A person re-identification method based on multi-dimensional feature relative metric entropy

Abstract

The application discloses a personnel re-identification method based on multi-dimensional feature relative metric entropy, and comprises the following steps: S1, decoding and extracting image frames from a video stream, positioning a personnel area through target detection and eliminating low-confidence detection, assigning a globally unique tracking identifier through motion trajectory correlation to establish a continuous tracking link; S2, performing scale normalization and histogram equalization preprocessing on the personnel bounding box area in the tracking link, eliminating low-quality images with motion blur and excessive occlusion, and constructing a multi-dimensional feature library composed of feature vectors and the like. The application covers personnel rotation, crouching and posture conversion processes by establishing a continuous tracking link, and introduces multi-dimensional feature relative metric entropy to quantify the geometric dispersion degree of the feature space, thereby solving the problem of dramatic feature distribution changes caused by posture offset of a traditional single-frame static template, significantly reducing the false recognition and missed recognition rates under different postures of the same identity, and improving the cross-mirror correlation robustness under posture changes.

Description

Technical Field

[0001] This invention belongs to the field of computer systems, specifically referring to a person re-identification method based on multidimensional feature relative metric entropy. Background Technology

[0002] Currently, Person Re-Identification (ReID) technology has wide applications in fields such as intelligent security, cross-camera tracking, and edge computing. Existing mainstream technical solutions can be mainly divided into the following three categories:

[0003] The first category is the single-frame static feature retrieval scheme. This scheme is typically deployed at the edge or center, using convolutional neural networks such as OSNet and ResNet as the backbone network to extract global features from a single frame image of a person. During cross-camera comparison, the system performs identity matching by calculating the cosine similarity or Euclidean distance between features. This type of method is simple to implement and computationally efficient, but its limitation lies in its reliance on a single-frame static template captured in a single snapshot, making it difficult to cover the pose changes that people experience in real-world scenarios (such as turning, crouching, occlusion, etc.). Research shows that the feature similarity between different poses of the same identity may be lower than the similarity between the same pose of different identities, leading to frequent misidentification and missed identification during cross-camera association. The root cause is the lack of an associated representation and dynamic supplementation mechanism for multi-dimensional pose features of the same identity, and the inability of single-frame templates to adapt to feature distribution shifts in complex cross-camera scenarios.

[0004] The second category: Simple multi-frame feature fusion schemes. To alleviate the insufficient pose coverage problem of single-frame methods, some schemes introduce multiple acquisitions of images from multiple cameras or multiple frames, extract multiple sets of features, and then merge them through temporal stitching, weighted averaging, or attention fusion before comparison. This type of method expands the feature sources to some extent, but still has the following problems: redundant features from repeated viewpoints are treated equally, and storage and computational costs increase linearly with the number of acquisitions; the fusion process lacks explicit quantification of the geometric dispersion of features in the feature space, making it impossible to determine whether the acquired features cover sufficient pose change dimensions, and pose blind spots persist. The root cause lies in the lack of a quantitative measurement mechanism for the balance of feature space distribution, and the lack of objective diagnostic basis for feature geometry in acquisition and fusion decisions.

[0005] The third category: Dynamic acquisition schemes driven by model confidence. This type of scheme attempts to dynamically control the acquisition termination condition by recognizing the network's prediction confidence or information entropy. When the system determines that the current feature set has reached a high confidence level, further acquisition is terminated. However, this method has a significant subjective blind spot: when acquired features are highly clustered in local regions of the feature space, the recognition model, due to its training bias towards intra-class clustering, may output excessively high confidence levels, thus misjudging that feature coverage is sufficient. At this point, a significant pose blind spot still exists in the feature space, which the model itself cannot perceive. This leads to a significant increase in the association failure rate when encountering uncovered pose changes in actual re-recognition tasks. The drawback of this type of scheme is that the acquisition closure relies on the model's subjective prediction state, rather than an independent measure of the objective dispersion of the feature library in the feature space, and it does not use the geometric integrity of the feature space as a rigid constraint for acquisition termination.

[0006] In summary, existing people re-identification technologies generally suffer from common problems such as incomplete pose coverage, redundant feature accumulation, and subjective blind spots in the model. There is an urgent need for a cross-camera re-identification acquisition mechanism capable of dynamically diagnosing feature distribution structure and autonomously supplementing pose dimensions to improve the robustness and generalization ability of people re-identification in complex scenarios. Summary of the Invention

[0007] The purpose of this invention is to overcome the technical defects of existing people re-identification technologies, such as incomplete pose coverage, redundant feature accumulation, and subjective blind spots in the model. It provides a cross-camera re-identification acquisition mechanism that can dynamically diagnose the feature distribution structure and autonomously supplement the pose dimension, and improve the robustness and generalization ability of people re-identification in complex scenarios. This is a people re-identification method based on multi-dimensional feature relative metric entropy.

[0008] To achieve the above objectives, the present invention employs the following technical solution: a person re-identification method based on multidimensional feature relative metric entropy, comprising the following steps:

[0009] S1. Decode the video stream to extract image frames, locate the personnel area through target detection and remove low-confidence detections, and assign a globally unique tracking identifier through motion trajectory association to establish a continuous tracking link;

[0010] S2. Perform scale normalization and histogram equalization preprocessing on the bounding box region of the personnel in the tracking link to remove low-quality images with motion blur and excessive occlusion, and construct a multi-dimensional feature library composed of feature vectors.

[0011] S3. Mark feature vector pairs with similarity higher than a set threshold as redundant feature groups, calculate the mean center of all feature vectors in each redundant feature group, and retain it as a representative feature, while deleting the remaining feature vectors in the group.

[0012] S4. Construct the distance matrix and obtain the absolute metric entropy and the relative metric entropy of the multidimensional features;

[0013] S5. Compare the relative entropy of the multidimensional feature with the set threshold. If it is higher than the threshold, the feature library is deemed to be sufficiently covered and marked as ready. If it is lower than the threshold, repeat steps S1 to S4 until the threshold requirement is met.

[0014] S6. Perform preprocessing and feature extraction on the current frame image to generate a query feature vector; calculate the similarity between this feature vector and representative features in all feature libraries that have been marked as ready, find the candidate feature library with the highest similarity, and output the tracking identifier corresponding to this feature library as the re-identification result.

[0015] Furthermore, in step S1, after decoding the video stream, image frames need to be extracted at fixed time intervals, and the timestamp and camera number corresponding to each frame are recorded synchronously.

[0016] The step S1, "assigning a globally unique tracking identifier through motion trajectory association to establish a continuous tracking link", specifically refers to associating the motion trajectory of the personnel area detected in consecutive frames, using a joint matching strategy of appearance features and spatial location to assign a globally unique tracking identifier to the personnel, and establishing a continuous tracking link for the personnel within the field of view of a single camera.

[0017] The tracking link continuously covers the target person's processes of turning, squatting, and changing posture.

[0018] Furthermore, the construction of the multidimensional feature library composed of feature vectors in step S2 includes the following steps:

[0019] S21. Input the preprocessed image into a deep feature extraction network. This network maps the input image into a high-dimensional feature vector through a cascaded structure of convolutional layers and fully connected layers, and outputs a fixed-dimensional feature representation.

[0020] S22. Using the tracking identifier as an index, cluster and store the feature vectors extracted from the same person under multiple cameras, multiple times, and multiple poses in chronological order to construct a multi-dimensional feature library exclusive to that person.

[0021] Step S3, which involves "marking feature vector pairs with similarity higher than a set threshold as redundant feature groups," specifically includes the following steps:

[0022] S31. Compare all feature vectors belonging to the same person in the constructed multidimensional feature library pairwise, calculate the cosine similarity between each pair of feature vectors, and construct a symmetric feature similarity matrix.

[0023] S32. Set a similarity threshold and mark the feature vector pairs corresponding to elements in the feature similarity matrix that are higher than the similarity threshold as redundant feature groups.

[0024] Step S4, "constructing the distance matrix and obtaining the absolute metric entropy and the multidimensional feature relative metric entropy," includes the following steps:

[0025] S41. For all representative features retained in the simplified multidimensional feature library, calculate the Euclidean distance between each pair and construct the distance matrix.

[0026] S42. Calculate the arithmetic mean of all off-diagonal elements of the distance matrix to obtain the absolute metric entropy.

[0027] S43. Calculate the total number of representative features in the simplified feature library and use the natural logarithm of this total number as a normalization factor.

[0028] S44. The ratio of absolute metric entropy to normalization factor is used as the relative metric entropy of multidimensional features.

[0029] The "similarity calculation" in step S6 uses the nearest neighbor retrieval strategy.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] (1) This invention establishes a continuous tracking link to cover the process of personnel turning, squatting and changing posture, and introduces multi-dimensional feature relative metric entropy to quantify the geometric dispersion of the feature space. This solves the problem of drastic changes in feature distribution caused by posture offset in traditional single-frame static templates, significantly reduces the false recognition and false recognition rate under different postures of the same identity, and can improve the robustness of cross-camera association under posture changes.

[0032] (2) This invention identifies redundant feature groups by constructing a feature similarity matrix and setting a threshold, calculates the mean center within the group as a representative feature and deletes the remaining redundant vectors, thus avoiding the storage overhead and computational burden caused by the linear accumulation of repeated viewpoint features in the multi-frame fusion scheme, and can achieve efficient simplification of the feature library and reduce feature storage and computational redundancy.

[0033] (3) The present invention uses multidimensional feature relative metric entropy based on feature space geometric distribution as an objective constraint condition for the termination of acquisition, replacing the traditional subjective judgment mechanism that relies on the prediction confidence or information entropy of the recognition network. This avoids the model misjudging the feature coverage due to the tendency of intra-class aggregation optimization, ensuring the true integrity of the feature library in the pose dimension, and effectively eliminating the subjective acquisition blind spot driven by model confidence.

[0034] (4) This invention compares the relative metric entropy with a preset threshold. When the entropy is lower than the threshold, it automatically triggers continuous acquisition and re-evaluation, forming a "acquisition-evaluation-supplementation" closed loop. This enables the system to autonomously identify posture blind spots and dynamically supplement missing features, avoiding insufficient or excessive acquisition. This not only improves the accuracy and scene adaptability of cross-camera personnel re-identification, but also enables autonomous diagnosis and closed-loop control of feature coverage sufficiency. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Example

[0038] like Figure 1 As shown, the person re-identification method based on multidimensional feature relative metric entropy in this embodiment includes the following steps S1 to S6, specifically:

[0039] S1. Decode the video stream to extract image frames, locate the personnel area through target detection and remove low-confidence detections, and assign a globally unique tracking identifier through motion trajectory association to establish a continuous tracking link.

[0040] This step is essentially a personnel tracking and detection step. After decoding the video stream, image frames need to be extracted at fixed time intervals, and the timestamp and camera number corresponding to each frame need to be recorded simultaneously. During the image frame extraction process, the personnel area also needs to be located through object detection processing, and the personnel bounding box coordinates and confidence scores are output. Detection results with confidence scores below a preset threshold are discarded.

[0041] The phrase "establishing a continuous tracking link by associating motion trajectories with globally unique tracking identifiers" specifically refers to associating the motion trajectories of personnel areas detected in consecutive frames, employing a joint matching strategy of appearance features and spatial location to assign globally unique tracking identifiers to personnel, and establishing a continuous tracking link for that personnel within the field of view of a single camera. This tracking link needs to continuously cover the target personnel's processes of turning, crouching, and posture changes.

[0042] This step enables stable and continuous identification of target individuals from the raw video stream, establishing a consistent identity association within the single camera's field of view. This provides a traceable and retrospective object carrier for subsequent multi-pose feature acquisition. This step achieves the following functions: First, identity anchoring: by assigning a globally unique tracking identifier, it ensures that the same person is identified as the same individual across different time frames and poses, avoiding identity drift and providing an index foundation for building a multi-dimensional feature library with "one person, one database." Second, low-quality filtering: eliminating low-confidence detection results reduces false detections from contaminating subsequent feature acquisition and improves the purity of the feature library. Third, continuous tracking capability: by jointly matching appearance features and spatial location, a continuous tracking link is established, enabling the system to cover the complete movement process of the target individual (including pose transitions such as turning and squatting), laying the foundation for subsequent multi-pose feature acquisition. Fourth, spatiotemporal information recording: synchronously recording timestamps and camera numbers provides traceable metadata for subsequent cross-camera association and feature temporal organization.

[0043] S2. Perform scale normalization and histogram equalization preprocessing on the bounding box region of the person in the tracking link to remove low-quality images with motion blur and excessive occlusion, and construct a multi-dimensional feature library composed of feature vectors.

[0044] This step essentially involves locating the person region within the extracted image frames through target detection, outputting the person's bounding box coordinates and confidence score, and discarding detection results with a confidence score below a preset threshold. The preprocessing of the person's bounding box region in the tracking link, including scale normalization and histogram equalization, and the removal of motion-blurred frames and low-quality images where the person's occlusion area exceeds a preset proportion, constitutes image preprocessing.

[0045] The construction of a multidimensional feature library composed of feature vectors includes the following steps:

[0046] S21. Input the preprocessed image into a deep feature extraction network. This network maps the input image into a high-dimensional feature vector through a cascaded structure of convolutional layers and fully connected layers, and outputs a fixed-dimensional feature representation.

[0047] S22. Using the tracking identifier as an index, cluster and store the feature vectors extracted from the same person under multiple cameras, multiple times, and multiple poses in chronological order to construct a multi-dimensional feature library exclusive to that person.

[0048] Specifically, scale normalization is performed on the bounding box regions of people. This is achieved by first outputting fixed-dimensional feature vectors from the deep feature extraction network, then performing L2 normalization on these vectors to ensure that all feature vectors have a uniform magnitude of 1. The normalized feature vectors have a more stable numerical range when calculating cosine similarity, avoiding interference from feature scale differences in similarity measurement and improving the accuracy of redundant feature recognition.

[0049] While extracting feature vectors, a lightweight pose estimation network (such as OpenPose) is introduced to predict the key points of the human skeleton in the current frame, generating pose descriptors (such as key point coordinate vectors or pose category labels). These pose descriptors and feature vectors are stored as key-value pairs. During subsequent relative entropy calculations, the pose labels can be used to analyze whether the distribution of the feature space covers predefined pose types, thus more accurately locating pose blind spots. Simultaneously, the deep feature extraction network employs a multi-scale feature fusion structure, prioritizing OSNet multi-branch convolutions. Before outputting the final feature vector, feature maps of different scales are adaptively fused, ensuring that the generated features simultaneously consider global contours and local details (such as clothing textures and accessories), improving the ability to discriminate occluded or low-resolution scenes.

[0050] During this process, a timestamp of the extraction time can be recorded for each feature vector, and a time window can be set. When the timestamp of a feature vector in the feature library exceeds the window range (e.g., more than 24 hours), its weight in the relative metric entropy calculation is automatically reduced or it is directly deleted to avoid interference from outdated features (e.g., people changing clothes) on pose coverage evaluation.

[0051] S3. Mark feature vector pairs with similarity higher than a set threshold as redundant feature groups, calculate the mean center of all feature vectors in each redundant feature group, and retain it as a representative feature, while deleting the remaining feature vectors in the group.

[0052] The main function of this step is to perform redundancy detection and simplification on the feature vectors in the multidimensional feature library of the same person, merge highly similar (i.e., very close in the feature space) feature vectors into a single representative feature, delete the remaining redundant vectors, and achieve compactness of the feature library.

[0053] The step of marking feature vector pairs with similarity higher than a set threshold as redundant feature groups specifically includes the following steps:

[0054] S31. Compare all feature vectors belonging to the same person in the constructed multidimensional feature library pairwise, calculate the cosine similarity between each pair of feature vectors, and construct a symmetric feature similarity matrix.

[0055] S32. Set a similarity threshold and mark the feature vector pairs corresponding to elements in the feature similarity matrix that are higher than the similarity threshold as redundant feature groups.

[0056] The calculation of the mean center of all feature vectors within each redundant feature group, and its retention as a representative feature, aims to eliminate feature redundancy and reduce storage and computational overhead. During continuous tracking of the same person, a large number of feature vectors with similar poses and viewpoints, or even repetitions, are extracted (e.g., consecutive frames during stable walking). These vectors have extremely high cosine similarity and carry highly overlapping discriminative information. By labeling redundant feature groups and retaining the mean center, the size of the feature library can be significantly reduced, avoiding linear storage expansion, while also reducing the computational complexity of subsequent relative metric entropy calculations and cross-camera retrieval. It should be further noted that this step differs from direct random downsampling or equal-interval frame extraction; this step determines redundancy based on the geometric similarity of the feature space. Vectors with similarity higher than a threshold are considered redundant, meaning they cluster in the same local region in the feature space. Retaining the mean center of this region as a representative feature removes duplicate samples while maintaining the position of the pose region in the feature space, without disrupting the original topological structure of the distribution.

[0057] Meanwhile, since step S4 requires calculating the relative metric entropy based on representative features, retaining a large number of redundant features would cause many similar vector pairs in the distance matrix to lower the absolute metric entropy, making the entropy value unable to truly reflect the breadth of pose coverage. By simplifying the redundancy and retaining only the center point of each feature cluster, the off-diagonal elements in the distance matrix better reflect the geometric separation degree between different pose regions. This allows the relative metric entropy to accurately quantify the true dispersion of the feature space and avoids the interference of redundant features on the evaluation results.

[0058] The similarity threshold set in this step directly adjusts the strictness of redundancy detection. A higher threshold (e.g., 0.95) means only extremely similar vectors are merged, preserving more subtle differences; a lower threshold (e.g., 0.85) results in a wider merging range and stronger simplification. This threshold can be flexibly configured by the system based on storage resources or scenario requirements, balancing efficiency and discriminative power.

[0059] S4. Construct the distance matrix and obtain the absolute metric entropy and the relative metric entropy of the multidimensional features.

[0060] Specifically, it includes the following steps:

[0061] S41. For all representative features retained in the simplified multidimensional feature library, calculate the pairwise Euclidean distance between each feature and construct a distance matrix. This distance matrix has zero diagonal elements, is symmetrically distributed, and the off-diagonal elements represent the geometric separation degree of features extracted under different poses, angles, and lighting conditions in the feature space.

[0062] S42. Calculate the arithmetic mean of all off-diagonal elements of the distance matrix to obtain the absolute metric entropy.

[0063] S43. Calculate the total number of representative features in the simplified feature library and use the natural logarithm of this total number as the normalization factor.

[0064] S44. The ratio of absolute metric entropy to the normalization factor is used as the relative metric entropy of multidimensional features. This ratio, or relative metric entropy of multidimensional features, eliminates the scale interference of differences in feature library capacity on dispersion metrics, making the feature richness between different individuals comparable across samples. A larger ratio indicates a higher average spatial separation per unit number of features, richer pose and viewpoint coverage, and higher feature completeness; a smaller ratio indicates that features are excessively clustered in local areas, with blind spots in coverage and insufficient feature completeness.

[0065] As described above, the core function of this step is to map the simplified representative feature set into a dimensionless quantitative index, namely, the multidimensional feature relative entropy. This index comprehensively reflects the average spatial separation degree per unit number of features in the feature library, that is, the coverage richness of dimensions such as pose, viewpoint, and illumination.

[0066] S5. Compare the relative entropy of the multidimensional feature with the set threshold. If it is higher than the threshold, the feature library is deemed to be sufficiently covered and marked as ready. If it is lower than the threshold, repeat steps S1 to S4 until the threshold requirement is met.

[0067] This step compares the multidimensional feature relative metric entropy calculated in step S4 with a preset threshold to form a closed-loop acquisition control decision. Its core is to use the geometric integrity of the feature space as a rigid objective constraint for terminating acquisition, replacing the subjective judgment relying on model confidence in traditional methods. This mechanism avoids the problem of insufficient acquisition caused by the model's blind confidence.

[0068] As a preferred approach, this step can also include setting a dynamic threshold adaptive mechanism, adaptive threshold correction based on acquisition history, acquisition direction guidance strategy, and dual constraints of feature quantity lower limit and entropy value.

[0069] The dynamic threshold adaptive mechanism means that the preset threshold is no longer fixed, but is dynamically adjusted based on environmental parameters such as camera installation height, field of view, and scene crowd density. For example, in high-crowd density scenes, the threshold can be appropriately lowered to speed up data acquisition, while in low-density scenes, the threshold can be increased to achieve higher accuracy. This mechanism can improve the system's generalization ability in different deployment environments.

[0070] Adaptive threshold correction based on data acquisition history involves recording the final relative entropy values ​​of multiple individuals who have reached the ready state under the same camera, statistically analyzing their distribution (such as mean and standard deviation), and automatically adjusting the threshold for judging new individuals in the same scene to the lower quartile of this distribution or the mean minus one standard deviation. This method allows the threshold to adaptively converge with scene features, reducing the cost of manual parameter tuning.

[0071] The acquisition direction guidance strategy refers to the system not only repeating acquisition when the entropy value of the multidimensional feature relative metric entropy is determined to be below a threshold, but also further analyzing the sparse regions of the current feature library in the feature space. Specifically, it involves clustering existing representative features, identifying the clustering direction with the fewest samples, or calculating the region with the greatest missing value on the convex hull of the feature space, and then sending a command to the video acquisition terminal to "prioritize acquiring the posture with the greatest difference from this region" (e.g., prompting the person to turn around or squat). This strategy can accelerate entropy convergence and reduce invalid acquisition rounds.

[0072] The dual constraint of feature quantity lower limit and entropy value refers to the requirement that, in order to avoid artificially high entropy values ​​caused by having only two representative features in the feature library but with a large distance between them (for example, two extreme poses are far apart, but the intermediate pose is completely missing), a feature quantity lower limit condition is added: the total number of representative features must reach at least a preset minimum value (e.g., 5), and the entropy value must be higher than a threshold before it can be marked as ready. This dual constraint can prevent misjudgments caused by insufficient samples.

[0073] S6. Perform preprocessing and feature extraction on the current frame image to generate a query feature vector; calculate the similarity between this feature vector and representative features in all feature libraries that have been marked as ready, find the candidate feature library with the highest similarity, and output the tracking identifier corresponding to this feature library as the re-identification result.

[0074] During execution, when the person to be identified appears in the field of view of the query camera, the same preprocessing and feature extraction operations as steps S1 to S3 are first performed on the current frame image of the person to generate a query feature vector. Then, the similarity between the query feature vector and the representative features in the feature library of all people marked as ready is calculated, and the nearest neighbor retrieval strategy is used to find the candidate feature library with the highest similarity. Finally, the tracking identifier corresponding to the candidate feature library is output as the re-identification result to complete the cross-camera person identity association.

[0075] The main function of this step is to perform similarity matching between the person to be identified captured in the field of view of the query camera and the multi-dimensional feature library of all persons that has been built (marked as ready), retrieve the candidate library with the most consistent identity, and output its globally unique tracking identifier, thus completing the association of person identity across camera scenarios.

[0076] This step achieves the following objectives: First, it establishes the final closed loop for cross-camera identity association. Following the previous steps, the system has established a comprehensive and compact representative feature set for each registered individual. This step utilizes these ready feature libraries as the retrieval base, comparing real-time query features with them and outputting matching tracking identifiers, thereby connecting the entire chain from data collection to identification and realizing "one-time database creation, cross-camera reuse."

[0077] Second, representative features are used to improve retrieval efficiency and accuracy. The retrieval target is not the original complete feature library, but a representative feature set after redundancy reduction. Compared with directly comparing all original features, the number of representative features is significantly reduced, lowering the computational complexity of online retrieval; at the same time, the spatial center of each pose clustering region is preserved, avoiding interference from local redundant features on similarity ranking and improving the robustness of retrieval.

[0078] Third, the quality of the retrieval database is guaranteed by an objective readiness status. In step S5, only feature libraries that pass the relative metric entropy threshold test are marked as "ready." This mechanism ensures that each feature library entering the retrieval database has sufficient geometric dispersion in the feature space (i.e., complete pose coverage). Compared to feature sets added to the database without verification in traditional methods, ready feature libraries have stronger generalization ability when encountering new poses from the querying user, effectively reducing missed identifications caused by pose blind spots.

[0079] Fourth, output a globally unique tracking identifier to complete identity binding. The final output tracking identifier (ID) is a unique identity code for personnel in the system, which can be directly used for subsequent business logic (such as trajectory drawing, control alarm deployment, behavior analysis, etc.), realizing the conversion from visual image to semantic identity.

[0080] As described above, the present invention can be well implemented.

Claims

1. A method for re-identifying people based on multidimensional feature relative metric entropy, characterized in that, Includes the following steps: S1. Decode the video stream to extract image frames, locate the personnel area through target detection and remove low-confidence detections, and assign a globally unique tracking identifier through motion trajectory association to establish a continuous tracking link; S2. Perform scale normalization and histogram equalization preprocessing on the bounding box region of the personnel in the tracking link to remove low-quality images with motion blur and excessive occlusion, and construct a multi-dimensional feature library composed of feature vectors. S3. Mark feature vector pairs with similarity higher than a set threshold as redundant feature groups, calculate the mean center of all feature vectors in each redundant feature group, and retain it as a representative feature, while deleting the remaining feature vectors in the group. S4. Construct the distance matrix and obtain the absolute metric entropy and the relative metric entropy of the multidimensional features; S5. Compare the relative entropy of the multidimensional feature with the set threshold. If it is higher than the threshold, the feature library is deemed to be sufficiently covered and marked as ready. If it is lower than the threshold, repeat steps S1 to S4 until the threshold requirement is met. S6. Perform preprocessing and feature extraction on the current frame image to generate a query feature vector; calculate the similarity between this feature vector and representative features in all feature libraries that have been marked as ready, find the candidate feature library with the highest similarity, and output the tracking identifier corresponding to this feature library as the re-identification result.

2. The person re-identification method based on multidimensional feature relative metric entropy according to claim 1, characterized in that, In step S1, after decoding the video stream, image frames need to be extracted at fixed time intervals, and the timestamp and camera number corresponding to each frame are recorded synchronously.

3. The person re-identification method based on multidimensional feature relative metric entropy according to claim 2, characterized in that, The step S1, "assigning a globally unique tracking identifier through motion trajectory association to establish a continuous tracking link", specifically refers to associating the motion trajectory of the personnel area detected in consecutive frames, using a joint matching strategy of appearance features and spatial location to assign a globally unique tracking identifier to the personnel, and establishing a continuous tracking link for the personnel within the field of view of a single camera.

4. The person re-identification method based on multidimensional feature relative metric entropy according to claim 3, characterized in that, The tracking link continuously covers the target person's processes of turning, squatting, and changing posture.

5. The person re-identification method based on multidimensional feature relative metric entropy according to claim 3, characterized in that, The construction of a multidimensional feature library composed of feature vectors in step S2 includes the following steps: S21. Input the preprocessed image into a deep feature extraction network. This network maps the input image into a high-dimensional feature vector through a cascaded structure of convolutional layers and fully connected layers, and outputs a fixed-dimensional feature representation. S22. Using the tracking identifier as an index, cluster and store the feature vectors extracted from the same person under multiple cameras, multiple times, and multiple poses in chronological order to construct a multi-dimensional feature library exclusive to that person.

6. The person re-identification method based on multidimensional feature relative metric entropy according to claim 5, characterized in that, Step S3, "marking feature vector pairs with similarity higher than a set threshold as redundant feature groups," specifically includes the following steps: S31. Compare all feature vectors belonging to the same person in the constructed multidimensional feature library pairwise, calculate the cosine similarity between each pair of feature vectors, and construct a symmetric feature similarity matrix. S32. Set a similarity threshold and mark the feature vector pairs corresponding to elements in the feature similarity matrix that are higher than the similarity threshold as redundant feature groups.

7. The person re-identification method based on multidimensional feature relative metric entropy according to claim 6, characterized in that, Step S4, "constructing the distance matrix and obtaining the absolute metric entropy and the multidimensional feature relative metric entropy," includes the following steps: S41. For all representative features retained in the simplified multidimensional feature library, calculate the Euclidean distance between each pair and construct the distance matrix. S42. Calculate the arithmetic mean of all off-diagonal elements of the distance matrix to obtain the absolute metric entropy. S43. Calculate the total number of representative features in the simplified feature library and use the natural logarithm of this total number as a normalization factor. S44. The ratio of absolute metric entropy to normalization factor is used as the relative metric entropy of multidimensional features.

8. The person re-identification method based on multidimensional feature relative metric entropy according to claim 7, characterized in that, The "similarity calculation" in step S6 uses the nearest neighbor retrieval strategy.

Images