A deep learning-based video event clue dynamic analysis method and system

Abstract

The application discloses a kind of video event clue dynamic analysis method and system based on deep learning, belong to intelligent video analysis and computer vision field, to solve the problem that main body clue is easy to break, misassociation and later evidence is difficult to backtrack historical relationship in the multi-channel video of asynchronous batch access, the application constructs double time event clue evidence graph, determines to be unlocked subgraph according to identity exclusion, physically inaccessible, carrying state is not continuous and event order conflict propagation, under the condition that relationship is unchanged outside lock boundary, identity, carrying state and event order are solved jointly, and version difference is updated clue chain and downstream index in orientation, realize the technical effect that local correction history clue is kept version traceable using new evidence.

Description

Technical Field

[0001] This invention relates to the fields of intelligent video analysis and computer vision, and in particular to a method and system for dynamic analysis of video event clues based on deep learning. Background Technology

[0002] With the increasing number of video cameras in scenarios such as public safety, park management, and traffic monitoring, multiple surveillance videos are typically sent to the analysis system asynchronously and in batches due to factors such as network bandwidth, storage scheduling, and device online status. Existing video analytics technologies can perform target detection, target tracking, cross-camera identity association, and event recognition on single-channel or short-term video, and use appearance features, motion trajectories, or camera topology to form clues about subjects and events.

[0003] However, the visual representation and relational state of the subject may become discontinuous after movement across cameras, brief obstruction, changes in lighting, alterations in clothing appearance, or the acquisition or placement of objects. This can easily lead to the same subject being split into multiple identities or different subjects being incorrectly merged. Adding identification results based solely on a single feature or a single video segment is insufficient to simultaneously constrain the subject's identity, the subject-object state, and the chronological relationship of events; erroneous associations accumulate as the chain of clues extends.

[0004] Furthermore, subsequently added video clips may provide more complete spatiotemporal evidence, thereby altering judgments about existing subject identities, state transitions, and event sequences. Existing one-way appending methods typically cannot pinpoint the historical scope of the actual impact of new evidence. A full recalculation would result in the reversal of irrelevant historical relationships, duplicate calculations, and repeated updates to downstream indexes, and it would be difficult to preserve the traceability between old conclusions and supporting evidence.

[0005] Therefore, there is a need for a dynamic analysis method and system for video event clues that can overcome the shortcomings of the existing technologies. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based method and system for dynamic analysis of video event clues. Addressing the problems of fragmented main clues, misassociations, and difficulty in correcting historical judgments due to later evidence caused by asynchronous batch access to videos, this invention proposes a technical solution combining dual-temporal observation, deep multi-semantic feature association, dual-temporal evidence graphs, conflict-driven influence domains, and version differential submission. The original observations and revisable relationships are separated and stored separately. Identity exclusivity, physical unreachability, discontinuous carried states, and event sequence conflicts are calculated and propagated along relational dependencies. Locking boundaries are formed at stable locations. Identity partitioning, carried state sequences, and event sequences are jointly solved only in the unlocked subgraphs. The clue chain and index are then updated in a targeted manner while retaining old versions and evidence sources. This invention possesses the technical effects of using newly added evidence to locally backtrack and correct historical relationships, suppressing irrelevant flips caused by weak evidence, and reducing redundant calculations across the entire graph and downstream redundant processing.

[0007] This invention provides a deep learning-based method for dynamic analysis of video event clues, including:

[0008] S1. Receive multiple video clips that are asynchronously accessed in batches, record the effective shooting time and the system access version time, perform target detection, local tracking and event detection, and generate trajectory clips, event clips and raw observations with source identifiers and time intervals;

[0009] S2. Utilize a deep feature coding network to extract the stable identity, variable appearance, subject-carrying object relationship, action event state features, and confidence level of the trajectory segments and event segments, and generate a candidate association set by combining the camera topology and time reachability relationship;

[0010] S3. Construct immutable evidence nodes based on the original observations, and construct revisable relational edges with start and end versions based on the identity attribution, state migration, event participation, and event sequence determined by the candidate association set, to form a dual-temporal event clue evidence graph.

[0011] S4. After a new batch of accesses, based on the newly added evidence and the evidence graph, calculate the conflict amount of identity exclusivity, physical inaccessibility, discontinuous carrying state and event sequence to form a conflict seed. Propagate the conflict amount along candidate associations, carrying state migration and event participation and sequence relationship to determine the subgraph to be unlocked and the locking boundary.

[0012] S5. Keep the relationship edges outside the locked boundary unchanged, and jointly solve the identity partition, subject-carrying item state sequence and event sequence in the subgraph to be unlocked by using association, state transition, event sequence and old version relationship change costs. Compare the solution results with the current version relationship to generate version difference, and update the event clue chain accordingly.

[0013] Optionally, S1 includes: determining the acquisition time interval written by the camera device or obtained by clock correction in the video segment as the effective shooting time, and determining the incrementing version number when the video segment is received and enters the analysis queue as the system access version time;

[0014] The detection boxes within the same camera are connected into trajectory segments based on inter-frame motion consistency and appearance feature similarity, and the start and end frames, camera identification, effective shooting time interval, and detection confidence of the trajectory segments are recorded.

[0015] Continuous frames with continuous motion states and boundary confidence scores not less than the event detection threshold are identified as event segments, and the event type, participating target, effective shooting time interval, and source video segment identifier are recorded.

[0016] Video clips lacking capture time or source identification are marked as evidence to be corrected and are prohibited from participating in cross-camera candidate associations until clock skew estimation or source verification is completed.

[0017] Optionally, S2 includes: performing temporal aggregation on multiple frame-level features of the trajectory segment using the identity encoding branch of the deep feature encoding network, and outputting the stable identity feature;

[0018] The appearance coding branch is used to encode the frame-level features corresponding to clothing color, texture and partial occlusion, and the variable appearance features are output.

[0019] The subject region, object region, and their relative positions are jointly encoded using relational encoding branches to output the subject-carrying object relationship features.

[0020] The action changes of the event segment are encoded using a temporal coding branch, and the action event state features are output.

[0021] The similarity of stable identity features, variable appearance features, subject-carrying object relationship features, and action event state features is multiplied by the confidence level of the corresponding feature, and the sum of the resulting products is divided by the sum of the confidence levels used in the calculation to obtain the fusion similarity.

[0022] The similarity threshold and confidence calibration value of each feature are determined based on the feature distribution of continuous trajectories from the same camera and the feature distribution of confirmed matching samples across cameras, and only two trajectory segments whose time interval is within the reachable interval of the corresponding camera topology and whose fusion similarity is not less than the similarity threshold are written into the candidate association set.

[0023] Furthermore, the training of the deep feature encoding network includes: constructing positive identity sample pairs based on cross-camera trajectory segments of the same subject; constructing negative identity sample pairs based on trajectory segments with overlapping effective shooting times and not satisfying the condition of the same subject; constructing relationship state training samples based on the acquisition, holding, putting down and transfer sequences of the subject and the carried object; and constructing action event training samples based on event segments with event type and start and end time labels.

[0024] The network parameters are updated by combining identity measurement loss, appearance attribute loss, subject-carrying object relationship classification loss, and action event temporal loss.

[0025] Using calibration samples that are independent of the training samples, the output scores of each branch are mapped to the confidence level. When some branches have missing input features, the confidence level of the corresponding branch is set to zero and the fusion weights are renormalized based on the remaining branches with input features. When all branches have missing input features, the corresponding original observation is marked as evidence to be corrected and is prohibited from participating in candidate association.

[0026] Optionally, S3 includes: saving the evidence identifier, source video segment identifier, camera identifier, effective shooting time interval, system access version time, feature summary and confidence level for each immutable evidence node;

[0027] For each revisable relation edge, store the relation type, starting version, ending version, supporting evidence identifier set, and relation confidence. The ending version of the currently valid relation edge is empty. When the relation is revoked, the version number of the revocation operation will be written to the ending version.

[0028] The events were queried in the order of occurrence according to the effective shooting time, and the identity attribution, carrying status and event relationship under any submitted version were reconstructed according to the system access version time.

[0029] If a relation edge lacks supporting evidence, or if the terminating version of the relation edge is not empty and its starting version is not earlier than the terminating version, the version submission of the relation edge is rejected and the current version before submission is retained.

[0030] Optionally, S4 includes: the value range of the four types of conflict quantities is all from zero to one, wherein, if the effective shooting time intervals of two trajectory segments associated with the same identity overlap, and the spatial distance between their detection areas is greater than the motion reachable distance determined by the inter-frame displacement distribution quantile of the confirmed continuous trajectory within the camera, or if the two trajectory segments are from different cameras and there is no topological reachable path between them within the overlapping time interval, it is determined that the identity exclusivity condition is met, and the confidence level of the trajectory segment that meets the identity exclusivity condition is determined as the identity exclusivity conflict quantity;

[0031] The normalized value of the time difference when the shooting time interval between two trajectory segments is less than the shortest passage time corresponding to the camera topology is determined as the physical inaccessibility conflict quantity.

[0032] The relationship confidence level when the relationship between adjacent subjects and carried objects does not conform to the preset state transition table of acquisition, holding, putting down or transferring is determined as the discontinuity conflict quantity of the carrying state.

[0033] The normalized value of time overlap when the end time of the preceding event required by the event dependency is later than the start time of the subsequent event is determined as the event order conflict quantity;

[0034] For any evidence node or relation edge whose conflict amount is not less than the corresponding conflict triggering threshold, the conflict seed is established. The conflict triggering threshold is determined based on the quantile value of the corresponding conflict amount distribution in the confirmed associated samples.

[0035] Furthermore, the propagation of conflict quantities along candidate associations, carrying state transitions and event participation and sequential relationships includes: using the conflict quantity of the conflict seed as the initial conflict energy, propagating it to adjacent nodes according to the product of the relationship confidence of adjacent relationship edges and the time decay coefficient determined by the monotonically decreasing mapping of the shooting time interval, and selecting the last conflict energy after arranging multiple conflict energies that reach the same node in ascending order of value.

[0036] The nodes with conflict energy not less than the propagation termination threshold and their connected edges are included in the subgraph to be unlocked. The propagation termination threshold is determined based on the propagation energy distribution quantile of the nodes that have not formed the four types of conflicts in the confirmed associated samples.

[0037] For nodes whose conflict energy is less than the propagation termination threshold, whose candidate associated members remain unchanged before and after the addition of new evidence, and whose sequential relationship between adjacent events remains unchanged, the relationship edge between the node and the subgraph to be unlocked is determined as the locking boundary.

[0038] When the propagated subgraph to be unlocked contains mutually disconnected connected components, solve for each connected component separately and submit version differences separately.

[0039] Optionally, S5 includes: using the identity label of the trajectory segment, the subject-carrying object relationship state, and the event order as variables to be solved, establishing identity uniqueness constraints, camera topology reachability constraints, carrying state transition constraints, and event dependency order constraints;

[0040] The association cost is determined based on stable identity features, variable appearance features and their confidence levels; the state transition cost is determined based on the degree of conformity between the candidate state and the preset state transition table; the event sequence cost is determined based on the time overlap caused by the order of candidate events; and the old version relationship change cost is determined based on the number of cancellations and additions of the relationship to be solved relative to the current version relationship.

[0041] The costs are weighted and summed to find the minimum objective value that satisfies the constraints. The weight of the cost of changing the old version relationship is determined based on the confidence level of the new evidence. The confidence level threshold is determined based on the intersection of the confidence level distributions of the new evidence of the confirmed relationship samples that should be changed and the confirmed relationship samples that should not be changed. The weight of the cost of changing the old version relationship is increased when the confidence level of the new evidence is less than the confidence level threshold.

[0042] Furthermore, the version difference includes old identity identifier, new identity identifier, revoked relationship edge, newly added relationship edge, event sequence before change, event sequence after change, confidence level before change, confidence level after change, and a set of supporting evidence identifiers;

[0043] When submitting the version difference, write the current submission version as the termination version for the revoked relationship edge, and write the current submission version as the starting version for the newly added relationship edge. Update the event thread chain and downstream index according to the identity identifier, object identifier and event identifier involved in the relationship edge.

[0044] When any relation edge write or downstream index update fails, an exception rollback is performed to undo the current write commit and keep the version before the commit valid.

[0045] After submission, each newly added relationship can be traced back to an immutable evidence node by verifying the supporting evidence identifier set.

[0046] On the other hand, the present invention also provides a deep learning-based video event clue dynamic analysis system, comprising:

[0047] The evidence generation module is used to execute step S1; the feature association module is used to execute step S2; the evidence graph construction module is used to execute step S3; the influence domain determination module is used to execute step S4; and the version solving and submission module is used to execute step S5.

[0048] The beneficial effects of this invention are:

[0049] 1. By simultaneously recording the effective shooting time and the system access version time, and separating and saving immutable evidence nodes from revisable relationship edges with start and end versions, it is possible to query clues by event time and reconstruct historical judgments by access version, so that the corrected relationship and its supporting evidence remain traceable.

[0050] 2. By forming conflict seeds from four types of video semantic conflicts and propagating them along identity candidates, state dependencies, and event dependencies, the subgraph to be unlocked and the locking boundary are determined. This allows the scope of historical correction to be determined by technical conflicts caused by new evidence, avoiding repeated solutions to unaffected relationships.

[0051] 3. By jointly solving identity, subject-carry state and event sequence under the condition that the relationship outside the locked boundary remains unchanged, and by introducing the old version relationship change cost and atomic version differential commit, it is possible to suppress irrelevant flips caused by low-confidence new evidence and reduce the repeated updates of event clue chain and downstream index. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is a flowchart of the video event clue dynamic analysis method based on deep learning according to the present invention.

[0054] Figure 2 This is a flowchart of the conflict-driven influence domain determination method described in S4 of the present invention. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0056] refer to Figures 1-2 A deep learning-based method for dynamic analysis of video event clues, comprising:

[0057] S1. Receive multiple video clips that are asynchronously accessed in batches, record the effective shooting time and the system access version time, perform target detection, local tracking and event detection, and generate trajectory clips, event clips and raw observations with source identifiers and time intervals;

[0058] S2. Utilize a deep feature coding network to extract the stable identity, variable appearance, subject-carrying object relationship, action event state features, and confidence level of the trajectory segments and event segments, and generate a candidate association set by combining the camera topology and time reachability relationship;

[0059] S3. Construct immutable evidence nodes based on the original observations, and construct revisable relational edges with start and end versions based on the identity attribution, state migration, event participation, and event sequence determined by the candidate association set, to form a dual-temporal event clue evidence graph.

[0060] S4. After a new batch of accesses, based on the newly added evidence and the evidence graph, calculate the conflict amount of identity exclusivity, physical inaccessibility, discontinuous carrying state and event sequence to form a conflict seed. Propagate the conflict amount along candidate associations, carrying state migration and event participation and sequence relationship to determine the subgraph to be unlocked and the locking boundary.

[0061] S5. Keep the relationship edges outside the locked boundary unchanged, and jointly solve the identity partition, subject-carrying item state sequence and event sequence in the subgraph to be unlocked by using association, state transition, event sequence and old version relationship change costs. Compare the solution results with the current version relationship to generate version difference, and update the event clue chain accordingly.

[0062] In this specific embodiment, S1 includes:

[0063] In this specific embodiment, the video event clue dynamic analysis system is deployed on the park's video analysis server. The receiving entry point establishes an access record based on the source video segment identifier, camera identifier, file verification value, encoding format, frame rate, device acquisition start and end time, and reception completion time. Cameras one through twelve acquire video at a resolution of 1920×1080 at 25 frames per second. The video segment length is set to 60 seconds. Video segments from different cameras are allowed to enter the analysis queue in different batches due to network caching. Within the same transaction, the system assigns a monotonically increasing system access version number to the video segments that have completed verification and entered the analysis queue, and writes the version number, reception batch number, and queue offset into the access record.

[0064] For segments where the camera device has its time synchronized with a trusted network, the system directly uses the device time of the first and last frames as the valid shooting time interval. For segments that have undergone clock correction, the system follows... Conversion number Frame validity period Indicates the first The effective capture time of the frame after correction, in milliseconds. Indicates the first The frame is the device time written by the camera device, in milliseconds. Indicates camera The clock offset relative to the analysis server's time base, in milliseconds. Indicates the frame number within the segment. This indicates the index corresponding to the camera identifier; the clock offset is determined by the 200 most recent certified bidirectional network time synchronization records for the same camera. Each record stores four timestamps: device sending, server receiving, server sending, and device receiving. Round-trip delay and offset estimates are first calculated based on these four timestamps. Only the offset estimate within the lowest quartile of the round-trip delay is retained, and the median is taken as the offset. This eliminates asynchronous transmission and buffering latency. When there are fewer than 50 valid samples or the median absolute deviation of the retained samples exceeds 120 milliseconds, the previous valid deviation version is frozen and the current segment is marked as evidence to be corrected.

[0065] The evidence generation module performs YOLOX object detection on the decoded frames, with the input tensor set to 640×640×3. The output includes the categories of people and objects carried, detection bounding boxes, and detection confidence scores. Local tracking is performed within the same camera using Kalman motion prediction and appearance cosine similarity. Compared with existing trajectories Connection cost according to calculate, Represents the detection box With trajectory The dimensionless connection cost takes values ​​between zero and one. This represents the intersection-union ratio of the two predicted bounding boxes. It is a dimensionless quantity with a value range of zero to one. The cosine similarity of the 128-dimensional local appearance features of the two entities, truncated to zero or one, is a dimensionless quantity. Indicates the index of the detection box in the current frame. Indicates the activity trajectory index. This represents the dimensionless weight of the cost term in the intersection-union ratio. The dimensionless weights of the appearance cosine cost terms are determined by minimizing the erroneous connectivity rate on the same confirmed continuous trajectory verification set of the same camera.

[0066] The local tracking module uses the Hungarian algorithm to find the match with the minimum connection cost. The existing trajectory is only written when the connection cost is no greater than 0.42 and the distance of the motion prediction center does not exceed the 99th percentile of the inter-frame displacement distribution of the corresponding camera's confirmed continuous trajectory. If the costs are the same, the detection box with the higher detection confidence is selected first, and then the unique match is determined by ascending order of the horizontal coordinate of the upper left corner of the detection box. The active trajectory that does not match for 15 consecutive frames is closed into a trajectory segment. The trajectory segment records the trajectory identifier, start and end frames, camera identifier, effective shooting time interval, frame-by-frame detection box summary, average detection confidence, and source video segment identifier. The source video segment identifier and the start frame number form an idempotent key to prevent duplicate writing.

[0067] The event detection module feeds each trajectory segment into a lightweight temporal convolutional network in a 16-frame sliding window. The network outputs the probabilities of six action states: acquisition, hold, release, transfer, entry, and exit, as well as the start and end boundary probabilities of the window. The training annotations are the start and end frames of actions verified by two people in the historical videos of the park. During inference, the probability of the same state in adjacent windows is first weighted and averaged. Then, consecutive frames with continuous action categories and a boundary confidence of not less than 0.70 are merged into event segments. The boundary threshold of 0.70 is determined by the minimum confidence on the independent validation set that makes the boundary error not exceed 8 frames. The merged event segment records the event identifier, event type, participating target identifier, start and end frames, effective shooting time interval, boundary confidence, and source video segment identifier.

[0068] The system writes the unexplained observations corresponding to each detection box, trajectory segment, and event segment into the raw observation table. The raw observation fields include evidence identifier, observation type, source video segment identifier, camera identifier, effective shooting time interval, system access version time, detection or event result, confidence level, and data verification value. The evidence identifier is generated by SHA-256 digest from the source video segment identifier, observation type, and frame interval. The system access version time uses the incrementing version number obtained when entering the analysis queue instead of the subsequent processing completion time to ensure that the order of later segments on the version axis can be reconstructed.

[0069] When a segment lacks a shooting time, source video segment identifier, camera identifier, or verification value, the evidence status of the access record is set to pending correction. The relevant trajectory segments, event segments, and original observations are still saved according to the current batch, but the cross-camera candidate qualification field is set to prohibited. After the clock deviation estimation is completed or the source verification is passed, a correction record is generated and the prohibition is lifted in the new system access version. The data before correction does not participate in the cross-camera candidate association calculation of S2.

[0070] After the evidence generation module completes the current batch, it writes the batch version, number of records, and verification summary of the trajectory fragment table, event fragment table, and original observation table into the batch list. Only when all three types of objects are successfully written will the batch status be changed to analyzable. If any writing fails, the current batch object will be rolled back and the access records will be retained for re-execution. The trajectory fragments, event fragments, and original observations with the status of analyzable will be used as inputs for S2 and S3.

[0071] In this specific embodiment, S2 includes:

[0072] The feature association module reads trajectory segments and event segments in S1 that are analyzable and whose cross-camera candidate qualifications are not prohibited. It extracts 16 frames evenly over time from each trajectory segment, expands the person bounding box by 10%, and scales it to a tensor of 256×128×3. At the same time, it retains the carrying frame, the relative center coordinates, area ratio, and occlusion ratio of the subject and the carrying frame. The event segment is sampled into an action sequence tensor with a length of 32 frames. If the length is insufficient, the last frame is repeated and the filling position is marked with a mask field.

[0073] The deep feature coding network uses a shared ResNet-50 backbone and four branches. The identity coding branch transforms the 2048-dimensional features output from each frame into 512-dimensional frame-level identity features through a fully connected layer. Form a time aggregation vector, and then... According to Perform L2 normalization. This represents a 512-dimensional dimensionless temporary identity aggregation vector representing a trajectory segment. This represents a 512-dimensional dimensionless stable identity feature vector of a trajectory segment after L2 normalization. Indicates the first The dimensionless attention weights of each valid frame, with the sum of the weights of all valid frames being one. Indicates the first A normalized 512-dimensional dimensionless identity feature vector of each valid frame. Indicates the valid frame index, Indicates the number of valid frames. This indicates that the vector within the parentheses is Euclidean L2; attention weights are generated by two layers of perceptrons from the occlusion ratio, detection confidence, and frame-level features; mask frame weights are reset to zero; no attention weights are generated when the L2 of the temporary identity aggregation vector is equal to zero. The confidence level of the identity branch is set to zero and fusion is completed by the remaining valid branches according to the missing branch rule described in S2.

[0074] The appearance encoding branch concatenates the clothing color histogram, texture convolution features, and local occlusion mask to output a 256-dimensional dimensionless variable appearance feature vector. The relation encoding branch feeds the main body region features, the carried object region features, and the relative position and area ratio into the graph attention layer to output a 128-dimensional dimensionless main body-carried object relation feature vector. The temporal encoding branch uses three layers of one-dimensional temporal convolution and bidirectional gated recurrent units to process the action sequence tensor and outputs a 128-dimensional dimensionless action event state feature vector as well as the probabilities of acquiring, holding, putting down, and transitioning states. The outputs of the four branches are all written into the feature record, which uses the trajectory segment identifier or event segment identifier and the network version number as a joint key.

[0075] The raw scores of each branch are temperature-scaled using an independent calibration set. This calibration set uses confirmed cross-camera matching trajectory pairs as source samples, forming a calibration sample set independent of the training samples. For each source sample, logarithmic score vectors for four branches—stable identity, variable appearance, subject-object relationship, and action event state—are extracted. Non-corresponding ends of other confirmed cross-camera matching trajectory pairs from the same batch are used as comparison candidates, with the actual cross-camera corresponding ends as the target category. This ensures that the calibration values ​​for all four branches are determined by the corresponding feature distribution of confirmed cross-camera matching samples. The branch pairs with the first The calibration confidence levels for each category are as follows: calculate, Indicates the first The branch pairs with the first The dimensionless calibration confidence level for each target category or matching result, with a value range from zero to one. and They represent the first Each branch corresponds to the target category. and summation categories The output is a dimensionless real logarithmic fraction. Indicates the first A dimensionless temperature parameter with each branch strictly greater than zero. This represents the four types of feature branch indexes. Indicates the target category index. This indicates the category index that participates in the same branch normalization. Indicates the first The number of categories in each branch. Represents an exponential function with the natural constant as the base and the input as a dimensionless quantity; each branch minimizes the negative log-likelihood over at least 200 of the aforementioned source samples and their contrasting candidates to determine... When the number of samples is less than 200, the previous effective temperature parameter version of the branch is used. When there is no effective version, the calibration confidence of the branch is set to zero and written to the calibration queue. The network version record saves the temperature parameters, the confirmed matching source sample version across cameras, the number of samples, the batch of comparison candidate generation, and the effective system access version by branch.

[0076] Network training samples are grouped by subject identifier and shooting date to form training, validation, and calibration sets. Trajectory segments of the same subject across cameras constitute positive identity samples; trajectory segments with overlapping effective shooting times and not belonging to the same subject after manual verification constitute negative identity samples; sequences of subject acquisition, holding, placing, and transferring of objects constitute relational state samples; event segments labeled with event type and start / end times constitute action event samples; and joint loss is calculated according to... Update network parameters. Indicates the dimensionless joint training loss. This represents the dimensionless identity measurement loss. This represents the dimensionless appearance attribute cross-entropy loss. This indicates a dimensionless subject-carrying object relationship classification loss. This represents the dimensionless loss obtained by summing the action event temporal classification loss and boundary regression loss after normalizing them according to the mean of their respective training sets. , , and These represent the dimensionless weights of the four losses, which are all set to 1 in this specific implementation and are confirmed by the validation set to have no degradation in each branch. The training stops and the network version is frozen when the joint loss of the validation set does not decrease for five consecutive rounds.

[0077] The candidate generator first retrieves the reachable interval between the cameras containing the two trajectory segments according to the camera topology table. The topology table fields include the starting camera, the ending camera, the shortest travel time, the longest travel time, the travel direction, the number of samples, the version number, and the effective time. The travel time is in seconds and is determined by the difference between the departure and entry time of the confirmed cross-camera matching trajectory according to the 1st percentile and 99th percentile values. In this specific embodiment, a record from camera 3 to camera 7 has a shortest travel time of 18 seconds, a longest travel time of 95 seconds, one-way reachability, a number of samples of 1260, and a version number of TP-202608. Trajectory pairs that do not match the topology key or whose time difference falls outside the reachable interval are not included in the similarity calculation.

[0078] Similarity of identity branches according to calculate, Representing trajectory segments and trajectory fragments Dimensionless stable identity feature similarity and These represent the 512-dimensional dimensionless stable identity feature vectors of the two trajectory segments after L2 normalization. This represents the vector transpose operation. and Indicates the index of the candidate trajectory segment. This means that the values ​​in parentheses are truncated to zero or one. The greater the similarity of stable identity features, the stronger the evidence of identity consistency.

[0079] Similarity of appearance branches according to calculate, Representing trajectory segments and trajectory fragments Dimensionless variable appearance feature similarity and These represent the 256-dimensional dimensionless variable appearance feature vectors of the two trajectory segments after being normalized by the L2 norm. and Indicates the index of the candidate trajectory segment. This represents the vector transpose operation. This means that the values ​​in parentheses are truncated to zero or one. Appearance similarity is only used as positive evidence of similarity and does not directly exclude identity association based on appearance differences, in order to allow for changes in clothing and partial occlusion.

[0080] Similarity of relation branches according to calculate, Representing trajectory segments and trajectory fragments The similarity of the dimensionless subject-carrier relationship features. and These represent the 128-dimensional dimensionless subject-carrier relationship feature vectors of the two trajectory segments after being normalized to L2. and Indicates the index of the candidate trajectory segment. This represents the vector transpose operation. This means that the values ​​within the parentheses are truncated to zero or one, and the calibration confidence of the branch is set to zero when there is a missing carry region, instead of encoding the empty region as a valid relation feature;

[0081] The similarity of action branches is based on calculate, Representing trajectory segments and trajectory fragments Dimensionless similarity of action-event state features of related events and These represent the 128-dimensional dimensionless action event state feature vectors of the two event segments after being normalized to L2. and Indicates the index of the candidate trajectory segment. This represents the vector transpose operation. This means that the value within the parentheses is truncated to between zero and one. When the event fragment does not exist or the boundary confidence is less than 0.70, the calibration confidence of this branch is set to zero.

[0082] For trajectory pairs that pass the time-reachability test, the system calculates the cosine similarity of stable identity, variable appearance, subject-carry relationship, and action event state features, and then follows the... Obtain the fusion similarity. Representing trajectory segments With trajectory fragments The dimensionless fusion similarity is truncated to zero to one. Indicates the first of the two The cosine similarity of class features is truncated to a value between zero and one. and Representing trajectory segments respectively With trajectory fragments No. The calibration confidence levels for class features are all dimensionless quantities ranging from zero to one. and Indicates the index of the candidate trajectory segment. This represents the four types of feature branch indexes. This represents a dimensionless smoothing quantity used for zero-value protection. In this specific implementation, it is set to one part per million. When the input of a certain branch is missing, the corresponding calibration confidence is set to zero and the weights are automatically renormalized by the other branches.

[0083] The similarity threshold table uses camera pairs, day / night time periods, and network version numbers as lookup keys, storing fusion similarity thresholds, four-branch calibration versions, the number of continuous trajectory samples, the number of negative samples, and the update time. For the two endpoint cameras and corresponding day / night time periods in the keys, the system collects pairs of consecutive trajectory segments belonging to the same subject and within the same camera, verified by manual review. The fusion similarity of each segment pair is calculated using the current four-branch calibration version. All dimensionless fusion similarities are sorted in ascending order, and the sample value corresponding to the sorted position multiplied by 5% and rounded up is taken as the fusion similarity. When there are parallel values ​​at this position, the same sample value is taken; This represents the dimensionless continuity benchmark threshold determined by the confirmed continuous trajectory feature distribution from the same camera. Indicates that it is never less than The first 0.1% grid point is formed with a step size of 0.001 and additional points are included. A finite, dimensionless set of candidate thresholds; the final threshold is determined according to... Calibration, This represents the dimensionless fusion similarity threshold to be written into the similarity threshold table. This represents the number of confirmed mismatched trajectory pairs under the same scene key, and is a positive integer. This indicates the sample index of the confirmed mismatched trajectory pair. Indicates the first Dimensionless fusion similarity of confirmed mismatched trajectory pairs This is an indicator function that takes a value of 1 if the condition within the parentheses is true and zero otherwise. Indicates the candidate threshold. This indicates that the minimum value is selected from the candidate thresholds that meet the conditions. This represents the upper limit of the acceptance ratio of dimensionless negative samples at the tail; in this specific implementation, it is set to one percent. Because... The lower boundary is given by the continuous trajectory distribution of the same camera. Furthermore, the minimum value that satisfies the negative sample tail constraint is selected above this boundary. Therefore, the continuous trajectory distribution provides a baseline lower bound, and the negative sample distribution only provides a safety constraint above it. The two do not generate parallel thresholds. Parallel negative samples at the relational statistical threshold are used in the formula, which is consistent with the not less than boundary used for candidate writing. When the number of samples of either the confirmed continuous trajectory pair or the confirmed mismatch trajectory pair from the same camera is less than 200, the previous effective threshold version is used. If there is no previous effective version or no threshold that meets the conditions in the candidate set, cross-camera candidates for the scene key are not generated and written into the calibration queue. In this specific embodiment, the fusion similarity threshold of a record of the third to seventh camera during the daytime and with the network version NET-04 is 0.73. When the search key is not matched, the threshold of the same camera pair for the entire time period is used. If it is still not matched, cross-camera candidates are not generated and the trajectory pair is written into the calibration queue.

[0084] Candidate eligibility according to And only when and Sure, Representing trajectory segments With trajectory fragments The candidate qualification flag has a value of zero or one. This indicates the time interval between the capture of two trajectory segments, in seconds. and These represent cameras. To the camera The shortest and longest travel times for topologically reachable intervals, both in seconds. Indicates dimensionless fusion similarity. This represents the dimensionless similarity threshold for key hits in the same scene. If the condition is not met, the candidate qualification flag is set to zero and the candidate association set is not written.

[0085] The candidate generator writes a candidate association record only when the fusion similarity is not less than the hit similarity threshold and the shooting time interval is within the topology reachable interval. The record fields include candidate association identifier, two trajectory segment identifiers, camera pair, shooting time interval, four types of similarity, four types of calibration confidence, fusion similarity, topology version and network version. If the fusion similarity is the same, they are sorted in ascending order by the absolute value of the difference between the time interval and the shortest travel time, and then the candidate association identifier is retained in ascending order.

[0086] The action event training samples are explicitly constructed from event segments labeled with event type and start and end times. The fields read during training include event type, start frame, end frame, participating target, and label verification status. Only event segments with a verification status of "pass" are included in the action event training set. The boundary regression target uses the normalized offset between the labeled start and end frames and the sampling window boundary. Network versions with a validation set boundary error exceeding 8 frames are not included in the calibration stage.

[0087] When all four branch input features are missing, the system changes the evidence status of the corresponding original observation to pending correction and prohibits it from participating in candidate association. When only some branches are missing, zero-confidence records are retained for traceability. When the same trajectory segment generates more than 50 candidates from the same target camera, only the 50 candidates with the highest fusion similarity are retained. The threshold version and deletion reason are written into the deletion record. Finally, the candidate association set, together with the feature record, network version and confidence calibration version, are handed over to S3 to construct revisable relationship edges.

[0088] In this specific embodiment, S3 includes:

[0089] The evidence graph construction module takes the original observations of S1 and the candidate association set of S2 as inputs, and uses an attribute graph to store the dual-temporal event clue evidence graph. The graph node table and the relation edge table use the evidence identifier and the relation edge identifier as primary keys, respectively. All writes are bound to the same version transaction through the batch version number, and the original observations are not allowed to be overwritten by the new identification results.

[0090] For each original observation, the module generates an immutable evidence node and saves the evidence identifier, source video segment identifier, camera identifier, start and end values ​​of the effective shooting time, system access version time, observation type, feature summary, confidence level, data verification value, and original object storage address. After the evidence node is written, only the verification status or external reference is allowed to be appended, and the original fields are not allowed to be modified. If the data verification values ​​corresponding to the same evidence identifier are different, the submission of this batch will be rejected and the conflict record will be written to the access exception queue.

[0091] The module generates candidate identity edges based on candidate association records, generates carrying status migration edges based on the subject-carrying item relationship status in adjacent event fragments, generates event participation edges based on the participating target identifier, and generates event sequence edges according to event dependency rules and valid shooting time. The event dependency rules come from a versioned rule library reviewed and released by two business personnel. Each rule stores the preceding event type, the succeeding event type, the consistency predicate of the participating subject or carrying item, the allowed overlap time, the rule version, the effective system access version, and the review status. Only rules with a review status of "passed" and effective in the proposed submission version are used. If no match is found, no event sequence edges are generated, and the event type pair is written to the pending verification queue. Each revisable relationship edge stores the relationship type, source node identifier, target node identifier, starting version, ending version, set of supporting evidence identifiers, relationship confidence, the version of the generated rule that was hit, and the commit transaction identifier. The ending version of the currently valid relationship is stored as an empty value.

[0092] The relationship confidence of identity attribution edges is taken as the S2 fusion similarity; the relationship confidence of edges carrying state transitions is taken as the minimum value of the relationship encoding branch calibration confidence and the probability of adjacent states; the relationship confidence of event participation edges is taken as the product of the event detection boundary confidence and the subject detection confidence; the relationship confidence of event sequence edges is taken as the minimum value of the confidence of the two event boundaries; all relationship confidences are dimensionless quantities between zero and one and retain the corresponding supporting evidence identifiers; relationship edges lacking at least one immutable evidence node as support are not included in the submission set.

[0093] Version query tool according to Determine the relationship edge in version Effective below Representing relation edges In system access version The valid flag below takes a value of zero or one. This indicates the index corresponding to the relation edge identifier. This represents the system access version number to be rebuilt, and is an incrementing integer. Representing relation edges The starting version number, Representing relation edges The final version number, A flag indicating that the database termination version field is empty. , and These represent logical AND, logical OR, and logical equivalence, respectively. When querying a historical version, the edges with parentheses are first filtered according to the rules for identity attribution, state migration, event participation, and event sequence. Then, the queryer consumes the filtered relation edges to reconstruct the identity attribution, state migration, and event relations under this version.

[0094] When querying the order of events, the queryer constructs a directed graph using the valid event sequence edges under the target system access version and performs a deterministic topological sort. Each time, from events with an in-degree of zero, it first selects the one with the earlier valid start time. If the start times are the same, it selects the one with the earlier valid end time. If they are still the same, it selects the one with the earlier lexicographical order of the event identifier. The event sequence edges are written by S5 after the solution order is transformed, and they are accompanied by the start version, end version, supporting evidence identifier, and version difference identifier. Therefore, only one replayable event sequence is generated for the same version. For historical compatible versions that have not yet saved versioned event sequence edges, the queryer directly sorts them according to the same three-level key of start time, end time, and event identifier, and writes the sorting source, time base version, event identifier sequence, event sequence edge summary, and query system access version into the query snapshot.

[0095] During the graph construction phase, the physical detection quantity is not re-estimated. The detection boxes, event boundaries, and confidence scores are all directly taken from the original observations already saved in S1. The coordinate range of the detection boxes is constrained by the video resolution, and the time range of the event boundaries is constrained by the effective time interval of the video segment to which they belong. If any detection box coordinates are out of bounds, the event end time is earlier than the start time, or the confidence score is not in the range of zero to one, the module will refuse the corresponding node and relation edge from entering the current version transaction.

[0096] Each historical version query outputs a set of identity attribution records, a set of subject-carrying item status records, and a set of event sequence records. The three sets are respectively accompanied by the query version number, the effective shooting time range, the number of records, and a content summary. The event sequence records also save the version difference identifier of the preceding and following events and the deterministic topology sorting summary. The calling end uses the content summary to verify the reconstruction results. If the start time or end time of the query range is missing, the query is rejected and the system access time is not used to replace the effective shooting time.

[0097] Before writing a relational edge, a version integrity check is performed. The check is considered to have failed if the evidence identifier set is empty, any identifier in the set cannot be resolved to an immutable evidence node, the starting version is not the current committed version, the ending version is not empty and the starting version is not earlier than the ending version, or the source node or target node does not exist. The module rejects the entire version transaction and retains the current version before the commit, while recording the failed edge identifier, rule version and check reason.

[0098] When a relationship is revoked by a subsequent version, the old edge is not deleted. Instead, in the same version transaction, the current committed version is written into the termination version of the old edge, and a new record with the new relationship edge identifier is added for the replacement relationship. The starting version of the new edge is the current committed version and the termination version is empty. The old edge and the new edge jointly store the version difference identifier. The same start and end version rules are used for the edges of the order of events, so that the order of events at the same effective shooting time can be replayed according to the system access version.

[0099] The module establishes indexes for identity identifiers, carried item identifiers, event types, valid shooting time intervals, and system access versions. Index entries only reference immutable evidence nodes or relation edges that satisfy version validity rules. The current batch nodes, candidate relation edges, and index entries are first written to a batch temporary storage area that is not visible to the query service and pass verification. The version status of the proposed graph is set to pending conflict evaluation, forming a read-only proposed dual-temporal evidence graph snapshot, proposed version number, and candidate relation edge set, which are then processed by S4 and S5. After S4 completes the connected component partitioning, the temporary storage manager uses immutable evidence nodes as the batch base record, splits the candidate relation edges and index entries involved in each connected component into independent submission units according to the component identifier, and assigns non-conflicting relations that do not belong to the component to be unlocked to the append submission unit. Only the corresponding unit becomes visible to the outside world after S5 completes version differential or append differential verification and atomically submits. Failed units only revoke their own temporary storage records and do not revoke other units that have already been submitted. Unsubmitted records do not enter the version validity query.

[0100] In this specific embodiment, S4 includes:

[0101] The influence domain determination module reads the read-only dual-temporal evidence graph snapshot of the proposed submission in the S3 state, which is pending conflict assessment. It adds the newly added evidence nodes in the current batch, the candidate related edges connected to them, the edges carrying state transitions, the event participation edges, and the event sequence edges to the conflict scanning queue. It also reads the four types of trigger thresholds corresponding to camera pairs, event types, and rule versions from the conflict threshold table. This snapshot has not yet been published to the query service. The S4 output is sent to S5 along with the proposed submission version number to avoid treating the same batch as both submitted and pre-submission states at the same time. The threshold table is established by the conflict distribution of confirmed related samples according to the 95th percentile value. The fields include conflict type, scene key, percentile value, number of samples, threshold, version number, and update time. All thresholds are dimensionless quantities between zero and one.

[0102] Conflict trigger threshold according to Calibration, Indicates the first The dimensionless trigger threshold for class conflicts ranges from zero to one. This represents the 95th percentile value. This indicates that the first confirmed related sample is... A dimensionless sample set of conflict quantities. The index represents four conflict types: identity exclusivity, physical unreachability, discontinuous carrying state, and event sequence. The corresponding threshold is updated only when the number of samples is not less than 200. When the number of samples is less than 200, the previous valid threshold version is used and the current scenario key is written to the pending calibration queue. When there is no previous valid threshold version, a versioned conservative default value of 0.20 is used for all four types of conflicts. The source of the default value, the scenario key, and the pending calibration status are recorded. The pending calibration status is removed after the scenario has accumulated not less than 200 confirmed samples and has been recalibrated.

[0103] Same camera The distance that can be reached by movement is according to Sure, Indicates camera The achievable distance of motion within a single frame interval, in pixels. This represents the 99th percentile value. This indicates the set of center displacement samples of adjacent frame detection boxes for a confirmed continuous trajectory by the camera, in pixels. This represents the camera index, with samples updated on a rolling basis over the last 30 days and abnormal displacements removed using a three-times median absolute deviation rule; only when... Update when containing no less than 200 valid samples If the sample size is insufficient or empty, the previous valid version will be used. If there is no previous valid version, the versioned conservative default value of 40 pixels will be used, and the camera will be... Write to the calibration queue;

[0104] Exclusivity of identity according to Sure, Representing trajectory segments With trajectory fragments The exclusive identifier, with a value of either zero or one. This indicates an overlap in the effective time intervals of the two images, and its value is either zero or one. and Representing trajectory segments respectively and Camera index, This represents the minimum spatial distance between the centers of the overlapping frame detection boxes, in pixels. This refers to the camera distance obtained according to the aforementioned formula for achievable distance and backtracking rules in S4. The distance that can be reached by movement, in pixels. This flag indicates the existence of a topologically reachable path across cameras during the overlapping time period; its value is either zero or one. , and These represent logical AND, logical OR, and logical equivalence, respectively. and Indicates the index of the trajectory segment;

[0105] For trajectory segments associated with the same identity and When the effective time intervals of two images overlap and they originate from the same camera, the module calculates the minimum spatial distance between the centers of the overlapping frame detection boxes and compares it with the reachable distance of the camera. If the spatial distance is greater, the identity exclusivity condition is met. When the two images originate from different cameras, the module queries whether a topologically reachable path exists within the overlapping time interval. If no path exists, the identity exclusivity condition is also met. The identity exclusivity conflict quantity is calculated according to... Obtain, Representing trajectory segments With trajectory fragments The dimensionless identity exclusive conflict quantity, and Let represent the mean detection confidence scores of the two trajectory segments, both being dimensionless quantities ranging from zero to one. This indicates that the minimum value is taken from the candidate values ​​within the parentheses. and This represents the index of a trajectory segment; the conflict count is set to zero when the identity exclusivity condition is not met.

[0106] Physically unreachable conflict quantities only apply to And in the topology table Effective cross-camera recording, according to calculate, Representing trajectory segments With trajectory fragments Dimensionless physical inaccessible conflict quantity Indicates from the camera to camera The shortest passage time, strictly positive, in seconds, derived from the S2 camera topology table. This represents the non-negative time interval from the end of the previous trajectory to the start of the next trajectory, in seconds. and These represent the camera indices for the two trajectory segments, respectively. This means truncating the calculation result to zero or one; the same camera trajectory does not use this formula, and when the shortest passage time is missing, zero, or negative, no physical unreachable conflict quantity is generated and the topology key is written into the queue to be corrected. When the time interval is not less than the shortest passage time, the conflict quantity is zero.

[0107] For subject-carrying object relationship states that are adjacent according to the effective shooting time, the module uses the subject identifier, carrying object identifier and the previous state as the lookup key for the state transition table. The state transition table stores the allowed subsequent states, minimum duration, maximum allowed interval, relationship type and version number between acquisition, holding, putting down and transfer. The duration and interval are in seconds. The records are statistically analyzed from the historical event sequence with state labels and manually verified. In this specific implementation, if the previous state is putting down and the next state is holding without any acquisition event in between, it is judged as not conforming to the state transition table. The carry state discontinuous conflict measure is the minimum value of the confidence of the two relationship edges before and after. If the lookup key is not matched, the state pair is written to the queue to be verified and the conflict seed is not established for the time being.

[0108] Carrying state discontinuity conflict quantity according to calculate, Indicates the relationship between adjacent subjects and their carriers. and Dimensionless carrying state discontinuity conflict quantity and Let represent the dimensionless confidence scores of the two relation edges in S3, respectively, with values ​​ranging from zero to one. This indicates that the minimum value is taken from the candidate values ​​within the parentheses. and This indicates the state index adjacent to the valid shooting time. This value is only used when the state transition table determines that the transition is not allowed. When the transition is allowed, the conflict amount is set to zero.

[0109] For event dependency rules, events are required. Prior to the event The relationship, the order of events, and the number of conflicts are determined by... calculate, Indicates an event With the event Dimensionless event order conflict quantity Indicates preceding events The effective end time of the shooting, in seconds. Indicates subsequent events The effective start time of the shooting, in seconds. This indicates the normalized baseline for allowed overlap time of the corresponding event type with respect to strictly greater than zero, in seconds. and Indicates the event node index, This means truncating the value within the parentheses to zero or one; The 99th percentile value is taken only when there are at least 200 positive overlap samples of confirmed correct event sequences. If the samples are empty, insufficient, or the percentile value is zero, the previous valid rule version is used. If there is no valid version, a conservative default value of 1 second is used and the event type pair is written into the calibration queue. The number of conflicts is zero when the numerator is not greater than zero.

[0110] All four types of conflict quantities are truncated to zero or one after calculation. The value range of the identity exclusive conflict quantity, the physical unreachable conflict quantity, the state discontinuity conflict quantity, and the event sequence conflict quantity is all between zero and one. When any of the conflict quantities is not less than the corresponding type trigger threshold, the module establishes a conflict seed for the evidence node or relation edge carrying the conflict. The conflict seed records the seed identifier, conflict type, initial conflict energy, supporting evidence identifier, trigger threshold version, and current system access version. If the same object triggers multiple conflicts, the details of each type are retained and the largest conflict quantity is used as the initial conflict energy for propagation.

[0111] Conflict propagation occurs along candidate association edges, edges carrying state transitions, event participation edges, and event sequence edges, with nodes... From adjacent nodes The received propagation energy according to calculate, Indicates from node propagation to nodes dimensionless conflict energy, Represents a node The dimensionless conflict energy currently stored, with values ​​ranging from zero to one. This represents the dimensionless confidence level of a relation derived from the corresponding edge in S3, with values ​​ranging from zero to one. This represents the absolute interval between the effective shooting times of two nodes, in seconds. This represents a time decay scale that is strictly greater than zero, measured in seconds, and is calibrated by the median of time intervals with conflict-free and confirmed relationships. In this specific implementation, it is set to 120 seconds. and Indicates the index of adjacent nodes. Indicates a value with the natural constant as the base and input It is a dimensionless, non-positive exponential function, and the exponential mapping decreases monotonically as the time interval increases;

[0112] Reaching the same node Multiple propagation energies according to polymerization, Represents a node Preserved dimensionless conflict energy Represents nodes The set of node indices that are adjacent and generate propagation values ​​in this round. Indicates adjacent nodes To the node Dimensionless propagation energy, This means that the propagation energy is sorted in ascending order and the last value is selected. The propagation queue is only updated when the new value is higher than the energy already stored in the node.

[0113] Propagation termination threshold according to Calibration, This represents the dimensionless propagation termination threshold, with a value ranging from zero to one. This represents the 99th percentile value. This represents the set of propagation energies of nodes in the confirmed associated samples that have not formed the four types of conflicts. All energies within this set are dimensionless quantities ranging from zero to one; only when... The threshold is updated when there are no less than 200 valid samples. When the samples are empty or insufficient, the previous valid threshold version is used. When there is no valid version, the versioned conservative default value of 0.18 is used and written to the calibration queue. When the node conflict energy is not less than the threshold, the node and its connected edges are included in the subgraph to be unlocked. When it is less than the threshold, energy propagation from the node is stopped.

[0114] The flag to be unlocked is included according to Sure, Represents a node The flag indicating whether the edges connected to it are included in the subgraph to be unlocked takes a value of zero or one. Represents a node dimensionless conflict energy, This represents the dimensionless propagation termination threshold. and These represent the nodes before and after the addition of new evidence, respectively. The set of candidate association identifiers, and These represent the nodes before and after the addition of new evidence, respectively. The set of state transition relationship identifiers carried. and These represent the nodes before and after the addition of new evidence, respectively. A set of event participation relationship identifiers and These represent the nodes before and after the addition of new evidence, respectively. A set of identifiers indicating the sequence of adjacent events. and They represent logical OR and logical equivalence, respectively. Indicates the node index; when the energy is below the threshold but any one of the four relation sets changes, it is directly included in the subgraph to be unlocked but energy continues to propagate from that node; when the energy is below the threshold and none of the four relation sets have changed, the flag is set to zero.

[0115] For nodes whose conflict energy is below the propagation termination threshold, the module compares the candidate association identifier set, the state transition relationship identifier set, the event participation relationship identifier set, and the adjacent event sequence relationship identifier set before and after adding new evidence. Only when all four sets are completely identical is the relationship edge between the node and the subgraph to be unlocked marked as a locked boundary. The locked boundary records the edge identifier, the node identifier outside the boundary, the current relationship type, the new and old summaries of the four sets, the current version, and the reason for locking. If any set changes, the node continues to be treated as a node to be unlocked without establishing a lock.

[0116] The boundary determination result is based on And only when , , , and Record, Indicates the node to be unlocked With neighboring nodes The locking flag for the edges between relationships can be either zero or one. Represents a node dimensionless conflict energy, and , and , and as well as and These represent the nodes before and after the addition of new evidence, respectively. A set of candidate associations, carrying state transitions, event participation, and adjacent event sequence identifiers. and This represents the index of adjacent nodes. If any set equality is false, the locking flag is set to zero and the node is locked. Included in the scope to be unlocked;

[0117] The module performs an undirected connected component search on the subgraph to be unlocked, generates a subgraph identifier for each connected component, and saves the set of nodes, the set of internal relation edges, the set of locked boundary edges, the set of conflict seeds, and the threshold version. Connected components that are disconnected from each other enter the solution queue of S5 and submit version differences respectively. When the subgraph to be unlocked is empty, the module generates an append commit task with batch version, empty subgraph identifier, zero-reversal relation set, newly added immutable evidence set, verified non-conflict newly added relation set, and index update set and sends it to S5. The branch is not recorded but discarded.

[0118] In this specific embodiment, S5 includes:

[0119] The version solving and submission module reads the connected components of the subgraph to be unlocked formed by S4 one by one, sets the identity label of each trajectory segment, the acquisition, maintenance, placement or transfer status of each subject-carrying object relationship, and the order position of each event node as variables to be solved, locks the boundary relationship edges and their identity, status and order as constants and writes them into the component constraint file, and does not allow modification of the boundary objects during the solving process;

[0120] Treat each trajectory segment within the unlock submap Establish a set of candidate identity tags This set consists of existing identity labels reachable from S2 candidate associated edges, identity labels allowed by the locked boundary, and a versioned new identity label used only for this trajectory, and includes... , Representing trajectory segments Whether to be assigned a candidate identity label binary variables, Indicates the index of the trajectory segment to be unlocked. Indicates the candidate identity label index. Representing trajectory segments A finite set of candidate identity labels, the equation ensuring that each trajectory segment acquires its existing identity or its uniquely created identity once and only once; an exclusive flag for each identity. trajectory fragments and and each common candidate identity tag join in Therefore, the same identity cannot be assigned to a pair of mutually exclusive trajectories at the same time;

[0121] For each candidate identity association edge generated by S2 Set a binary choice variable For each common candidate identity label Set a binary auxiliary variable and join , , as well as ; Indicate candidate edges Whether it was selected in the identity-related results Indicates whether both ends obtain identity tags simultaneously. , and Indicates the index of the candidate trajectory segment. Indicates the index of common candidate identity labels. The intersection of sets is represented by summation, and summation represents the accumulation of all common candidate identity labels; this is known from the one-and-only assignment constraint of each trajectory. The value is either zero or one, and is one if and only if both ends obtain the same candidate identity label. The trajectory of zero candidate associated edges still forms a definite singleton identity partition through its exclusive newly created identity label. The boundary identity constant is locked by... Delete the label field within the label constraint boundary that conflicts with this constant;

[0122] Treating the unlocking of connected components Set an integer order variable for each event node. And for any different events and join in , and Representing events respectively and events The order of positions This represents the total number of event nodes within a component. and Indicates the event node index, Indicates from one to The module establishes a time sorting key for each event, consisting of the valid start time of shooting, the valid end time of shooting, and the event identifier. (This refers to any event.) The time sort key is lexicographically earlier than the event. Join at time Hitting the S3 versioned event dependency rule and requiring Prior to The same strict inequality is also added; the position of events outside the boundary is locked as a constant in the current version, and the existing dependencies and time ordering relationships between events inside the boundary and this constant are converted into strict inequalities in the same direction. If the dependency rule is opposite to the deterministic time ordering key, the component is judged as a constraint infeasible and transferred to manual review. The optimization result is not used to cover the effective shooting time order.

[0123] The average cost of candidate identity association is based on calculate, Let $\mathbf{ ... This represents the set of candidate identity-related edges generated by S2 and that have passed the input validity check. This indicates the number of edges in the set. Indicate candidate edges Binary choice variables, This represents the dimensionless fusion similarity obtained from S2, with values ​​ranging from zero to one. and Indicates the index of the candidate trajectory segment. This represents the summation of all candidate edges in the set; the cost of selecting an edge varies. The cost of rejecting an edge increases and decreases monotonically. Increases and monotonically increases, when Directly stipulated Instead of performing division, a unique singleton identity partition is output by a newly created identity label for each trajectory;

[0124] The state transition cost is set to zero for candidate states that are allowed in the state transition table and satisfy the duration range. For candidate states that are allowed but whose duration exceeds the calibrated range, the cost is the ratio of the excess to the maximum allowed interval that is strictly greater than zero, and then truncated to zero or one. State transitions that are not allowed are set as infeasible. Each subject-carrier state transition in the candidate solution where the preceding and following states, durations, and the state transition table keys used are all valid is considered as a valid state cost term. These terms form a set. and in accordance with Calculate the mean state transition cost. This represents the mean of the dimensionless state transition cost. Denotes the set of effective state cost items. This indicates the number of items in the set. Indicates the index of the effective state cost item. This represents the state transition rule obtained according to the aforementioned rules. Dimensionless cost This indicates the summation of all items within the specified set; when When it is an empty set, it is directly defined The value is zero and no division is performed; the event ordering cost does not reuse the conflict amount of S4 for the current relation, but rather for the events in each candidate solution. and Recalculate according to candidate order , Indicates events in candidate solutions Prior to the event The dimensionless order cost of time, and These represent the end time of the candidate preceding event and the start time of the candidate subsequent event, respectively, in seconds. This represents the normalized baseline obtained by S4 for the same event type and rule version, which is strictly greater than zero, in seconds. and Indicates the event node index, This means truncating the values ​​within the parentheses to zero or one; after arranging the candidate solutions in order, each pair of adjacent events with complete effective start and end times is considered as a valid event sequence cost term, and these ordered event pairs form a set. and in accordance with Calculate the mean cost of event ordering. This represents the mean of the dimensionless event order costs. Represents a set of validly ordered event pairs. This represents the number of ordered event pairs in the set, as shown in the formula. and Using the aforementioned event node index, Using the dimensionless order cost already defined in the same paragraph above, The meaning of accumulating a specified set is retained; when When it is an empty set, it is directly defined The cost of the relationship change in the old version is obtained by dividing the sum of the number of edges canceled and the number of edges added relative to the current version of the candidate solution by the number of currently valid edges inside the subgraph to be unlocked. When the number of currently valid edges inside the subgraph to be unlocked is zero, one is used as the dimensionless counting benchmark. All four types of costs are normalized to zero to one.

[0125] The confidence level of newly added evidence is taken as the maximum confidence level of the new batch of evidence nodes in the current component, and the weight of the cost of relationship change in the old version is based on... Sure, This represents the dimensionless cost weight for changing relationships in the old version. This represents the dimensionless basic weight, which is set to 0.25 in this specific implementation. This indicates the magnitude of the dimensionless low confidence level adjustment, set to 0.75. This represents the dimensionless change confidence threshold that is strictly greater than zero and not greater than one. This represents the dimensionless maximum confidence level of newly added evidence within a component, with a value range from zero to one. This indicates that the maximum value is taken from the candidate values ​​in parentheses. The threshold is determined by the intersection of the confidence distributions of newly added evidence for confirmed relationship changes and confirmed relationship changes that should not be changed. When there are multiple intersections, the intersection that minimizes the weighted classification error rate of the independent validation set is selected. When there are ties, the one closest to 0.5 is selected. When there are no intersections or the candidate intersection is zero, the previous effective threshold version is used. When there is no effective version, 0.5 is used and written into the calibration queue. The final threshold is limited to 0.05 to 0.95 and recorded in the solution parameter version. Therefore, the divisor in the formula is always positive and the weight is adjusted upward as the confidence of newly added evidence decreases when it is lower than the threshold.

[0126] The solver is Find the minimum objective value that satisfies all constraints for the objective function. This represents the dimensionless joint cost of candidate solutions to the subgraph to be unlocked. This represents the dimensionless mean of the association cost among the candidate solutions. The dimensionless mean of the state transition cost. The dimensionless mean of the cost of event sequence. This represents the cost of changing relationships in the dimensionless old version. , and The dimensionless weights representing the costs of association, state transition, and event sequence are set to 0.40, 0.25, and 0.35 respectively in this specific implementation and are selected based on the accuracy of relationships on historically confirmed revised samples. Using the aforementioned weights that change with the confidence level of new evidence, the solver employs 0-1 integer programming and limits the solution time for a single connected component to 30 seconds;

[0127] When multiple minimum solutions with the same objective value exist, the module sequentially selects the solution with fewer relationship changes in the old version, higher sum of supporting evidence confidence, lexicographically first identity tag sequence obtained by ascending connection of trajectory identifiers, lexicographically first subject-carrying object state sequence, and lexicographically first event identifier sequence obtained by connecting in order of position as the sole output; the solver records five termination states: proven optimal, feasible solution within the time limit but not proven optimal, infeasible, no feasible solution obtained, and solution error. Only solutions proven optimal within 30 seconds enter version difference; solutions with feasible solutions within the time limit... If a solution is proposed but not proven to be optimal, the candidate solution is not submitted. The current feasible objective value, global bound, optimality gap, solver status, and parameter version are recorded, and the component is set to await manual review. If the solution is infeasible, no feasible solution is obtained, or the solution is incorrect, it is also not submitted, and the corresponding diagnostic information is recorded. All branches that are not submitted continue to use all valid relation edges, event thread chains, and downstream indexes of the current version before submission. The solution temporary results are not exposed to the query service. Only when the optimal solution is proven is the identity partition, subject-carrying item state sequence, event order, objective value, and solution parameter version written into the solution result record.

[0128] The module compares the proven optimal solution with the current version's valid relationships item by item, generating version difference records. Fields include old identity identifier, new identity identifier, set of revoked relationship edge identifiers, source and target nodes and relationship types of newly added relationship edges, event order before change, event order after change, confidence level before change, confidence level after change, set of supporting evidence identifiers, identifier of subgraph to be unlocked, and current submission version number. The module generates versioned event precedence edges for adjacent events according to the changed order, and writes the old event precedence edges that need to be terminated and the event precedence edges that need to be added into the version difference. Relationships that have not changed are not included in the version difference, so that the downstream processing scope is consistent with the actual modified objects.

[0129] Version commits use a single database transaction. First, the current commit version is written as the termination version for each revoked identity-attribution edge, state-transition edge, event-participation edge, and event sequence edge. Then, the current commit version is written as the starting version for each newly added relation edge, while keeping the termination version empty. Subsequently, the event thread chain, identity inverted index, carried-item state index, and event time index are updated according to the identity identifier, object identifier, and event identifier involved in the difference. The index update task carries the version difference identifier to ensure that repeated execution will not produce duplicate entries. The event time index also saves the summary of the event sequence edge after the change, so that the queryer can read and consume the order relationship of the same transaction commit according to the target version.

[0130] Before a transaction is committed, the module parses immutable evidence nodes one by one along the set of supporting evidence identifiers in the version difference, and verifies that there is corresponding evidence for the nodes at both ends of the newly added relation edge, the relation type, and the relation confidence. If any relation edge writing fails, index update fails, or evidence tracing fails, the current transaction is rolled back, the newly added edge of the current transaction is deleted, and the writing of the terminated version of the old edge is undone, so that the version before the commit remains valid. At the same time, the failure stage, edge identifier, and difference identifier are written to the exception queue.

[0131] S5 does not retrain the deep feature encoding network. Instead, it reads the fusion similarity of the trained, independently calibrated, and frozen version of S2. The sampling objects used by S5 are the trajectory segments and event segments formed by S1. The measurement input includes the effective time interval of the shooting, detection confidence, boundary confidence, and S2 fusion similarity. The time is uniformly converted to seconds. The confidence and similarity are uniformly limited to zero to one. Any candidate relationship with a negative time interval, out-of-bounds confidence, or missing similarity is marked as invalid and removed from the feasible solution set before solving. The removal results, the original measurement fields, the S2 network version, the confidence calibration version, and the similarity threshold version are all saved in the solution input snapshot.

[0132] The triggering conditions for abnormal rollback are any failure to write a relation edge, failure to update a downstream index, failure to trace evidence, or inconsistency in the digest verification within 20 seconds after the transaction ends. The rollback parameter record includes the version number before the commit, the snapshot address of the relation edge, the snapshot address of the index, and the recovery deadline. After the rollback is triggered, the transaction manager immediately restores the version pointer before the commit and replays the snapshot within a 60-second recovery period. The recovery threshold is set to ensure that the number of relation edges, the event thread chain digest, and the downstream index digest are all consistent with the snapshot at the start of the transaction.

[0133] When all three checks reach the recovery threshold, the security status is recorded as recovered and the read-only restriction of the query service is lifted. If any one of them is still inconsistent at the end of the 60-second recovery period, the query service remains read-only, the current version before the commit is used as the security parameter set, and the snapshot recovery task before the commit is written to the high-priority exception queue. During the recovery period, it is prohibited to start new solution transactions for the same unopened subgraph.

[0134] After all relation edges and downstream indexes are successfully updated, the module sets the version difference status to committed and atomically updates the current version pointer to the committed version. The event clue query request reads the new identity attribution, subject-carrying item status sequence, and the valid event sequence edges of the target version, and obtains the unique event order according to the deterministic topology sorting rules of S3. At the same time, the old clues can still be reconstructed according to the historical system access version. If S4 contains multiple disconnected connected components, each component is committed sequentially using independent commit units, independent database transactions, and independent version difference identifiers. If a single component fails, only its own temporary relations and indexes are revoked, without changing the trace records of other successful components. The query service only reads the committed units.

[0135] When S5 receives the empty subgraph append commit task from S4, it does not start integer programming. Instead, it generates an append-type version difference with zero undo. The module first verifies the data check value of the newly added immutable evidence, the evidence nodes at both ends of each non-conflicting newly added relationship, the relationship type, the relationship confidence, and the supporting evidence set. If there is no committable relationship, the newly added immutable evidence and the batch completion status are still retained. Then, in an append commit transaction, the new relationship that has passed the verification, the event sequence edges, and the corresponding indexes are written and the current version pointer is atomically updated. If any evidence verification, relationship writing, or index update fails, the entire append commit transaction is rolled back. The newly added relationship and index entries of the transaction are deleted, the version pointer before the commit is kept valid, and the failed objects are written to the exception queue, so that even non-conflicting batches have a definite commit and rollback path.

[0136] The version solver establishes a set of topologically inaccessible pairs of trajectories to be unlocked based on the same topology version used when S4 calculates physical inaccessibility conflicts, the valid capture time, and the direction of travel. For any and any common candidate identity tags All of them reuse the common label mutual exclusion constraint of the aforementioned identity exclusive trajectory pairs. This represents the set of trajectory segment pairs where there is no valid directed topological path and the time interval between shots is less than the versioned shortest travel time or greater than the versioned longest travel time. This represents a pair of trajectory segment indices. This indicates that both are candidate identity labels. and These represent the acquisition of identity labels from two trajectory segments. binary variables, and Let each represent a finite set of candidate identity labels. Represents the intersection of sets; when the trajectory When the trajectory corresponding to the identity constant outside the locked boundary is unreachable under the same topological snapshot, from Delete the constant label. If the topological key or the effective shooting time is missing, mark the corresponding trajectory as pending correction and prohibit merging with any existing identity, but retain its exclusive new identity label to maintain feasibility.

[0137] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0138] This invention uses stable identity, variable appearance, subject-carrying object relationship and action event state features generated by deep feature encoding, together with camera topology and time reachability, for candidate association, and uses four types of conflict quantities to determine the influence domain of new evidence on the evidence map, so that identity, state and event order in cross-camera, occlusion and carrying state change scenarios can be coordinated and corrected in the same constraint model.

[0139] This invention improves the traditional one-way addition structure of clues by versioning and local solving: immutable evidence nodes preserve the source of observation, start and end version relationship edges preserve historical judgments, locking boundaries limit the solution range, and version difference drives downstream directional updates, thereby achieving local backtracking correction of historical clues by later evidence while preserving the traceability of old versions.

Claims

1. A method for dynamic analysis of video event clues based on deep learning, characterized in that, include: S1. Receive multiple video clips that are asynchronously accessed in batches, record the effective shooting time and the system access version time, perform target detection, local tracking and event detection, and generate trajectory clips, event clips and raw observations with source identifiers and time intervals; S2. Utilize deep feature coding networks to extract stable identities, variable appearances, subject-carrying object relationships, action event state features, and confidence levels from trajectory segments and event segments. Combine these features with camera topology and time reachability to generate a candidate association set. S3. Construct immutable evidence nodes based on original observations, and construct revisable relational edges with start and end versions based on the identity attribution, state migration, event participation, and event sequence determined by the candidate association set, forming a dual-temporal event clue evidence graph. S4. After a new batch of accesses, based on the newly added evidence and evidence graph, calculate the conflict quantities of identity exclusivity, physical inaccessibility, discontinuous carrying state, and event sequence to form conflict seeds. Propagate the conflict quantities along candidate associations, carrying state migrations, and event participation and sequence relationships to determine the subgraph to be unlocked and the locking boundary. S5. Keep the relationship edges outside the locked boundary unchanged, and jointly solve the identity partition, subject-carrying item state sequence and event sequence in the subgraph to be unlocked by association, state transition, event sequence and old version relationship change cost. Compare the solution results with the current version relationship to generate version difference, and update the event clue chain accordingly.

2. The method for dynamic analysis of video event clues based on deep learning according to claim 1, characterized in that, S1 includes: The acquisition time interval obtained by writing to or clock-correcting the video clip is determined as the effective shooting time. The incrementing version number when the video clip is received and enters the analysis queue is determined as the system access version time. Detection boxes within the same camera are connected into trajectory segments based on inter-frame motion consistency and appearance feature similarity, and the start and end frames, camera identifier, effective shooting time interval, and detection confidence of the trajectory segments are recorded. Continuous frames with continuous motion states and boundary confidence not less than the event detection threshold are determined as event segments, and the event type, participating target, effective shooting time interval, and source video clip identifier are recorded. Video segments lacking shooting time or source identifiers are marked as evidence to be corrected, and are prohibited from participating in cross-camera candidate association until clock deviation estimation or source verification is completed.

3. The method for dynamic analysis of video event clues based on deep learning according to claim 1, characterized in that, S2 includes: using the identity encoding branch of the deep feature encoding network to perform temporal aggregation on multiple frame-level features of the trajectory segment, and outputting the stable identity feature; using the appearance encoding branch to encode the frame-level features corresponding to clothing color, texture, and partial occlusion, and outputting the variable appearance feature; using the relationship encoding branch to jointly encode the subject region, object region, and their relative positions, and outputting the subject-carrying object relationship feature; using the temporal encoding branch to encode the action changes of the event segment, and outputting the action event state feature; multiplying the similarity of the stable identity feature, the similarity of the variable appearance feature, the similarity of the subject-carrying object relationship feature, and the similarity of the action event state feature by the confidence level of the corresponding feature, and dividing the sum of the resulting products by the sum of the confidence levels involved in the calculation, to obtain the fusion similarity; determining the similarity threshold and the confidence level calibration value of each feature according to the feature distribution of the continuous trajectory of the same camera and the feature distribution of the confirmed matching samples across cameras, and only writing two trajectory segments whose time interval is located in the reachable interval of the corresponding camera topology and whose fusion similarity is not less than the similarity threshold into the candidate association set.

4. The method for dynamic analysis of video event clues based on deep learning according to claim 1, characterized in that, S3 includes: saving the evidence identifier, source video segment identifier, camera identifier, effective shooting time interval, system access version time, feature summary, and confidence level for each immutable evidence node; saving the relationship type, starting version, ending version, supporting evidence identifier set, and relationship confidence level for each revisable relationship edge, wherein the ending version of the currently valid relationship edge is empty, and when the relationship is revoked, the version number of the revocation operation is written into the ending version; querying the occurrence order of events according to the effective shooting time, and reconstructing the identity attribution, carrying status, and event relationship under any submitted version according to the system access version time; when a relationship edge lacks supporting evidence, or the ending version of the relationship edge is not empty and its starting version is not earlier than the ending version, rejecting the version submission of the relationship edge and retaining the current version before submission.

5. The method for dynamic analysis of video event clues based on deep learning according to claim 1, characterized in that, In step S4, the values ​​of the four types of conflict quantities are all between zero and one. Specifically, if the effective shooting time intervals of two trajectory segments associated with the same identity overlap, and the two trajectory segments originate from the same camera, the spatial distance between their detection areas is greater than the motion reachable distance determined by the inter-frame displacement distribution quantile of the confirmed continuous trajectories within that camera; or if the two trajectory segments originate from different cameras, there is no topologically reachable path between them within the overlapping time interval, then the identity exclusivity condition is satisfied. The confidence level of the trajectory segment satisfying the identity exclusivity condition is determined as the identity exclusivity conflict quantity. The shooting time interval between the two trajectory segments is... The time difference normalized value when it is less than the shortest passage time corresponding to the camera topology is determined as the physical inaccessibility conflict quantity; the relationship confidence when the relationship state of adjacent subjects and objects does not conform to the preset state transition table of acquisition, holding, putting down or transfer is determined as the carrying state discontinuity conflict quantity; the time overlap normalized value when the end time of the preceding event required by the event dependency relationship is later than the start time of the subsequent event is determined as the event sequence conflict quantity; the conflict seed is established for any evidence node or relationship edge whose conflict quantity is not less than the corresponding conflict triggering threshold, and the conflict triggering threshold is determined according to the corresponding conflict quantity distribution quantile value in the confirmed associated samples.

6. The method for dynamic analysis of video event clues based on deep learning according to claim 5, characterized in that, The propagation of conflict quantities along candidate associations, carrying state transitions, and event participation and sequence relationships includes: using the conflict quantity of the conflict seed as the initial conflict energy, propagating it to adjacent nodes according to the product of the relationship confidence of adjacent relationship edges and the time decay coefficient determined by the monotonically decreasing mapping of the shooting time interval, and selecting the last conflict energy after arranging multiple conflict energies reaching the same node in ascending order of value; including nodes with conflict energy not less than the propagation termination threshold and their connected relationship edges in the subgraph to be unlocked, the propagation termination threshold being determined based on the quantile value of the propagation energy distribution of nodes that have not formed the four types of conflicts in the confirmed association samples; for nodes with conflict energy less than the propagation termination threshold, whose candidate association members remain unchanged before and after adding new evidence and whose adjacent event sequence relationships remain unchanged, determining the relationship edge between them and the subgraph to be unlocked as the locking boundary; when the subgraph to be unlocked after propagation contains mutually disconnected connected components, solving for each connected component separately and submitting version differences separately.

7. The method for dynamic analysis of video event clues based on deep learning according to claim 6, characterized in that, S5 includes: using the identity label of the trajectory segment, the subject-carrying object relationship status, and the event order as variables to be solved, establishing identity uniqueness constraints, camera topology reachability constraints, carrying state transition constraints, and event dependency order constraints; determining the association cost based on stable identity features, variable appearance features, and their confidence levels; determining the state transition cost based on the degree of conformity between the candidate state and the preset state transition table; determining the event sequence cost based on the time overlap caused by the candidate event order; determining the old version relationship change cost based on the number of cancellations and additions of the relationship to be solved relative to the current version relationship; weighting and summing the costs and finding the minimum objective value that satisfies the constraints, wherein the weight of the old version relationship change cost is determined based on the confidence level of the new evidence, and the change confidence threshold is determined based on the intersection of the confidence distribution of new evidence of the confirmed relationship to be changed samples and the confirmed relationship not to be changed samples, so that the weight of the old version relationship change cost is increased when the confidence level of the new evidence is less than the change confidence threshold.

8. The method for dynamic analysis of video event clues based on deep learning according to claim 7, characterized in that, The version difference includes old identity identifiers, new identity identifiers, revoked relationship edges, newly added relationship edges, the order of events before the change, the order of events after the change, the confidence level before the change, the confidence level after the change, and a set of supporting evidence identifiers. When submitting the version difference, the current submission version is written as the termination version for revoked relationship edges and as the starting version for newly added relationship edges. The event clue chain and downstream index are updated according to the identity identifiers, object identifiers, and event identifiers involved in the relationship edge. If any relationship edge writing or downstream index update fails, an abnormal rollback is performed, the current submission is revoked, and the version before the submission remains valid. After the submission is completed, the set of supporting evidence identifiers is used to verify that each newly added relationship can be traced back to an immutable evidence node.

9. The method for dynamic analysis of video event clues based on deep learning according to claim 3, characterized in that, The training of the deep feature encoding network includes: constructing positive identity sample pairs based on cross-camera trajectory segments of the same subject; constructing negative identity sample pairs based on trajectory segments with overlapping effective shooting times and not satisfying the condition of the same subject; constructing relationship state training samples based on the acquisition, holding, putting down, and transfer sequences of the subject and the object; and constructing action event training samples based on event segments with event type and start and end time labels. The network parameters are updated by jointly using identity measurement loss, appearance attribute loss, subject-object relationship classification loss, and action event temporal loss. Using calibration samples that are independent of the training samples, the output scores of each branch are mapped to the confidence level. When some branch input features are missing, the confidence level of the corresponding branch is set to zero and the fusion weights are re-normalized based on the remaining branches with input features. When all branch input features are missing, the corresponding original observation is marked as evidence to be corrected and prohibited from participating in candidate association.

10. A deep learning-based video event clue dynamic analysis system, used to execute the deep learning-based video event clue dynamic analysis method according to any one of claims 1 to 9, characterized in that, include: The evidence generation module is used to execute step S1; The feature association module is used to execute step S2; The evidence graph construction module is used to execute step S3; The influence domain determination module is used to execute step S4; the version solution and submission module is used to execute step S5.

Images