Faceless identity recognition and confirmation method and system applied to intelligent camera video monitoring scene

Abstract

The application discloses a faceless identity recognition and confirmation method and system applied to an intelligent camera video monitoring scene, and comprises the following steps: acquiring continuous video frames containing target personnel, and performing target tracking on the target personnel to form a track segment corresponding to the same target personnel; extracting multi-modal body shape features in a preset number of continuous video frames corresponding to the track segment, wherein the multi-modal body shape features at least include human appearance attributes and behavior posture attributes; respectively calculating frame weights reflecting imaging quality for each video frame, wherein the frame weights are determined at least according to definition, occlusion rate and angle integrity, and the multi-modal body shape features are weighted and fused according to the frame weights to obtain track-level semantic description; and dividing features in the track-level semantic description into stable attributes and changeable attributes, and respectively obtaining stable attribute vectors and changeable attribute vectors through semantic embedding coding.

Description

Technical Field

[0001] This invention relates to the field of camera technology, and in particular to a faceless identity recognition and verification method and system for use in intelligent camera video surveillance scenarios. Background Technology

[0002] With the development of smart home and video surveillance technologies, devices such as home smart cameras, doorbell cameras, indoor monitoring cameras, and outdoor security cameras have been widely used in scenarios such as home security, personnel monitoring, visitor identification, and abnormal event alerts. In these scenarios, cameras typically need to identify the person entering the monitored area in order to push notifications to the user for events such as "family member arrives home," "visitor arrives," and "stranger appears."

[0003] Current identity verification methods primarily rely on facial recognition technology, which involves capturing frontal or near-frontal facial images of the target individual, extracting facial features, and comparing them with facial templates in an identity database. However, in actual use of home security cameras, the target individual often cannot consistently provide clear and complete facial images. For example, the person may be in a side profile, facing away from the camera, looking down, wearing a hat, wearing a mask, in low light at night, with partial obstruction, at a distance from the camera, or with motion blur. In these situations, the camera struggles to acquire facial features that meet the recognition requirements, leading to recognition failures or a significant decrease in accuracy for traditional facial recognition methods.

[0004] Besides facial recognition, existing technologies also include identity matching methods based on pedestrian re-identification or single-frame appearance features. These methods typically utilize features such as human appearance, clothing color, and body shape to match individuals. However, in home camera scenarios, single-frame images are easily affected by factors such as instantaneous pose, partial occlusion, changes in lighting, changes in viewing angle, and motion blur, leading to instability in the extracted body shape or clothing features. Furthermore, appearance information such as clothing, accessories, and carried items is highly variable, and relying solely on clothing color or single-frame appearance features can easily result in misidentification between similar individuals.

[0005] Meanwhile, existing identity recognition systems typically output recognition results directly, lacking detailed assessments of the reliability of those results. When the confidence level of the system's recognition results is insufficient, existing technologies often cannot distinguish between different processing states such as "can be directly confirmed," "needs temporary storage and observation," and "requires active user confirmation," which can easily lead to problems such as low-confidence results being directly pushed, high false alarm rates, or users being unable to correct errors in a timely manner.

[0006] Furthermore, existing home security camera identification systems typically lack an effective user feedback loop. Even if a user discovers an error in the identification result, the system struggles to translate the user's correction into constraints for subsequent identifications, failing to continuously reduce the probability of misidentification between similar individuals. Therefore, current technology still requires an identification and verification scheme that can combine multi-frame body shape features, stable and volatile attribute modeling, confidence-based decision-making, and a user feedback loop without relying on facial features, in order to improve the stability and reliability of identification in faceless scenarios. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a faceless identity recognition and verification method and system applicable to smart camera video surveillance scenarios.

[0008] To achieve the above objectives, the present invention provides a faceless identity recognition and verification method applied to intelligent camera video surveillance scenarios, comprising the following steps: Step S1: Acquire consecutive video frames containing the target person, and perform target tracking on the target person to form a trajectory segment corresponding to the same target person; Step S2: Extract multimodal body features from a preset number of consecutive video frames corresponding to the trajectory segment. The multimodal body features include at least human appearance attributes and behavioral posture attributes. Step S3: Calculate the frame weights reflecting the imaging quality for each video frame. The frame weights are determined based on at least sharpness, occlusion rate, and viewpoint integrity. Then, the multimodal body features are weighted and fused according to the frame weights to obtain a trajectory-level semantic description. Step S4: Divide the features in the trajectory-level semantic description into stable attributes and volatile attributes, and obtain stable attribute vectors and volatile attribute vectors respectively through semantic embedding encoding; Step S5: Based on the stable attribute vector, the volatile attribute vector, trajectory consistency and negative sample constraint, match with the candidate identity vector in the identity database to obtain the matching score of at least one candidate identity, wherein the matching does not depend on the participation of facial features in the calculation; Step S6: Determine the recognition confidence level Conf based on attribute completeness, vector matching degree, and historical stability; Step S7: Based on the comparison results of the identification confidence Conf with the high confidence threshold T_high and the low confidence threshold T_low, perform a three-level threshold decision: when Conf≥T_high, directly confirm the candidate identity; when T_low≤Conf<T_high, store the identification result as the result to be confirmed; when Conf<T_low, initiate an identity confirmation request to the user terminal. Step S8: Receive the confirmation result returned by the user terminal, and update the identity vector corresponding to the candidate identity when the confirmation result indicates that the identification is correct, and update the negative sample constraint when the confirmation result indicates that the identification is incorrect, so as to be called for subsequent identification.

[0009] Preferably, in steps S2 and S3, the preset number of consecutive video frames is 8 to 32 frames, more specifically 16 frames; in the preset number of consecutive video frames, multimodal body features are extracted for each video frame, and the clarity score, occlusion rate score, and viewpoint integrity score of the video frame are calculated respectively; the frame weight of the corresponding video frame is determined according to the clarity score, occlusion rate score, and viewpoint integrity score, and the frame weight is normalized and used to perform weighted fusion of the multimodal body features extracted from each video frame to generate a trajectory-level semantic description corresponding to the same trajectory segment.

[0010] Preferably, in step S4, the stable attributes include at least height, body shape, gait, and shoulder-to-waist ratio, and the variable attributes include at least upper garment color and / or style, lower garment color and / or style, accessory category, and carried item category; the trajectory-level semantic description is converted into standardized text descriptions according to the stable attributes and variable attributes respectively, and the standardized text descriptions are input into the semantic embedding model to generate stable attribute vectors and variable attribute vectors respectively, wherein the stable attribute vector is a 256-dimensional vector and the variable attribute vector is a 256-dimensional vector.

[0011] Preferably, in step S5, the matching score is determined according to the following formula: MatchScore=λ1×cos(Vs,LibVs)+λ2×cos(Vv,LibVv)+λ3×TrackConsistency λ4×NegativePenalty; Where Vs represents the stable attribute vector corresponding to the current trajectory segment, Vv represents the volatile attribute vector corresponding to the current trajectory segment, LibVs represents the library stable attribute vector corresponding to the candidate identity, LibVv represents the library volatile attribute vector corresponding to the candidate identity, TrackConsistency represents the consistency parameter of features of multiple frames within the same trajectory segment, and NegativePenalty represents the negative sample penalty term determined by the negative sample constraint set. λ1 represents the weight of stable attributes, λ2 represents the weight of volatile attributes, λ3 represents the weight of trajectory consistency, and λ4 represents the weight of negative sample penalty. λ1 > λ2, λ1 = 0.45, λ2 = 0.30, λ3 = 0.15, and λ4 = 0.10.

[0012] Preferably, in steps S6 and S7, the recognition confidence level is determined through confidence level calibration and a three-level threshold decision mechanism, and the recognition confidence level is calculated according to the following formula: Conf = α × Cs + β × Cv + γ × Ch; Wherein, Conf represents the recognition confidence, Cs represents the attribute completeness, which is determined based on the ratio of the number of multimodal body shape attributes detected in the current trajectory segment to the total number of preset attributes; Cv represents the vector matching degree, which is obtained by normalizing the matching score; Ch represents the historical stability, which is determined based on the confidence statistics of the candidate identity in the historical recognition record; α, β and γ represent the weights of attribute completeness, vector matching degree and historical stability, respectively, and α=0.25, β=0.40 and γ=0.35; The recognition confidence level is calibrated using a temperature scaling method, with a temperature parameter T=1.5; The three-level threshold decision mechanism includes: setting a high confidence threshold T_high=0.85 and a low confidence threshold T_low=0.60; when Conf≥T_high, the candidate identity is determined as a high confidence identity and directly confirmed; when T_low≤Conf<T_high, the candidate identity is determined as a medium confidence identity and the recognition result is stored as a pending confirmation result; when Conf<T_low, the candidate identity is determined as a low confidence identity and an identity confirmation request is initiated to the user terminal.

[0013] Preferably, in step S8, the confirmation result is used to update the identity database and subsequent identification process, specifically including: When the confirmation result indicates that the identification is correct, the identity vector corresponding to the candidate identity is updated using an exponential sliding method, with the update formula as follows: LibVector_new = 0.7×LibVector_old + 0.3×NewVector; Wherein, LibVector_new represents the updated identity vector, LibVector_old represents the original identity vector, and NewVector represents the new identity vector generated for the current trajectory segment; If the confirmation result indicates an identification error, the stable attribute vector, volatile attribute vector, candidate identity identifier, and confirmation result corresponding to the current trajectory segment are written into the negative sample constraint set. In the subsequent identity matching process, based on the similarity between the stable attribute vector and / or volatile attribute vector corresponding to the current trajectory segment and the negative sample vector in the negative sample constraint set, a negative sample penalty term is determined, and the negative sample penalty term is used to reduce the matching score of the corresponding candidate identity. In the subsequent identity matching process, the weights of the stable attribute vector and the volatile attribute vector are adaptively adjusted based on the occlusion rate, view integrity, and consistency with historical clothing features of the current trajectory segment. When clothing changes are detected to exceed the preset clothing change threshold or the occlusion rate exceeds the preset occlusion threshold, the weights of the stable attribute vector are increased and the weights of the volatile attribute vector are decreased. In the subsequent three-level threshold decision process, auxiliary judgment is also made based on the matching score interval between the first candidate identity and the second candidate identity and the identity consistency of multiple consecutive trajectory segments; when the matching score interval is lower than the preset interval threshold, or the identity consistency of multiple consecutive trajectory segments is lower than the preset consistency threshold, the recognition result enters the active confirmation state or the pending confirmation state.

[0014] The present invention also provides a faceless identity recognition and verification system for smart camera video surveillance scenarios, including a camera device, a trajectory generation module, a feature extraction module, a semantic aggregation module, a dual vector encoding module, an identity matching module, a confidence assessment module, a verification interaction module, a feedback update module, an identity database storage unit, a processor, a memory, and a communication interface; The camera device is used to acquire continuous video frames containing the target person; The trajectory generation module is used to perform target tracking on the target person to form trajectory segments corresponding to the same target person; The feature extraction module is used to extract multimodal body features from a preset number of consecutive video frames corresponding to the trajectory segment. The multimodal body features include at least human appearance attributes and behavioral posture attributes. The semantic aggregation module is used to determine frame weights based on sharpness, occlusion rate, and viewpoint integrity, and to perform weighted fusion of the multimodal body features according to the frame weights to obtain a trajectory-level semantic description. The dual-vector encoding module is used to divide the features in the trajectory-level semantic description into stable attributes and volatile attributes, and obtain stable attribute vectors and volatile attribute vectors respectively through semantic embedding encoding; The identity matching module is used to match the stable attribute vector, the volatile attribute vector, trajectory consistency and negative sample constraints with the candidate identity vector in the identity database to obtain the matching score of at least one candidate identity, wherein the matching does not depend on the participation of facial features in the calculation. The confidence assessment module is used to determine the identification confidence Conf based on attribute completeness, vector matching degree, and historical stability. The confirmation interaction module is used to directly confirm the candidate identity when Conf≥T_high, store the recognition result as a result to be confirmed when T_low≤Conf<T_high, and initiate an identity confirmation request to the user terminal when Conf<T_low, based on the comparison results of the recognition confidence Conf with the high confidence threshold T_high and the low confidence threshold T_low. The feedback update module is used to receive the confirmation result returned by the user terminal, and update the identity vector corresponding to the candidate identity when the confirmation result indicates that the identification is correct, and update the negative sample constraint when the confirmation result indicates that the identification is incorrect. The memory stores a computer program that, when executed by the processor, performs the method as described in any one of claims 1 to 6.

[0015] Preferably, the identity database storage unit stores at least a stable attribute vector, a volatile attribute vector, historical stability parameters, and negative sample constraint data for each registered identity; the communication interface is used to send an identity confirmation request to the user terminal when Conf < T_low, and to receive confirmation results returned by the user through the application push interface and / or voice interaction interface.

[0016] Preferably, the processor is configured to perform at least a portion of target tracking, multimodal body feature extraction, and trajectory-level semantic aggregation at the camera end, and at least a portion of identity database matching, confidence assessment, and feedback update at the server end which is connected to the camera end; when determining the candidate identity ranking result, the processor calculates the similarity of stable attribute vectors and volatile attribute vectors using cosine similarity, and determines the final matching result by combining trajectory consistency and negative sample constraints.

[0017] Preferably, the feedback update module is used to construct a negative sample prototype library and a confusion relationship graph stored by identity pairs. The negative sample prototype library is used to store trajectory-level vectors that have been confirmed as misidentified by the user. The confusion relationship graph is used to record the misidentification relationships between different candidate identities and update the penalty weights according to the number of misidentifications, negative sample similarity, and time decay coefficient. The initial matching process of the dual-vector encoding module and the identity matching module is executed on the camera end processor. The data uploaded to the server includes at least encrypted stable attribute vectors, volatile attribute vectors, and identification confidence, without uploading the original video frames.

[0018] The technical solution of this invention has the following beneficial effects: This invention performs target tracking on a person and generates trajectory segments. It extracts multimodal body features from multiple consecutive video frames, then determines frame weights based on sharpness, occlusion rate, and viewpoint integrity, and performs weighted fusion to form a trajectory-level semantic description. Compared to recognition methods that rely solely on single-frame images, this invention utilizes the temporal information of the same person across consecutive video frames, reducing the impact of single-frame occlusion, motion blur, local missing data, and incomplete viewpoints on the recognition results, thereby improving the stability of identity recognition in faceless scenarios.

[0019] This invention divides body shape features in trajectory-level semantic descriptions into stable attributes and volatile attributes, and generates stable attribute vectors and volatile attribute vectors respectively. Stable attributes include at least height, body shape, gait, and shoulder-to-waist ratio, while volatile attributes include at least upper garment color and / or style, lower garment color and / or style, accessory category, and carried item category. Through separate modeling of stable and volatile attributes, this invention can utilize both long-term, relatively stable identity features such as height, body shape, gait, and shoulder-to-waist ratio, and short-term appearance information such as clothing, accessories, and carried items for auxiliary identification, thereby improving the accuracy and adaptability of identity matching.

[0020] This invention does not rely on facial features for calculation during identity matching. Instead, it performs candidate identity matching based on stable attribute vectors, volatile attribute vectors, trajectory consistency, and negative sample constraints. Therefore, even when the target person is in profile, facing away from the camera, has their face obscured by a hat or mask, is in low light at night, or has insufficient facial image quality, it can still achieve identity recognition through multimodal body features, expanding the applicability of smart cameras in real-world home security scenarios.

[0021] This invention integrates stable attribute similarity, volatile attribute similarity, trajectory consistency, and a negative sample penalty term through a matching score formula. The weight of stable attributes is greater than that of volatile attributes, making identity matching more focused on long-term stable physical features, while using volatile attributes as auxiliary criteria. Simultaneously, trajectory consistency reflects the stability of features across multiple frames within the same trajectory segment, and the negative sample penalty term reduces the matching score of candidate identities corresponding to historical misidentified samples, thereby reducing false identification between individuals with similar body shapes or clothing.

[0022] This invention determines the recognition confidence level based on attribute completeness, vector matching degree, and historical stability, and calibrates the confidence level using temperature scaling to make it more suitable for subsequent classification judgment. By setting a high-confidence threshold T_high and a low-confidence threshold T_low, this invention can classify the recognition results into three processing states: high-confidence, medium-confidence, and low-confidence. High-confidence results are directly confirmed, medium-confidence results are stored as pending confirmation, and low-confidence results actively initiate an identity verification request to the user terminal. This avoids the direct confirmation of low-confidence recognition results and improves the reliability of the system's recognition result output.

[0023] This invention introduces an active verification mechanism. When the recognition confidence level falls below the low confidence threshold, the system can proactively initiate an identity verification request through the user terminal, instead of passively waiting for the user to discover the recognition error. This approach allows users to participate in the verification or correction of low confidence results in a timely manner, which helps reduce the false alarm rate and improves the interactive experience of the smart camera's identity recognition function.

[0024] This invention utilizes user confirmation results to form a feedback loop. When the user confirms a correct identification, the system updates the identity vector of the corresponding candidate identity using an exponential sliding method, gradually bringing the identity center vector in the identity database closer to the actual observed features. When the user confirms an incorrect identification, the system writes the stable attribute vector, volatile attribute vector, candidate identity identifier, and confirmation result corresponding to the current trajectory segment into the negative sample constraint set, forming a negative sample penalty term in subsequent identity matching. Through the above positive and negative feedback mechanisms, the system can continuously optimize the identity database and reduce the probability of repeated misidentification.

[0025] This invention can also adaptively adjust the weights of stable and volatile attribute vectors based on the occlusion rate, viewpoint integrity, and consistency with historical clothing features of the current trajectory segment. When significant clothing changes or severe occlusion are detected, the weights of stable attribute vectors are increased and those of volatile attribute vectors are decreased, making the recognition process more reliant on stable attributes such as height, body shape, gait, and shoulder-to-waist ratio, thus avoiding instability in recognition results due to clothing changes or partial occlusion.

[0026] This invention can construct a negative sample prototype library and a confusion relationship graph stored by identity pairs. By recording the misidentification relationship between different candidate identities and updating the penalty weight by combining the number of false confirmations, negative sample similarity and time decay coefficient, the system can continuously constrain similar people who are easily confused, thereby further reducing the risk of misidentification between similar family members, similar visitors or similarly dressed people.

[0027] This invention supports collaborative processing between the device and the server. The dual-vector encoding and initial matching process can be executed on the camera's processor. The data uploaded to the server includes at least encrypted stable attribute vectors, volatile attribute vectors, and identification confidence scores, without uploading the original video frames. This ensures identity verification capabilities while reducing the transmission of raw video data, thus enhancing user privacy protection. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a flowchart of the trajectory-level multi-frame semantic aggregation process of the present invention; Figure 3 This is a schematic diagram of the stable / volatile property dual-vector modeling of the present invention; Figure 4 This is a schematic diagram illustrating identity matching and matching score calculation according to the present invention; Figure 5 This is a flowchart of the confidence calibration and three-level threshold decision-making process of the present invention; Figure 6 This is a flowchart illustrating the active confirmation and feedback closed-loop update process of the present invention. Figure 7 A schematic diagram of the structure of the negative sample prototype library and the obfuscation relationship graph; Figure 8 System architecture diagram of the present invention; Figure 9 Diagram of collaborative processing architecture between client and server sides. Detailed Implementation

[0029] The technical solution of the present invention will be clearly and completely described below with reference to specific embodiments. It should be understood that the following embodiments are only used to illustrate the technical concept of the present invention and are not intended to limit the scope of protection of the present invention. Without departing from the technical concept of the present invention, those skilled in the art can make equivalent substitutions or adaptive adjustments to some of the steps, parameters, models, module deployment methods or data processing methods.

[0030] Reference Figures 1 to 9 This invention provides a faceless identity recognition and verification method and system for smart camera video surveillance scenarios, applicable to smart video acquisition devices such as indoor cameras, outdoor cameras, doorbell cameras, and security cameras. Unlike face recognition methods that rely on clear facial images, this invention does not rely on facial features for calculation during identity matching. Instead, it performs identity recognition based on multimodal body features of the target person in the video, including height, body shape, gait, shoulder-to-waist ratio, clothing, accessories, carried items, and behavioral postures. Combined with confidence calibration, three-level threshold decision-making, and user confirmation feedback, it achieves identity recognition and continuous optimization in faceless scenarios.

[0031] This embodiment provides a faceless identity recognition and verification method for smart camera video surveillance scenarios, including the following steps.

[0032] Step S1: Acquire consecutive video frames and form a trajectory segment: The intelligent camera acquires continuous video frames containing the target person. When the target person appears in the continuous video frames, the system first performs human detection on the target person, obtaining the human detection box, target center point, target size, human contour information and / or human key point information of the target person in each video frame.

[0033] After detecting a target person, the system performs target tracking to generate trajectory segments corresponding to the same target person. Target tracking can employ DeepSORT, ByteTrack, Kalman filter-based multi-target tracking algorithms, appearance feature-based target association algorithms, or other target tracking algorithms that can maintain the continuity of the same target identity in consecutive video frames.

[0034] The trajectory segment is used to represent the temporal movement of the same target person in consecutive video frames. The trajectory segment may include a trajectory identifier, frame number, human detection bounding box, target position, target size, direction of movement, key points of the human body, and the image region corresponding to the target person. By forming trajectory segments, the system can continuously analyze the features of the same person in different video frames, avoiding the instability caused by relying solely on a single frame image.

[0035] For example, a door camera continuously captures images of a person P entering the doorway area. The system detects this person in frames 1 through 20 and associates them with the same target using a target tracking algorithm, thus forming a trajectory segment. This trajectory segment is used for feature extraction in subsequent multi-frame sequences and trajectory-level semantic aggregation.

[0036] Step S2: Extract multimodal body shape features: Within a predetermined number of consecutive video frames corresponding to the trajectory segment, the system extracts multimodal body features of the target person. These multimodal body features include at least physical appearance attributes and behavioral posture attributes.

[0037] Among these, human appearance attributes can include gender estimation, age group estimation, height estimation, body shape estimation, shoulder width, waist width, shoulder-to-waist ratio, clothing color, clothing style, bottom color, bottom style, shoe color, hat, glasses, backpack, hand-held items, umbrella, trolley, and means of transportation. Behavioral posture attributes can include standing, walking, turning, bending over, facing away from the camera, walking sideways, entering an area, leaving an area, gait rhythm, stride variation, and body orientation.

[0038] In one embodiment, the preset number of consecutive video frames N is 8 to 32 frames, and more specifically 16 frames. When the length of a trajectory segment is less than the preset number, all valid frames in the trajectory segment can be selected; when the length of a trajectory segment is greater than the preset number, the preset number of valid frames can be selected by continuous window sampling, fixed interval sampling, or sorting based on frame quality.

[0039] In one embodiment, the preset number of continuous video frames can be adaptively determined based on the trajectory segment length, camera frame rate, target person's movement speed, and the occlusion status of the current scene. Specifically, when the target person moves quickly or the trajectory segment is short, a smaller number of video frames can be selected for processing; when the target person moves slowly, the trajectory segment is long, or the current scene has partial occlusion, low illumination, or the target person is facing away from the camera, a larger number of video frames can be selected for processing.

[0040] By extracting multimodal body features from multiple consecutive video frames, the adverse effects of momentary occlusion, abnormal posture, motion blur, low illumination, or missing local human body regions in a single frame image on identity recognition results can be reduced. Compared to single-frame recognition methods, trajectory-level multi-frame feature extraction can utilize the continuous appearance and behavioral information of the target person across multiple video frames, thus providing a more stable data foundation for subsequent semantic aggregation and identity matching.

[0041] Step S3: Calculate frame weights and generate trajectory-level semantic descriptions: For each video frame, the system calculates a frame weight that reflects the image quality. The frame weight is determined based on at least sharpness, occlusion rate, and viewpoint integrity.

[0042] Specifically, the sharpness score can be determined based on the edge strength, gradient energy, Laplacian variance, target region resolution, or image sharpness of the target person's image region. The occlusion rate score can be determined based on the proportion of occluded areas within the human detection bounding box, the proportion of missing human keypoints, the proportion of effective human body areas, or the number of visible body parts. The viewpoint integrity score can be determined based on the target person's front, side, or back orientation, the completeness of the human outline, the number of human keypoints, and the completeness of visible body parts.

[0043] In one specific embodiment, for the i-th frame, the system calculates a sharpness score Ci, an occlusion score Oi, and a view integrity score Vi, and determines a frame quality score Qi according to the following method: Qi = a × Ci + b × (1 Oi)+c×Vi; Where a, b, and c are preset weight coefficients. Since a higher occlusion rate generally indicates fewer usable human body regions for identification in the video frame, a weight of 1 is used. Oi indicates the positive contribution of occlusion factors to frame quality.

[0044] Subsequently, the system normalizes the quality scores of each frame to obtain the frame weight wi for each video frame: wi=Qi / ΣQi.

[0045] The system performs weighted fusion of the multimodal body features extracted from each video frame based on the normalized frame weight wi, generating a trajectory-level semantic description corresponding to the same trajectory segment.

[0046] For continuous attributes, such as height, shoulder-to-waist ratio, and gait rhythm, a weighted average method can be used for fusion; for discrete attributes, such as top color, bottom color, hat, and backpack, a weighted voting method can be used for fusion; for behavioral posture attributes, a trajectory-level behavioral description can be generated based on the posture change results in multiple video frames.

[0047] For example, a trajectory segment may consist of 20 video frames. Frame 1 has high clarity and low occlusion, resulting in a higher frame weight; Frame 5 has partial occlusion, resulting in a lower frame weight; and Frame 12 is positioned with its back to the camera, resulting in lower viewpoint completeness, further reducing its frame weight. During fusion, the system ensures that high-quality frames contribute more to the trajectory-level semantic description, thereby improving the stability of the trajectory-level body shape description.

[0048] In a set of validation samples, when the preset number of consecutive video frames N is 4 frames, the recognition stability is relatively low due to the limited available temporal information. When N is 8 to 16 frames, the trajectory-level semantic aggregation effect is significantly improved. When N is further increased to 32 frames, the recognition effect can still be improved, but the marginal gain is small relative to the increase in computational cost. Therefore, in one embodiment, N is set to 16 frames to achieve a better balance between recognition stability and computational overhead.

[0049] Step S4: Divide stable attributes into stable and volatile attributes and generate bi-vectors. The system divides the features in the trajectory-level semantic description into stable attributes and volatile attributes, and obtains stable attribute vectors and volatile attribute vectors through semantic embedding encoding respectively.

[0050] The stable attributes refer to identity-related attributes that are relatively stable over a relatively long period of time and are not easily changed by daily clothing. Stable attributes include at least height, body type, gait, and shoulder-to-waist ratio. In other embodiments, stable attributes may also include shoulder width, leg length ratio, body contour proportions, walking rhythm, and movement posture habits.

[0051] The variable attribute refers to an appearance attribute that may change with date, scene, season, or personal attire. Variable attributes include at least the color and / or style of clothing, the color and / or style of bottoms, accessory categories, and carried item categories. In other embodiments, variable attributes may also include shoe color, hat, glasses, backpack, handbag, umbrella, vehicle, etc.

[0052] In its implementation, the system converts the trajectory-level semantic description into standardized text descriptions based on stable and volatile attributes. For example, a standardized text description of stable attributes could be: "Male, young man, approximately 175 cm tall, medium build, stable gait, shoulder-to-waist ratio approximately 1.42." A standardized text description of volatile attributes could be: "Wearing a black jacket, blue jeans, a baseball cap, and carrying a black backpack."

[0053] The system inputs the standardized text descriptions mentioned above into the semantic embedding model, generating stable attribute vectors Vs and volatile attribute vectors Vv. Furthermore, both the stable and volatile attribute vectors can be 256-dimensional vectors. After generating the vectors, the system can normalize them to facilitate subsequent cosine similarity calculations.

[0054] By adopting a modeling approach that separates stable attribute vectors and volatile attribute vectors, the system can utilize both long-term stable features such as height, body shape, gait, and shoulder-to-waist ratio, as well as short-term auxiliary features such as clothing, accessories, and carried items during identity recognition, thereby improving the stability of identity recognition in faceless scenarios.

[0055] Step S5: Identity matching based on bi-vector, trajectory consistency, and negative sample constraints. The system matches candidate identity vectors in the identity database based on the stable attribute vector, the volatile attribute vector, trajectory consistency, and negative sample constraints, obtaining a matching score for at least one candidate identity. The matching does not rely on facial features in the calculation.

[0056] For each registered identity, the identity database contains at least a stable attribute vector LibVs, a volatile attribute vector LibVv, historical stability parameters, and negative sample constraint data. The system calculates the similarity between the stable attribute vector Vs and the volatile attribute vector Vv corresponding to the current trajectory segment and the stable attribute vector LibVs and the volatile attribute vector LibVv of each candidate identity in the identity database.

[0057] In one embodiment, the match score is determined according to the following formula: MatchScore = λ1 × cos(Vs, LibVs) + λ2 × cos(Vv, LibVv) + λ3 × TrackConsistency λ4×NegativePenalty; Where Vs represents the stable attribute vector corresponding to the current trajectory segment, Vv represents the volatile attribute vector corresponding to the current trajectory segment, LibVs represents the library stable attribute vector corresponding to the candidate identity, LibVv represents the library volatile attribute vector corresponding to the candidate identity, TrackConsistency represents the consistency parameter of features across multiple frames within the same trajectory segment, and NegativePenalty represents the negative sample penalty term determined by the negative sample constraint set.

[0058] λ1 represents the weight of stable attributes, λ2 represents the weight of volatile attributes, λ3 represents the weight of trajectory consistency, and λ4 represents the weight of negative sample penalty, with λ1 > λ2. Further, λ1 = 0.45, λ2 = 0.30, λ3 = 0.15, and λ4 = 0.10.

[0059] Because stable attributes better reflect the long-term identity characteristics of a target person, their weight is greater than that of volatile attributes. Volatile attributes are used to assist in identity recognition. Track consistency is used to characterize whether the features of multiple frames within the same trajectory segment are stable. The negative sample penalty term is used to reduce the matching score of the candidate identity corresponding to historical misidentified samples. TrackConsistency can be determined based on the consistency of attributes across frames within the same trajectory segment. For example, when attributes such as height estimation, body shape estimation, shoulder-to-waist ratio, gait rhythm, and clothing color change little across multiple frames, TrackConsistency is high; when the attributes of multiple frames within the same trajectory segment differ significantly, TrackConsistency is low. NegativePenalty can be determined based on the similarity between the current trajectory segment vector and the negative sample vectors in the negative sample constraint set. For example, if the current trajectory segment is highly similar to a historical misidentified sample of a candidate identity, NegativePenalty increases, thereby reducing the MatchScore of that candidate identity. The system can calculate the matching score for multiple candidate identities in the identity database separately and output the candidate identities in descending order of matching score.

[0060] Furthermore, during subsequent identity matching, the system can adaptively adjust the weights of stable and volatile attribute vectors based on the occlusion rate, viewpoint integrity, and consistency with historical clothing features of the current trajectory segment. When clothing changes exceed a preset clothing change threshold or the occlusion rate exceeds a preset occlusion threshold, the system increases the weights of stable attribute vectors and decreases the weights of volatile attribute vectors. This makes identity recognition more reliant on stable attributes such as height, body shape, gait, and shoulder-to-waist ratio, avoiding over-reliance on short-term clothing features.

[0061] Step S6: Determine the recognition confidence level: After obtaining the matching scores of candidate identities, the system determines the recognition confidence level (Conf) based on attribute completeness, vector matching degree, and historical stability.

[0062] In one embodiment, the identification confidence level is calculated according to the following formula: Conf = α × Cs + β × Cv + γ × Ch; Where Conf represents recognition confidence, Cs represents attribute completeness, Cv represents vector matching degree, and Ch represents historical stability; α, β, and γ represent the weights of attribute completeness, vector matching degree, and historical stability, respectively, with α=0.25, β=0.40, and γ=0.35. Cs represents attribute completeness, which can be determined based on the ratio of the number of multimodal body shape attributes detected in the current trajectory segment to the total number of preset attributes. For example, if the preset attributes include eight categories: height, body shape, gait, shoulder-to-waist ratio, upper garment color, lower garment color, accessories, and carried items, and if six of these categories are effectively detected in the current trajectory segment, then Cs can be 6 / 8.

[0063] Cv represents the vector matching degree, which can be obtained by normalizing the matching score. For example, the MatchScore of a candidate identity can be mapped to the interval between 0 and 1 as Cv. Ch represents historical stability, which can be determined based on the average confidence level of the candidate identity in historical recognition records, the number of historical confirmations, the number of false confirmations, or the recognition stability over a recent period.

[0064] Furthermore, to make the output recognition confidence score more consistent with the true probability distribution, the system can perform temperature scaling calibration on the recognition confidence score, with a temperature parameter T=1.5. In one embodiment, the system can first obtain the uncalibrated confidence score Conf_raw, and then perform calibration in the following manner: Conf = sigmoid(logit(Conf_raw) / T); where sigmoid represents the Sigmoid function and logit represents the log-odds function. Temperature scaling can reduce the risk of misjudgments caused by model overconfidence or low confidence, making subsequent three-level threshold decisions more stable.

[0065] Step S7: Perform a three-level threshold decision: The system performs a three-level threshold decision based on the comparison between the identification confidence level (Conf) and the high confidence threshold (T_high) and low confidence threshold (T_low). In one embodiment, the high confidence threshold (T_high) = 0.85 and the low confidence threshold (T_low) = 0.60. When Conf ≥ T_high, the system classifies the candidate identity as a high-confidence identity and directly confirms it. In this case, the system can send a confirmation notification to the user terminal, such as "Mom has returned home" or "Identity A has visited." When T_low ≤ Conf < T_high, the system classifies the candidate identity as a medium-confidence identity and stores the identification result as a pending confirmation result. In this case, the system can silently record the identification result without immediately initiating an active confirmation request to the user, or it can display a "Possibly an identity" prompt when the user views historical events, video recordings, or identity recognition records. When Conf < T_low, the system classifies the candidate identity as a low-confidence identity and initiates an identity confirmation request to the user terminal. For example, the system can send a prompt to the user, "A person has been detected. Do you want to confirm their identity?", through application push, pop-up window, voice interaction or message notification, and provide confirmation entry points such as candidate identity selection, stranger marking, and manual correction.

[0066] Furthermore, when performing the three-level threshold decision, the system can also combine the matching score interval between the first and second candidate identities and the identity consistency of multiple consecutive trajectory segments for auxiliary judgment. If the matching score interval between the first and second candidate identities is lower than a preset interval threshold, it indicates that there is a risk of confusion among multiple candidate identities; if the identity consistency of multiple consecutive trajectory segments is lower than a preset consistency threshold, it indicates that the current recognition result is not stable enough. At this time, the system can put the recognition result into an active confirmation state or a pending confirmation state to reduce the risk of misidentification.

[0067] Step S8: Receive confirmation results and provide feedback updates: The system receives the confirmation result returned by the user terminal and updates the identity database and subsequent identification process based on the confirmation result.

[0068] When the confirmation result indicates that the identification is correct, the system updates the identity vector corresponding to the candidate identity using an exponential sliding method. The update formula is as follows: LibVector_new=0.7×LibVector_old+0.3×NewVector; Here, LibVector_new represents the updated identity vector, LibVector_old represents the original identity vector, and NewVector represents the new identity vector generated for the current trajectory segment. Through exponential sliding updates, the identity center vector in the identity database can gradually approach the actual observed features confirmed by the user, thereby improving the accuracy of subsequent recognition.

[0069] When the confirmation result indicates an identification error, the system writes the stable attribute vector, volatile attribute vector, candidate identity identifier, and confirmation result corresponding to the current trajectory segment into the negative sample constraint set. In subsequent identity matching, the system determines a negative sample penalty term based on the similarity between the stable attribute vector and / or volatile attribute vector corresponding to the current trajectory segment and the negative sample vectors in the negative sample constraint set, and uses this negative sample penalty term to reduce the matching score of the corresponding candidate identity.

[0070] For example, if the system previously misidentifies "Identity B" as "Identity A," and the user corrects the error, the system will write the vector corresponding to the current trajectory segment into the negative sample constraint set for "Identity A." Subsequently, when a new trajectory segment that is highly similar to this negative sample appears, the system will add NegativePenalty when calculating the matching score for "Identity A," thereby reducing the probability of misidentifying it as "Identity A" again.

[0071] This embodiment also provides a faceless identity recognition and verification system applied to intelligent camera video surveillance scenarios. The system includes a camera device, a trajectory generation module, a feature extraction module, a semantic aggregation module, a dual-vector encoding module, an identity matching module, a confidence assessment module, a verification interaction module, a feedback update module, an identity database storage unit, a processor, a memory, and a communication interface.

[0072] The camera device is used to acquire continuous video frames containing the target person. The camera device can be an indoor camera, an outdoor camera, a doorbell camera, or other intelligent devices with video capture capabilities.

[0073] The trajectory generation module is used to perform target tracking on target personnel in order to generate trajectory segments corresponding to the same target personnel.

[0074] The feature extraction module is used to extract multimodal body features from a preset number of consecutive video frames corresponding to the trajectory segment. The multimodal body features include at least human appearance attributes and behavioral posture attributes.

[0075] The semantic aggregation module is used to determine frame weights based on sharpness, occlusion rate, and viewpoint integrity, and to perform weighted fusion of multimodal body features according to the frame weights to obtain a trajectory-level semantic description.

[0076] The dual-vector encoding module is used to divide the features in the trajectory-level semantic description into stable attributes and volatile attributes, and then obtain stable attribute vectors and volatile attribute vectors through semantic embedding encoding respectively.

[0077] The identity matching module is used to match candidate identity vectors in the identity database based on stable attribute vectors, volatile attribute vectors, trajectory consistency, and negative sample constraints, to obtain a matching score for at least one candidate identity. The matching does not rely on facial features in the calculation.

[0078] The confidence assessment module is used to determine the recognition confidence Conf based on attribute completeness, vector matching degree, and historical stability, and can perform temperature scaling calibration on the recognition confidence Conf.

[0079] The confirmation interaction module performs a three-level threshold decision based on the comparison results between the recognition confidence level (Conf) and the high confidence threshold (T_high) and low confidence threshold (T_low). When Conf ≥ T_high, the confirmation interaction module directly confirms the candidate identity, such as matching the target candidate identity with the highest score or meeting the preset sorting conditions; when T_low ≤ Conf < T_high, the recognition result is stored as a pending confirmation result; when Conf < T_low, an identity confirmation request is initiated to the user terminal through the communication interface.

[0080] The feedback update module is used to receive the confirmation result returned by the user terminal, and update the identity vector corresponding to the candidate identity when the confirmation result indicates that the identification is correct, and update the negative sample constraint when the confirmation result indicates that the identification is incorrect.

[0081] The identity database storage unit stores at least stable attribute vectors, volatile attribute vectors, historical stability parameters, and negative sample constraint data for each registered identity. Furthermore, the identity database storage unit can also store identity names, identity tags, historical identification records, positive sample trajectory vectors, negative sample trajectory vectors, and obfuscated identity relationships.

[0082] The communication interface is used to communicate with the user terminal. When Conf < T_low, it sends an identity verification request to the user terminal and receives the confirmation result returned by the user through the application push interface and / or voice interaction interface.

[0083] In one embodiment, the system employs a collaborative processing approach between the edge and the server. The camera-side processor performs at least a portion of target tracking, multimodal body feature extraction, trajectory-level semantic aggregation, dual-vector encoding, and initial matching; the server performs at least a portion of identity database matching, confidence assessment, feedback updates, or multi-device identity database collaboration.

[0084] To enhance privacy protection, after dual-vector encoding and initial matching are completed at the camera end, the data uploaded to the server includes at least encrypted stable attribute vectors, volatile attribute vectors, and identification confidence scores, instead of uploading the original video frames. The server can then use the encrypted vector data to update the identity database, sort candidate identities, or perform cross-device collaboration, thereby reducing the privacy risks associated with the external transmission of the original video data.

[0085] In a further embodiment, the feedback update module constructs a negative sample prototype library and a confusion relationship graph stored by identity pairs. The negative sample prototype library is used to store trajectory-level vectors that have been confirmed as misidentified by the user. For each registered identity, the system can maintain a corresponding set of negative sample prototypes. Each negative sample prototype includes at least a stable attribute vector, a volatile attribute vector, a misidentified candidate identity, a true confirmation result, the number of misconceptions, and the last update time.

[0086] A confusion graph is used to record the misidentification relationships between different candidate identities. Nodes in the graph represent registered identities, and edges represent instances of confusion or misidentification between two identities. Edge weights can be determined based on the number of false positives, negative sample similarity, and a time decay coefficient. The more false positives and the higher the negative sample similarity, the greater the edge weight; as time progresses, if no further false positives occur, the edge weight can decrease according to the time decay coefficient. In subsequent identification processes, when the current trajectory segment is highly similar to a negative sample prototype of a certain identity, the system increases the NegativePenalty corresponding to that identity, thereby reducing the matching score of that identity. In this way, the system can gradually reduce misidentification between people with similar body shapes and clothing. In one specific embodiment, an outdoor smart camera detects a target person P entering a doorway area. The target person P's face is obscured by a hat, making effective identity recognition through facial features impossible.

[0087] The system first tracks the target person P, forming a trajectory segment T of 20 frames. Then, the system extracts multimodal body features of the target person P from the trajectory segment T, such as gender, age group, height and body type, clothing color, clothing style, hat, backpack, items carried, and behavioral posture, and calculates the sharpness score, occlusion rate score, and view integrity score for each frame.

[0088] The system performs weighted fusion of features from each frame based on frame weights to obtain a trajectory-level semantic description. This trajectory-level semantic description is divided into stable attributes and volatile attributes. The stable attributes include "male, young man, height approximately 175 cm, medium build, shoulder-to-waist ratio approximately 1.42", while the volatile attributes include "black jacket, blue jeans, baseball cap, black backpack".

[0089] The system generates a stable attribute vector Vs and a volatile attribute vector Vv, respectively, and compares them with the database vectors of candidate identities "Identity A" and "Identity B". If the similarity of the stable attribute with candidate identity "Identity A" is 0.92, the similarity of the volatile attribute is 0.78, the trajectory consistency is 0.95, and the negative sample penalty term is 0, then the matching score is: MatchScore=0.45×0.92+0.30×0.78+0.15×0.95 0.10 × 0 = 0.7905.

[0090] The system further calculates attribute completeness Cs, vector matching degree Cv, and historical stability Ch. For example, if Cs=0.85, Cv=0.79, and Ch=0.88, then the identification confidence is: Conf=0.25×0.85+0.40×0.79+0.35×0.88=0.8365.

[0091] Since T_low ≤ Conf < T_high, the system classifies the candidate identity as a medium-confidence identity and stores the recognition result "possibly identity A" as a pending confirmation result. If the user subsequently reviews the record and confirms that the person is identity A, the system updates the identity vector corresponding to identity A using an exponential sliding method; if the user confirms that the person is not identity A, the system writes the current trajectory segment into the negative sample constraint set corresponding to identity A to reduce the probability of subsequent misidentification.

[0092] Experimental verification and performance evaluation: To verify the technical effectiveness of the faceless identity recognition and verification method described in this invention, a comparative test was conducted on the method of this invention and a face feature-based identity recognition method in a set of test samples. The test samples included various common monitoring scenarios of smart cameras, such as clear frontal view, side view, face occlusion, back to the camera, and low-light nighttime conditions. During the test, the method of this invention did not rely on face features for calculation, but instead performed identity recognition based on trajectory-level multi-frame semantic aggregation, stable attribute vectors, volatile attribute vectors, trajectory consistency, confidence calibration, and feedback loop closure.

[0093] In one test embodiment, the recognition accuracy of the method of the present invention and the baseline face recognition method in different scenarios is compared as shown in Table 1 below: Table 1

[0094] The test results above show that the baseline face recognition method has a high recognition accuracy in clear frontal scenes. However, in faceless or weak face scenes such as side profile, face occlusion, back to the camera, and low light at night, the method of this invention can complete identity recognition based on multimodal body features such as height, body shape, gait, shoulder-to-waist ratio, clothing, accessories, carried items, and behavioral posture, thereby improving the recognition stability in faceless scenes.

[0095] Furthermore, to verify the optimization effect of the active confirmation mechanism and feedback loop on the recognition results, a comparative test was conducted on a set of test samples, comparing the recognition scheme without the active confirmation mechanism with the recognition scheme with the active confirmation mechanism enabled. The test results are shown in Table 2 below: Table 2

[0096] The test results above show that by setting a high-confidence threshold T_high and a low-confidence threshold T_low, and proactively initiating an identity verification request to the user terminal when Conf < T_low, the number of cases where low-confidence recognition results are directly confirmed as identity results can be reduced. Furthermore, user confirmation results can be further used to update the identity vector or negative sample constraint set, allowing the identity database to be continuously optimized during subsequent use, thereby reducing the probability of misidentification between people with similar body shapes and clothing.

[0097] The above embodiments can support the technical content of this invention regarding trajectory-level multi-frame semantic aggregation, stable / volatile attribute dual-vector modeling, matching score calculation, confidence calibration and three-level threshold decision-making, active confirmation, positive and negative feedback closed loop, end-side processing, and privacy-preserving upload.

[0098] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A faceless identity recognition and verification method applied to intelligent camera video surveillance scenarios, characterized in that, Includes the following steps: Step S1: Acquire consecutive video frames containing the target person, and perform target tracking on the target person to form a trajectory segment corresponding to the same target person; Step S2: Extract multimodal body features from a preset number of consecutive video frames corresponding to the trajectory segment. The multimodal body features include at least human appearance attributes and behavioral posture attributes. Step S3: Calculate the frame weights reflecting the imaging quality for each video frame. The frame weights are determined based on at least sharpness, occlusion rate, and viewpoint integrity. Then, the multimodal body features are weighted and fused according to the frame weights to obtain a trajectory-level semantic description. Step S4: Divide the features in the trajectory-level semantic description into stable attributes and volatile attributes, and obtain stable attribute vectors and volatile attribute vectors respectively through semantic embedding encoding; Step S5: Based on the stable attribute vector, the volatile attribute vector, trajectory consistency and negative sample constraint, match with the candidate identity vector in the identity database to obtain the matching score of at least one candidate identity, wherein the matching does not depend on the participation of facial features in the calculation; Step S6: Determine the recognition confidence level Conf based on attribute completeness, vector matching degree, and historical stability; Step S7: Based on the comparison results of the identification confidence Conf with the high confidence threshold T_high and the low confidence threshold T_low, perform a three-level threshold decision: when Conf≥T_high, directly confirm the candidate identity; when T_low≤Conf<T_high, store the identification result as the result to be confirmed; when Conf<T_low, initiate an identity confirmation request to the user terminal. Step S8: Receive the confirmation result returned by the user terminal, and update the identity vector corresponding to the candidate identity when the confirmation result indicates that the identification is correct, and update the negative sample constraint when the confirmation result indicates that the identification is incorrect, so as to be called for subsequent identification.

2. The method according to claim 1, characterized in that, In steps S2 and S3, the preset number of consecutive video frames is 8 to 32 frames. In the preset number of consecutive video frames, multimodal body features are extracted for each video frame, and the clarity score, occlusion rate score, and viewpoint integrity score of the video frame are calculated respectively. The frame weight of the corresponding video frame is determined according to the clarity score, occlusion rate score, and viewpoint integrity score. The frame weight is normalized and then used to perform weighted fusion of the multimodal body features extracted from each video frame to generate a trajectory-level semantic description corresponding to the same trajectory segment.

3. The method according to claim 1, characterized in that, In step S4, the stable attributes include at least height, body shape, gait, and shoulder-to-waist ratio, and the variable attributes include at least the color and / or style of the upper garment, the color and / or style of the lower garment, the category of accessories, and the category of carried items. The trajectory-level semantic description is converted into standardized text descriptions according to the stable attributes and the variable attributes, and the standardized text descriptions are input into the semantic embedding model to generate stable attribute vectors and variable attribute vectors, respectively. The stable attribute vector is a 256-dimensional vector, and the variable attribute vector is a 256-dimensional vector.

4. The method according to claim 1, characterized in that, In step S5, the matching score is determined according to the following formula: MatchScore=λ1×cos(Vs,LibVs)+λ2×cos(Vv,LibVv)+λ3×TrackConsistency λ4×NegativePenalty; Where Vs represents the stable attribute vector corresponding to the current trajectory segment, Vv represents the volatile attribute vector corresponding to the current trajectory segment, LibVs represents the library stable attribute vector corresponding to the candidate identity, LibVv represents the library volatile attribute vector corresponding to the candidate identity, TrackConsistency represents the consistency parameter of features of multiple frames within the same trajectory segment, and NegativePenalty represents the negative sample penalty term determined by the negative sample constraint set. λ1 represents the weight of stable attributes, λ2 represents the weight of volatile attributes, λ3 represents the weight of trajectory consistency, and λ4 represents the weight of negative sample penalty. λ1 > λ2, λ1 = 0.45, λ2 = 0.30, λ3 = 0.15, and λ4 = 0.

10.

5. The method according to claim 4, characterized in that, In steps S6 and S7, the recognition confidence level is determined through confidence level calibration and a three-level threshold decision mechanism, and the recognition confidence level is calculated according to the following formula: Conf = α × Cs + β × Cv + γ × Ch; Wherein, Conf represents the recognition confidence, Cs represents the attribute completeness, which is determined based on the ratio of the number of multimodal body shape attributes detected in the current trajectory segment to the total number of preset attributes; Cv represents the vector matching degree, which is obtained by normalizing the matching score; Ch represents the historical stability, which is determined based on the confidence statistics of the candidate identity in the historical recognition record; α, β and γ represent the weights of attribute completeness, vector matching degree and historical stability, respectively, and α=0.25, β=0.40 and γ=0.35; The recognition confidence level is calibrated using a temperature scaling method, with a temperature parameter T=1.5; The three-level threshold decision mechanism includes: setting a high confidence threshold T_high=0.85 and a low confidence threshold T_low=0.60; when Conf≥T_high, the candidate identity is determined as a high confidence identity and directly confirmed; when T_low≤Conf<T_high, the candidate identity is determined as a medium confidence identity and the recognition result is stored as a pending confirmation result; when Conf<T_low, the candidate identity is determined as a low confidence identity and an identity confirmation request is initiated to the user terminal.

6. The method according to claim 1, characterized in that, In step S8, the confirmation result is used to update the identity database and subsequent identification process, specifically including: When the confirmation result indicates that the identification is correct, the identity vector corresponding to the candidate identity is updated using an exponential sliding method, with the update formula as follows: LibVector_new=0.7×LibVector_old+0.3×NewVector; Wherein, LibVector_new represents the updated identity vector, LibVector_old represents the original identity vector, and NewVector represents the new identity vector generated for the current trajectory segment; If the confirmation result indicates an identification error, the stable attribute vector, volatile attribute vector, candidate identity identifier, and confirmation result corresponding to the current trajectory segment are written into the negative sample constraint set. In the subsequent identity matching process, based on the similarity between the stable attribute vector and / or volatile attribute vector corresponding to the current trajectory segment and the negative sample vector in the negative sample constraint set, a negative sample penalty term is determined, and the negative sample penalty term is used to reduce the matching score of the corresponding candidate identity. In the subsequent identity matching process, the weights of the stable attribute vector and the volatile attribute vector are adaptively adjusted based on the occlusion rate, view integrity, and consistency with historical clothing features of the current trajectory segment. When clothing changes are detected to exceed the preset clothing change threshold or the occlusion rate exceeds the preset occlusion threshold, the weights of the stable attribute vector are increased and the weights of the volatile attribute vector are decreased. In the subsequent three-level threshold decision process, auxiliary judgment is also made based on the matching score interval between the first candidate identity and the second candidate identity and the identity consistency of multiple consecutive trajectory segments; when the matching score interval is lower than the preset interval threshold, or the identity consistency of multiple consecutive trajectory segments is lower than the preset consistency threshold, the recognition result enters the active confirmation state or the pending confirmation state.

7. A faceless identity recognition and verification system applied to intelligent camera video surveillance scenarios, characterized in that, It includes a camera device, a trajectory generation module, a feature extraction module, a semantic aggregation module, a dual vector encoding module, an identity matching module, a confidence assessment module, a confirmation interaction module, a feedback update module, an identity database storage unit, a processor, a memory, and a communication interface; The camera device is used to acquire continuous video frames containing the target person; The trajectory generation module is used to perform target tracking on the target person to form trajectory segments corresponding to the same target person; The feature extraction module is used to extract multimodal body features from a preset number of consecutive video frames corresponding to the trajectory segment. The multimodal body features include at least human appearance attributes and behavioral posture attributes. The semantic aggregation module is used to determine frame weights based on sharpness, occlusion rate, and viewpoint integrity, and to perform weighted fusion of the multimodal body features according to the frame weights to obtain a trajectory-level semantic description. The dual-vector encoding module is used to divide the features in the trajectory-level semantic description into stable attributes and volatile attributes, and obtain stable attribute vectors and volatile attribute vectors respectively through semantic embedding encoding; The identity matching module is used to match the stable attribute vector, the volatile attribute vector, trajectory consistency and negative sample constraints with the candidate identity vector in the identity database to obtain the matching score of at least one candidate identity, wherein the matching does not depend on the participation of facial features in the calculation. The confidence assessment module is used to determine the identification confidence Conf based on attribute completeness, vector matching degree, and historical stability. The confirmation interaction module is used to directly confirm the candidate identity when Conf≥T_high, store the recognition result as a result to be confirmed when T_low≤Conf<T_high, and initiate an identity confirmation request to the user terminal when Conf<T_low, based on the comparison results of the recognition confidence Conf with the high confidence threshold T_high and the low confidence threshold T_low. The feedback update module is used to receive the confirmation result returned by the user terminal, and update the identity vector corresponding to the candidate identity when the confirmation result indicates that the identification is correct, and update the negative sample constraint when the confirmation result indicates that the identification is incorrect. The memory stores a computer program that, when executed by the processor, performs the method as described in any one of claims 1 to 6.

8. The system according to claim 7, characterized in that, The identity database storage unit stores at least a stable attribute vector, a volatile attribute vector, historical stability parameters, and negative sample constraint data for each registered identity; the communication interface is used to send an identity confirmation request to the user terminal when Conf < T_low, and to receive the confirmation result returned by the user through the application push interface and / or voice interaction interface.

9. The system according to claim 7, characterized in that, The processor is configured to perform at least a portion of target tracking, multimodal body feature extraction, and trajectory-level semantic aggregation at the camera end, and at least a portion of identity database matching, confidence assessment, and feedback update at the server end which is in communication with the camera end; when determining the candidate identity ranking result, the processor calculates the similarity of stable attribute vectors and volatile attribute vectors using cosine similarity, and determines the final matching result by combining trajectory consistency and negative sample constraints.

10. The system according to claim 7, characterized in that, The feedback update module is used to construct a negative sample prototype library and a confusion relationship graph stored by identity. The negative sample prototype library is used to store trajectory-level vectors that have been confirmed as misidentified by the user. The confusion relationship graph is used to record the misidentification relationship between different candidate identities and update the penalty weight according to the number of misconfirmations, negative sample similarity and time decay coefficient. The initial matching process of the dual-vector encoding module and the identity matching module is executed on the camera end processor. The data uploaded to the server includes at least encrypted stable attribute vectors, volatile attribute vectors, and identification confidence scores, without uploading the original video frames.

Images