High-density group physical training testing system based on facial feature recognition
By dividing the face into local regions to extract features and generate a credibility vector, identity anchors and temporal identity maps are established, solving the problem of continuity in face recognition and target tracking in high-density group physical training tests, and achieving accurate identity recognition and trajectory maintenance under occlusion conditions.
Patent Information
- Application Number
- CN202611292858.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-25
- Publication Date
- 2026-11-10
AI Technical Summary
In high-density group physical training tests, existing technologies struggle to maintain the continuity of face recognition and target tracking when multiple people are moving in parallel, close to each other, or overlapping, leading to identity recognition errors and trajectory confusion. In particular, when a face is partially occluded, it is difficult to accurately determine whether the candidate's face before and after the occlusion belongs to the same person.
The face is divided into multiple local facial regions, local facial features are extracted and region confidence vectors are generated, identity anchors are established, and multiple candidate identity chains are retained in the occluded area through the temporal identity graph module. After re-exposure, the face recognition template is jointly filtered according to identity occupancy, adjacency topology and test phase status.
It effectively reduces identity swapping caused by the highest similarity or nearest position in a single frame, maintains the temporal correspondence between personnel identity and trajectory, preserves the identity purity and feature differentiation of the face recognition template, avoids misidentified features entering the template, and ensures the accuracy of high-density group physical training tests.
Smart Images

Figure CN123071897A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and biometric recognition technology, specifically a high-density group physical testing system based on facial feature recognition. Background Technology
[0002] Current high-density group physical training tests typically employ continuous video capture within a fixed camera area. Before personnel enter the testing area or before the test begins, face detection, facial feature extraction, and identity verification are performed, establishing an initial correspondence between the identified personnel ID and the personnel bounding boxes in the video. During the test, the system continuously tracks the personnel bounding boxes, maintaining trajectory numbers based on the overlap of bounding boxes in adjacent image frames, the distance between their center points, movement direction, or appearance features. The number of actions, completion time, or test phase are then recorded in the corresponding trajectory. The face recognition component often uses complete face images to form a unified feature vector. This unified feature vector is compared with pre-registered face templates for similarity, and the highest similarity or a fixed acceptance criterion determines the personnel's identity. If a person briefly leaves the field of vision or face detection fails, the system typically relies on the original trajectory to continue transmitting identity. When a person re-enters the identifiable state, the current facial features are re-matched with the registered template to restore the association between the personnel ID and the trajectory.
[0003] In situations involving multiple people moving in parallel, approaching each other, or crossing paths, conventional techniques typically handle face recognition and target tracking separately. The face detector outputs the visible face region in each image frame, the face recognizer generates feature vectors based on the complete facial region, and the target tracker assigns trajectory numbers based on the spatiotemporal continuity of the bounding boxes. When a face is obscured by an arm, hair, or adjacent person, or when head turning or motion blur results in the loss of complete facial features, the face recognizer may stop outputting identity information or provide a low-confidence matching result based on the remaining facial information. To maintain trajectory continuity, the target tracker usually connects the bounding boxes before and after occlusion based on the nearest position, predicted displacement, or box overlap. During periods of multiple people crossing paths, the positional relationships of the bounding boxes change rapidly, and relying solely on distance and overlap can easily lead to trajectory swapping. When a face is subsequently detected again, the system often directly overwrites the existing trajectory with the current recognition result, lacking joint verification of the relative positions of the group before and after occlusion and the continuity of the testing process.
[0004] To adapt to facial changes during motion, some existing systems update person templates during testing. This is done by incorporating facial features with acceptable similarity into the existing template, or by replacing the initial template with the average of features from the most recent frames. While this approach can supplement samples from profile views, expressions, and areas of partial blur, template updates typically rely solely on the similarity between the current frame and existing templates. They fail to differentiate the visibility of local areas such as the forehead, eye area, bridge of the nose, cheekbones, and mouth area, and do not preserve information about identity uncertainty during occlusion. If adjacent individuals have similar facial features, or if erroneous trajectories have linked other individuals' facial regions to the current identity, mismatched features may be written into the template, causing the template center to gradually deviate from the initial registered identity. Subsequent recognition continues to use the affected template for comparison, easily leading to erroneous recognition results participating in a cycle of updates and perpetuating identity mismatches from one occlusion period into subsequent testing phases.
[0005] Therefore, in high-density group physical training tests, when multiple individuals experience continuous occlusion, cross-positioning, and repeated exposure of partial faces, the complete facial features, individual trajectories, and single-frame spatial positions cannot independently maintain reliable identity continuity. The system struggles to determine whether the candidate faces before and after occlusion belong to the same individual. The technical reason for this problem is that conventional solutions typically perform local optimal matching within each image frame and, when the recognition result is uncertain, still solidify the identity based on the nearest location or highest similarity. This fails to incorporate the credibility of local facial regions, changes in adjacency topology between individuals, and the reachability relationships during the physical training test phase into the same temporal constraints. Once identities are exchanged within the occlusion zone, erroneous results may be written into the face template, turning short-term mismatches into continuous mismatches. Ultimately, this disrupts the correspondence between individual identities, video trajectories, and physical training test data, and makes it impossible to perform local retrospective correction of the solidified identity based on the re-exposed facial evidence in subsequent image frames. Summary of the Invention
[0006] The purpose of this invention is to provide a high-density group physical forging test system based on facial feature recognition, which can solve the problems in the background art mentioned above.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A high-density group physical training test system based on facial feature recognition includes a facial sequence acquisition module, a local feature representation module, an identity anchor point module, a temporal identity graph module, and an identity parsing module. The face sequence acquisition module is used to obtain the continuous face regions of each candidate from the physical training test video; The local feature representation module is used to extract local facial features according to facial partitions and generate region confidence vectors corresponding to occlusion, pose and sharpness. The identity anchor module is used to establish personnel identity anchors based on highly reliable and complete faces. The temporal identity graph module is used to write candidate identities, local feature matching relationships, personnel adjacency topology, and test phase status into cross-frame identity nodes and transition edges. The identity resolution module is used to retain multiple candidate identity chains in the occluded area, and after the person is re-exposed, the target identity chain is determined according to the mutual exclusion of identity occupancy, the continuity of adjacency topology and the reachability constraints during the test phase. Only the local facial features that meet the constraints are used to update the facial recognition template of the corresponding person.
[0008] Preferably, the local feature representation module divides the candidate face into the forehead, eye area, bridge of the nose, bilateral cheekbone area and mouth area, generates a visibility mask for each facial area, and constructs a region confidence vector based on the continuity of visible pixels, the integrity of region boundaries, the degree of pose projection compression and the consistency of motion blur direction. The local feature representation module also independently matches each local facial feature with the anchor sub-feature of the corresponding facial region in the identity anchor point, performs sparse weighting on the matching results according to the region confidence vector, and blocks the facial region from participating in the generation of the candidate identity chain when the visibility mask of any facial region and the anchor sub-feature do not satisfy the geometric correspondence.
[0009] Preferably, the personnel adjacency topology is composed of the relative orientation, adjacency order and local density relationship between the candidate face regions in the same image frame. The temporal identity graph module establishes topological corresponding edges for the personnel adjacency topology in consecutive image frames, and records the adjacency sequence before occlusion, the occlusion entry order and the direction of disappearance of the visible area for candidates who have cross occlusion. After the candidate is re-revealed, the temporal identity graph module establishes a topological recovery relationship based on the relative orientation sequence of the re-revealed personnel and the adjacent sequence before occlusion, and marks the identity transfer edges that violate the occlusion entry order, the adjacent order maintenance condition, or the short-term displacement continuity condition as unselectable edges.
[0010] Preferably, the temporal identity graph module generates stage state nodes according to the preparation stage, action execution stage and test end stage, and configures identity transition edges corresponding to the stage states for candidate identity nodes in adjacent image frames. The identity resolution module retains multiple candidate identity chains that satisfy the local face feature matching relationship within the occlusion interval. It applies joint constraints to each candidate identity chain, namely, that the same identity occupies only one face region at the same time, that the same face region belongs to only one identity, and that the stage state cannot jump backward. Based on the combined cost of local feature deviation, topology restoration deviation, and stage transition deviation, it selects the target identity chain with the minimum total cost from the candidate identity chains.
[0011] Preferably, the local feature representation module establishes an anchor point sub-feature set for each facial region, and divides the anchor point sub-feature set into a forward visible subset, a side visible subset, and an occlusion boundary subset according to the pose change direction of the candidate in consecutive image frames; When generating the region confidence vector, the local feature representation module selects the corresponding anchor point sub-feature subset according to the pose projection direction of the candidate face, calculates the intra-class dispersion of the local face features within the anchor point sub-feature subset and the inter-class interval relative to the corresponding subsets of other persons, and sets the facial regions with increased intra-class dispersion and decreased inter-class interval as frozen regions, so that the frozen regions only participate in candidate identity elimination and do not participate in identity anchor point update.
[0012] Preferably, the temporal identity graph module extracts at least one layer of adjacent personnel centered on the candidate who has cross-occlusion, forming a local topological segment containing the central candidate, adjacent personnel and relative directional relationships, and converts the local topological segment before and after occlusion into an adjacency sequence sorted by azimuth angle. When the number of visible persons before and after occlusion is inconsistent, the temporal identity graph module constructs a sequence correspondence between two adjacent sequences that allows missing nodes but does not allow the confirmed adjacent order to be reversed. It also uses the identity anchor points of adjacent persons that are not occluded to fix the sequence boundary and deletes candidate identity chains that exceed the sequence boundary or cause the confirmed adjacent order to be reversed from the temporal identity graph.
[0013] Preferably, the temporal identity graph module expands each candidate identity node into a face observation layer node, an identity hypothesis layer node, and a test phase layer node. The face observation layer node records the local face features and the region confidence vector. The identity hypothesis layer node records the correspondence between candidate identities and identity anchor points. The test phase layer node records the test phase that the candidate is in at the corresponding time. The identity resolution module establishes cross-layer consistent edges between adjacent layer nodes and performs conflict edge removal on graph structures that connect multiple fixed identity assumption layer nodes to the same face observation layer node, connect multiple face observation layer nodes to the same identity assumption layer node at the same time, and connect the test phase layer nodes in reverse order.
[0014] Preferably, the identity parsing module configures a stable feature region and a feature region to be verified for each identity anchor point. The local face features obtained after the occlusion ends and corresponding to the target identity chain are first written into the feature region to be verified. In subsequent consecutive image frames, the consistency between the local face features and the stable feature region, the consistency between the target identity chain and the personnel adjacency topology, and the continuity between the current test stage and the previous test stage are verified respectively. Only when all three consistency checks are satisfied, the local facial features in the feature region to be verified are merged into the stable feature region. If any consistency check is not satisfied, the corresponding feature to be verified is deleted and the identity anchor point state before writing is restored.
[0015] Preferably, the identity resolution module sets forward fixed boundaries and backward fixed boundaries in the temporal identity graph according to the start and end times of the occlusion interval, and sets the identity assumption layer nodes that have been fixed before the forward fixed boundary and the identity assumption layer nodes that have been confirmed by a complete face after the backward fixed boundary as irreplaceable nodes. The identity resolution module performs a bidirectional path search between two non-replaceable nodes, generating a forward candidate path from the forward fixed boundary to the backward fixed boundary and a reverse candidate path from the backward fixed boundary to the forward fixed boundary. It retains only the intersection of paths that correspond to each other in terms of identity occupancy relationship, cross-layer consistent edges, and test phase order, and determines the target identity chain after removing conflicting edges in the intersection of the paths.
[0016] Preferably, the feature regions to be verified are set up with feature queues according to local facial regions, and each feature queue records the local facial features, the corresponding region confidence vector, and the target identity chain identifier associated with the writing. Before performing the identity anchor point merging, the identity resolution module performs a closed-loop review of the local facial features in each feature queue in chronological order, completes the first face recognition matching between the newly written features and the stable feature region, reconstructs the candidate identity chain for the corresponding time period using the matching results, and compares the consistency between the reconstructed target identity chain and the target identity chain identifier associated with the writing. Merging is only performed on local facial features where the two identity chains are consistent and the feature interval between the corresponding person and the adjacent identity anchor point has not been compressed.
[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The candidate face is divided into multiple local facial regions. Local facial features are extracted from each region, and region confidence vectors are formed corresponding to occlusion, pose, and clarity. High-confidence complete faces are used to establish identity anchors. Then, candidate identities, local feature matching relationships, personnel adjacency topology, and test phase states are written into a cross-frame temporal identity graph. During occlusion, a single recognition result is not immediately fixed, but multiple candidate identity chains are retained. After the person is re-revealed, the candidate identity chains are jointly filtered according to identity occupancy mutual exclusion, adjacency topology continuity, and test phase reachability relationships. Local facial features retain usable identity information in the occluded state, personnel adjacency topology restricts relative position changes before and after crossover, and test phase states exclude identity transfers inconsistent with the test process. These three types of information jointly constrain the same target identity chain, which can reduce identity swaps caused by the highest similarity or nearest position in a single frame and maintain the temporal correspondence between personnel identity and trajectory during continuous movement of high-density groups.
[0018] 2. A verification path is set up for template updates that is integrated with identity resolution. Local facial regions are selected for matching, candidate exclusion, or frozen updates based on visibility and inter-class distinction status, preventing low-confidence regions from directly changing identity anchors. Local topological fragments before and after occlusion are bounded by adjacency sequences, occlusion entry order, and unoccluded adjacent individuals, ensuring that identity transfer edges that violate adjacency order are deleted before resolution. In the face observation, identity hypothesis, and testing phases, hierarchical nodes are used to establish cross-layer consistency relationships, and occluded intervals undergo bidirectional path verification through forward and backward fixed boundaries. Local facial features obtained after re-exposure first enter the feature area to be verified, and are then merged into the stable feature area after verification by stable feature consistency, topological consistency, phase continuity, and identity chain closed loop. This prevents misidentified features from directly entering the template and participating in subsequent comparisons, maintaining the identity purity of the face recognition template and the feature distinction relationship between different individuals. Attached Figure Description
[0019] Figure 1 A flowchart illustrating the overall process of continuous identity recognition in a high-density group physical forging test system based on facial feature recognition. Figure 2 Flowchart for local facial feature extraction, region confidence calculation, and frozen region determination; Figure 3 Flowchart for constructing the personnel adjacency topology, parsing the hierarchical temporal identity graph, and determining the target identity chain; Figure 4 This is a flowchart for closed-loop verification of local facial features to be verified and for updating the facial recognition template. Detailed Implementation
[0020] In one embodiment, please refer to Figure 1A high-density group physical training testing system based on facial feature recognition includes a face sequence acquisition module, a local feature representation module, an identity anchor point module, a temporal identity graph module, and an identity parsing module. The face sequence acquisition module obtains continuous face regions of each candidate from the physical training test video. The local feature representation module extracts local face features according to facial partitions and generates region confidence vectors corresponding to occlusion, pose, and clarity. The identity anchor point module establishes personnel identity anchor points based on highly reliable complete faces. The temporal identity graph module integrates candidate identities, local feature matching relationships, personnel adjacency topology, and test order. The segment state is written to the cross-frame identity node and transition edge; the identity resolution module is used to retain multiple candidate identity chains in the occluded area, and after the person is re-revealed, the target identity chain is determined according to the mutual exclusion of identity occupancy, the continuity of adjacency topology, and the reachability constraints of the test phase. Only the local facial features that meet the constraints are used to update the facial recognition template of the corresponding person; in the specific implementation, each module consists of program instructions deployed in the same computing environment. The program instructions share candidate index, local feature record, stage state record and template version record according to the video frame timestamp, so that the same candidate uses consistent data reference relationship in different processing stages.
[0021] After receiving the continuous body forging test video, the face sequence acquisition module performs face region detection and cross-frame candidate association for each image frame. The detection results record at least the face region coordinates, detection time, visible contour, head rotation direction, and sharpness descriptor. Cross-frame candidate association does not directly output the final identity, but assigns a temporary candidate index to the current face region and establishes a short-term observation chain based on the positional continuity relationship in the previous image frame. When a face region temporarily disappears due to occlusion, the short-term observation chain enters a waiting-to-continue state. The waiting-to-continue state only retains the observation position before disappearance, movement trend, local visible facial area, and adjacent personnel relationship, and does not directly equate the position prediction result with the personnel identity. When multiple newly appearing face regions may correspond to the same waiting-to-continue observation chain, candidate continuation relationships are established separately and uniformly parsed by the temporal identity graph module, avoiding the premature solidification of target tracking results into identity results from the data entry point.
[0022] Please refer to Figure 2The local feature representation module establishes a normalized facial coordinate system for each candidate face. The normalized facial coordinate system determines the rotation and scale relationships based on the visible eye center line, the main direction of the bridge of the nose, and the stable detectable parts of the outer contour of the face. For side faces or partially occluded faces, it is not required that all key points exist at the same time. Instead, local positioning points that can form non-collinear relationships are selected based on the current visible area to complete local registration. After registration, the candidate face is divided into forehead, periorbital, nasal bridge, bilateral cheekbone and periorbital regions, and local image blocks, visibility masks and local facial features are generated respectively. Each position in the visibility mask is used to indicate whether the corresponding image content is formed by the real facial region. Occlusion positions caused by arms, hair, adjacent people's faces and the boundary of the image do not participate in local feature aggregation. Missing positions in the local image block are skipped by mask convolution instead of fixed grayscale filling.
[0023] The region credibility vector is calculated according to the following formula: , ; in, Let R represent the region confidence vector of the candidate face at time t, and let R represent the number of facial regions. This represents the confidence level of the r-th facial region. Indicates the continuity of visible pixels. Indicates the integrity of the region boundary. Indicates the fidelity of attitude projection. Indicates motion sharpness, , , , The calculation weights are pre-normalized and have a sum greater than zero. For example, the visible pixel continuity of the peri-eye region is 0.9, the region boundary integrity is 0.8, the pose projection fidelity is 0.7, and the motion sharpness is 0.6. When all four calculation weights are 1, the confidence of the peri-eye region is 0.75. During local feature matching, this region is retained and its contribution ratio to the peri-eye features of the complete frontal face is reduced. If the visible pixel continuity and region boundary integrity of the peri-mouth region are close to 0 due to arm occlusion, the corresponding confidence is reduced and it is excluded from the template update process.
[0024] The identity anchor module establishes an identity anchor when a person enters the test area and their face is in a stable visible state. The stable visible state is determined by the proportion of the complete facial region in consecutive image frames, the amplitude of pose changes, and the consistency of local features. The identity anchor is not a single feature vector, but consists of the complete face anchor, anchor sub-features of each facial region, facial pose description at the time of establishment, and template version identifier. For multiple high-confidence face samples obtained consecutively from the same person, deviation samples are first eliminated based on the mutual distance between local features, and then the anchor center and acceptable discrete range are generated for the retained samples. After the identity anchor is established, it remains in a read-only state. New local facial features added during the test must not directly cover the initial anchor, but are instead entered into an independent record to be verified, so as to restore the unupdated state when the identity chain parsing backtracks.
[0025] The temporal identity graph module generates cross-frame identity nodes based on the candidate face regions in each image frame. Each cross-frame identity node is associated with at least the candidate face index, local face features, region confidence vector, candidate identity set, personnel adjacency topology, and test stage state. Mutually exclusive identity relationships are established between different nodes in the same image frame, and transition edges are established between nodes in adjacent image frames that have the possibility of positional continuation, local feature continuation, or topological continuation. Each transition edge simultaneously records local feature deviation, displacement continuity deviation, adjacency topology deviation, and stage state deviation, and does not determine whether the edge is retained based on a single similarity. When the face is completely invisible, placeholder nodes without face observations can be established within a limited occlusion range. Placeholder nodes only inherit the candidate identity set and adjacency boundary at the occlusion entrance and do not generate new features that can be written into the face recognition template.
[0026] In this embodiment, the personnel adjacency topology is composed of the relative orientation, adjacency order, and local density relationship between candidate face regions within the same image frame. The temporal identity graph module establishes topological corresponding edges for personnel adjacency topologies in consecutive image frames and records the pre-occlusion adjacency sequence, occlusion entry order, and visible area disappearance direction for candidates that experience cross-occlusion. After a candidate re-reveals, the temporal identity graph module establishes a topological recovery relationship based on the relative orientation sequence of the re-revealed personnel and the pre-occlusion adjacency sequence, and marks identity transfer edges that violate the occlusion entry order, adjacency order maintenance condition, or short-term displacement continuity condition as unselectable edges. The relative orientation is represented by the directed relationship from the current candidate face center to the center of the neighboring face, the adjacency order is formed according to the orientation around the current candidate, and the local density relationship is expressed by the number of candidate faces and the relative spacing order within a predetermined neighborhood, so that the topological information remains stable under the overall image translation and local scale changes.
[0027] The combined cost of identity transfer edges is calculated using the following formula: ; in, Denotes the combined cost of transition edge e. This indicates the local facial feature bias after weighting by regional confidence level. Indicates continuous deviation of short-time displacement. Indicates the topological deviation of personnel adjacency. This indicates the transfer deviation during the testing phase. This indicates the amount of identity-related conflicts. , , , , The coefficients are non-negative and at least three are greater than zero; for example, if the local facial feature deviations of two candidate transition edges are similar, where the first edge maintains the adjacency order before and after occlusion and there is no duplicate identity occupation, and the second edge causes the adjacency order of the two people to be reversed and connected to the already occupied identity node, then the second edge's... and The cost of the combination increases, and it becomes higher than that of the first edge, thus being excluded when solving the candidate identity chain.
[0028] The temporal identity graph module generates stage state nodes according to the preparation stage, action execution stage, and test completion stage, and configures identity transition edges corresponding to the stage states for candidate identity nodes in adjacent image frames. The stage states are not set individually based on the personnel's identity, but are read from the test process record corresponding to the current personnel's continuous observation. The preparation stage allows entry into the action execution stage, and the action execution stage allows maintaining the original state or entering the test completion stage. Personnel who have entered the test completion stage will not return to the preparation stage. For occlusion intervals where detection results are temporarily missing, the stage state nodes are maintained according to the occlusion entry state or migrated along the legal direction, without being replaced based on the stage states of adjacent personnel. When a candidate identity chain needs to be connected to the nodes before and after the occlusion through a stage reversal jump, the corresponding transition edge is deleted to avoid connecting personnel who have completed the test to the trajectory of personnel who are still in the preparation stage simply because of local facial similarity.
[0029] The identity resolution module retains multiple candidate identity chains that satisfy local face feature matching relationships within the occlusion region. It applies joint constraints to each candidate identity chain: the same identity occupies only one face region at any given time, the same face region belongs to only one identity, and the stage state cannot jump backwards. Based on the combined cost of local feature deviation, topology restoration deviation, and stage transition deviation, it selects the target identity chain with the minimum total cost from the candidate identity chains. Multiple candidate identity chains start from nodes confirmed before occlusion and end at nodes where a complete face or a highly reliable local face is obtained again after occlusion, allowing the passage of placeholder nodes and low-reliability observation nodes. During the solution process, the mutual exclusion relationship of identities is transformed into conflict constraints between paths. Paths sharing the same identity or the same face observation at the same time cannot be retained simultaneously. Then, through layer-by-layer expansion and conflict pruning, a set of candidate paths satisfying all constraints is obtained.
[0030] In a preferred embodiment, the local feature representation module divides the candidate face into forehead, periorbital, nasal bridge, bilateral cheekbone, and perioral regions, generates a visibility mask for each facial region, and constructs a region credibility vector based on visible pixel continuity, region boundary integrity, pose projection compression degree, and motion blur direction consistency. The local feature representation module also independently matches each local facial feature with the anchor point sub-features of the corresponding facial region in the identity anchor points, performs sparse weighting on the matching results according to the region credibility vector, and blocks the facial region from participating in the generation of the candidate identity chain when the visibility mask and anchor point sub-feature of any facial region do not satisfy the geometric correspondence. The geometric correspondence is determined by the relative order and projection range of local positioning points in the normalized facial coordinate system. If the right cheekbone image block is incorrectly mapped to the left cheekbone position due to side face flipping, it will not enter the identity comparison of the corresponding region even if the local texture is similar.
[0031] Local face matching uses region-level distance calculation followed by sparse aggregation. The specific calculation formula is as follows: ; in, This represents the comprehensive local feature distance between the candidate face at time t and the j-th person identity anchor point. This indicates whether the r-th facial region satisfies the geometric correspondence; a value of 1 indicates that it satisfies the correspondence, and a value of 0 indicates that it does not. Using regional credibility, This represents the normalized distance between the current local facial features and the anchor point sub-feature corresponding to the j-th person. To prevent positive numbers with a denominator of zero; for example, the periorbital and nasal bridge regions satisfy a geometric correspondence and have confidence levels of 0.8 and 0.7 respectively, while the periorbital region, due to occlusion, results in... When the value is 0, the overall distance is determined only by the area around the eyes and the bridge of the nose; occlusion textures around the mouth will not shorten the matching distance for incorrect identities.
[0032] Furthermore, the local feature representation module establishes an anchor point sub-feature set for each facial region, and divides the anchor point sub-feature set into a forward visible subset, a lateral visible subset, and an occlusion boundary subset according to the pose change direction of the candidate in consecutive image frames. When generating the region confidence vector, the local feature representation module selects the corresponding anchor point sub-feature subset according to the pose projection direction of the candidate's face, calculates the intra-class dispersion of the local facial features within the anchor point sub-feature subset and the inter-class interval relative to the corresponding subsets of other people, and sets the facial regions with increased intra-class dispersion and decreased inter-class interval as frozen regions, so that the frozen regions only participate in candidate identity elimination and not in identity anchor point update. The pose change direction is determined according to the deflection trend of the main direction of the bridge of the nose relative to the normalized facial coordinate system. Instead of directly switching subsets using the pose results of a single frame, it requires that the deflection trend in continuous observation remain consistent before switching, so as to avoid frequent changes in the matching benchmark near the critical pose.
[0033] In this embodiment, the forward visible subset stores anchor point sub-features formed by complete or near-complete facial regions, the lateral visible subset stores local features that still have stable geometric correspondences according to leftward and rightward deviations, and the occlusion boundary subset stores local features where the facial region is adjacent to the occlusion edge but the visible contour is continuous. The intra-class dispersion is calculated by the distance distribution from the current local feature to multiple anchor point sub-features in the selected subset, and the inter-class interval is calculated by the difference between the minimum distance from the current local feature to the self-identity subset and the minimum distance to the corresponding subsets of other individuals. When a local region exhibits a large dispersion in the self-identity subset and the distance difference between it and adjacent identities decreases, the local region is recorded in the freeze flag table. The freeze flag table is bound to the template version, and the corresponding freeze flag can only be removed after a complete face is obtained and the identity is reconfirmed.
[0034] In a preferred embodiment, please refer to Figure 3 The temporal identity graph module extracts at least one layer of adjacent personnel centered on the candidate personnel involved in the cross-occlusion, forming a local topological fragment containing the central candidate personnel, adjacent personnel, and relative directional relationships. It then converts the local topological fragments before and after occlusion into adjacency sequences ordered by azimuth angle. When the number of visible personnel before and after occlusion is inconsistent, the temporal identity graph module constructs a sequence correspondence between the two adjacency sequences that allows missing nodes but does not allow the reversal of the confirmed adjacency order. It uses the identity anchor points of unoccluded adjacent personnel to fix the sequence boundaries and deletes candidate identity chains that exceed the sequence boundaries or cause the confirmed adjacency order to be reversed from the temporal identity graph. At least one layer of adjacent personnel refers to the set of personnel that have a direct adjacency relationship with the central candidate personnel. When the local density is high, secondary adjacent personnel connected to directly adjacent personnel can also be added to form a local topological reference that can resist the short-term absence of a single person.
[0035] The corresponding costs of adjacent sequences before and after occlusion are calculated using the following formula: ; in, This represents the sequence correspondence cost between a local topological fragment U before occlusion and a local topological fragment V after occlusion. This indicates that the node correspondence relationship maintains the confirmed adjacency order. This indicates the set of nodes that participate. and These represent the relative azimuth angles of the corresponding nodes before and after occlusion. and They represent the normalized adjacency distances, This indicates the inconsistency between the identity anchor points of the corresponding nodes. Indicates the number of missing nodes. , , , The coefficient is non-negative; for example, if one unconfirmed person is missing after occlusion but the positional order of two confirmed adjacent persons is maintained, passage is allowed. If a record is missing but the corresponding relationship is retained, and the order of two confirmed adjacent individuals is reversed, then the corresponding... It does not enter the minimized range.
[0036] When unobstructed adjacent person identity anchor points are used to fix the boundary of the adjacency sequence, the temporal identity graph module sets the corresponding nodes of the confirmed adjacent persons as boundary nodes, and records which side of the line connecting each boundary node the central candidate person is located on and the insertion interval in the orientation sorting. If the new candidate face that appears after occlusion is outside the boundary node's defined interval, the corresponding identity transfer edge does not participate in the subsequent path expansion. When there is only one available boundary node, a semi-open boundary is formed by combining the disappearance direction at the occlusion entrance and the relative order of the unconfirmed adjacent nodes on the other side. When all directly adjacent persons are occluded at the same time, the boundary is expanded outward to the secondary adjacent persons, and an alternative boundary is established with the nearest person who can maintain the identity anchor point confirmation before and after occlusion, so as to avoid mistaking distant persons in the global picture as local topology references.
[0037] In a preferred embodiment, the temporal identity graph module expands each candidate identity node into a face observation layer node, an identity hypothesis layer node, and a test phase layer node. The face observation layer node records local face features and region confidence vectors, the identity hypothesis layer node records the correspondence between candidate identities and identity anchors, and the test phase layer node records the test phase in which the candidate is located at the corresponding time. The identity resolution module establishes cross-layer consistent edges between adjacent layer nodes and performs conflict edge removal on graph structures that connect multiple fixed identity hypothesis layer nodes to the same face observation layer node, connect multiple face observation layer nodes to the same identity hypothesis layer node at the same time, and connect test phase layer nodes in reverse order. Fixed identity hypothesis layer nodes refer to identity nodes that are confirmed by complete face anchors or fixed boundaries before and after and are not allowed to be replaced within the current occlusion range. Unfixed nodes are allowed to retain multiple candidate identity connections.
[0038] The target path cost of the layered graph is calculated using the following formula: ; in, This represents the total cost of candidate path P. Using the identity transfer edge combination cost, This indicates the number of cross-layer identity conflicts in the path. Indicates the number of reverse connections in the stage. This indicates the number of paths that do not conform to the fixed topological boundaries. , , This is a non-negative penalty coefficient; for example, if a candidate path has low local facial feature bias, but the same identity assumption connects two facial observations at the same time and the stage states are reversed, then... and If all values are greater than 0, the total cost is higher than that of candidate paths without conflicts. Therefore, the identity resolution module does not use the simple face similarity result as the final identity chain.
[0039] Conflict edge removal is performed by propagating from fixed nodes to unfixed nodes. The identity resolution module first locks the fixed identity assumption layer nodes in the forward fixed boundary and backward fixed boundary, and then deletes the cross-layer consistent edges that have the same identity occupied by the fixed node at the same time. The unique attribution relationship of the face observation layer nodes is checked in the time direction for the remaining graph structure. If the same observation node is connected to multiple unfixed identity assumption nodes, it is not directly deleted, but candidates are retained based on the local feature deviation and topological correspondence of the corresponding path. The reverse connection of the test phase layer node is directly deleted before the path search, because it can be judged as an illegal connection without relying on face similarity. The graph structure after removal still retains the necessary uncertainty during the occlusion period, but will not retain paths that violate identity uniqueness and phase monotonicity.
[0040] In a preferred embodiment, please refer to Figure 4 The identity resolution module configures a stable feature region and a feature region to be verified for each identity anchor point. The local facial features obtained after the occlusion ends and corresponding to the target identity chain are first written into the feature region to be verified. In subsequent consecutive image frames, the consistency between the local facial features and the stable feature region, the consistency between the target identity chain and the personnel adjacency topology, and the continuity between the current test phase and the previous test phase are verified respectively. Only when all three consistency checks are satisfied, the local facial features in the feature region to be verified are merged into the stable feature region. If any consistency check is not satisfied, the corresponding feature to be verified is deleted and the identity anchor point state before writing is restored. The stable feature region retains feature records that have passed complete face confirmation or closed-loop review. The feature region to be verified is saved separately according to the writing time, facial region, target identity chain, and template version, and is not used to calculate the template center together with the stable feature region.
[0041] Three consistency checks are performed within the same verification window. The consistency between local facial features and stable feature regions is determined by the feature distance and pose subset correspondence within the same region. The consistency between the target identity chain and the personnel adjacency topology is determined by whether the adjacency sequence within the verification window maintains interpretable changes. The continuity between the current test phase and the previous test phase is determined by whether the phase state migrates along a legal direction. When the target identity chain corresponding to any feature to be verified is replaced within the verification window, the pending merging state of that feature is immediately canceled. When multiple local regions are written to the feature region to be verified at the same time, region-level verification is performed separately. The periorbital region is allowed to pass, while the periorbital region continues to remain in the pending verification state due to occlusion. Synchronous merging is not forced on a complete face as a unit to prevent a single contaminated region from driving other trusted regions to update together.
[0042] The merging gating of the features to be verified is calculated according to the following formula: ; in, This represents the merged gate value for the u-th record with the unverified feature. This represents the normalized average distance from the stable feature region to the region with the same regional features. This represents the topological consistency of the target identity chain within the verification window. Represents a continuous quantity representing a stage state. This represents the feature interval maintained between the user's identity and its nearest neighbor's identity. , , , They are respectively represented as , , and The corresponding non-negative coefficients; for example, when the distance between the periorbital feature to be verified and the stable feature area is small, the topological relationship is continuous, the stage state is legal and the feature interval between it and the adjacent identity is not reduced, all four quantities remain within the range that can be merged. If the topological consistency quantity drops to 0 due to identity chain backtracking, the corresponding record will not enter the stable feature area.
[0043] Identity anchor state restoration is accomplished through template version records. When a feature to be verified is written, the version identifier and region index of the current stable feature area are recorded. The merging operation generates a new template version without overwriting the old version. When subsequent consistency checks fail or the target identity chain is replaced, the version is rolled back to the version before writing based on the previous version identifier in the record to be verified, and the anchor sub-features derived from that record to be verified are deleted. If other independently verified local features have been merged after the same version, only the affected records are revoked according to the region index and derivation relationship, and the unrelated merging results are not revoked. Parent-child relationships are preserved between template versions, so that multiple local updates within the occlusion range can be partially rolled back according to the identity chain parsing results.
[0044] In a preferred embodiment, the identity resolution module sets forward fixed boundaries and backward fixed boundaries in the temporal identity graph according to the start and end times of the occlusion interval. The identity assumption layer nodes that have been solidified before the forward fixed boundary and the identity assumption layer nodes that have been confirmed by complete faces after the backward fixed boundary are set as non-replaceable nodes. The identity resolution module performs bidirectional path search between the two non-replaceable nodes, generating forward candidate paths from the forward fixed boundary to the backward fixed boundary and reverse candidate paths from the backward fixed boundary to the forward fixed boundary, respectively. Only the intersection of paths that correspond to each other in terms of identity occupancy relationship, cross-layer consistent edges, and test phase order is retained, and the conflicting edges in the path intersection are removed to determine the target identity chain. The start time of the occlusion interval is the first time when continuous face observations change from a stable state to a missing or low-confidence state, and the end time is the start time when face observations that can be confirmed by identity are obtained again and the consistency is continuously satisfied.
[0045] Forward path search inherits confirmed identities and occluded entry adjacency sequences from the forward fixed boundary, expanding to the backward fixed boundary in an increasing time direction; reverse path search inherits reconfirmed face identities and occluded exit adjacency sequences from the backward fixed boundary, backtracking to the forward fixed boundary in a decreasing time direction; when the identity assumptions, face observation affiliations, and stage states corresponding to the two types of paths are consistent at the same time, the relevant nodes and edges enter the path intersection; if the forward path relies on positional continuity to connect to a certain person, while the reverse path points to another person based on the reappearing complete face, the inconsistent part remains unconfirmed and the suboptimal path continues to be checked, without directly solidifying the identity using the unilateral result of either direction; after the path intersection is determined, mutual exclusion checks are performed on the shared nodes between multiple identity chains.
[0046] In a preferred embodiment, the feature regions to be verified are set up with feature queues according to local facial regions. Each feature queue records local facial features, the corresponding region confidence vector, and the target identity chain identifier associated with the write operation. Before performing identity anchor point merging, the identity parsing module performs closed-loop verification of the local facial features in each feature queue in chronological order. The newly written features are matched with the stable feature regions for the first time. The matching results are then used to reconstruct the candidate identity chain for the corresponding time period. The reconstructed target identity chain is compared with the target identity chain identifier associated with the write operation. Merging is performed only on local facial features where the two identity chains are consistent and the feature interval between the corresponding person and the adjacent identity anchor point is not compressed. The feature queues maintain a sequential relationship according to the write time. If the previous record has not completed the closed-loop verification, the subsequent record can continue to be enqueued but will not skip the previous record to complete the merging.
[0047] The consistency determination of closed-loop review is calculated according to the following formula: ; in, This indicates whether the u-th feature record to be verified has passed the closed-loop review. This is an instruction operation that takes a value of 1 if the condition is true and 0 if the condition is false. Indicates the target identity chain identifier associated with the write operation. This indicates that the features to be verified are used to reconstruct the target identity chain identifier after the first face recognition match. This represents the feature interval between the user's identity before merging and the nearest neighboring identity anchor. This indicates the feature interval after the assumption is merged; for example, if the two identity chains have the same identifier but the feature interval is assumed to be smaller after merging, the second indicator operation is set to 0. A value of 0 indicates that the corresponding feature remains in the feature region to be verified or is deleted instead of entering the stable feature region.
[0048] The first face recognition matching is performed only between the feature to be verified and the stable feature region, without using subsequent records in the same feature queue as a reference, to avoid multiple unconfirmed local features supporting each other and forming an erroneous closed loop. When reconstructing the candidate identity chain for the corresponding time period, a local time period covering the occlusion entry, writing time, and subsequent confirmation time is selected with the writing time of the feature to be verified as the center. The local feature matching relationship, personnel adjacency topology, and test phase state are recalculated, and the face observation layer node and template version saved at that time are used. If the reconstructed identity chain is consistent with the identifier at the time of writing, but the dependent adjacency topology edge has been marked as a conflict edge in the subsequent parsing, the closed loop review is still judged to fail.
[0049] In this embodiment, when multiple people in the test area start physical training tests simultaneously, the face sequence acquisition module establishes a temporary candidate index for each person, the identity anchor module establishes identity anchors using the highly reliable complete faces when entering the test area, and the temporal identity graph module synchronously records the adjacency topology of the people in the preparation phase. After the people enter the action execution phase, if two people cross each other and one of them has their mouth and cheekbones obscured, the local feature representation module can still extract local facial features from the eye area, forehead, and bridge of the nose, and mask the obscured areas according to the region confidence vector. The temporal identity graph module retains two sets of candidate identity chains and records the order of the obscuration entry. After the people are re-exposed, the target identity chain is determined by combining the adjacency sequence recovery relationship and the stage state reachability relationship. The relevant local features first enter the feature area to be verified and are not directly written into the stable feature area.
[0050] Specifically, if the first visible face after occlusion has a close local feature distance to both of the two person identity anchor points, the identity resolution module does not immediately update the identity based on the smaller distance. Instead, it checks the impact of the two candidate paths on the identity occupancy relationship at the same time. If one path causes the same identity to connect to two face observation layer nodes simultaneously, or requires another person to return from the test end stage to the action execution stage, then the corresponding cross-layer consistent edge or stage connection is cut off. The remaining path must also satisfy the sequence correspondence of local topological segments before and after occlusion. If the re-revealed face is located outside the sequence boundary defined by the confirmed adjacent persons, then the corresponding identity transfer edge will no longer participate in the solution. The target identity chain formed after the above processing is used to restore the correspondence between the person identity and the video trajectory.
[0051] Furthermore, during the test video processing, the system saves the intermediate data required to form the temporal identity graph, including the local image index, visibility mask summary, region confidence vector, candidate identity distance, adjacency sequence, stage status, and template version identifier for each face observation layer node. The intermediate data is organized into a traceable record according to the occlusion interval. When a new complete face confirmation result appears at the backward fixed boundary, only the local graph structure intersecting with the corresponding occlusion interval is re-parsed, and the entire test video is not reprocessed. During the local backtracking, the confirmed identity nodes before the forward fixed boundary and after the backward fixed boundary remain unchanged. The template update is revoked or merged according to the derivation relationship of the feature record to be verified, so that the identity chain correction will not change the personnel identity records that are unrelated to the current occlusion event.
[0052] Preferably, the construction of the personnel adjacency topology does not rely on a fixed personnel number order, but uses the adjacent personnel whose identities can be confirmed in the current image frame as anchor points, and represents unconfirmed personnel as directional intervals and adjacency orders relative to the anchor points; when the group as a whole moves in the same direction, the absolute coordinates of each personnel change, but the local adjacency order remains comparable within a short time range; when an individual personnel leaves the current neighborhood, the sequence correspondence allows for missing nodes and records the cost of the missing node, without directly interpreting the missing node as an identity exchange; when a new personnel enters the neighborhood, the new node can only establish a topological corresponding edge at the insertion position that does not disrupt the confirmed adjacency order, thereby making the topological constraints applicable to test images where the number of personnel changes finitely before and after occlusion.
[0053] Preferably, the test phase state is generated from the state record during the physical training test. The state record includes at least the temporal relationship between entering the test area, entering the preparation state, starting the action, continuing the action, ending the action, and leaving the test area. The preparation phase corresponds to entering the preparation state until the start of the action, the action execution phase corresponds to the start of the action until the end of the action, and the test end phase corresponds to the end of the action. The state record only serves as a temporal constraint for the identity chain and does not replace facial features to complete identity verification. For personnel whose specific action state cannot be determined but whose test has not ended, the phase state can be retained as a candidate set of the preparation phase and the action execution phase. Path search only allows the candidate set to converge along a valid direction and does not allow regression from the test end phase to the candidate set.
[0054] Preferably, the stable feature region is maintained separately according to complete face anchors and local region anchors. Complete face anchors are used to quickly narrow down the candidate identity set in the unobstructed state, while local region anchors are used to form region-level matching in the partially obstructed state. When merging local features, only the anchor sub-feature set of the corresponding facial region is updated, and the anchor sub-features of regions unrelated to it are not recalculated. When a local region is frozen multiple times in a row, the initial anchor sub-features remain unchanged, and the intra-class dispersion and inter-class interval of the region are re-evaluated after the subsequent complete face confirmation results. Frozen regions can still exclude candidate identities through obvious inconsistencies with other identity anchors, but their matching results are not used as the basis for adding personal template samples.
[0055] In another preferred embodiment, the identity resolution module can establish interconnected temporal identity subgraphs for multiple simultaneously occurring occlusion intervals. Each subgraph is bounded by the corresponding local group of people and the occlusion time range. Multiple subgraphs share confirmed identity anchor nodes but do not share unconfirmed observation nodes. When the same person enters two consecutive occlusion intervals, the backward fixed boundary of the previous subgraph serves as the forward fixed boundary of the next subgraph. Only after the identity resolution of the previous subgraph is completed will the relevant identity nodes in the next subgraph become fixed. If the previous subgraph undergoes backtracking correction, the next subgraph is re-searched along the shared fixed boundary without modifying other subgraphs unrelated to the shared identity, thus maintaining the consistency of the identity chain in multiple consecutive occlusion scenarios.
[0056] In another preferred embodiment, the local feature representation module can use the partitioned output of the same face feature extraction network as each local face feature, or it can perform feature extraction separately on each normalized facial region. However, the outputs of different regions need to be mapped to a local feature space that can be independently compared. When using partitioned output, region aggregation is performed on the feature map according to the visibility mask. The feature response corresponding to the occlusion position is not included in the aggregation result. When extracting separately, each region is registered according to the pose subset in the anchor point sub-feature set, and the output dimension remains consistent within the same region. Regardless of the implementation method, the comprehensive local feature distance is sparsely aggregated according to the region credibility and geometric correspondence, and missing regions are not replaced with zero vectors and then used for averaging.
[0057] In this embodiment, the personnel identity results output by the system are determined by the identity assumption layer nodes of the target identity chain in each image frame. The association between test data and personnel identity is established based on the solidified identity nodes. Nodes in the occluded area that have not yet completed bidirectional path verification remain in a pending confirmation state. The relevant test data can be temporarily stored under the corresponding temporary candidate index and written into the personnel identity record after the target identity chain is determined. If the subsequent closed-loop review causes a partial replacement of the identity chain, the personnel identity index and test data reference relationship within the time range are synchronously corrected according to the node time range in the temporal identity graph, without modifying the data records outside the fixed boundary. This forms a complete data closed loop from face sequence, local features, identity anchors, temporal identity graph, target identity chain to template update record.
Claims
1. A high-density group physical training testing system based on facial feature recognition, characterized in that, It includes a face sequence acquisition module, a local feature representation module, an identity anchor point module, a temporal identity graph module, and an identity parsing module; The face sequence acquisition module is used to obtain the continuous face regions of each candidate from the physical training test video; The local feature representation module is used to extract local facial features according to facial partitions and generate region confidence vectors corresponding to occlusion, pose and sharpness. The identity anchor module is used to establish personnel identity anchors based on highly reliable and complete faces. The temporal identity graph module is used to write candidate identities, local feature matching relationships, personnel adjacency topology, and test phase status into cross-frame identity nodes and transition edges. The identity resolution module is used to retain multiple candidate identity chains in the occluded area, and after the person is re-exposed, the target identity chain is determined according to the mutual exclusion of identity occupancy, the continuity of adjacency topology and the reachability constraints during the test phase. Only the local facial features that meet the constraints are used to update the facial recognition template of the corresponding person.
2. The high-density group physical forging test system based on facial feature recognition according to claim 1, characterized in that, The local feature representation module divides the candidate face into the forehead, eye area, bridge of the nose, bilateral cheekbone area and mouth area, generates a visibility mask for each facial area, and constructs a region confidence vector based on the continuity of visible pixels, the integrity of region boundaries, the degree of pose projection compression and the consistency of motion blur direction. The local feature representation module also independently matches each local facial feature with the anchor sub-feature of the corresponding facial region in the identity anchor point, performs sparse weighting on the matching results according to the region confidence vector, and blocks the facial region from participating in the generation of the candidate identity chain when the visibility mask of any facial region and the anchor sub-feature do not satisfy the geometric correspondence.
3. The high-density group physical forging test system based on facial feature recognition according to claim 1, characterized in that, The personnel adjacency topology is composed of the relative orientation, adjacency order and local density relationship between the candidate face regions in the same image frame. The temporal identity graph module establishes topological corresponding edges for the personnel adjacency topology in consecutive image frames, and records the adjacency sequence before occlusion, the occlusion entry order and the disappearance direction of the visible area for candidates who have cross occlusion. After the candidate is re-revealed, the temporal identity graph module establishes a topological recovery relationship based on the relative orientation sequence of the re-revealed personnel and the adjacent sequence before occlusion, and marks the identity transfer edges that violate the occlusion entry order, the adjacent order maintenance condition, or the short-term displacement continuity condition as unselectable edges.
4. The high-density group physical forging test system based on facial feature recognition according to claim 2, characterized in that, The temporal identity graph module generates stage state nodes according to the preparation stage, action execution stage and test end stage, and configures identity transition edges corresponding to the stage states for candidate identity nodes in adjacent image frames. The identity resolution module retains multiple candidate identity chains that satisfy the local face feature matching relationship within the occlusion interval. It applies joint constraints to each candidate identity chain, namely, that the same identity occupies only one face region at the same time, that the same face region belongs to only one identity, and that the stage state cannot jump backward. Based on the combined cost of local feature deviation, topology restoration deviation, and stage transition deviation, it selects the target identity chain with the minimum total cost from the candidate identity chains.
5. The high-density group physical forging test system based on facial feature recognition according to claim 2, characterized in that, The local feature representation module establishes an anchor point sub-feature set for each facial region, and divides the anchor point sub-feature set into a forward visible subset, a side visible subset, and an occlusion boundary subset according to the direction of the candidate's pose change in consecutive image frames. When generating the region confidence vector, the local feature representation module selects the corresponding anchor point sub-feature subset according to the pose projection direction of the candidate face, calculates the intra-class dispersion of the local face features within the anchor point sub-feature subset and the inter-class interval relative to the corresponding subsets of other persons, and sets the facial regions with increased intra-class dispersion and decreased inter-class interval as frozen regions, so that the frozen regions only participate in candidate identity elimination and do not participate in identity anchor point update.
6. The high-density group physical forging test system based on facial feature recognition according to claim 4, characterized in that, The temporal identity graph module extracts at least one layer of adjacent personnel centered on the candidate personnel who have experienced cross-occlusion, forming a local topological fragment containing the central candidate personnel, adjacent personnel, and relative directional relationships, and converts the local topological fragments before and after occlusion into an adjacency sequence sorted by azimuth angle. When the number of visible persons before and after occlusion is inconsistent, the temporal identity graph module constructs a sequence correspondence between two adjacent sequences that allows missing nodes but does not allow the confirmed adjacent order to be reversed. It also uses the identity anchor points of adjacent persons that are not occluded to fix the sequence boundary and deletes candidate identity chains that exceed the sequence boundary or cause the confirmed adjacent order to be reversed from the temporal identity graph.
7. The high-density group physical forging test system based on facial feature recognition according to claim 6, characterized in that, The temporal identity graph module expands each candidate identity node into a face observation layer node, an identity hypothesis layer node, and a test phase layer node. The face observation layer node records the local face features and the region confidence vector. The identity hypothesis layer node records the correspondence between candidate identities and identity anchor points. The test phase layer node records the test phase that the candidate is in at the corresponding time. The identity resolution module establishes cross-layer consistent edges between adjacent layer nodes and performs conflict edge removal on graph structures that connect multiple fixed identity assumption layer nodes to the same face observation layer node, connect multiple face observation layer nodes to the same identity assumption layer node at the same time, and connect the test phase layer nodes in reverse order.
8. The high-density group physical forging test system based on facial feature recognition according to claim 7, characterized in that, The identity resolution module configures a stable feature region and a feature region to be verified for each identity anchor point. The local face features obtained after the occlusion ends and corresponding to the target identity chain are first written into the feature region to be verified. In subsequent consecutive image frames, the consistency between the local face features and the stable feature region, the consistency between the target identity chain and the personnel adjacency topology, and the continuity between the current test phase and the previous test phase are verified respectively. Only when all three consistency checks are satisfied, the local facial features in the feature region to be verified are merged into the stable feature region. If any consistency check is not satisfied, the corresponding feature to be verified is deleted and the identity anchor point state before writing is restored.
9. The high-density group physical forging test system based on facial feature recognition according to claim 8, characterized in that, The identity resolution module sets forward fixed boundaries and backward fixed boundaries in the temporal identity graph according to the start and end times of the occlusion interval, and sets the identity assumption layer nodes that have been solidified before the forward fixed boundary and the identity assumption layer nodes that have been confirmed by a complete face after the backward fixed boundary as irreplaceable nodes. The identity resolution module performs a bidirectional path search between two non-replaceable nodes, generating a forward candidate path from the forward fixed boundary to the backward fixed boundary and a reverse candidate path from the backward fixed boundary to the forward fixed boundary. It retains only the intersection of paths that correspond to each other in terms of identity occupancy relationship, cross-layer consistent edges, and test phase order, and determines the target identity chain after removing conflicting edges in the intersection of the paths.
10. The high-density group physical forging test system based on facial feature recognition according to claim 8, characterized in that, The feature regions to be verified are set up with feature queues according to local facial regions. Each feature queue records the local facial features, the corresponding region confidence vector, and the target identity chain identifier associated with the writing. Before performing the identity anchor point merging, the identity resolution module performs a closed-loop review of the local facial features in each feature queue in chronological order, completes the first face recognition matching between the newly written features and the stable feature region, reconstructs the candidate identity chain for the corresponding time period using the matching results, and compares the consistency between the reconstructed target identity chain and the target identity chain identifier associated with the writing. Merging is only performed on local facial features where the two identity chains are consistent and the feature interval between the corresponding person and the adjacent identity anchor point has not been compressed.
