A vehicle re-identification method and system for cross-camera imaging degradation characterization self-calibration

Abstract

The application discloses a vehicle re-identification method and system for cross-camera imaging degradation representation self-calibration. Firstly, vehicle detection, target cropping and quality screening are performed on a multi-camera road monitoring video, an identity-independent region set is constructed from the vehicle detection frame edge, the road background strip and the stable background area after removing the vehicle identity main body; then color response, high-frequency residual spectrum, compression block effect and blur kernel parameters are extracted, and a camera imaging degradation representation which can be updated online is formed through robust aggregation; further, a shared appearance encoder and a reversible appearance normalization adapter are used to decompose the vehicle basic feature into identity normalized features and degradation residuals, and the vehicle identity details are maintained through feature reconstruction and cross-degradation re-injection constraint; finally, an identity normalization library and a degradation anchor point library are established, candidate vehicles are recalled according to the identity normalized features, and reordering is completed by fusing identity similarity, re-injection similarity and degradation residual gate score.

Description

Technical Field

[0001] This invention belongs to the fields of intelligent traffic video analysis, computer vision, vehicle re-identification and cross-camera target retrieval technology, and specifically relates to a vehicle re-identification method and system for self-calibration of cross-camera imaging degradation characterization. Background Technology

[0002] Vehicle re-identification refers to the retrieval of the same vehicle from multiple road surveillance cameras with non-overlapping fields of view. It is widely used in traffic operation analysis, vehicle trajectory association, and abnormal vehicle investigation. Existing technologies typically use convolutional neural networks, visual Transformers, or local attention models to extract vehicle appearance features, and can combine attributes such as vehicle color, model, logo, headlights, and windows, or reorder candidate results using camera topology, time intervals, road rules, and driving direction.

[0003] Different cameras vary in installation height, resolution, bitrate, white balance, exposure gain, lighting method, lens aging, and compression encoding method. The same vehicle may exhibit noticeable color shifts, blocky compression noise, motion blur, overexposure, and loss of detail when viewed from different cameras. Directly comparing vehicle depth features may lead the model to misinterpret image degradation as differences in vehicle identity, or it may misinterpret similar degradation patterns observed in different vehicles under the same camera as similar identities.

[0004] Existing domain adaptation and style transfer methods typically treat camera differences as domain biases to be eliminated, or convert the entire vehicle image to the style of the target camera. They lack image degradation variables that can be estimated online, updated independently, and used in consistency verification during the retrieval phase. Whole-image style transfer can also alter fine-grained identity cues such as headlights, stickers, wheel rims, window reflections, and local textures.

[0005] Existing spatiotemporal reordering methods typically rely on fixed camera topology, vehicle trajectories, road directions, or license plate recognition results. In scenarios involving temporary camera access, unknown topology, cross-regional collaborative searches, invisible license plates, or privacy restrictions, these priors are difficult to obtain reliably, and new camera access often requires retraining or extensive annotation. Therefore, it is necessary to explicitly model camera imaging degradation to achieve cross-camera vehicle appearance calibration and reliable retrieval without relying on precise trajectories, fixed topology, and license plate information. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a vehicle re-identification method and system for self-calibration of cross-camera imaging degradation representations. The aim is to estimate online-updateable camera imaging degradation representations from identity-independent regions, perform reversible appearance normalization and cross-degradation re-injection verification in the vehicle feature space, thereby reducing the interference of camera domain bias on vehicle identity features and improving the accuracy, stability, and interpretability of cross-camera vehicle matching under complex imaging conditions.

[0007] The first aspect of the present invention provides a vehicle re-identification method, including multi-camera video data processing, identity-independent region sampling, construction of camera imaging degradation characterization, reversible normalization of vehicle appearance features, cross-degradation exchange training, dual-database index establishment, and dual-database consistency retrieval.

[0008] According to one embodiment of the present invention, the camera imaging degradation characterization consists of color response components, spectral residual components, compression block effect components, and blur kernel components, and is updated online through a time sliding window and a robust aggregator, so that it is adjusted with changes in day and night, weather, bit rate, and exposure conditions, while not fluctuating drastically with changes in a single vehicle identity.

[0009] According to one embodiment of the present invention, the reversible appearance normalization adapter decomposes the basic features of the vehicle into identity normalization features and degradation residuals, and preserves identity details such as headlights, stickers, window outlines, wheel hubs and local textures of the body through residual reconstruction constraints.

[0010] According to one embodiment of the present invention, an identity normalization library is used to quickly recall candidate vehicles, a degradation anchor library is used for cross-degradation re-injection verification and degradation residual gating, and finally the ranking result is output by fusing identity similarity, re-injection similarity and degradation residual gating scores.

[0011] A second aspect of the present invention provides a vehicle re-identification system, including a video access module, an identity-independent region sampling module, a camera imaging degradation characterization module, a shared appearance encoding module, a reversible appearance normalization module, a training module, a dual-database indexing module, and a retrieval output module.

[0012] This invention can explicitly describe and update camera imaging degradation online, reducing the annotation and retraining costs when new cameras are added; it implements reversible normalization in the feature space to avoid style transfer of the whole image destroying vehicle identity details; it balances rapid candidate recall and degradation interpretability verification through identity normalization library and degradation anchor point library; and it does not rely on license plates, precise trajectories or fixed camera topologies, making it suitable for low light, low bit rate, rain and fog, backlight and same color different vehicle scenarios. Attached Figure Description

[0013] Figure 1This is a flowchart of the overall process for self-calibrating vehicle re-identification based on cross-camera imaging degradation characterization, as disclosed in an embodiment of the present invention. Figure 2 This is a schematic diagram of a camera imaging degradation characterization construction module disclosed in an embodiment of the present invention; Figure 3 This is a flowchart of dual-library consistency retrieval and cross-degradation re-injection disclosed in an embodiment of the present invention. Detailed Implementation

[0014] Unless otherwise expressly stated, the term "comprising" and its variations in this specification and claims should be understood to include the listed elements or steps, without excluding other elements or steps. The following embodiments are used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention.

[0015] like Figure 1 As shown in the figure, this embodiment discloses a vehicle re-identification method for self-calibration of cross-camera imaging degradation characterization, which specifically includes the following steps.

[0016] The vehicle re-identification method in this embodiment includes the following steps: S1. Process the collected multi-camera road surveillance videos, including video decoding, vehicle detection, vehicle target cropping, low-quality image removal, camera number recording, and division of training set, query set, and candidate library.

[0017] S2. Construct an identity-independent region sampling module. This module does not directly use the vehicle body region, but extracts the vehicle bounding box edge, road background strip, guardrail sky region, and stable background region after low-confidence target removal from consecutive frames of the same camera, in order to estimate the camera's own imaging degradation.

[0018] S3. Construct a camera imaging degradation characterization module. This module calculates color response statistics, high-frequency residual spectrum, compression block effect, and blur kernel estimation results, which are then used by a robust aggregator to form a low-dimensional degradation characterization for each camera. .

[0019] S4. Construct a shared appearance encoder and a reversible appearance normalization adapter. The shared appearance encoder extracts the vehicle's basic representation. ; Reversible normalized adapter according to and Output identity normalization features and degraded residuals .

[0020] S5. Design a cross-degradation exchange training strategy. Incorporate vehicle identity features... Degradation characteristics of another camera The system combines these features to generate a re-injected representation of the vehicle under the condition of target camera degradation, thereby constraining the separation of identity features and degradation features.

[0021] S6. Establish a standardized identity database and degraded anchor library . Save the de-degenerate identity representation for rapid recall; Save the degradation residuals and camera degradation characterization for degradation consistency verification during candidate reordering.

[0022] S7. Input the vehicle image to query, first through... Recall candidate vehicles, perform cross-degeneracy re-injection verification, and finally fuse identity similarity, re-injection similarity, and degradation residual gating score to output Rank-k vehicle re-identification result.

[0023] In step S1, the data processing procedure includes the following: S1-1. Decode the video stream of each camera into a fixed frame rate image sequence, and record the camera number, timestamp, resolution, bit rate and exposure parameters for each frame; if the above parameters cannot be provided directly by the device, they are estimated by image statistics.

[0024] S1-2. Use a general vehicle detector to obtain the vehicle target box, and expand it according to the vehicle box expansion ratio. Generate an expanded bounding box. The expanded bounding box is used to preserve information on vehicle edge reflections, contour diffusion, and lens degradation, while the main vehicle bounding box is used for subsequent identity representation learning.

[0025] S1-3. Eliminate low-quality images based on sharpness, brightness, occlusion ratio, and detection confidence. For the same camera, retain samples covering daytime, nighttime, rain / fog, low bitrate, backlighting, and fill light variations to ensure that degradation characteristics reflect the main imaging conditions of the camera.

[0026] S1-4. Establish training set, query set and candidate library according to camera non-overlapping retrieval protocol; if there are known vehicle identity labels, they are used for supervised training; if there are no identity labels, high confidence pseudo-labels can be used or only the camera degradation representation can be updated.

[0027] (1) in, The sample quality score. Indicates the detection confidence level. Indicates sharpness, Indicates the occlusion ratio. This indicates a lack of exposure.

[0028] In step S2, the identity-independent region sampling module is constructed as follows: S2-1. Expand the vehicle detection frame to obtain the expanded frame. and will and The difference set is used as the vehicle edge degradation sampling region. This region preserves reflectivity, noise, compression, and blur information near the vehicle body outline, but minimizes vehicle brand, color, and license plate information.

[0029] S2-2. In each frame of the image, select a background strip area consisting of roads, guardrails, sky, road edges and fixed background, and remove pedestrians, non-motorized vehicles and temporary obstructions by judging the inter-frame stability.

[0030] S2-3. Remove the low-confidence vehicle detection frame, the strong occlusion frame, and the overexposed headlight area from the sampling mask to obtain the set of identity-independent regions. .

[0031] S2-4, to Time-sliding sampling is used to enable camera degradation characteristics to be updated online with changes in day / night, weather, and compression bitrate, but without drastic fluctuations with changes in a single vehicle's identity.

[0032] (2) in, Indicates the background strip area. This indicates an area that may contain details about the vehicle's identity.

[0033] In step S3, the definition and calculation process of camera imaging degradation characterization are as follows: Camera image degradation is characterized by four components: color response component, spectral residual component, compression block effect component, and blur kernel component. The color response component describes white balance shift, brightness distribution, and gamma response; the spectral residual component describes high-frequency noise and sharpening artifacts; the compression block effect component describes block boundary energy caused by video coding; and the blur kernel component describes the direction of motion blur or defocus blur.

[0034] (3) (4) (5) (6) in, Indicates a robust aggregator. , and These represent the color mean, variance, and gamma response estimates, respectively. Represents a spectrum histogram. This indicates the compression block effect. This represents the fuzzy kernel parameter.

[0035] In step S4, the structure of the shared appearance encoder and the reversible appearance normalization adapter is as follows: S4-1, Shared Appearance Encoder Receive vehicle cropped image Output basic appearance features The encoder can use a CNN, Transformer, or a visual language pre-trained model as its backbone.

[0036] S4-2, Degradation Condition Normalization Layer based on camera degradation characteristics Generate channel-level scaling and bias parameters, for Conditional calibration is performed on components related to camera imaging.

[0037] S4-3, the dual-branch gating layer decomposes the calibrated features into identity normalization features. and degraded residuals . Used for vehicle identification retrieval. Used to save the degradation interpretation of the image under the current camera.

[0038] S4-4, Residual Reconstruction Layer , and Recombined into reconstructed features To ensure that the standardization process does not damage the vehicle's distinctive features such as headlights, stickers, window outlines, wheels, and body textures.

[0039] (7) (8) (9) In step S5, the training connection relationships of the reversible appearance normalization adapter are as follows: S5-1: The identity branch uses cross-camera samples from the same vehicle as positive samples and samples from different vehicles as negative samples, optimizing through comparative learning and ranking learning. The ability to discern identity.

[0040] S5-2, Degeneracy branches use camera IDs or camera degradation cluster IDs as supervision signals, making It can express the imaging degradation of the current image, and at the same time, it can make the gradient inversion Try not to carry camera category information.

[0041] S5-3, Reconstructing Branch Minimum and The distance between them avoids excessive elimination of vehicle details due to the over-standardization of identity features.

[0042] S5-4, Cross-degradational exchange branch will Degradation characteristics of target camera Combining, generating The requirement is that the samples from the same vehicle should still have features close to those of the target image even under conditions of degradation of the target camera.

[0043] (10) (11) In steps S6 and S7, the dual-database consistency retrieval process is as follows: S6-1, Identity Standardization Library Store the identity normalization features of each vehicle image in the candidate library. And a nearest neighbor index is built to quickly recall the top K candidate vehicles.

[0044] S6-2, Degraded Anchor Point Library Store the degradation residuals of candidate vehicles Basic characteristics and candidate camera degradation characteristics It is used to calculate the degeneracy interpretability during the reordering phase.

[0045] S7-1, Query vehicle q, input into the model to obtain... , and The system first The system recalls a candidate set based on identity similarity. .

[0046] S7-2, For each candidate vehicle ,Will Degradation characteristics of candidate cameras Combine and generate re-injected features for the queried vehicle under candidate camera imaging conditions. .

[0047] S7-3, combining identity similarity, re-injection similarity, and degenerate residual gating scores to obtain the final ranking score. .

[0048] (12) (13) (14) (15) The overall calculation process for the loss function design in step S5 is as follows: (16) (17) (18) (19) (20) in, This indicates a gradient inversion layer, making Insensitive to camera type, making Sensitive to camera degradation. Recommended weight: =0.5, =0.2, =0.5, =1.0, =0.07, =0.3. The above value is an example and can be adjusted based on the actual dataset.

[0049] The second aspect of this invention discloses a vehicle re-identification system, the system structure and deployment method of which are as follows: The system can be deployed as a video access module, an identity-independent sampling module, a degradation characterization module, a reversible normalization module, a dual-database indexing module, and a retrieval output module. Each module can be deployed in an edge computing box, a central server, or a cloud-edge collaborative environment, and maintain a unified data management process through camera IDs, timestamps, and image quality scores.

[0050] The video access module receives video streams from multiple cameras, outputs frame sequences and camera numbers, and uniformly records timestamps, bitstream parameters, and camera metadata, providing basic data for subsequent online updates of degradation characteristics.

[0051] The identity-independent sampling module takes video frames and vehicle detection boxes as input and outputs background stripes, vehicle outer edges, and stable background samples after removing low-confidence vehicles. The module's purpose is to remove as much identity information as possible, such as vehicle brand, color, and license plate, retaining only statistical clues that describe camera image degradation.

[0052] The degradation characterization module takes identity-independent samples as input and outputs camera imaging degradation characterization. This module extracts degradation vectors such as color response, spectral residual, compression block effect, and blur direction, and updates them over time using a robust aggregator.

[0053] The reversible normalization module uses vehicle basic features and camera degradation characteristics As input, output normalized identity features With degraded residuals This module is used to separate vehicle identity information from camera degradation information, while avoiding excessive removal of vehicle details through reconstruction constraints.

[0054] Dual-database index module receives , and Maintain separate identity standardization databases With Degenerate Anchor Library . Used for rapid recall of candidate vehicles. Used to verify the interpretability of degradation of candidate images during the reordering stage.

[0055] The retrieval output module receives the query image and candidate set, and outputs the Rank-k candidate from the same vehicle, its confidence level, and the explainable degradation difference. For low-confidence candidates, the system can return a description of the camera degradation difference to assist in manual review.

[0056] Alternative implementation methods The degradation characterization of camera imaging can be composed of artificial statistical features such as color, spectrum, compression and blur, or it can be learned by a small self-supervised network. As long as the characterization originates from an identity-independent region and is used for subsequent reversible normalization and cross-degradation re-injection verification, it is an equivalent implementation of the present invention.

[0057] The reversible appearance normalization adapter can be implemented using a multilayer perceptron, conditional normalization, low-rank adapter, gated residual module, or lightweight Transformer; as long as it can output identity normalization features and degenerate residuals, and maintain vehicle identity details through reconstruction constraints, the purpose of this invention can be achieved.

[0058] The dual-database index can be implemented using vector databases, inverted indexes, hash indexes, or graph indexes; candidate recall and degradation consistency reordering can be executed serially or in parallel. Each module can be deployed on edge computing boxes, central servers, or cloud-edge collaborative environments.

[0059] The foregoing embodiments are merely illustrative of the technical principles of the present invention. Equivalent substitutions, modifications, or combinations made by those skilled in the art without departing from the spirit and essence of the present invention should all fall within the protection scope of the claims of the present invention.

Claims

1. A vehicle re-identification method for self-calibration of degradation characterization across camera imaging, characterized in that, The method includes: S1: Acquire road surveillance video from multiple cameras, perform video decoding, vehicle detection, vehicle target cropping, and image quality filtering on the road surveillance video, and record the camera identifiers corresponding to the vehicle images; S2: Construct a set of identity-independent regions from the outer edge of the vehicle detection box, the road background strip, and the stable background region after removing the vehicle's identity subject in consecutive video frames from the same camera; S3: Extract color response components, spectral residual components, compression block effect components, and blur kernel components from the identity-independent region set, and obtain the imaging degradation characterization of each camera through robust aggregation; S4: Extract vehicle basic features using a shared appearance encoder, and decompose the vehicle basic features into identity normalization features and degradation residuals using a reversible appearance normalization adapter conditioned on the imaging degradation characterization. S5: Train the shared appearance encoder and the reversible appearance normalization adapter by employing identity discrimination constraints, degradation separation constraints, feature reconstruction constraints, and cross-degradation re-injection constraints; S6: Establish an identity normalization library for storing identity normalization features, and a degradation anchor point library for storing degradation residuals, vehicle basic features, and camera imaging degradation representations; S7: Perform candidate recall on the query vehicle image, and re-rank the candidate vehicles by fusing identity similarity, cross-degradation re-injection similarity and degradation residual gating score, and output the vehicle re-identification result.

2. The vehicle re-identification method according to claim 1, characterized in that, In step S1, the sample quality score is determined based on the vehicle detection confidence, image clarity, occlusion ratio, and underexposure degree, and vehicle images with sample quality scores below a preset threshold are removed; vehicle images covering daytime, nighttime, rain / fog, low bit rate, backlight, and fill light changes are retained for the same camera.

3. The vehicle re-identification method according to claim 1, characterized in that, Step S2 includes: The vehicle detection frame is expanded outward according to a preset expansion ratio, and the difference between the expanded frame and the vehicle detection frame is used as the vehicle edge degradation sampling area. Background strip regions are identified from roads, guardrails, sky, road edges and fixed backgrounds, and pedestrians, non-motorized vehicles and temporary obstructions are removed by inter-frame stability judgment; The identity-independent region set is obtained by removing low-confidence vehicle detection frames, strongly occluded areas, and overexposed headlight areas from the vehicle edge degradation sampling area and the background strip area.

4. The vehicle re-identification method according to claim 1, characterized in that, In step S3, the color response component is used to characterize white balance shift, brightness distribution and gamma response, the spectral residual component is used to characterize high-frequency noise and sharpening artifacts, the compression block effect component is used to characterize the block boundary energy generated by video coding, and the blur kernel component is used to characterize the direction and intensity of motion blur or defocus blur.

5. The vehicle re-identification method according to claim 1, characterized in that, The reversible appearance normalization adapter includes a degradation condition normalization layer, a dual-branch gating layer, and a residual reconstruction layer. The degradation condition normalization layer generates channel-level scaling parameters and bias parameters based on the camera imaging degradation characterization. The dual-branch gating layer outputs identity normalization features and degradation residuals. The residual reconstruction layer reconstructs the vehicle's basic features based on the identity normalization features, the degradation residuals, and the camera imaging degradation characterization.

6. The vehicle re-identification method according to claim 1, characterized in that, Step S5 includes: Use cross-camera samples of the same vehicle as positive samples and samples of different vehicles as negative samples to train the identity branch; The degenerate branch is trained using camera identifiers or camera degenerate clustering identifiers, and the identity normalization features are made insensitive to camera category by gradient inversion. The vehicle identity normalization features are combined with the target camera imaging degradation features to generate cross-degradation re-injection features, and the cross-degradation re-injection features are constrained to be close to the basic vehicle features of the same vehicle under the target camera.

7. The vehicle re-identification method according to claim 1, characterized in that, In step S6, the identity normalization library establishes a nearest neighbor index based on the identity normalization features of the candidate vehicles, which is used to recall the top K candidate vehicles; the degradation anchor point library is used to calculate the degradation interpretability of the candidate vehicles under the corresponding camera imaging conditions during the candidate reordering stage.

8. The vehicle re-identification method according to claim 1, characterized in that, In step S7, the identity normalization features of the queried vehicle and the identity normalization features of the candidate vehicle are used to calculate the identity similarity. The identity normalization features of the queried vehicle and the candidate camera imaging degradation features are combined to obtain re-injection features and the re-injection similarity is calculated. The degradation residual gating score is determined based on the distance between the degradation residual of the candidate vehicle and the predicted degradation residual. The identity similarity, the re-injection similarity and the degradation residual gating score are then weighted and fused.

9. A vehicle re-identification system for self-calibration of degradation characterization across camera imaging, characterized in that, It includes a video access module, an identity-independent region sampling module, a camera imaging degradation characterization module, a shared appearance coding module, a reversible appearance normalization module, a training module, a dual-database indexing module, and a retrieval output module; each module is configured to execute the vehicle re-identification method according to any one of claims 1 to 8.

10. The vehicle re-identification system according to claim 9, characterized in that, The camera imaging degradation characterization module uses a time sliding window to update the imaging degradation characterization of each camera online; the dual-database index module is deployed on edge computing devices, central servers, or cloud-edge collaborative environments, and maintains the identity normalization database and degradation anchor database through camera identifiers, timestamps, and image quality scores.

Images