current image frame and historical frames, targets with high far-field similarity tags are identified, and whether they simultaneously possess at least one reflection tag (first reflection tag, second reflection tag, or third reflection tag) is used as the triggering condition; this condition is mainly applicable to the long-term persistence of image ghosting caused by glass reflection. When a candidate object shows suspected reflection features in multiple viewpoints, it can be regarded as having a potential risk of false alarm, thus entering the security protection path in advance;

[0076] Dual-label joint triggering condition: Based on the dual-label joint rule mechanism, when a candidate object has both the first reflection label and the second reflection label, it has been clearly classified into the reflection candidate set. The system regards it as a highly suspected double-glass reflection virtual image target and no longer performs redundant verification. It directly enters the reflection confirmation path and subsequent blocking path execution process to identify reflection artifacts in a timely manner and avoid misidentifying them as intruders or real foreground targets.

[0077] Tag occlusion coordination triggering condition: Combining the reflection tag information of the candidate object with the occlusion consistency verification result, if the candidate object has at least one reflection tag, and after comparing the occlusion direction with the double-layer reflection parameter constraint, a physical contradiction is generated and a third reflection tag is generated, then the system determines that the object is a high-risk marked object of "reflection artifact mistakenly entering the foreground" and forces it to enter the high-priority security protection alarm process.

[0078] In this embodiment, the priority of the security alarm process is specifically set as follows:

[0079] High-priority security protection alarm process: Applicable to candidate objects that meet the tag occlusion collaborative triggering conditions, that is, there is a reflective tag and it fails the occlusion consistency check at the same time, indicating that it has a clear spatial physical contradiction relationship and is very likely to be an abnormal target that is mistakenly identified into the foreground by a reflection artifact; the system will immediately trigger the high-priority alarm process, including forcibly terminating foreground recognition, activating the security shielding mechanism and pushing the alarm signal to the main control server;

[0080] Medium-priority security protection alarm process: Applicable to candidate objects that meet the dual-tag joint triggering conditions, that is, those that clearly have typical dual-tag characteristics of double-layer glass reflection but have no risk of obstruction or conflict. Such objects are judged as stable reflection artifact targets. The system will trigger the standard protection path, execute the blocking identification, area freezing and dynamic observation mechanism, and enter the candidate list to maintain status tracking.

[0081] Low-priority security protection alarm process: Applicable to candidate objects that only meet the joint triggering conditions of far-field matching tags, indicating that although the target may be a false alarm object due to reflection, the current frame features do not have sufficient evidence of reflection contradictions; the system will adopt a delayed judgment and secondary verification mechanism to include it in the low-priority observation state.

[0082] In this embodiment, if the preset gating trigger condition is not triggered, it is determined that the candidate object does not pose a risk of interference from double-layer glass reflection, and the candidate object is allowed to pass.

[0083] In this embodiment, when the preset gating trigger condition is not triggered, the system assumes that the current candidate object does not have significant reflection artifact features. It does not reach the gating threshold in far-field matching, reflection label status, and occlusion consistency test results. Therefore, it is assumed that the candidate object does not have the risk of double-glass reflection interference and is determined to be a valid target object, which can directly enter the target recognition subsequent processing flow.

[0084] Example 2: The adaptive scene perception security protection system proposed in this invention is applied to the adaptive scene perception security protection method proposed in Example 1. It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the adaptive scene perception security protection method in Example 1.

[0085] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. An adaptive scene-aware security protection method, characterized in that, Includes the following steps: S1. Collect background images of the area outside the double-layer glass, establish a set of far-field mapping entries and a far-field static texture database, and combine the candidate objects and their neighboring background blocks based on the target detection output as the candidate object region of interest. Perform far-distance scene feature mapping matching, and calculate the far-field similarity label and crosstalk weighting coefficient corresponding to the candidate object. S2. In the region of interest of the candidate object, extract the edge gradient response along the glass normal and determine the bimodal structure to generate the first reflection label; Calculate the image displacement ratio of candidate objects in adjacent frames. When the image displacement ratio falls within the preset double-layer reflection displacement range, generate a second reflection label. Based on the first and second reflection labels, use the dual-label joint rule to determine whether the candidate object conforms to the reflection candidate set. In S2, the image displacement ratio of the candidate object is the ratio of the pixel displacement of the candidate object in adjacent frame images to the expected disparity displacement calculated by the camera pose estimation; the preset double-layer reflection displacement range is the range of pixel displacement ratio values ​​calculated by physical modeling based on the camera imaging parameters, the distance between the front and rear surfaces of the double-layer glass, the glass normal direction, and the camera shooting angle. The second reflection tag is a logical marker generated based on the matching result of the displacement ratio of the candidate object image in adjacent frame images and the preset double-layer reflection displacement interval. It is used to identify whether the candidate object conforms to the parallax motion characteristics of double-layer glass reflection. S3. Establish an occlusion relationship graph between candidate objects and other objects. For all candidate objects, perform consistency verification based on the occlusion direction and the double-layer reflection parameter constraints. When the occlusion direction of a candidate object in the occlusion relationship graph is inconsistent with the double-layer reflection parameter constraints, generate a third reflection label. In S3, the occlusion relationship graph is a directed graph structure constructed based on the spatial position, edge overlap area and boundary direction of the candidate object and other objects in the image frame. The occlusion relationship graph is constructed as follows: calculate the edge overlap area, occlusion boundary direction and depth ranking index between the candidate object and other objects, and establish a directed graph structure representing the occlusion relationship between targets in the two-dimensional image plane. The nodes in the graph are the candidate object and other objects, the edge direction represents the front and back occlusion relationship, and the edge weight represents the occlusion confidence level. In S3, the occlusion direction refers to the spatial relative hierarchical direction of the candidate object in the occlusion relationship diagram; the double-layer reflection parameter constraint means that the reflected object corresponding to the virtual image reflected by the double-layer glass can only appear in space at the relatively far end position of a specific reflection layer. The consistency verification based on the occlusion direction and the double-layer reflection parameter constraints specifically includes: for candidate objects in the reflection candidate set, comparing whether the occlusion direction of the candidate object in the occlusion relationship diagram is consistent with the double-layer reflection parameter constraints; when inconsistency occurs, the candidate object is marked as physically inconsistent with reflection, and a third reflection label is generated. S4. Generate an event determination vector based on far-field similarity markers, crosstalk weighting coefficients, and three reflection tags, and match the event determination vector with preset gating trigger conditions to determine whether the candidate object enters the security protection alarm process.

2. The adaptive scene-aware security protection method according to claim 1, characterized in that: In S1, long-distance scene feature mapping matching refers to matching the image texture features of the region of interest of the candidate object and the multi-scale texture vector of the neighboring background block with the far-field mapping entries stored in the far-field static texture database. Long-distance scene feature mapping matching adopts a dual-constraint matching strategy of histogram of directional gradients and image brightness consistency. The far-field similarity marker is a texture similarity index between the candidate object's region of interest and the matched far-field entry; the crosstalk weighting coefficient is a dynamic adjustment factor calculated based on the gradient perturbation ratio between the candidate object's central region and edge region.

3. The adaptive scene-aware security protection method according to claim 2, characterized in that: In step S2, extracting the edge gradient response along the glass normal and determining the bimodal structure includes the following steps: Within the region of interest of the candidate object, local edge image regions are extracted according to the preset glass normal direction, and the edge gradient response function along the normal direction in the local edge image region is calculated. In the edge gradient response function, the distribution of local maxima in the normal direction is detected. If there are two main peak responses within a preset spatial distance range, and the amplitudes of the two main peak responses both exceed the background response threshold, it is determined that a bimodal structure exists. Based on the geometric symmetry of the bimodal structure and the response amplitude ratio, the first reflection label of the candidate object is generated. The first reflection label is an identifier data object generated based on the bimodal structure of the edge gradient response extracted along the glass normal, used to identify candidate objects in the image that have the edge bimodal structure response features caused by double-layer glass reflection.

4. The adaptive scene-aware security protection method according to claim 3, characterized in that: In S2, the dual-label joint rule is a joint judgment mechanism triggered when a candidate object simultaneously meets the generation conditions of the first reflection label and the second reflection label. It is used to improve the accuracy of double-sided glass reflection recognition and avoid misjudgment by a single label. The dual-label joint rule is used to determine whether a candidate object fits the reflection candidate set, as follows: When a candidate object has both a first reflection label and a second reflection label, the candidate object is directly included in the reflection candidate set and marked as a highly suspected virtual image object. The reflection candidate set is a collection of objects classified as highly likely to be reflection virtual images, serving as the basis for consistency verification.

5. The adaptive scene-aware security protection method according to claim 4, characterized in that: In S4, the preset gating trigger condition is a gating determination mechanism used to determine whether a candidate object triggers the security protection alarm process. The preset gating trigger conditions include the following three types: Far-field matching tag joint triggering condition: Identify candidate objects with high far-field similarity tags and at least one reflection tag, and determine them as candidate objects with potential reflection false alarm risk; Dual tag joint triggering condition: Identify candidate objects in the reflection candidate set, directly regard them as highly suspected double-layer glass reflection virtual image targets, and trigger entry into the reflection confirmation path and blocking path; Tag occlusion collaborative triggering condition: Identify candidate objects that simultaneously have at least one reflection tag and whose occlusion consistency verification fails, as the criterion for reflection artifacts entering the foreground, and trigger the high-priority security protection alarm process.

6. The adaptive scene perception security protection method according to claim 5, characterized in that: If the preset gating trigger condition is not triggered, it is determined that there is no risk of double-glass reflection interference from the candidate object, and the candidate object is allowed to pass.

7. An adaptive scene-aware security protection system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes a computer program to implement the adaptive scene-aware security protection method as described in any one of claims 1-6.

Images