Visual identification method and system based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a visual identification method and system based on artificial intelligence. The method comprises the following steps: acquiring a scene identifier, a visual probe list and installation geometric parameters thereof, generating a session configuration packet and performing version registration, synchronously collecting image frames of each visual probe, constructing a quality label set and forming an evidence credibility table; executing target detection reasoning and target tracking reasoning, assembling an evidence slot position record set according to a task slot position template, performing cross-probe consistency verification and anti-verification, and generating an abnormal conclusion record and a risk grading mark; and a closed-loop feedback record is generated in combination with a rechecking result, and the abnormal threshold parameter and the slot weight parameter are updated. According to the method, a closed-loop link with version penetration, evidence admission and anti-evidence linkage is configured, so that structured organization and consistency judgment of multi-probe visual evidence are realized, and the stability, traceability and continuous evolution capability of a visual identification process are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a visual recognition method and system based on artificial intelligence. Background Technology

[0002] In the field of artificial intelligence technology, existing solutions for visual recognition based on scene identification and visual probe lists typically involve acquiring image frames from each visual probe and performing target detection or target tracking inference. These solutions suffer from limitations such as a lack of unified registration standards for session configuration packet hashes and version numbers, a lack of consistent quantitative descriptions of image frame quality differences, and a lack of structured merging of evidence credibility. Existing methods often rely on image frames acquired in a single acquisition to directly enter the inference and output process. When occlusion ratios, jitter indicators, and time synchronization deviations coexist, issues arise such as a lack of traceable separation between the set of admission evidence frames and the set of frames to be acquired, and a lack of stable correlation between the quality label set and the evidence credibility table. This makes it difficult to reliably achieve the recording of abnormal conclusions and risk classification labeling. For the joint processing of session configuration packages, quality tag sets, evidence slot record sets, consistency scoring tables, and counter-evidence result tables, existing technologies generally have insufficient integration in the stages of cross-probe consistency verification, abnormal threshold condition triggering, counter-evidence verification calling the corresponding supplementary evidence to be collected and reasoning. It is difficult to form a consistent process in application scenarios based on the access and review results of the handling record: collection—entry control—assembling evidence slot record sets—generating consistency scoring tables—outputting counter-evidence result tables—generating handling records—generating closed-loop feedback records—generating threshold weight update packages. As a result, it is difficult to form a verifiable closed-loop correspondence between the update version numbers of abnormal threshold parameters and slot weight parameters and false alarm markers, missed alarm markers, and uncertain markers. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a visual recognition method based on artificial intelligence, comprising: S100: Obtain scene identifier, visual probe list, installation geometry parameters of each visual probe, synchronization trigger parameters, task slot template version, generate session configuration package and register session configuration package hash and version number; S200: Based on the session configuration package, synchronously acquire image frames from each visual probe, calculate sharpness index, exposure index, occlusion ratio, jitter index, and time synchronization deviation, and generate a quality label set; generate an evidence credibility table according to the quality label set and perform access control, outputting an access evidence frame set and a supplementary acquisition set. S300. Based on the set of admission evidence frames, perform target detection inference and target tracking inference to generate a candidate target set; assemble the candidate target set into an evidence slot record set according to the task slot template version; the evidence slot record set includes probe number, time window index, target feature summary, evidence credibility, and slot gap marker; S400: Based on the evidence slot record set, perform cross-probe consistency verification and generate a consistency score table; trigger rebuttal verification for records in the consistency score table that meet the abnormal threshold conditions; the rebuttal verification calls the supplementary evidence to be collected and performs re-reasoning, and outputs the rebuttal result table. S500: Generate abnormal conclusion records and risk classification markers based on the consistency scoring table and the counter-evidence result table; match the disposal strategy library version according to the risk classification markers and generate disposal instruction sets; execute the disposal instruction sets to generate disposal records. S600. Generate a closed-loop feedback record based on the access review results of the handling record; the closed-loop feedback record includes false alarm markers, missed alarm markers, and uncertain markers; update the abnormal threshold parameter and slot weight parameter according to the closed-loop feedback record, generate a threshold weight update package and register the update version number.

[0004] Furthermore, the session configuration package includes a probe role labeling table, a field of view coverage relationship table, a unified time reference field, a task slot template version field, and a disposal strategy library version field.

[0005] Furthermore, the access control adopts a hard threshold rule, which includes an upper limit threshold for the occlusion ratio and an upper limit threshold for the time synchronization deviation. Image frames that do not meet the hard threshold rule are written into the gap mark of the acquisition and merging registration slot.

[0006] Furthermore, the evidence credibility table is generated by aggregating the quality tag set according to a weight vector, which is given by the quality weight version field in the session configuration package.

[0007] Furthermore, the candidate target set includes target bounding box coordinates, category labels, confidence scores, appearance feature vectors, and trajectory numbers; the target tracking inference uses trajectory numbers and appearance feature vectors to perform cross-frame association and outputs trajectory continuity labels.

[0008] Furthermore, the task slot template version defines identity consistency slots, temporal continuity slots, geometric consistency slots, and environmental confidence slots; the evidence slot record set stores slot values, slot source probe numbers, slot source time window indexes, and slot gap markers according to slot numbers.

[0009] Furthermore, the cross-probe consistency verification generates identity consistency score, location continuity score, geometric constraint residual, and event causal consistency score, and generates a consistency score table according to the score fusion rules.

[0010] Furthermore, the counter-evidence verification includes a counter-evidence time window index and a counter-evidence probe number. The counter-evidence verification performs the same target detection reasoning and evidence slot assembly on the counter-evidence as S300. The counter-evidence result table includes a counter-evidence consistency score and a counter-evidence pass mark.

[0011] Furthermore, the handling instruction set includes alarm instructions, supplementary sampling instructions, parameter adjustment instructions, and review instructions; the handling record includes the handling strategy library version, evidence slot record set hash, abnormal conclusion record identifier, and risk classification marker; the threshold weight update package includes the abnormal threshold parameter update amount and slot weight parameter update amount; the abnormal threshold parameter update amount and slot weight parameter update amount are generated by the closed-loop feedback record grouped and statistically analyzed according to false alarm markers, missed alarm markers, and uncertain markers.

[0012] Furthermore, an artificial intelligence-based visual recognition system, applied to any of the methods described above, includes: The configuration and version management module is used to obtain scene identifiers, visual probe list, installation geometric parameters of each visual probe, synchronization trigger parameters, task slot template version, and disposal strategy library version, and generate session configuration packages. The data acquisition and quality assessment module synchronously acquires image frames from each visual probe based on the session configuration package, calculates sharpness index, exposure index, occlusion ratio, jitter index, and time synchronization deviation, and generates a quality label set. The reasoning and slot assembly module performs target detection reasoning and target tracking reasoning to generate a candidate target set; The consistency verification and disproven verification module performs cross-probe consistency verification and generates a consistency scoring table. The handling and recording module generates abnormal conclusion records and risk classification markers based on the consistency scoring table; The closed-loop feedback and parameter update module generates closed-loop feedback records based on the handling records and the review results.

[0013] The following are its main beneficial effects: (1) To address the problem of difficulty in verifying configuration due to the lack of a unified registration standard for session configuration package hash and version number in the existing scheme, by registering the session configuration package hash and version number and using the session configuration package as a consistent input carrier, the geometric parameters, synchronization trigger parameters and task slot template versions of each visual probe are kept consistent in the operation chain of synchronous acquisition of image frames, generation of quality tag set, generation of evidence credibility table and assembly of evidence slot record set. The session configuration package forms a verifiable association record with the evidence slot record set, the consistency scoring table and the disposal record, which is suitable for scenarios where multiple visual probes are accessed collaboratively and the task slot template version evolves.

[0014] (2) To address the problem that existing solutions lack a consistent quantitative description of image frame quality differences, resulting in a lack of traceable separation between the set of admission evidence frames and the set of images to be acquired, the quality label set is constructed based on the sharpness index, exposure index, occlusion ratio, jitter index, and time synchronization deviation. The quality label set generates the evidence credibility table to drive the admission control, so that the set of admission evidence frames entering the target detection inference and target tracking inference has the same evidence credibility caliber as the evidence credibility table. At the same time, image frames that do not pass the admission control are written into the set of images to be acquired, forming a closed-loop diversion link of acquisition-evaluation-admission-acquisition around the session configuration package. This is suitable for engineering sites where occlusion ratio, jitter index, and time synchronization deviation coexist and acquisition quality fluctuates.

[0015] (3) To address the problem that the lack of integration between cross-probe consistency verification, abnormal threshold triggering and counter-evidence verification in the existing scheme makes it difficult to review abnormal conclusion records and risk classification markers, the evidence slot record set is used to structure probe number, time window index, target feature summary, evidence credibility and slot gap marker. This allows cross-probe consistency verification to form a traceable scoring basis in the consistency scoring table. When the abnormal threshold condition is met, counter-evidence verification is triggered to call the supplementary evidence to be collected and to output the counter-evidence result table. This allows the abnormal conclusion record and the risk classification marker to be supported by the consistency scoring table and the counter-evidence result table and to be carried by the disposal record. After the disposal record is connected to the review result, a closed-loop feedback record containing false alarm markers, missed alarm markers and uncertain markers is generated. Based on this, the abnormal threshold parameters and slot weight parameters are updated to generate a threshold weight update package and the update version number is registered. This forms a closed-loop link corresponding to the lack of consistency process in the background technology: collection-access-assembly-consistency verification-counter-evidence verification-disposal-feedback-update. This is suitable for scenarios where there are conflicts in multi-probe evidence or frequent occurrences of slot gap markers. Attached Figure Description

[0016] Figure 1 A flowchart illustrating an artificial intelligence-based visual recognition method provided in an embodiment of this application; Figure 2 This is a structural block diagram of an artificial intelligence-based visual recognition system provided in an embodiment of this application. Detailed Implementation

[0017] Example 1: Refer to Figure 1 This is a flowchart illustrating a visual recognition method based on artificial intelligence provided in an embodiment of the present invention. The process may include at least steps S100-S600: S100: Obtain scene identifier, visual probe list, installation geometry parameters of each visual probe, synchronization trigger parameters, task slot template version, generate session configuration package and register session configuration package hash and version number; S200: Based on the session configuration package, synchronously acquire image frames from each visual probe, calculate sharpness index, exposure index, occlusion ratio, jitter index, and time synchronization deviation, and generate a quality label set; generate an evidence credibility table according to the quality label set and perform access control, outputting an access evidence frame set and a supplementary acquisition set. S300. Based on the set of admission evidence frames, perform target detection inference and target tracking inference to generate a candidate target set; assemble the candidate target set into an evidence slot record set according to the task slot template version; the evidence slot record set includes probe number, time window index, target feature summary, evidence credibility, and slot gap marker; S400: Based on the evidence slot record set, perform cross-probe consistency verification and generate a consistency score table; trigger rebuttal verification for records in the consistency score table that meet the abnormal threshold conditions; the rebuttal verification calls the supplementary evidence to be collected and performs re-reasoning, and outputs the rebuttal result table. S500: Generate abnormal conclusion records and risk classification markers based on the consistency scoring table and the counter-evidence result table; match the disposal strategy library version according to the risk classification markers and generate disposal instruction sets; execute the disposal instruction sets to generate disposal records. S600. Generate a closed-loop feedback record based on the access review results of the handling record; the closed-loop feedback record includes false alarm markers, missed alarm markers, and uncertain markers; update the abnormal threshold parameter and slot weight parameter according to the closed-loop feedback record, generate a threshold weight update package and register the update version number.

[0018] To address the problems in existing technologies where scene identifiers and visual probe lists are accessed from multiple sources, have inconsistent field definitions, and lack a unified timeline and geometric definition for the installation geometric parameters and synchronization trigger parameters of each visual probe, making it difficult to reuse and trace across the cross-probe acquisition, alignment, and judgment links under the same session baseline, this invention completes the structured access of scene identifiers and visual probe lists during the session startup phase in step S100, and solidifies the installation geometric parameters and synchronization trigger parameters of each visual probe into a unified timeline and geometric definition. Simultaneously, the task slot template version, quality weight version, and disposal strategy library version fields are assembled into the session configuration package to form a versioned registration definition. A cross-step traceable input baseline is formed by registering the session configuration package hash and version number, allowing subsequent steps to consistently call the session configuration package. Specifically, this includes: S100: Obtain scene identifier, visual probe list, installation geometry parameters of each visual probe, synchronization trigger parameters, task slot template version, generate session configuration package and register session configuration package hash and version number; Specifically, S100 is executed by the session configuration module of the visual recognition system. Its input sources include the scene identifier on the scene side, the list of visual probes on the deployment side, the installation geometric parameters of each visual probe on the installation side, the synchronization trigger parameters on the acquisition side, and the task slot template version on the configuration side. The scene identifier is used to characterize the engineering scene boundary and management scope of a visual recognition session. The scene identifier includes a scene type field, a scene area field, a scene time period field, and a scene instance number field. The scene type field is used to distinguish the scene categories in engineering embodiments such as park entrances and exits, indoor passages, production workstations, vending machine aisles, drone flight paths, and mobile robot operation areas. The scene area field is used to record the spatial description information of the scene boundary. The scene time period field is used to record the time description information of the session's effective period. The scene instance number field is used to establish a traceable association between multiple sessions. The visual probe list is used to characterize the set of data sources participating in this session. The visual probe list includes probe number field, probe type field, probe installation location field, probe field of view parameter field, and probe interface parameter field. The probe number field is used to bind the source of image frame data, quality label set, and evidence slot record set in subsequent steps. The probe type field is used to identify types such as fixed camera, panoramic camera, gimbal camera, vehicle-mounted camera, UAV-mounted camera, or robot-mounted camera. The probe installation location field is used to record the installation point description of the probe in the scene coordinate system. The probe field of view parameter field is used to record imaging constraints such as field of view range, resolution, focal length, distortion parameters, and frame rate limit. The probe interface parameter field is used to record the access method and communication parameters of the acquisition link. The geometric parameters installed on each visual probe are used to solidify the geometric correspondence between the probe and the scene. Each visual probe installation geometric parameter includes probe extrinsic parameter fields and probe intrinsic parameter fields. The probe extrinsic parameter fields include a position vector field and an attitude angle field. The position vector field is used to describe the spatial position of the probe in the scene coordinate system, and the attitude angle field is used to describe the attitude relationship between the probe's line of sight and the scene coordinate system. The probe intrinsic parameter fields include a principal point parameter field, a focal length parameter field, and a distortion coefficient field. The distortion coefficient field is used to unify the geometric correction caliber in subsequent image frame data processing links. The synchronization trigger parameters are used to solidify the timing constraints of multi-probe synchronous acquisition. The synchronization trigger parameters include a unified time base field, a trigger source identifier field, a trigger period field, a trigger jitter upper limit field, and an alignment window field. The unified time base field is used to establish a unified time axis across probes. The trigger source identifier field is used to identify the trigger source of hardware triggering, network time synchronization, or master control triggering. The trigger period field is used to describe the period configuration of synchronous acquisition. The trigger jitter upper limit field is used to limit the drift range of acquisition triggering in the time dimension. The alignment window field is used to limit the window size of subsequent time window index construction.The task slot template version is used to solidify the assembly template of the evidence slot record set. The task slot template version includes a template version number field, a slot definition table field, and a slot mandatory marker field. The slot definition table field is used to describe the slot number, slot value structure, and value source constraints of identity consistency slots, temporal continuity slots, geometric consistency slots, and environmentally credible slots. The slot mandatory marker field is used to identify the minimum set of slots required to form an anomaly threshold judgment, thereby providing consistent structural constraints for the subsequent S200 entry control and S300 evidence slot assembly.

[0019] Furthermore, after receiving the above input, the session configuration module performs configuration solidification processing, which includes field validation, caliber alignment, version assembly, and exception logging. Field validation includes uniqueness verification of the scene instance number field of the scene identifier, duplicate verification of the probe number field of the visual probe list, integrity verification of the probe extrinsic and intrinsic parameter fields of the installation geometry parameters of each visual probe, legality verification of the unified time base field and trigger cycle field of the synchronization trigger parameters, and consistency verification of the template version number field of the task slot template version and the slot definition table field. Calibration alignment includes establishing a binding relationship between the probe field of view parameter fields in the visual probe list and the probe intrinsic parameter fields in the installation geometry parameters of each visual probe, writing the unified time base field and alignment window field in the synchronization trigger parameters into the unified time axis configuration substructure, and writing the slot definition table fields in the task slot template version into the slot template configuration substructure, thereby ensuring that subsequent steps for image frame data acquisition, alignment, quality label set generation, and evidence slot record set assembly all follow the same caliber. The version assembly includes writing the template version number field of the task slot template version and the disposal strategy library version field into the version field area of ​​the session configuration package. The disposal strategy library version field is used as a component field of the session configuration package for subsequent risk classification labeling and matching of the disposal strategy library version in S500. Simultaneously, the quality weight version field is written into the quality weight field area of ​​the session configuration package. This quality weight version field is used as a weight vector reference in the subsequent S200 generation of the evidence credibility table based on the quality label set. The anomaly record includes generating a configuration anomaly record and binding the scenario instance number field and probe number field when field verification or caliber alignment results in missing or conflicting items. The configuration anomaly record is written into the anomaly record field area of ​​the session configuration package and registers the anomaly type field and anomaly timestamp field, thus forming a traceable configuration anomaly chain. Even with configuration anomaly records, the session configuration module still outputs the session configuration package and uses the anomaly type field as one of the gating conditions input for subsequent S200 access control, constraining the collection and assembly process of low-credibility sessions.

[0020] Furthermore, after completing the configuration solidification process, the session configuration module performs session configuration package hash generation and session configuration package version number registration. The session configuration package hash is used to perform consistency verification on the content of the session configuration package. Specifically, it serializes and concatenates the scene identifier, visual probe list, installation geometric parameters of each visual probe, synchronization trigger parameters, task slot template version, quality weight version field, handling strategy library version field, and exception record field in the session configuration package to form a hash input string. Then, it performs digest calculation to obtain the session configuration package hash and writes the session configuration package hash into the hash field of the session configuration package. The session configuration package version number is used to identify the configuration generation of the session configuration package in the closed-loop evolution. Specifically, it is generated by combining the template version number field, quality weight version field, and disposal strategy library version field of the task slot template version, and then writing the session configuration package version number into the version number field of the session configuration package. When the session configuration module detects a change in the task slot template version corresponding to the same scenario instance number field, it triggers the increment registration of the session configuration package version number and generates a version change record. The version change record is written into the version record field area of ​​the session configuration package and bound to the session configuration package hash to support the tracing of the version chain when the threshold weight update package of S600 is registered to update the version number. Understandably, the above-mentioned session configuration package hash and session configuration package version number constitute the minimum audit field set of the session configuration package. The minimum audit field set, together with the scenario instance number field and probe number field, forms the primary key reference across steps, thereby enabling the subsequent generation of the quality tag set and evidence credibility table by S200 to have traceable input.

[0021] In the engineering implementation, taking the implementation of the scene type field as the park entrance / exit as an example, the scene identifier is written and distributed to the session configuration module by the park management system before entering the monitoring period. The visual probe list includes the probe number field and interface parameter field of the fixed camera and the side pan-tilt camera above the entrance / exit channel. The installation geometric parameters of each visual probe are generated by the deployment calibration process and distributed through the configuration library. The synchronization trigger parameters are configured by the main control acquisition gateway and provide a unified time reference field and an alignment window field. The task slot template version is provided by the template repository and includes the slot definition table fields of identity consistency slots and geometric consistency slots. After the session configuration module completes the field verification, it generates a session configuration package and writes the session configuration package hash and session configuration package version number. The probe number field and unified time reference field of the session configuration package are called by the synchronization acquisition process in the subsequent S200. The task slot template version is called by the evidence slot assembly process in the subsequent S300. The disposal strategy library version field is called by the risk classification mark matching process in the subsequent S500. The quality weight version field is called by the evidence credibility table generation process in the subsequent S200.

[0022] The output of S100 is a session configuration package, which includes scene identifier, visual probe list, installation geometric parameters of each visual probe, synchronization trigger parameters and task slot template version, and registers session configuration package hash and version number in the session configuration package; the session configuration package is called as input to S200 to constrain the unified standard for synchronously acquiring image frames of each visual probe, generating quality label set and executing access control.

[0023] Summary of the technical effects of this step: This step completes the structured access of scene identifiers and visual probe lists during the session startup phase, and solidifies the installation geometric parameters and synchronization trigger parameters of each visual probe into a unified timeline and geometric caliber; This step assembles the task slot template version, quality weight version field, and disposal strategy library version field into the session configuration package and forms a versioned registration caliber; This step forms a cross-step traceable input baseline by registering the session configuration package hash and version number and provides it for S200 to call.

[0024] To address the problems in existing technologies where multi-visual probe acquisition often lacks session-level synchronization constraints and windowed organization, resulting in unstable correlation between image frame data and time base, and a lack of structured quantitative descriptions and unified entry criteria for quality differences such as occlusion, jitter, and exposure, leading to unstable evidence in the inference chain under low-quality frame input and a lack of diversion criteria for supplementary acquisition and verification, this invention completes cross-probe synchronous acquisition and windowed organization under session configuration package constraints in step S200, and binds image frame data to a unified time axis to form a traceable acquisition record; by constructing a quality label set and an evidence credibility table, the multi-dimensional quality status is structured and introduced into the entry control criteria, thereby outputting the set of entry evidence frames and the collection to be supplemented, forming a diversion input baseline for subsequent inference and consistency verification, specifically including: S200: Based on the session configuration package, synchronously acquire image frames from each visual probe, calculate sharpness index, exposure index, occlusion ratio, jitter index, and time synchronization deviation, and generate a quality label set; generate an evidence credibility table according to the quality label set and perform access control, outputting an access evidence frame set and a supplementary acquisition set. Specifically, S200 is executed by the synchronous acquisition and quality assessment module of the visual recognition system, and its input source is the session configuration package generated in S100. The visual probe list, installation geometric parameters of each visual probe, and synchronous trigger parameters carried in the session configuration package constitute the minimum input set of the synchronous acquisition link. The task slot template version, quality weight version field, and disposal strategy library version field are loaded together as session-level configuration constraints to solidify the gating criteria for subsequent access control and the convergence criteria of the evidence credibility table. The synchronous acquisition and quality assessment module enters the running state when the session initiation trigger condition is met. The session initiation trigger condition is jointly defined by the trigger source identifier field and the trigger period field in the synchronous trigger parameters. When the trigger source identifier field is a hardware trigger, the synchronous acquisition and quality assessment module accesses an external trigger pulse through the trigger interface and forms a frame sampling time record under the constraint of the unified time reference field. When the trigger source identifier field is a master control trigger, the synchronous acquisition and quality assessment module sends an acquisition command to the acquisition port corresponding to the visual probe list at the period time corresponding to the trigger period field and registers the timestamp. In the above running state, the unified time reference field serves as the reference time for time alignment, and the alignment window field serves as the boundary of the same time window index, providing a fixed standard for subsequent time synchronization deviation calculation and access control.

[0025] Furthermore, the process of synchronously acquiring image frames from each visual probe includes probe access, frame capture, timestamp binding, and buffer orchestration. Probe access is driven by the probe interface parameter fields in the visual probe list. These fields include an access protocol field, a link address field, and a bitstream parameter field. The synchronous acquisition and quality assessment module establishes an acquisition session channel based on the access protocol field and performs connectivity verification on the link address field. Probe numbers that fail the connectivity verification are written into the acquisition anomaly record and bound to the scene instance number field in the scene identifier. Frame capture includes extracting image frame data from each probe channel according to the trigger beat and generating a frame sequence number field and a frame size field for the image frame data. Timestamp binding includes recording an acquisition timestamp field and a trigger timestamp field for each image frame data, and deriving a reference timestamp field from a unified time reference field. The acquisition timestamp field is used to identify the actual arrival time on the probe side, the trigger timestamp field is used to identify the trigger time of this acquisition, and the reference timestamp field is used for unified alignment across probes. The buffer orchestration includes writing image frame data from each probe within the same window into a window buffer based on the alignment window field, and generating a time window index for the window buffer. This time window index serves as one of the sources of the time window index field for the evidence slot record set in subsequent steps S300. Understandably, the probe number field and time window index formed during the synchronous acquisition phase constitute a common key for the subsequent quality label set, evidence credibility table, admission evidence frame set, and supplementary acquisition set, used for consistent referencing of source and timing during cross-step transitions.

[0026] Furthermore, the calculation of sharpness index, exposure index, occlusion ratio, jitter index, and time synchronization deviation is completed by the quality assessment sub-link. The quality assessment sub-link performs index calculations on each probe image frame data in the window buffer and generates quality label entries. The sharpness index is a numerical label characterizing the intensity of image details. Specifically, the synchronous acquisition and quality assessment module performs grayscale conversion and edge response extraction on the image frame data, generates a sharpness index based on the edge response distribution, and writes the sharpness index into the sharpness index field of the quality label entry. The exposure index is a numerical label characterizing brightness distribution and saturation state. Specifically, brightness histogram statistics are performed on the image frame data, and the proportion of saturated pixels and dark pixels is extracted to generate an exposure index, which is written into the exposure index field of the quality label entry. The occlusion ratio is a label characterizing the proportion of the effective field of view that is occluded. Specifically, the effective field of view area is determined based on the probe field of view parameter field in the visual probe list, and the effective field of view area is... The domain performs foreground occlusion detection and calculates the occlusion ratio, which is written into the occlusion ratio field of the quality label entry. The jitter index is a numerical label characterizing the viewpoint disturbance between adjacent frames. Specifically, feature points are extracted between consecutive image frames corresponding to the same probe number field, and inter-frame matching is performed. The jitter index is generated based on the matching displacement distribution and written into the jitter index field of the quality label entry. The time synchronization deviation is a numerical label characterizing the cross-probe time alignment error. Specifically, using the reference timestamp field as a benchmark, the deviation between the acquisition timestamp field and the trigger timestamp field of each probe image frame data within the same time window index is calculated to obtain the time synchronization deviation, which is written into the time synchronization deviation field of the quality label entry. The above quality label entries are further bound to the probe number field, time window index, and frame sequence number field to form a traceable record and written into the quality label set. When a probe is missing image frame data within a certain time window index, the synchronous acquisition and quality assessment module generates a missing marker field in the quality label set and binds the probe number field and the time window index for subsequent access control gap determination and generation of supplementary acquisition data.

[0027] Furthermore, the process of generating the evidence credibility table based on the quality tag set is executed by the credibility aggregation sub-link. This sub-link extracts the sharpness index field, exposure index field, occlusion ratio field, jitter index field, and time synchronization deviation field from the quality tag set, performs weight aggregation, generates evidence credibility, and writes it into the evidence credibility table. The weight input for the weight aggregation comes from the quality weight version field in the session configuration package. The weight vector corresponding to the quality weight version field is loaded into a quality weight configuration sub-structure and bound to the scene type field in the scene identifier, thus ensuring consistency in the evidence credibility calculation caliber under different scene type fields. Specifically, the credibility aggregation sub-link normalizes the quality label entries and generates normalized quality component records. These records include normalized sharpness, exposure, occlusion, jitter, and synchronization components. Subsequently, the normalized quality component records are weighted and aggregated according to a weight vector to generate an evidence credibility field. This field, along with the probe number, time window index, and frame sequence number, is written into the evidence credibility table. Understandably, the evidence credibility table serves as one of the sources of evidence credibility fields in the subsequent evidence slot record set in S300. Furthermore, the evidence credibility table acts as a direct input for the access control process in this step, ensuring a traceable field chain for the access control operation.

[0028] Furthermore, the implementation process of the access control is executed by the gating decision sub-link. The gating decision sub-link takes the quality label set and the evidence credibility table as input, and combines them with the synchronization trigger parameters and the anomaly record field in the session configuration package to form a gating condition set. The gating condition set includes hard threshold rules and supplementary acquisition trigger rules. The hard threshold rules include an upper limit threshold for occlusion ratio and an upper limit threshold for time synchronization deviation. The upper limit threshold for occlusion ratio and the upper limit threshold for time synchronization deviation are loaded as gating configuration fields and bound to the quality weight version field. The gating decision sub-link reads the occlusion ratio field and the time synchronization deviation field for each quality label entry and compares them with the upper limit threshold for occlusion ratio and the upper limit threshold for time synchronization deviation. Image frame data that meets the hard threshold rules is marked as an access candidate frame, and image frame data that does not meet the hard threshold rules is marked as a non-access frame and written into the supplementary acquisition package. The supplementary acquisition triggering rule is based on a joint determination of the missing marker field, the evidence credibility field, and the acquisition anomaly record. Specifically, when the missing marker field exists or the evidence credibility field is lower than the credibility threshold field, the gating determination sub-link generates a supplementary acquisition trigger record and writes the supplementary acquisition trigger record into the acquisition set to be supplemented. The supplementary acquisition trigger record includes a probe number field, a time window index, a trigger reason field, and a trigger timestamp field. When an acquisition anomaly record appears in the anomaly record field area of ​​the session configuration packet and matches the current probe number field, the gating determination sub-link performs a weight reduction mark on the window buffer corresponding to the probe number field and writes the weight reduction mark into the additional mark field of the evidence credibility table, so that the source of weight reduction can be identified when the S300 calls the access evidence frame set.

[0029] Furthermore, the implementation process of the output admission evidence frame set and the supplementary acquisition set includes set construction, field binding, and cross-step reference registration. Specifically, the gating judgment sub-link groups the admission candidate frames that meet the hard threshold rule according to the probe number field and time window index and constructs an admission evidence frame set. The admission evidence frame set includes image frame data, probe number field, time window index, frame sequence number field, acquisition timestamp field, reference timestamp field, and evidence credibility field, wherein the evidence credibility field is backfilled by the evidence credibility table; at the same time, the non-admission frames, missing records corresponding to the missing marker field, and supplementary acquisition trigger records are aggregated to construct the supplementary acquisition set. The supplementary acquisition set includes probe number field, time window index, slot gap marker, and supplementary acquisition trigger record field, wherein the slot gap marker is generated and registered by the missing marker field and the hard threshold rule judgment result. Understandably, the set of admission evidence frames is called as input to S300 in this patent process to perform target detection inference and target tracking inference and generate a candidate target set. The set of evidence to be collected is called as input to the counter-evidence verification in S400 in this patent process to trigger supplementary evidence collection and re-inference when the abnormal threshold condition is met. Therefore, when this step outputs, the time window index and probe number fields of the set of admission evidence frames are registered together with the session configuration package hash as a session-level reference record, and the session-level reference record is written into the acquisition and gating record field area. The acquisition and gating record field area is referenced by the closed-loop feedback record of the subsequent S600 to support the version traceability of the threshold weight update package.

[0030] In an engineering embodiment, taking the vending machine aisle as the scenario type field as an example, the session configuration package is loaded by the cabinet controller at the start of the operation period and sent to the synchronous acquisition and quality assessment module. The visual probe list includes probe number fields and bitstream parameter fields for the fixed cameras inside the cabinet and the side cameras on the cabinet door. The synchronous triggering parameter adopts master control triggering and the acquisition cycle is limited by the trigger period field. The synchronous acquisition and quality assessment module writes image frame data to the window buffer and generates a time window index at each triggering cycle. Then, within the same time window index, it calculates the sharpness index and exposure index. The occlusion ratio, jitter index, and time synchronization deviation are written into the quality label set, and then the weight vector corresponding to the quality weight version field is aggregated to generate the evidence credibility table. The gating judgment sub-link completes the admission judgment and generates the admission evidence frame set based on the upper limit threshold of occlusion ratio and the upper limit threshold of time synchronization deviation. At the same time, the image frame data corresponding to the abnormal occlusion ratio caused by the cabinet door opening is written into the gap mark of the slot to be supplemented collection and registration. The admission evidence frame set is called by S300 for candidate target set generation, and the slot to be supplemented collection is called by S400 for supplementary collection evidence re-inference for counter-evidence verification.

[0031] S200 takes the session configuration package as input, and outputs an admission evidence frame set and a supplementary collection set through synchronous acquisition, index calculation, quality label set generation, evidence credibility table aggregation and admission control determination. The admission evidence frame set includes a probe number field, a time window index and an evidence credibility field, and is called as the input position of the "admission evidence frame set" in S300. The supplementary collection set includes a probe number field, a time window index and a slot gap mark, and is called as the input position of the "supplementary collection set" in S400.

[0032] Summary of the technical effects of this step: This step completes cross-probe synchronous acquisition and windowed organization under the constraints of the session configuration package, and binds image frame data with a unified time axis to form a traceable acquisition record; This step structures the multi-dimensional quality status through a quality label set and an evidence credibility table and introduces the admission control caliber; This step outputs the admission evidence frame set and merges it with the acquisition to be supplemented to form a split input baseline for S300 and S400.

[0033] To address the problem that existing technologies often rely on single-frame or single-probe results for target detection and tracking inference, lacking a unified inference entry point and trajectory-level primary key organization under the constraints of admission evidence, and lacking a templated assembly mechanism between the inference results and the business evidence structure, making it difficult to precipitate evidence and describe gap states with fixed field calibers when referencing across steps, this invention completes target detection and tracking inference under the constraints of the admission evidence frame set in step S300, forming a candidate target set containing a trajectory number field; and assembles the candidate target information into an evidence slot record set based on the task slot template version, completing the binding of the evidence credibility field; simultaneously, the gap state to be supplemented is introduced into the evidence slot record set through the slot gap marker field, providing a referable basis for subsequent consistency verification and rebuttal verification triggering from the evidence structure level, specifically including: S300. Based on the set of admission evidence frames, perform target detection inference and target tracking inference to generate a candidate target set; assemble the candidate target set into an evidence slot record set according to the task slot template version; the evidence slot record set includes probe number, time window index, target feature summary, evidence credibility, and slot gap marker; Specifically, step S300 is executed by the inference and slot assembly module of the visual recognition system. Its input source is the set of admission evidence frames output by step S200, and it synchronously references the task slot template version field, the installation geometric parameters of each visual probe, and the evidence credibility field in the evidence credibility table from the session configuration package. The set of admission evidence frames serves as the inference input carrier in this step, including image frame data, probe number field, time window index, frame sequence number field, acquisition timestamp field, reference timestamp field, and evidence credibility field. Among them, the image frame data serves as the original input for target detection inference, the probe number field and time window index serve as the primary key for cross-probe and cross-window organization, and the evidence credibility field serves as one of the weight inputs for subsequent slot assembly. When the inference and slot assembly module enters the running state, it first performs a consistency check on the session configuration package hash and the session configuration package version number, and writes a pass flag to the inference session record. When the consistency check fails, the inference and slot assembly module generates an inference anomaly record and binds it to the probe number field and time window index, writes the inference anomaly record to the inference session record, and continues to execute the inference process, thus forming a traceable abnormal operation chain. Understandably, the minimum input set of this step consists of the admission evidence frame set, the task slot template version field, and the evidence credibility field. The installation geometric parameters of each visual probe are loaded as preferred extended inputs to provide a geometric caliber for the spatial constraint description of the candidate target set and the generation of cross-probe slot consistency fields.

[0034] Furthermore, the target detection inference is executed by a detection inference submodule, which includes a model loading unit, a preprocessing unit, an inference execution unit, and a post-processing unit. The model loading unit reads the recognition model version parameter group from the session configuration package and loads the target detection model, while simultaneously registering the model version record field and binding it to the session configuration package version number field. The target detection model is a network model composed of a feature extraction layer, a feature fusion layer, and a detection head. The feature extraction layer extracts multi-scale features from the image frame data, the feature fusion layer fuses the multi-scale features and outputs a fused feature map, and the detection head performs candidate box regression and category determination on the fused feature map and outputs intermediate detection results. The preprocessing unit performs size normalization, distortion correction, and color space normalization on the image frame data. Distortion correction is performed based on the probe intrinsic parameter field and distortion coefficient field in the geometric parameters of each visual probe installation, ensuring that the image frame data corresponding to different probe number fields maintain geometric consistency. When the geometric parameters of each visual probe installation are unavailable, the preprocessing unit writes a distortion correction missing marker into the inference session record and continues to execute the remaining preprocessing steps. The inference execution unit inputs the preprocessed image frame data into the target detection model and generates intermediate detection results. These intermediate results include a candidate box set, a candidate box confidence score set, and a class label set. The post-processing unit performs candidate box merging and overlap resolution on the intermediate detection results, and generates target box coordinate fields, class label fields, and confidence score fields for each candidate box. Simultaneously, it backfills the probe number field, time window index, and frame sequence number field into each candidate box record, forming a frame-level detection result record. This frame-level detection result record is organized as a trajectory candidate input packet as input for subsequent target tracking inference. The trajectory candidate input packet registers its input reference information and is bound to a session configuration packet hash in the inference session record.

[0035] Furthermore, the target tracking inference is executed by a tracking inference submodule, which includes a feature extraction unit, an association determination unit, a trajectory management unit, and a trajectory output unit. The feature extraction unit extracts the target bounding box coordinate field from the frame-level detection result record and extracts appearance feature vectors within the target bounding box region. These appearance feature vectors are vectorized descriptive information representing the target's appearance. The appearance feature vectors and the target bounding box coordinate field are jointly written into the target candidate feature record. The association determination unit uses the target candidate feature record as input and performs association determination between consecutive frame sequence number fields corresponding to the same probe number field. Association determination includes position neighborhood gating and appearance similarity gating. Position neighborhood gating determines candidate association relationships based on the displacement change of the target bounding box coordinate field, while appearance similarity gating determines candidate association relationships based on the similarity of appearance feature vectors. When the time window index changes, the association determination unit performs cross-window association on the first and last frames of adjacent time window indices based on the temporal continuity constraint formed by the alignment window field, thereby ensuring that the trajectory number continues between adjacent time window indices. The trajectory management unit manages the candidate association relationships output by the association determination through trajectory merging and trajectory segmentation. Trajectory merging is used to combine the association results of the same target in multiple frames into a single trajectory number. Trajectory segmentation is used to terminate the existing trajectory number and generate a new trajectory number when the association determination is not satisfied. The trajectory management unit generates a trajectory event record each time it merges or segments. The trajectory event record includes a trajectory number field, an event type field, an association confidence field, and a time window index field, and is written into the inference session record. The trajectory output unit generates a trajectory continuity marker after the trajectory number stabilizes. The trajectory continuity marker is used to characterize the stability of the trajectory in consecutive frames and consecutive time window indices. The trajectory continuity marker and the trajectory number field are written together into the trajectory result record. Understandably, the trajectory number field and trajectory continuity marker output by the target tracking inference are one of the necessary fields in the candidate target set of this step. The target box coordinate field, category label field, confidence score field, and appearance feature vector are the other necessary fields in the candidate target set. Together, they constitute the input basis for assembling the evidence slot record set according to the task slot template version.

[0036] Furthermore, the process of generating the candidate target set is executed by the candidate set orchestration submodule. This submodule aligns the frame-level detection result records and trajectory result records using primary keys and generates candidate target records. Each candidate target record includes a target bounding box coordinate field, a category label field, a confidence score field, an appearance feature vector, a trajectory number field, and a trajectory continuity label. It also carries a probe number field and a time window index to form a cross-probe convergent candidate target index. The candidate set orchestration submodule further performs deduplication and conflict resolution on candidate target records within the same time window index. Conflict resolution is selected based on the overlap relationship of the target bounding box coordinate field and the confidence score field. The resolution process is written to a conflict resolution record, which is then bound to the probe number field and the time window index and written to the inference session record. For candidate target records from different probe number fields but falling within the same time window index, the candidate set orchestration submodule establishes a cross-probe geometric mapping relationship by referencing the probe extrinsic parameter fields in the installation geometry parameters of each visual probe, and generates a geometric consistency auxiliary field. This geometric consistency auxiliary field is one of the input sources for the geometric constraint residuals in the subsequent S400 cross-probe consistency verification. When the probe extrinsic parameter field is unavailable, the candidate set orchestration submodule writes a geometric consistency missing flag into the inference session record and continues to output the candidate target set. Finally, the candidate set orchestration submodule aggregates the candidate target records by time window index and outputs the candidate target set. The candidate target set hash is registered in the inference session record and bound to the session configuration package version number field.

[0037] Furthermore, the process of assembling the candidate target set into an evidence slot record set according to the task slot template version is executed by the slot assembly submodule. The slot assembly submodule uses the slot definition table field corresponding to the task slot template version field as the assembly rule input, and the candidate target set and the evidence credibility field as the assembly data input. The slot definition table fields contain the slot numbers and value structures of identity consistency slots, temporal continuity slots, geometric consistency slots, and environmental confidence slots. The slot assembly submodule extracts and aggregates fields from the candidate target set according to the slot numbers, generates slot values, and writes them into the evidence slot record. Specifically, for identity consistency slots, the slot assembly submodule extracts the category marker field, appearance feature vector, and trajectory number field from the candidate target record, aggregates them according to the trajectory number field to generate an identity feature summary, and writes it into the target feature summary field; for temporal continuity slots, the slot assembly submodule generates a temporal summary from the trajectory continuity marker and frame sequence number fields and merges it into the target feature summary field; for geometric consistency slots, the slot assembly submodule generates a geometric summary from the geometric consistency auxiliary field and merges it into the target feature summary field; for environmental confidence slots, the slot assembly submodule generates an environmental summary from the evidence credibility field bound to the candidate target record and the occlusion ratio field and jitter index field of the quality label entry, and merges it into the target feature summary field. The target feature summary field is a summary structure of the candidate target record under the slot semantics, which includes an identity summary subfield, a temporal summary subfield, a geometric summary subfield, and an environmental summary subfield. These subfields are written as structured fields into the evidence slot record set and are called by the consistency scoring table generation process in subsequent S400.

[0038] Furthermore, when generating each evidence slot record, the slot assembly submodule writes the probe number field and time window index as the primary key into the evidence slot record set, and fills the evidence credibility field back into the evidence credibility field of the evidence slot record set, thereby binding evidence quality with slot evidence. If a candidate target set is missing under a certain probe number field and a certain time window index, the slot assembly submodule reads the slot gap mark from the supplementary acquisition set output by S200 and writes it into the slot gap mark field of the evidence slot record set. For slots marked as mandatory in the slot template version field, if their corresponding slot value is missing, the slot assembly submodule generates a slot gap reason field and writes it into the inference session record, while simultaneously setting the slot gap mark field to a gap state. This allows S400 to directly reference this gap state record and locate the supplementary acquisition trigger record in the supplementary acquisition set when triggering rebuttal verification. Understandably, the minimum set of fields in the evidence slot record set consists of the probe number field, time window index, target feature summary field, evidence credibility field, and slot gap marker field. This minimum set of fields, together with the session configuration packet hash, forms a cross-step reference caliber, which is used by the S400 to perform primary key alignment and score fusion when performing cross-probe consistency verification.

[0039] In an engineering embodiment, taking the implementation of scene type field as an example, the set of admission evidence frames is output by the vehicle multi-camera acquisition link and includes probe number fields and time window indexes for the front and side cameras. The inference and slot assembly module loads the recognition model version parameter group in each time window index and performs target detection inference on the image frame data to generate frame-level detection result records containing category marker fields such as vehicles, pedestrians, and traffic signs. Subsequently, target tracking inference is performed, and a trajectory number field is formed based on appearance feature vectors and location neighborhood gating, and a trajectory continuity marker is generated. The candidate set arrangement submodule generates a geometric consistency auxiliary field for candidate target records across cameras and writes it into the candidate target set. The slot assembly submodule assembles the candidate target set into an evidence slot record set according to the task slot template version field and writes the evidence credibility field and slot gap marker field into the corresponding records. The evidence slot record set is called as input to S400 in this patent process to generate a consistency scoring table and trigger rebuttal verification on records that meet the abnormal threshold conditions. The slot gap mark field and the evidence credibility field participate in the abnormal threshold condition determination and rebuttal verification input positioning of S400.

[0040] S300 takes the set of admission evidence frames as input, generates a candidate target set through target detection inference and target tracking inference, and assembles the candidate target set into an evidence slot record set under the constraint of the task slot template version; wherein, the evidence slot record set includes a probe number field, a time window index, a target feature summary field, an evidence credibility field, and a slot gap marker field, and is called as the input position of the "evidence slot record set" in S400.

[0041] This step's technical effects can be summarized as follows: Under the constraints of the admission evidence frame set, this step completes detection and tracking reasoning and forms a candidate target set containing the trajectory number field; this step assembles the candidate target information into an evidence slot record set based on the task slot template version and completes the binding of the evidence credibility field; this step introduces the gap state to be collected into the evidence slot record set through the slot gap marker field and provides a reference basis for the S400's rebuttal verification trigger.

[0042] To address the common issues in existing technologies where cross-probe consistency verification of multi-probe evidence often lacks primary key alignment rules and scoring fusion criteria, the scoring source and version reference relationship are unclear, and there is a lack of a rebuttal verification link linked with supplementary evidence after anomaly detection, resulting in a lack of traceable evidence and rebuttal chains in the anomaly detection process and difficulty in forming a verifiable closed loop for anomaly threshold triggering conditions, this invention completes cross-probe primary key alignment and multi-dimensional consistency score fusion under the evidence slot record set caliber in step S400, and registers the scoring source and version reference relationship; when the anomaly threshold condition is met, the supplementary evidence in the pending supplementary collection set is retrieved and the established inference link is reused for re-inference, outputting a rebuttal result table; thus forming a dual-table structure of consistency score table and rebuttal result table, providing traceable evidence reference links and rebuttal link inputs for subsequent anomaly conclusion record generation, specifically including: S400: Based on the evidence slot record set, perform cross-probe consistency verification and generate a consistency score table; trigger rebuttal verification for records in the consistency score table that meet the abnormal threshold conditions; the rebuttal verification calls the supplementary evidence to be collected and performs re-reasoning, and outputs the rebuttal result table. Specifically, S400 is executed by the consistency verification and rebuttal verification module of the visual recognition system. Its input source is the evidence slot record set output by S300, and it synchronously references the unified time reference field, probe role labeling table, field of view coverage relationship table, and task slot template version field in the supplementary acquisition package output by S200 and the session configuration package output by S100. The evidence slot record set serves as the structured input for consistency verification in this step, including the probe number field, time window index field, target feature summary field, evidence credibility field, and slot gap mark field. Among them, the probe number field and time window index field constitute the primary key caliber for cross-probe alignment, the target feature summary field is the summary carrier of identity consistency slots, temporal continuity slots, geometric consistency slots, and environmentally credible slots, the evidence credibility field serves as one of the weight inputs during scoring fusion, and the slot gap mark field serves as one of the trigger location inputs for rebuttal verification. When the consistency verification and disproving verification module enters the running state, it first reads the session configuration package hash and the session configuration package version number and generates a verification session identifier. The verification session identifier is then bound to the time window index field and written to the consistency verification log. When the version number corresponding to the verification session identifier is inconsistent with the task slot template version field, the consistency verification and disproving verification module generates a version inconsistency record and associates this record with the evidence slot record set hash, writing it to the consistency verification log, thereby forming an auditable version reference chain. Understandably, the evidence slot record set, the supplementary acquisition set, and the unified time reference field constitute the minimum input set for S400 operation. The probe role labeling table and the field of view coverage relationship table are loaded as preferred extended inputs for selecting consistency scoring items and pruning cross-probe pairing ranges.

[0043] Furthermore, the cross-probe consistency verification is performed by the consistency score generation submodule, which includes a primary key alignment unit, a score item construction unit, a constraint residual calculation unit, a score fusion unit, and a table generation unit. The primary key alignment unit uses the probe number field and the time window index field as index keys to establish a cross-probe candidate pairing set within the same time window index. The generation rules for the candidate pairing set are jointly defined by the probe role labeling table and the field of view coverage relationship table. The probe role labeling table is used to limit the range of probe numbers participating in the consistency verification and label their role types, while the field of view coverage relationship table is used to limit probe number combinations with overlapping fields of view. The primary key alignment unit writes pairing exclusion records for probe number combinations not in the field of view coverage relationship table and binds them to the time window index field to write them into the consistency verification log. The scoring item construction unit parses the target feature summary field on the candidate pairing set, extracting the identity summary subfield, time sequence summary subfield, geometric summary subfield, and environment summary subfield respectively, generating candidate inputs for identity consistency scoring, location continuity scoring, and event causal consistency scoring, and writing the evidence credibility field as a weighted input for each scoring item into the scoring item input package. The constraint residual calculation unit performs constraint residual calculation on the geometric consistency scoring item, with its input being the geometric summary subfield and the installation geometric parameters of each visual probe in the session configuration package. The constraint residual calculation unit establishes a geometric mapping relationship for cross-probe candidate target records within the same time window index and generates a geometric constraint residual field; when the installation geometric parameters of each visual probe are missing or do not meet the geometric mapping constraints, the constraint residual calculation unit generates a geometric residual unavailable flag and binds the flag to the probe number field and the time window index field and writes it into the consistency verification log. The scoring fusion unit fuses identity consistency scores, location continuity scores, geometric constraint residuals, and event causal consistency scores according to scoring fusion rules. The scoring fusion rules are pre-registered rule structures, including scoring item weight fields, evidence credibility weight fields, and gap reduction fields. The gap reduction field is invoked when the slot gap marker field is in a gap state. During the fusion process, the scoring fusion unit generates a consistency score field and a score source record field for each candidate pair. The score source record field specifies the scoring items participating in the fusion and their weight reference information. The table generation unit writes the consistency score field, score source record field, probe number field pair, time window index field, slot gap marker field, and evidence credibility field into the consistency score table, and generates a consistency score table hash field and registers the hash reference relationship in the consistency verification log, so that the consistency score table can be traced and referenced in subsequent steps. Understandably, in this step, the "consistency score table" is a structured output product; its field definitions are fixed in this step and it is invoked as the input location of the S500's "consistency score table."

[0044] Furthermore, the triggering of rebuttal verification for records meeting the abnormal threshold conditions in the consistency scoring table is executed by the rebuttal triggering and complex reasoning submodule. This submodule includes a threshold determination unit, a rebuttal location unit, a supplementary evidence retrieval unit, a complex reasoning execution unit, and a result table generation unit. The threshold determination unit takes the consistency scoring table as input and reads the abnormal threshold condition record structure. This structure includes a consistency scoring threshold field, a geometric constraint residual threshold field, an evidence credibility threshold field, and a gap triggering condition field. The threshold determination unit performs threshold comparison on each record in the consistency scoring table and generates a rebuttal triggering mark field for records meeting the abnormal threshold conditions. Simultaneously, it generates a rebuttal time window index field and a rebuttal probe number field. The rebuttal time window index field is taken from the time window index field and combined with a unified time reference field for window positioning. The rebuttal probe number field is determined by the probe role labeling table to be a probe number that has a mutual verification relationship with the trigger record probe number field. The counter-evidence localization unit inputs the counter-evidence time window index field and the counter-evidence probe number field into the acquisition set to be supplemented for index retrieval. The acquisition set to be supplemented contains candidate records for supplementation registered by S200 and their corresponding slot gap marker fields. Based on this, the counter-evidence localization unit generates supplementation evidence index records and writes them into the counter-evidence retrieval log. When the acquisition set to be supplemented does not contain matching records, the counter-evidence localization unit generates a missing supplementation record and binds this record with the trigger record identifier of the consistency scoring table and writes it into the counter-evidence retrieval log. The supplementation evidence retrieval unit retrieves supplementation evidence from the supplementation cache or acquisition queue based on the supplementation evidence index records. The supplementation evidence is a combination of image frame data and its corresponding probe number field, time window index field, acquisition timestamp field, frame sequence number field, and quality label entry summary field. The supplementation evidence retrieval unit performs integrity verification on the retrieved supplementation evidence and registers the supplementation evidence hash field. The supplementation evidence hash field is written into the counter-evidence retrieval log and establishes a reference relationship with the hash field of the consistency scoring table.

[0045] Furthermore, the complex inference execution unit performs the same target detection inference and evidence slot assembly as S300 on the supplementary evidence. Its runtime reuses the assembly rules of the model loading unit, preprocessing unit, inference execution unit, and slot assembly submodule of the inference and slot assembly module. The model version record field and the task slot template version field are still provided by the session configuration package and matched with the version reference information in the consistency scoring table. During the complex inference process, the complex inference execution unit inputs the image frame data from the supplementary evidence into the target detection inference to obtain complex inference candidate target records, and assembles and generates rebuttal evidence slot records according to the task slot template version field. The complex inference execution unit aligns the rebuttal evidence slot records with the evidence slot records corresponding to the trigger records, and inputs them into the scoring item construction unit and constraint residual calculation unit of the consistency scoring generation submodule to generate a rebuttal consistency scoring field. The complex inference execution unit further generates a counter-evidence pass marker field. This counter-evidence pass marker field is a tokenized representation of the threshold comparison result between the counter-evidence consistency score field and the anomaly threshold condition record structure. This marker is then bound to the counter-evidence time window index field and the counter-evidence probe number field and written into the counter-evidence operation log. The result table generation unit writes the counter-evidence consistency score field, counter-evidence pass marker field, counter-evidence time window index field, counter-evidence probe number field, supplementary evidence hash field, and trigger record identifier field into the counter-evidence result table, and generates a counter-evidence result table hash field, which is registered in the counter-evidence operation log, enabling the counter-evidence result table to have a traceable input-output link. Understandably, the counter-evidence result table, as one of the output products of S400, is called as the input location of the "counter-evidence result table" in subsequent S500, and participates in the generation of anomaly conclusion records together with the consistency score table.

[0046] In an engineering implementation example, taking the implementation of smart park visual recognition as the scene type field, park entrances and exits, building passages, and key areas correspond to different probe role labeling tables. The field of view coverage relationship table limits the bidirectional probes at entrances and exits to form mutually verifying combinations with the cross-probe probes in passages. After the evidence slot record set output by S300 enters S400, the primary key alignment unit generates a cross-probe candidate pairing set within the same time window index, and constructs a scoring item input package for the identity digest subfield and the time sequence digest subfield. At the same time, it calculates the geometric constraint residual field based on the installation geometric parameters of each visual probe. The scoring fusion unit outputs the consistency scoring field according to the scoring fusion rules and writes it into the consistency scoring table. When the consistency score field and the geometric constraint residual field satisfy the threshold comparison relationship in the anomaly threshold condition record structure, the threshold determination unit generates a counter-evidence time window index field and a counter-evidence probe number field, and retrieves supplementary collection candidate records in the supplementary collection list. The supplementary collection evidence retrieval unit retrieves the supplementary collection evidence and registers the supplementary collection evidence hash field. The re-inference execution unit reuses the target detection inference and evidence slot assembly of S300 to generate counter-evidence evidence slot records and outputs the counter-evidence consistency score field and the counter-evidence pass mark field. The result table generation unit generates a counter-evidence result table and registers the counter-evidence result table hash field. The consistency score table and the counter-evidence result table are passed to S500 in this patent process and are called as the input positions of the "consistency score table" and the "counter-evidence result table" of S500, respectively, to generate anomaly conclusion records and match the handling strategy library version.

[0047] The S400 takes the evidence slot record set as input, performs cross-probe consistency verification, and outputs a consistency score table. For records in the consistency score table that meet the abnormal threshold conditions, it triggers a counter-evidence verification. The counter-evidence verification calls the supplementary evidence to be collected and performs re-inference to output a counter-evidence result table. The consistency score table includes identity consistency score, location continuity score, geometric constraint residual, event causal consistency score, and consistency score fields, and together with the counter-evidence result table, it serves as the input of the S500.

[0048] This step's technical effects can be summarized as follows: This step completes cross-probe primary key alignment and multi-dimensional consistency score fusion under the evidence slot record set caliber, and registers the score source and version reference relationship; This step retrieves supplementary evidence from the pending supplementary collection set under the triggering of abnormal threshold conditions and reuses the S300 inference link to output the counter-evidence result table; This step forms a dual-table structure of consistency score table and counter-evidence result table, enabling subsequent abnormal conclusion records to generate traceable evidence links and counter-evidence links.

[0049] To address the problems in existing technologies where abnormal conclusion generation is often based solely on a single score or reasoning result, lacking an assembly mechanism that integrates consistency scores and counter-evidence verification with a consistent caliber, and where risk grading and handling strategy matching lack version field constraints and evidence citation registration, resulting in a lack of verifiable mapping between the handling chain and the session configuration package, and difficulty in providing a stable input caliber for subsequent review and parameter updates, this invention, through step S500, completes abnormal candidate screening, counter-evidence constraint fusion, and abnormal conclusion record assembly under the dual-table constraints of the consistency score table and the counter-evidence result table, and registers the evidence citation chain under the same session configuration package version number; subsequently, the abnormal conclusion record is mapped to a risk grading marker, and a handling instruction set is generated and stored in the database as a handling record under the version field constraint of the handling strategy library, ensuring that the handling record is consistent with the preceding evidence chain in terms of structure and version caliber, thereby providing an aligned input object for subsequent closed-loop feedback records and version updates, specifically including: S500: Generate abnormal conclusion records and risk classification markers based on the consistency scoring table and the counter-evidence result table; match the disposal strategy library version according to the risk classification markers and generate disposal instruction sets; execute the disposal instruction sets to generate disposal records. Specifically, step S500 is executed by the anomaly conclusion generation and handling orchestration module of the visual recognition system. Its input sources are the consistency scoring table and the counter-evidence result table output by S400. It also synchronously references the handling strategy library version field, task slot template version field, and unified time base field from the session configuration package output by S100. Furthermore, it references the evidence slot record set hash output by S300 and the consistency scoring table hash and counter-evidence result table hash registered by S400 as link tracing fields. The consistency scoring table serves as the main input data structure for anomaly determination in this step, containing identity consistency scoring fields, location continuity scoring fields, geometric constraint residual fields, event causality consistency scoring fields, consistency scoring fields, probe number field pairs, time window index fields, slot gap marker fields, and scoring source record fields. The counter-evidence result table serves as the counter-evidence constraint input data structure in this step, containing counter-evidence time window index fields, counter-evidence probe number fields, counter-evidence consistency scoring fields, and counter-evidence pass marker fields. It is associated with the records in the consistency scoring table through a trigger record identifier field. When the abnormal conclusion generation and handling orchestration module enters the running state, it reads the session configuration package hash and the session configuration package version number to generate a handling session identifier, binds the handling session identifier to the time window index field and writes it into the handling log, and writes the handling strategy library version field into the version reference area of ​​the handling log, so that the handling strategy reference link in this step and the preceding collection, inference and verification links complete the closed registration under the same session configuration package version number.

[0050] Furthermore, the generation of abnormal conclusion records based on the consistency scoring table and the counter-evidence result table is executed by the abnormality judgment and conclusion assembly submodule. This submodule includes a threshold loading unit, a record alignment unit, an abnormal candidate generation unit, a counter-evidence constraint fusion unit, a conclusion assembly unit, and a hierarchical marker generation unit. The threshold loading unit reads the abnormal threshold condition record structure and the threshold reference information of the scoring fusion rules. The abnormal threshold condition record structure includes a consistency scoring threshold field, a geometric constraint residual threshold field, an identity consistency scoring threshold field, a location continuity scoring threshold field, and an event causal consistency scoring threshold field, and also includes a gap triggering condition field associated with the slot gap marker field. The threshold loading unit binds the threshold reference information with the handling session identifier and writes it to the handling log. The record alignment unit establishes a candidate abnormal index for the consistency scoring table based on the time window index field and the probe number field, and aligns the counter-evidence records in the counter-evidence result table with the candidate abnormal index based on the trigger record identifier field, generating an abnormality judgment alignment package. When a counter-evidence record is missing, the record alignment unit generates a counter-evidence missing marker field and binds it with the candidate abnormal index, writing it to the handling log. The anomaly candidate generation unit performs a threshold comparison on a record-by-record basis for the anomaly determination alignment package. The threshold comparison adopts a hard threshold rule determination method, which is given by the anomaly threshold condition record structure. The anomaly candidate generation unit generates an anomaly candidate mark field when the consistency score field meets the consistency score threshold field and the geometric constraint residual field meets the geometric constraint residual threshold field, and generates a gap-driven anomaly mark field when the slot gap mark field meets the gap trigger condition field. The anomaly candidate generation unit writes the anomaly candidate mark field, the gap-driven anomaly mark field, the time window index field, and the probe number field into the anomaly candidate table and registers the anomaly candidate table hash.

[0051] Furthermore, the counter-evidence constraint fusion unit uses the counter-evidence result table as constraint input to perform fusion verification on the abnormal candidate table. Its fusion method is based on the counter-evidence pass flag field to perform state constraints: when the counter-evidence pass flag field is in the pass state, the counter-evidence constraint fusion unit rewrites the abnormal candidate flag field to the counter-evidence not supported flag field and writes the counter-evidence consistency score field into the counter-evidence score field of the abnormal candidate table; when the counter-evidence pass flag field is in the fail state, the counter-evidence constraint fusion unit retains the abnormal candidate flag field and writes the counter-evidence consistency score field into the counter-evidence score field; when the counter-evidence missing flag field exists, the counter-evidence constraint fusion unit generates a counter-evidence missing handling flag field and writes it into the abnormal candidate table. The conclusion assembly unit generates an abnormal conclusion record on records where the abnormal candidate flag field is in a valid state. The abnormal conclusion record includes an abnormal conclusion record identifier field, a time window index field, a probe number field pair, a conclusion type flag field, an evidence citation hint field, and a conclusion source version field. The conclusion type flag field is obtained by mapping the threshold trigger combination of the identity consistency score field, the location continuity score field, the geometric constraint residual field, and the event causality consistency score field. The evidence citation hint field records the combined reference relationship between the consistency score table hash, the disproving result table hash, and the evidence slot record set hash. The conclusion source version field records the session configuration package version number and the disposal strategy library version field. Understandably, the abnormal conclusion record identifier field serves as one of the primary key fields for subsequent disposal records, used to bind the disposal instruction set to the abnormal conclusion record.

[0052] Furthermore, the generation of risk grading labels is performed by a grading label generation unit. This unit takes abnormal conclusion records as input and reads the risk grading rule structure. The risk grading rule structure includes grading dimension fields, dimension weight fields, and grading mapping fields. The grading dimension fields include a consistency score field, a counter-evidence score field, an evidence credibility field, and a slot / gap label field. The evidence credibility field is introduced from the consistency score table and, together with the score source record field, defines its value. The slot / gap label field is used to grade and correct the gap-driven abnormal label field. The grading label generation unit calculates the grading feature vector for the abnormal conclusion records and performs grading mapping to generate risk grading label fields. It then binds these risk grading label fields to the abnormal conclusion record identifier field and writes them into the abnormal conclusion record. When a grading dimension field is missing, the grading label generation unit generates a grading uncertainty label field and writes it into the abnormal conclusion record. Simultaneously, it writes the grading uncertainty label field into the handling log, ensuring that subsequent review and closed-loop feedback records have an entry field in S600. At the end of this step, the abnormal conclusion record and the risk classification marker field generate an abnormal conclusion record hash and register the hash reference relationship in the disposal log, so that the generation process of the abnormal conclusion record has an auditable version and evidence caliber.

[0053] Furthermore, the process of matching the risk-level label to the disposal strategy library version and generating a disposal instruction set is executed by the disposal strategy matching and instruction generation submodule. This submodule includes a strategy library loading unit, a hierarchical matching unit, an instruction template rendering unit, an instruction orchestration unit, and an instruction output unit. The strategy library loading unit loads the disposal strategy library version record structure based on the disposal strategy library version field in the session configuration package. This record structure includes a strategy entry index field, an applicable risk-level label field, an action type field, an action parameter field, and an instruction template version field. The strategy library loading unit writes the disposal strategy library version field and the instruction template version field to the disposal log. The hierarchical matching unit takes the abnormal conclusion record and the risk-level label field as input, retrieves the matching strategy entry index field from the disposal strategy library version record structure, and binds the abnormal conclusion record identifier field to the strategy entry index field to generate a strategy matching record. When no matching strategy entry is found, the hierarchical matching unit generates a strategy missing label field and writes it to the disposal log, while simultaneously generating a review instruction candidate label field and writing it to the strategy matching record. The instruction template rendering unit renders policy entries based on the action type field and the instruction template version field. The action type field includes alarm instructions, supplementary sampling instructions, parameter adjustment instructions, and review instructions. The rendering fields for alarm instructions include alarm channel marker field, alarm content summary field, and associated abnormal conclusion record identifier field. The rendering fields for supplementary sampling instructions include supplementary sampling probe number field, supplementary sampling time window index field, and supplementary sampling priority marker field. The rendering fields for parameter adjustment instructions include parameter item marker field, parameter adjustment amount field, and effective range marker field. The rendering fields for review instructions include review queue marker field, evidence citation prompt field, and risk classification marker field. The instruction orchestration unit performs deduplication and orchestration on multiple policy matching records within the same time window index field. Deduplication uses the combined key of the abnormal conclusion record identifier field and the action type field. Orchestration sorts the probe number field for supplementary sampling based on the probe role labeling table and merges the windows of the supplementary sampling time window index field based on the unified time base field to generate a disposal instruction set. The disposal instruction set includes an instruction set identifier field, an instruction sequence number field, an action type field, an action parameter field, an associated abnormal conclusion record identifier field, and a disposal strategy library version field, and generates a disposal instruction set hash field which is written to the disposal log. Understandably, the disposal instruction set is one of the key output products of this step, and its action type field and action parameter field constitute the minimum output set of disposal orchestration in this patent process.

[0054] Furthermore, the generation of disposal records from the execution disposal instruction set is performed by the instruction execution and record archiving submodule, which includes an execution scheduling unit, an external interface adaptation unit, an execution receipt collection unit, and a record archiving unit. The execution scheduling unit queues the disposal instruction set according to the action type field and selects the scheduling priority flag field based on the risk level flag field to write to the execution queue. The external interface adaptation unit maps alarm instructions to message payloads of alarm channels and writes them to the alarm sending log; maps supplementary sampling instructions to supplementary sampling tasks on the acquisition side and writes the supplementary sampling probe number field and supplementary sampling time window index field to the supplementary sampling task sheet; maps parameter adjustment instructions to parameter service change requests and writes the parameter item flag field and parameter adjustment amount field to the parameter change sheet; and maps review instructions to task entries in the review queue and writes the evidence citation prompt field to the review task sheet. The execution receipt collection unit collects receipts from various external interfaces and generates a receipt status flag field and a receipt timestamp field. The receipt status flag field is bound to the instruction sequence number field and written into the execution receipt table. When the receipt status flag field is in a failure state, the execution receipt collection unit generates a failure reason summary field and writes it into the execution receipt table, and also writes the failure reason summary field into the handling log. The record archiving unit aggregates the handling instruction set, the execution receipt table, and the abnormal conclusion record to generate a handling record. The handling record includes a handling strategy library version field, an evidence slot record set hash field, an abnormal conclusion record identifier field, a risk classification flag field, an instruction set identifier field, a receipt status flag field, and a handling record timestamp field, and generates a handling record hash field to be registered in the handling log. The handling record is called as the input position for S600's "generating a closed-loop feedback record based on the handling record access review result," where the abnormal conclusion record identifier field and the risk classification flag field in the handling record are used to align and merge the review results in S600.

[0055] In an engineering implementation, taking a visual recognition scenario at an urban road intersection as an example, multi-directional camera probes at the intersection form probe number field pairs and generate the same time window index field under a unified time reference field. After the consistency score table output by S400 enters S500, the anomaly candidate generation unit generates anomaly candidate marker fields based on the consistency score threshold field and the geometric constraint residual threshold field, and introduces the counter-evidence pass marker field from the counter-evidence result table through the trigger record identifier field; the counter-evidence constraint fusion unit generates a counter-evidence not supported marker field or retains the anomaly candidate marker field accordingly, and the conclusion assembly unit generates anomaly conclusion record identifier field and registers the evidence citation prompt field. The graded marker generation unit generates risk graded marker fields according to the risk graded rule structure, and the graded matching unit retrieves the strategy entry index field in the disposal strategy library version record structure accordingly and renders alarm instructions or supplementary collection instructions. The instruction execution and record archiving submodule collects execution receipts and generates disposal records, so that the disposal records contain the disposal strategy library version field, the evidence slot record set hash field, the anomaly conclusion record identifier field, and the risk graded marker field, and serve as the input product of S600 to complete the cross-step connection.

[0056] S500 generates abnormal conclusion records and risk classification markers using the consistency scoring table and the counter-evidence result table as inputs. It then generates a disposal instruction set by matching the disposal strategy library version with the risk classification markers and executes the disposal instruction set to generate a disposal record. The abnormal conclusion record includes an abnormal conclusion record identifier field and an evidence citation prompt field. The risk classification marker field and the disposal strategy library version field together limit the citation scope of the disposal strategy. The disposal record includes the disposal strategy library version, the evidence slot record set hash, the abnormal conclusion record identifier, and the risk classification marker, and is transmitted to S600 as input.

[0057] In summary, the technical effects of this step are as follows: Under the dual constraints of the consistency scoring table and the counter-evidence result table, this step completes the screening of anomaly candidates, the fusion of counter-evidence constraints, and the assembly of anomaly conclusion records, and registers the evidence reference links under the same session configuration package version number; this step maps the anomaly conclusion records to risk classification markers and generates a set of disposal instructions under the constraints of the disposal strategy library version field, and stores them in the database as disposal records; the disposal records output by this step provide input criteria for subsequent closed-loop feedback record alignment and review results and version updates.

[0058] To address the issues in existing technologies where there is a lack of primary key alignment and error type solidification mechanisms between processing and review results, making it difficult to feed back error attribution such as false positives, missed positives, and uncertainties to the parameter updates of the anomaly threshold and slot weights, and where parameter updates lack registration of reference relationships with the session configuration package version number and policy version fields, leading to easy drift in version caliber and difficulty in forming an auditable and consistent parameter loading chain during system iteration, this invention generates a closed-loop feedback record and solidifies the error type marker based on primary key alignment between the processing record and the review result in step S600; groups and statistically analyzes the closed-loop feedback record to form the anomaly threshold parameter update volume and slot weight parameter update volume, and assembles them into a threshold weight update package; and registers the update version number of the threshold weight update package, establishing its reference relationship with the session configuration package version number and policy version field, so that subsequent steps have a consistent version caliber when loading parameters and maintain the continuous traceability of the closed-loop chain. Specifically, this includes: S600: Generate a closed-loop feedback record based on the access review results of the handling record; the closed-loop feedback record includes false alarm markers, missed alarm markers, and uncertain markers; update the abnormal threshold parameter and slot weight parameter according to the closed-loop feedback record, generate a threshold weight update package and register the update version number; Specifically, S600 is executed by the closed-loop feedback and parameter evolution module of the visual recognition system. Its input source is the handling record output by S500. During the access phase, it synchronously references the unified time base field, task slot template version field, and handling strategy library version field from the session configuration package output by S100. It also references the evidence slot record set hash output by S300 and the consistency scoring table hash and the counter-evidence result table hash output by S400 as evidence chain alignment fields. In this step, the handling record serves as the data carrier structure aligned to the primary key. It includes at least the handling strategy library version field, the evidence slot record set hash field, the abnormal conclusion record identifier field, the risk classification marker field, the instruction set identifier field, the receipt status marker field, and the handling record timestamp field. Among them, the abnormal conclusion record identifier field is used to perform primary key-level association with the review result, the risk classification marker field is used for grouping and merging in the closed-loop statistics, and the handling strategy library version field is used to limit the strategy scope targeted by this closed-loop feedback. The closed-loop feedback and parameter evolution module generates a feedback session identifier upon startup and binds the feedback session identifier with the session configuration package version number, session configuration package hash, and handling strategy library version fields, writing them into the feedback log to ensure that the reference to the preceding handling link in this step remains consistent in terms of version.

[0059] Furthermore, the access review results are executed by the review access and alignment submodule, which includes a review result acquisition unit, a review result standardization unit, and a processing alignment unit. The review result acquisition unit accesses the review result data stream from the review workbench, manual inspection terminal, or third-party business system. This data stream includes a review task identifier field, a review conclusion marker field, a review timestamp field, and a review evidence reference field; the review evidence reference field points to the hash field of the evidence slot record set or the abnormal conclusion record identifier field. The review result standardization unit performs enumeration mapping on the review conclusion marker field, mapping the review conclusions of external systems to a unified review conclusion marker field, and converts the review timestamp field to the time caliber defined by the unified time reference field, generating a standardized review result table. The standardized review result table includes an abnormal conclusion record identifier field, a review conclusion marker field, a review timestamp field, a review task identifier field, and a review evidence reference field. The disposal alignment unit uses the disposal records as the main table and the standardized review result table as the subordinate table. It performs one-to-one or one-to-many associations based on the abnormal conclusion record identifier field. In the one-to-many case, the latest record using the review timestamp field is designated as the primary review record, and the remaining records are written to the review history appendix and its hash is registered. When the standardized review result table lacks the abnormal conclusion record identifier field, the disposal alignment unit performs a backtracking match between the review evidence reference field and the evidence slot record set hash field. The backtracking match uses hash consistency comparison and the time window constraint of the disposal record timestamp field to generate candidate matching pairs, which are then written to the alignment confirmation table and their alignment uncertainty flag field is registered. Understandably, the abnormal conclusion record identifier field and the review evidence reference field constitute the minimum input set for achieving closed-loop alignment in this step, and the alignment uncertainty flag field is used to generate uncertainty flags in subsequent closed-loop feedback records.

[0060] Furthermore, the generation of closed-loop feedback records is executed by the feedback mark generation and archiving submodule, which includes an error type discrimination unit, a feedback record assembly unit, and a feedback archiving unit. The error type discrimination unit extracts the receipt status mark field, the review conclusion mark field, and the risk classification mark field from the alignment result output by the handling alignment unit, and generates an error type discrimination input package by combining the abnormal conclusion record identifier field with the handling type context in the handling chain. The handling type context is obtained by tracing back from the instruction set identifier field to the instruction execution receipt table, which contains at least an action type field and a receipt status mark field. The error type discrimination unit compares the review conclusion marker field with the conclusion status of the handling link. The conclusion status is obtained by tracing back from the abnormal conclusion record identifier field to the abnormal conclusion record. The abnormal conclusion record includes a conclusion type marker field and a risk classification marker field. During the comparison process, when the abnormal conclusion record shows an anomaly while the review conclusion marker field shows no anomaly, the error type discrimination unit generates a false alarm marker field and sets it to a valid state. When the abnormal conclusion record shows no anomaly while the review conclusion marker field shows an anomaly, the error type discrimination unit generates a missed alarm marker field and sets it to a valid state. When the alignment uncertainty marker field is valid or the review conclusion marker field is in an undeterminable state, the error type discrimination unit generates an uncertainty marker field and sets it to a valid state. Furthermore, when generating the false alarm marker field or the missed alarm marker field, the error type discrimination unit simultaneously generates an error cause attribution field. The error cause attribution field is obtained by mapping the slot gap marker field, the counter-evidence pass marker field, and the consistency score field, and this field is written to the feedback log so that subsequent parameter updates have a traceable error context.

[0061] Furthermore, the feedback record assembly unit assembles closed-loop feedback records using the anomaly conclusion record identifier field as the primary key. These closed-loop feedback records include false alarm flag fields, missed alarm flag fields, uncertain flag fields, risk classification flag fields, handling strategy library version fields, task slot template version fields, review timestamp fields, evidence slot record set hash fields, and consistency scoring table hash fields. A closed-loop feedback record hash field is then generated and written to the feedback archiving unit. The feedback archiving unit associates the closed-loop feedback record hash field, the session configuration package version number, and the handling record hash field with the previous session configuration package hash, evidence slot record set hash, and consistency scoring table hash, and registers this association in the version log. This ensures that the closed-loop feedback record forms a complete link reference relationship with the preceding session configuration package hash, evidence slot record set hash, and consistency scoring table hash. Understandably, the false alarm flag field, missed alarm flag field, and uncertain flag field in the closed-loop feedback record are core fields output in this step, while the remaining fields are version and evidence caliber fields supporting subsequent parameter updates.

[0062] Furthermore, the update of the anomaly threshold parameter and slot weight parameter based on the closed-loop feedback record is executed by the parameter update and version registration submodule. This submodule includes a statistical grouping unit, an update quantity calculation unit, a constraint checking unit, an update package configuration unit, and a version registration unit. The statistical grouping unit takes the closed-loop feedback record as input and performs group statistics based on the false alarm marker field, the missed alarm marker field, and the uncertainty marker field. Within each group, it performs hierarchical statistics based on the conclusion type marker field, the risk classification marker field, and the role field of the probe role labeling table, generating a feedback statistics table. The feedback statistics table includes a statistical window marker field, a group type marker field, a slot number field, a scoring item marker field, a sample count field, and an error count field. The slot number field and the scoring item marker field are obtained from the field mapping relationship between the evidence slot record set and the consistency scoring table. The update quantity calculation unit reads the current anomaly threshold parameters and the current slot weight parameters. The current anomaly threshold parameters include the consistency score threshold field, the geometric constraint residual threshold field, the identity consistency score threshold field, the location continuity score threshold field, and the event causal consistency score threshold field. The current slot weight parameters include the identity consistency slot weight field, the temporal continuity slot weight field, the geometric consistency slot weight field, and the environmental confidence slot weight field. In the false alarm group, the update quantity calculation unit performs a rollback calculation on the score item marker field that triggers the threshold to generate the anomaly threshold parameter update quantity field. In the missed alarm group, it performs a forward calculation on the score item marker field that does not trigger the threshold to generate the anomaly threshold parameter update quantity field. At the same time, in the uncertain group, it generates an uncertainty suppression coefficient field to limit the update magnitude. Furthermore, the update quantity calculation unit performs responsibility attribution on the slot number field based on the error cause attribution field, and generates the slot weight parameter update quantity field by lowering the weight of slots with higher false alarm contributions and raising the weight of slots with higher false alarm contributions. The update quantity field is then bound to the task slot template version field, so that the slot weight parameter update quantity field can be separated when the template version changes.

[0063] Furthermore, the constraint checking unit applies boundary constraints and consistency constraints to the abnormal threshold parameter update quantity field and the slot weight parameter update quantity field. Boundary constraints include threshold value range constraints and weight normalization constraints. Consistency constraints include threshold monotonicity constraints under the same scoring fusion rule and weight sparsity constraints under the same task slot template version field. When the constraint check fails, the constraint checking unit generates a rollback flag field and writes it to the feedback log. Simultaneously, it writes the corresponding update quantity to the update set to be reviewed and registers the hash of the update set to be reviewed for subsequent manual review or policy rollback process invocation. The update package assembly unit assembles the threshold weight update package when the constraint check passes. The threshold weight update package includes the abnormal threshold parameter update quantity field and the slot weight parameter update quantity field, and also includes an update applicable version field and an update effective window field. The update applicable version field is generated by combining the disposal strategy library version field and the task slot template version field, and the update effective window field is generated by combining the statistical window flag field and the unified time base field. The version registration unit generates an update version number for the threshold weight update package and establishes a reference relationship between the update version number and the session configuration package version number, the handling strategy library version field, the task slot template version field, and the closed-loop feedback record hash field, and registers it in the version log. At the same time, the version registration unit writes the update version number into the version index table of the parameter service, so that the abnormal threshold condition record structure of S400 and the weight vector of S200 can load the corresponding parameter set based on the update version number in subsequent sessions, and complete the versioned connection across steps.

[0064] In an engineering implementation example, taking a multi-probe visual recognition scenario for park perimeter security as an example, the handling record is generated by S500 and associated with the review task identifier field of the review workbench through the abnormal conclusion record identifier field. The review access and alignment submodule maps the review conclusion to the review conclusion marker field and aligns it with the handling record. The error type discrimination unit generates a false alarm marker field or a missed alarm marker field based on the review conclusion marker field and the conclusion type marker field of the abnormal conclusion record, and generates an uncertain marker field when the match is incomplete. The statistical grouping unit generates a feedback statistical table by grouping the false alarm marker field, the missed alarm marker field, and the uncertain marker field. The update quantity calculation unit generates the abnormal threshold parameter update quantity field and the slot weight parameter update quantity field accordingly. The update package assembly unit assembles the threshold weight update package and the version registration unit registers the update version number, so that the threshold weight update package is loaded in subsequent sessions and participates in the iterative reference of the abnormal threshold condition record structure of S400 and the evidence slot record set assembly caliber of S300.

[0065] The S600 takes the processing record output by the S500 as input to access the verification result and generates a closed-loop feedback record. The closed-loop feedback record includes a false alarm flag field, a missed alarm flag field, and an uncertain flag field, and is archived under the constraint of the version caliber field. Based on this, this step updates the anomaly threshold parameter and slot weight parameter according to the closed-loop feedback record, generates a threshold weight update package and registers the update version number. The threshold weight update package includes an anomaly threshold parameter update amount field and a slot weight parameter update amount field. The update version number is written to the version log and used for subsequent sessions to load and call.

[0066] This step's technical effects can be summarized as follows: Based on the primary key alignment of the handling records and review results, this step generates closed-loop feedback records and solidifies error type markers; this step groups and statistically analyzes the closed-loop feedback records to form the update volume of the abnormal threshold parameter and the update volume of the slot weight parameter, and assembles them into a threshold weight update package; this step registers the update version number of the threshold weight update package and establishes a reference relationship with the session configuration package version number and policy version field, so that subsequent steps have a consistent version caliber when loading parameters.

[0067] Example 2: Figure 2 A structural block diagram of an artificial intelligence-based visual recognition system according to an embodiment of the present invention is shown. Figure 2 As shown, the structure may include: The configuration and version management module 01 is used to acquire scene identifiers, a list of visual probes, installation geometric parameters of each visual probe, synchronization trigger parameters, task slot template versions, and handling strategy library versions, generate session configuration packages, and register the session configuration package hash and version number. Specifically, the configuration and version management module receives scene identifiers and a list of visual probes written by the external deployment side, and reads the installation geometric parameters, synchronization trigger parameters, task slot template versions, and handling strategy library versions of each visual probe. The scene identifier represents the scene category and boundary range of the monitoring area or work unit; the visual probe list represents the set of probe numbers participating in the acquisition and their connection port mappings; the installation geometric parameters of each visual probe include probe installation pose, field of view direction, focal length calibration results, and distortion correction parameters; the synchronization trigger parameters include trigger source identifier, trigger cycle, trigger tolerance, and lost trigger handling parameters; the task slot template version represents the set of slot numbers in the evidence slot record set and slot value constraints; and the handling strategy library version represents the risk. The mapping rule version from the hierarchical marker to the disposal instruction set; the configuration and version management module performs field normalization, missing item verification, and version conflict judgment on the aforementioned input objects. When a version conflict or field missing is detected, an exception is recorded and the system reverts to the most recent available version number. At the same time, the existing session configuration package is reused under the condition that the version remains unchanged; the configuration and version management module encapsulates the normalized input objects into a session configuration package, performs hash calculation on the session configuration package to obtain the session configuration package hash, writes the session configuration package hash and version number into the version registration area, and generates an index record; the session configuration package is output to the data acquisition and quality assessment module as the configuration input for synchronous acquisition. At the same time, the session configuration package hash and version number are written into the record objects of subsequent modules, so that the inference and slot assembly module, consistency verification and disproving verification module, disposal and recording module, and closed-loop feedback and parameter update module can complete the version association when writing the evidence credibility table, evidence slot record set, consistency score table, disproving result table, disposal record, and threshold weight update package.

[0068] The data acquisition and quality assessment module 02 is used to synchronously acquire image frames from each visual probe based on the session configuration package, calculate sharpness index, exposure index, occlusion ratio, jitter index, and time synchronization deviation, generate a quality label set, and generate an evidence credibility table according to the quality label set. After executing access control, it outputs the set of access evidence frames and the set to be supplemented for acquisition. Specifically, the data acquisition and quality assessment module receives the session configuration package from the configuration and version management module, establishes a probe connection session according to the visual probe list and loads the synchronization trigger parameters, issues a synchronization acquisition command to each visual probe, and... The local acquisition buffer registers probe numbers and frame arrival timestamps. When a connection is interrupted, a frame is lost, or a timeout occurs, an acquisition anomaly flag is written, and the corresponding probe number is added to the probe queue for supplementary acquisition. The data acquisition and quality assessment module calculates a sharpness index for each visual probe image frame. This sharpness index is formed by edge intensity statistics, contrast distribution, and blur kernel estimation. The exposure index is formed by the saturation area ratio of the brightness histogram and grayscale distribution offset. The occlusion ratio is formed by the proportion of the visible foreground area and statistics of the connected components at the occlusion boundary. The jitter index is calculated from adjacent frames. The global motion estimation residual and high-frequency displacement fluctuation statistics are used to form the time synchronization deviation, which is obtained by aligning the frame arrival timestamp with the triggering benchmark corresponding to the synchronization triggering parameter. The data acquisition and quality assessment module writes each indicator into a quality label set, which is organized according to probe number and time window index, and registers indicator missing and abnormal indicators in the quality label set. The data acquisition and quality assessment module performs admission control according to the hard threshold rules in the session configuration package. The hard threshold rules include the upper limit threshold of occlusion ratio and the upper limit threshold of time synchronization deviation. Image frames that do not meet the hard threshold rules are written into the supplementary acquisition and merged registration slot gap mark. At the same time, the data acquisition and quality assessment module aggregates and generates an evidence credibility table according to the quality label set and writes the evidence credibility into the frame metadata of the corresponding image frame. The data acquisition and quality assessment module outputs the admission evidence frame set to the inference and slot assembly module as input to the target detection inference and target tracking inference, and outputs the supplementary acquisition set to the consistency verification and disproven verification module for disproven verification call. At the same time, the evidence credibility table is written into shared storage for the inference and slot assembly module to read the evidence credibility field when assembling the evidence slot record set.

[0069] The inference and slot assembly module 03 is used to perform target detection inference and target tracking inference based on the admission evidence frame set to generate a candidate target set, and assemble the candidate target set into an evidence slot record set according to the task slot template version. Specifically, the inference and slot assembly module receives the admission evidence frame set and evidence confidence table from the data acquisition and quality assessment module, reads the task slot template version in the session configuration package, and loads the admission evidence frame set in segments according to the probe number and time window index to form an inference input batch and register the batch number and the corresponding session configuration package hash. In the target detection inference stage, the inference and slot assembly module performs candidate region generation, feature extraction, category determination and position regression on each image frame, outputs the candidate target set and writes the target box coordinates, category label and confidence score into the candidate target set. In the target tracking inference stage, the inference and slot assembly module performs cross-frame association based on the trajectory number and appearance feature vector. The appearance feature vector is generated by the target region feature extraction result, and the cross-frame association includes a time sequence gate. The system performs control, spatial neighborhood constraint, and appearance similarity matching, outputs trajectory continuity markers, and writes them back to the candidate target set. The inference and slot assembly module assembles the candidate target set into slots according to the task slot template version. Slot assembly includes selecting candidate target fields based on slot numbers, aggregating multi-frame trajectory fragments based on time window indexes, writing evidence confidence fields based on the evidence confidence table, and writing slot gap markers for missing slots. The inference and slot assembly module forms an evidence slot record set, which includes probe numbers, time window indexes, target feature summaries, evidence confidence, and slot gap markers. The target feature summaries are generated by compressing target bounding box coordinates, category labels, confidence scores, appearance feature vectors, and trajectory numbers, and are associated with trajectory continuity markers. The evidence slot record set is output to the consistency verification and disproven verification module as input for cross-probe consistency verification. At the same time, the inference and slot assembly module writes the hash of the evidence slot record set into shared storage for the disposal and recording module to associate and call when generating disposal records.

[0070] The consistency verification and rebuttal verification module 04 is used to perform cross-probe consistency verification based on the evidence slot record set and generate a consistency score table. It triggers rebuttal verification for records in the consistency score table that meet the abnormal threshold conditions, calls the supplementary evidence corresponding to the supplementary acquisition set, and outputs a rebuttal result table after re-inference. Specifically, the consistency verification and rebuttal verification module receives the evidence slot record set from the inference and slot assembly module and the supplementary acquisition set from the data acquisition and quality assessment module. It aggregates multi-probe evidence slot records according to the same time window index, performs cross-probe consistency verification, and generates a consistency score table. The cross-probe consistency verification includes the generation of identity consistency score, position continuity score, geometric constraint residual, and event causality consistency score. The identity consistency score is obtained by similarity matching of multi-probe target feature summaries. The position continuity score is obtained by spatial consistency after mapping the multi-probe target box coordinates through installation geometric parameters. The geometric constraint residual is obtained by statistical analysis of the field of view coverage relationship table and the geometric projection residual under installation geometric parameter constraints. The event causality consistency score is derived from the temporal relationship between trajectory continuity markers and time window indexes. The verification module writes the aforementioned scores into the consistency score table according to the scoring fusion rules, and registers the abnormal threshold condition matching results in the consistency score table. The abnormal threshold conditions are provided by the abnormal threshold parameters maintained by the closed-loop feedback and parameter update module. When the abnormal threshold parameter reading fails, it falls back to the abnormal threshold parameter corresponding to the most recent updated version number in the version registration area. When there is a record in the consistency score table that meets the abnormal threshold conditions, the consistency verification and counter-verification module triggers counter-verification verification and generates a counter-verification time window index and a counter-verification probe number. It calls the corresponding supplementary evidence in the supplementary collection and enters the re-inference process. The re-inference process calls the reasoning and slot assembly module to perform the same target detection reasoning and slot assembly action to obtain the counter-verification evidence slot record and performs consistency score calculation on the counter-verification evidence slot record. The consistency verification and counter-verification module outputs a counter-verification result table, which contains the counter-verification consistency score and counter-verification pass mark. The consistency score table and the counter-verification result table are output together to the handling and recording module as input for the generation of abnormal conclusion records. At the same time, the call record of the counter-verification process is written to the shared storage for the closed-loop feedback and parameter update module to trace and read.

[0071] The handling and recording module 05 is used to generate abnormal conclusion records and generate risk classification tags based on the consistency scoring table and the counter-evidence result table. It then matches the handling strategy library version with the risk classification tags to generate a handling instruction set and a handling record. Specifically, the handling and recording module receives the consistency scoring table and the counter-evidence result table from the consistency verification and counter-evidence verification module, reads the handling strategy library version from the session configuration package, aligns the consistency scoring table and the counter-evidence result table according to the time window index, writes an abnormal handling tag to records where the counter-evidence pass mark is invalid or missing, and registers a pending review identifier. The handling and recording module generates abnormal conclusion records based on the consistency scoring table and the counter-evidence result table, and registers the abnormal conclusion record identifier, associated probe number, time window index, and evidence slot record set hash in the abnormal conclusion record. Simultaneously, it generates risk classification tags based on the abnormal threshold condition matching results and the counter-evidence consistency score. Risk classification labeling; the handling and recording module generates a handling instruction set by matching the handling strategy library version according to the risk classification label. The handling instruction set includes alarm instructions, supplementary data collection instructions, parameter adjustment instructions, and review instructions. The alarm instructions are written to the external alarm channel release queue, the supplementary data collection instructions are written back to the pending supplementary data collection and scheduling entry of the data acquisition and quality assessment module, the parameter adjustment instructions are written to the parameter change request entry of the closed-loop feedback and parameter update module, and the review instructions are written to the review task queue and the review task identifier is registered. The handling and recording module generates a handling record, which includes the handling strategy library version, evidence slot record set hash, abnormal conclusion record identifier, and risk classification label. The handling record is output to the closed-loop feedback and parameter update module as input for the generation of closed-loop feedback records. At the same time, the release status of the handling instruction set is written to the status field of the handling record for alignment when the subsequent review results are accessed.

[0072] The closed-loop feedback and parameter update module 06 is used to generate closed-loop feedback records based on the handling records and the review results. These closed-loop feedback records include false alarm markers, missed alarm markers, and uncertain markers. The module updates the anomaly threshold parameters and slot weight parameters according to the closed-loop feedback records, generates a threshold weight update package, and registers the update version number. Specifically, the closed-loop feedback and parameter update module receives handling records from the handling and recording module and reads the review results associated with the handling records from the review task queue. It aligns the review results with the anomaly conclusion record identifier, risk classification marker, and evidence slot record set hash using primary keys to form a closed-loop feedback record. False alarm markers, missed alarm markers, and uncertain markers are written into the closed-loop feedback record. A false alarm marker corresponds to the review result negating the anomaly conclusion record; a missed alarm marker corresponds to the review result pointing to a missing anomaly conclusion record; and an uncertain marker corresponds to the review result failing to complete a deterministic judgment. The closed-loop feedback and parameter update module updates the anomaly threshold parameters and slot weight parameters according to the closed-loop feedback records. The update process includes updating the parameters according to the false alarm markers, missed alarm markers, and uncertain markers. The system marks groups for statistics, generates update volumes for abnormal threshold parameters and slot weight parameters, writes them to the version registration area, and completes version conflict judgment. When concurrent update conflicts occur, a new update version number is generated based on the version number increment rule, and conflict branch records are retained. The closed-loop feedback and parameter update module generates a threshold weight update package, which includes update volumes for abnormal threshold parameters and slot weight parameters, and registers the update version number. The threshold weight update package is written back to the consistency verification and rebuttal verification module as a parameter input for abnormal threshold condition judgment. At the same time, the update volume of slot weight parameters is written back to the data acquisition and quality assessment module for the aggregation weight update of the evidence credibility table, and the update version number is synchronized to the version registration area of ​​the configuration and version management module for session configuration package version association.

Claims

1. A visual recognition method based on artificial intelligence, characterized in that, include: S100: Obtain scene identifier, visual probe list, installation geometry parameters of each visual probe, synchronization trigger parameters, task slot template version, generate session configuration package and register session configuration package hash and version number; S200: Based on the session configuration package, synchronously acquire image frames from each visual probe, calculate sharpness index, exposure index, occlusion ratio, jitter index, and time synchronization deviation, and generate a quality label set; generate an evidence credibility table according to the quality label set and perform access control, outputting an access evidence frame set and a supplementary acquisition set. S300. Based on the set of admission evidence frames, perform target detection inference and target tracking inference to generate a candidate target set; assemble the candidate target set into an evidence slot record set according to the task slot template version; the evidence slot record set includes probe number, time window index, target feature summary, evidence credibility, and slot gap marker; S400: Based on the evidence slot record set, perform cross-probe consistency verification and generate a consistency score table; For records in the consistency scoring table that meet the abnormal threshold conditions, a counter-evidence verification is triggered. The counter-evidence verification calls the supplementary evidence to be collected and performs re-reasoning, and outputs the counter-evidence result table. S500: Generate abnormal conclusion records and risk classification markers based on the consistency scoring table and the counter-evidence result table; match the disposal strategy library version according to the risk classification markers and generate disposal instruction sets; execute the disposal instruction sets to generate disposal records. S600. Generate a closed-loop feedback record based on the access review results of the handling record; the closed-loop feedback record includes false alarm markers, missed alarm markers, and uncertain markers; Update the anomaly threshold parameters and slot weight parameters according to the closed-loop feedback record, generate the threshold weight update package and register the update version number.

2. The method according to claim 1, characterized in that, The session configuration package includes a probe role labeling table, a field of view coverage relationship table, a unified time base field, a task slot template version field, and a disposal strategy library version field.

3. The method according to claim 1, characterized in that, The entry control adopts a hard threshold rule, which includes an upper limit threshold for the occlusion ratio and an upper limit threshold for the time synchronization deviation. Image frames that do not meet the hard threshold rule are written into the gap mark of the slot to be acquired and merged.

4. The method according to claim 1, characterized in that, The evidence credibility table is generated by aggregating the quality tag set according to a weight vector, which is given by the quality weight version field in the session configuration package.

5. The method according to claim 1, characterized in that, The candidate target set includes target bounding box coordinates, category labels, confidence scores, appearance feature vectors, and trajectory numbers; the target tracking inference uses trajectory numbers and appearance feature vectors to perform cross-frame association and outputs trajectory continuity labels.

6. The method according to claim 1, characterized in that, The task slot template version defines identity consistency slots, temporal continuity slots, geometric consistency slots, and environmental confidence slots; the evidence slot record set stores slot values, slot source probe numbers, slot source time window indexes, and slot gap markers according to slot number.

7. The method according to claim 1, characterized in that, The cross-probe consistency verification generates identity consistency score, location continuity score, geometric constraint residual, and event causal consistency score, and generates a consistency score table according to the score fusion rules.

8. The method according to claim 1, characterized in that, The counter-evidence verification includes a counter-evidence time window index and a counter-evidence probe number. The counter-evidence verification performs the same target detection reasoning and evidence slot assembly as S300 on the counter-evidence. The counter-evidence result table includes a counter-evidence consistency score and a counter-evidence pass mark.

9. The method according to claim 1, characterized in that, The handling instruction set includes alarm instructions, supplementary sampling instructions, parameter adjustment instructions, and review instructions; the handling record includes the handling strategy library version, evidence slot record set hash, abnormal conclusion record identifier, and risk classification marker; the threshold weight update package includes the abnormal threshold parameter update quantity and slot weight parameter update quantity; the abnormal threshold parameter update quantity and slot weight parameter update quantity are generated by the closed-loop feedback record grouped and statistically analyzed according to false alarm marker, missed alarm marker, and uncertain marker.

10. A visual recognition system based on artificial intelligence, applied to the method of any one of claims 1-9, characterized in that, include: The configuration and version management module is used to obtain scene identifiers, visual probe list, installation geometric parameters of each visual probe, synchronization trigger parameters, task slot template version, and disposal strategy library version, and generate session configuration packages. The data acquisition and quality assessment module synchronously acquires image frames from each visual probe based on the session configuration package, calculates sharpness index, exposure index, occlusion ratio, jitter index, and time synchronization deviation, and generates a quality label set. The reasoning and slot assembly module performs target detection reasoning and target tracking reasoning to generate a candidate target set; The consistency verification and disproven verification module performs cross-probe consistency verification and generates a consistency scoring table. The handling and recording module generates abnormal conclusion records and risk classification markers based on the consistency scoring table; The closed-loop feedback and parameter update module generates closed-loop feedback records based on the handling records and the review results.

Images