A vehicle trajectory generation method, medium and device for traffic management

Abstract

The application discloses a vehicle trajectory generation method, medium and equipment for traffic management, aiming to improve the accuracy and robustness of vehicle re-identification. The method discloses a method for traffic management, medium and equipment for traffic management, aiming to improve the accuracy and robustness of vehicle re-identification. The method discloses a method for traffic management, medium and equipment for traffic management, aiming to improve the accuracy and robustness of vehicle re-identification. The method discloses a method for traffic management, medium and equipment for traffic management, aiming to improve the accuracy and robustness of vehicle re-identification. The method discloses a method for traffic management, medium and equipment for traffic management, aiming to improve the accuracy and robustness of vehicle re-identification. The method discloses a method for traffic management, medium and equipment for traffic management, aiming to improve the accuracy and robustness of vehicle re-identification. The method discloses a method for traffic management, medium and equipment for traffic management, aiming to improve the accuracy and robustness of vehicle re-identification. The method discloses a method for traffic management, medium and equipment for traffic management, aiming to improve the accuracy and robustness of vehicle re-identification. The method discloses a method for traffic management, medium and equipment for traffic management, aiming to improve the accuracy and robustness of vehicle re-identification. The method discloses a method for traffic management, medium and equipment for traffic management, aiming to improve the accuracy and robustness of vehicle re-identification. The method discloses a method for traffic management, medium and equipment for traffic management, aiming to improve the accuracy and robustness of vehicle re-identification.

Description

Technical Field

[0001] This invention relates to the field of vehicle trajectory generation, and in particular to a method, medium, and device for generating vehicle trajectories for traffic management. Background Technology

[0002] In intelligent traffic monitoring systems within industrial parks, logistics bases, or closed management areas, cross-camera vehicle re-identification (ReID) is a key technology for vehicle trajectory tracking. Existing solutions typically combine ReID features of the vehicle's overall appearance with OCR recognition results of the license plate number for identity matching. However, in practical applications, due to factors such as long shooting distance, poor lighting conditions, damaged license plates, or tilted angles, license plate images are often blurry, leading to OCR recognition failures or incorrect outputs. Forcibly relying on unreliable OCR information for matching will result in serious misjudgments; while completely discarding license plate area information results in the loss of potential local discriminative clues. Furthermore, when there are many vehicles with highly similar appearances within the park (such as a fleet from the same company), relying solely on ReID features is insufficient to effectively distinguish individuals. Therefore, there is an urgent need for a robust vehicle re-identification method that does not rely on OCR text but can adaptively utilize visual features of the license plate area and integrate other viewpoint-specific local identifiers. Summary of the Invention

[0003] To address one of the aforementioned technical problems, the present invention adopts the following technical solution:

[0004] According to one aspect of the present invention, a method for generating vehicle trajectories for traffic management is provided, comprising the following steps:

[0005] Based on the target vehicle image and the comparison vehicle image, obtain the overall ReID feature similarity SimReID and the local effective comparison region similarity SimPlatei between the target vehicle and the comparison vehicle;

[0006] Based on SimReID and SimPlatei, generate the similarity S between the two vehicles, where S satisfies the following condition:

[0007] S = A × SimReID + Bi × SimPlatei.

[0008] Where A is the overall ReID feature weight, Bi is the weight of the local effective comparison region, and i is the level number of the local effective comparison region, i=1, 2 or 3, and satisfies A+Bi=1;

[0009] The method for determining Bi and the local effective comparison region is as follows:

[0010] Based on the license plate area, define three progressively expanding areas around the license plate area, after removing the license plate characters;

[0011] The image features of three regions in the target vehicle tracking image are extracted and compared with the similarity features of the corresponding standard template image. The standard template image features are the image features of the area surrounding the license plate of the corresponding vehicle model in the standard state.

[0012] If the similarity between the image features of the outer region surrounding a license plate in the target vehicle image and the features of the standard template is lower than a preset difference threshold, then the region is identified as a locally effective comparison region, and Bi is determined; where Bi is inversely proportional to the level number of the locally effective comparison region.

[0013] Based on S, it is determined whether the vehicle in the vehicle image to be compared is the same vehicle as the target vehicle being tracked, and an identity ID corresponding to the vehicle in the vehicle image to be compared is assigned. The vehicle record points of the continuously identified vehicles with the same identity ID are connected to generate the vehicle driving trajectory, which is used for traffic flow monitoring, abnormal behavior early warning and traffic dispatch control in parks, roads or specific areas.

[0014] According to a second aspect of the present invention, a non-transitory computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the above-described method for generating vehicle trajectories for traffic management.

[0015] According to a third aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for generating vehicle trajectories for traffic management.

[0016] This invention has at least one of the following beneficial effects:

[0017] This invention effectively overcomes the problems of low accuracy and frequent misjudgments in traditional OCR methods under adverse conditions such as insufficient lighting, damaged license plates, or tilted angles by integrating the overall ReID feature similarity SimReID with the local effective comparison region similarity SimPlatei. Traditional vehicle re-identification methods rely on the OCR recognition results of the license plate number for matching. However, in practical applications, the success rate of OCR recognition drops significantly due to the aforementioned factors, leading to frequent mismatches or failures to match. The method proposed in this invention abandons the reliance on OCR text information and instead compares the overall ReID feature similarity between the target tracking vehicle image and the image of the vehicle to be identified, and integrates the visual feature similarity based on the extended area around the license plate for comprehensive judgment. Since the area around the license plate may exhibit different image specificities due to damage or bending, and large vehicles often have bumpers or monitoring equipment installed around the license plate, the image in this area has significant distinguishing features. Therefore, this solution achieves more robust vehicle identification by combining overall ReID features and local specific features. This method not only avoids the inherent limitations of OCR technology, but also fully extracts and utilizes the local identification information provided by the license plate area, thereby significantly improving the accuracy and reliability of recognition.

[0018] Meanwhile, to address the potential blurring of the license plate area, this invention defines three progressively expanding regions around the license plate, each with the license plate characters removed. This adaptively utilizes the visual features of these regions to assist in the vehicle re-identification process. Specifically, by extracting the image features of each of these three regions from the target-tracking vehicle image and comparing their similarity with the corresponding standard template image features, localized differences caused by license plate damage, defacement, or modification can be effectively identified. Once a significant difference is detected between the image features of a license plate's peripheral expansion region and the standard template features, that region is identified as a locally valid comparison region and used to calculate the overall similarity S. This method ensures that even when license plate information is unreliable, accurate identity matching can still be achieved through other reliable localized features, further enhancing the system's adaptability and reliability.

[0019] Furthermore, in industrial parks, logistics bases, or closed management areas, there are often a large number of vehicles with highly similar appearances, which poses a challenge to vehicle re-identification relying solely on ReID features. The solution proposed in this invention not only considers the overall ReID features of the vehicle but also introduces the similarity of locally effective comparison regions based on the area extending outward from the license plate. This provides an additional effective means of distinguishing vehicles with similar appearances. For example, in a fleet of vehicles from the same company, although the vehicles may appear extremely similar, the personalized decorations, wear and tear, or unique markings around each vehicle may differ. By taking these subtle differences into account, this invention can more accurately distinguish individual vehicles, thereby improving the effectiveness of vehicle re-identification.

[0020] Furthermore, since the weight Bi is inversely proportional to the level number of the locally effective comparison area, the first-level area, which is closest to the license plate edge and has the smallest coverage, mainly includes highly sensitive details such as the license plate frame, mounting bracket, and adjacent paint. These areas are easily affected by personalized modifications (such as stickers, scratches, stains, bent license plate edges, or differences in brackets), resulting in significant visual differences and strong discriminative power. The second and third-level areas expand outwards, gradually incorporating large, common body parts such as bumpers and hoods. These areas are highly consistent across vehicles of the same model, diluting personalized information and significantly reducing discriminative power. Therefore, smaller areas are more likely to carry discriminative features similar to "individual fingerprints" and should be given higher weights, while larger areas, due to their greater commonality and diluted individual features, have correspondingly lower weights. This design prioritizes the most discriminative, detailed local information, effectively suppressing common interference while improving matching accuracy, significantly enhancing the accuracy and robustness of vehicle re-identification in highly similar scenarios. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart of a vehicle trajectory generation method for traffic management provided in an embodiment of the present invention.

[0023] Figure 2 A flowchart of a vehicle traffic control method based on a queue pool is provided for an embodiment of the present invention. Detailed Implementation

[0024] 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.

[0025] As one possible embodiment of the present invention, such as Figure 1 As shown, a method for generating vehicle trajectories for traffic management is provided, including the following steps:

[0026] S100: Based on the target vehicle image and the comparison vehicle image, obtain the overall ReID feature similarity SimReID and the local effective comparison region similarity SimPlatei between the target vehicle and the comparison vehicle.

[0027] In this invention, the target vehicle being tracked corresponds to the query vehicle in a ReID scenario, while the vehicle to be compared corresponds to the candidate vehicle or reference vehicle. By calculating the similarity between the two, it can be determined whether two images captured at different times or by different cameras belong to the same vehicle.

[0028] Specifically, in this invention, the image information of the target vehicle is typically acquired when the vehicle enters the park. A dedicated registration area is usually set up at the park entrance where vehicles briefly stop, facilitating the system's collection of high-quality data. Simultaneously, multiple cameras are typically deployed at the entrance, providing comprehensive coverage and excellent imaging conditions (such as sufficient lighting, suitable distance, and no obstructions), enabling the stable capture of clear images of the entire vehicle's exterior and high-precision images of the license plate area. The license plate number is then reliably extracted using OCR. Therefore, this invention uses the moment of vehicle entry as the initial identity anchor point, ensuring that the target vehicle information acquired at this stage (including identity ID, ReID features, license plate image, etc.) possesses high accuracy and completeness.

[0029] The vehicle images to be compared are video frames captured by various surveillance cameras on internal roads, parking lots or other areas after the vehicle enters the park.

[0030] By acquiring the overall ReID (Re-Identification) feature similarity SimReID between the target tracking vehicle and the vehicle to be compared, as well as the similarity SimPlatei of the local effective comparison regions, basic data support is provided for subsequent vehicle identity verification.

[0031] First, SimReID calculation relies on a deep learning model to extract global appearance feature vectors from two images. This process considers not only the overall appearance information of the vehicle, such as model and color, but also subtle visual differences. This step can quickly locate potentially matching targets among a large number of vehicles; however, due to the significant appearance similarity between different vehicles, relying solely on SimReID may lead to a high false positive rate. Therefore, additional discriminative information needs to be introduced to improve recognition accuracy.

[0032] Against this backdrop, SimPlatei was proposed to evaluate the similarity of locally effective comparison regions of vehicles, particularly specific areas around the license plate. These regions exhibit high variability and individual differences, providing stronger discriminative power. By extracting features and calculating similarity for these local regions, different vehicles, even those that are very similar in appearance, can be effectively identified without relying on OCR (Optical Character Recognition). Furthermore, a multi-level local comparison mechanism (i.e., i=1, 2, 3) in subsequent steps further enhances the robustness and adaptability of the system.

[0033] S200: Based on SimReID and SimPlatei, generate the similarity S between the two vehicles, where S satisfies the following condition:

[0034] S = A × SimReID + Bi × SimPlatei.

[0035] Where A is the overall ReID feature weight, Bi is the weight of the local effective comparison region, and i is the level number of the local effective comparison region, i=1, 2 or 3, and satisfies A+Bi=1.

[0036] The method for determining Bi and the local effective comparison region is as follows:

[0037] S201: Based on the license plate area, define three progressively expanding areas around the license plate, with the license plate characters removed.

[0038] Specifically, the three progressively expanding areas around the license plate, after removing the license plate characters, can include:

[0039] The first outer expansion area is the smallest area that is right next to the outer edge of the license plate and after the license plate characters have been removed by masking.

[0040] The second extended region is the region formed by extending the first extended region outward by a first preset pixel value.

[0041] The third expansion area is the area formed by expanding outward by a second preset pixel value based on the second expansion area.

[0042] In practical use, these three levels of external areas can be set up as shown in the following example:

[0043] The first outer expansion area covers the license plate mounting frame and the vehicle body paint adjacent to the four edges of the license plate, but does not include the license plate character area, and extends outward by no more than 5 pixels.

[0044] The second extended area further covers the upper edge of the bumper or a partial area of ​​the tailgate, extending outward by a total distance of 6 to 15 pixels, based on the first extended area.

[0045] The third outward expansion area extends beyond the second outward expansion area to include part of the lower edge of the front grille, the hood seam line, or the trunk opening seam area, with a total outward extension distance of 16 to 30 pixels.

[0046] In this process, each outer region excludes the area containing the license plate characters using a masking method, retaining only the visual context information of the non-text surrounding the license plate.

[0047] In industrial parks, logistics bases, or closed management areas, there are often a large number of vehicles with highly similar appearances, which poses a challenge to vehicle re-identification relying solely on ReID features. The solution proposed in this invention not only considers the overall ReID features of the vehicle but also introduces the similarity of locally effective comparison regions based on the area extending outward from the license plate. This provides an additional effective means of distinguishing vehicles with similar appearances. For example, within the same company's fleet, although vehicles may appear extremely similar, each vehicle may differ in its personalized decorations, wear and tear, or unique markings within a small area. By taking these subtle differences into account, this invention can more accurately distinguish individual vehicles, thereby improving the effectiveness of vehicle re-identification.

[0048] Furthermore, since the weight Bi is inversely proportional to the level number of the locally effective comparison area, the first-level area, which is closest to the license plate edge and has the smallest coverage, mainly includes highly sensitive details such as the license plate frame, mounting bracket, and adjacent paint. These areas are easily affected by personalized modifications (such as stickers, scratches, stains, bent license plate edges, or differences in brackets), resulting in significant visual differences and strong discriminative power. The second and third-level areas expand outwards, gradually incorporating large, common body parts such as bumpers and hoods. These areas are highly consistent across vehicles of the same model, diluting personalized information and significantly reducing discriminative power. Therefore, smaller areas are more likely to carry discriminative features similar to "individual fingerprints" and should be given higher weights, while larger areas, due to their greater commonality and diluted individual features, have correspondingly lower weights. This design prioritizes the most discriminative, detailed local information, effectively suppressing common interference while improving matching accuracy, significantly enhancing the accuracy and robustness of vehicle re-identification in highly similar scenarios.

[0049] S202: Extract the image features of each of the three regions in the target vehicle tracking image, and compare their similarity with the corresponding standard template image features. The standard template image features are the image features of the license plate periphery of the corresponding vehicle model in its standard state. Specifically, the standard template image features are the pre-stored image features of the three license plate periphery regions corresponding to the target vehicle model in its unmodified state.

[0050] S203: If the similarity between the image features of the area surrounding a license plate in the target vehicle image and the features of the standard template is lower than a preset difference threshold, then this area is determined as a locally effective comparison area, and Bi is determined. Bi is inversely proportional to the level number of the locally effective comparison area. For example, this inverse relationship can specifically satisfy:

[0051] When i=1, Bi=0.6. When i=2, Bi=0.3. When i=3, Bi=0.1.

[0052] S203 includes:

[0053] S213: When determining the local effective comparison area, starting from the first outer expansion area, the image features of the corresponding area in the target tracking vehicle image are compared with the features of the standard template image.

[0054] S223: If the similarity of a certain extended region is lower than the preset difference threshold, then the region is determined as a local effective comparison region, and the comparison of larger extended regions is stopped.

[0055] S233: If the similarity of the three outer regions is higher than the preset difference threshold, it is determined that there is no personalized difference around the license plate, and Bi is set to zero.

[0056] S300: Based on S, determine whether the vehicle in the vehicle image to be compared is the same vehicle as the target vehicle being tracked, and assign the corresponding identity ID to the vehicle in the vehicle image to be compared for use in vehicle trajectory drawing.

[0057] This invention aims to address the problem of OCR technology's difficulty in recognizing license plates due to blurriness, damage, or poor shooting angles, and proposes a vision-based, rather than text-based, comparison mechanism. This mechanism works by defining three progressively expanding peripheral regions that do not include license plate characters (the first layer ≤ 5 pixels, covering the border and adjacent paint; the second layer 6–15 pixels, covering the bumper or tailgate; the third layer 16–30 pixels, extending to the grille or hood seam), thus preserving only the vehicle's non-textual visual information.

[0058] The similarity of features of each region of the target vehicle is compared with that of a standard template, starting with the most refined first region and proceeding step by step. If a region has a significant difference from the template (below a preset threshold), it is marked as a "locally valid comparison region" and further searching is stopped; if all three regions are highly consistent with the template, local features are not used for comparison (Bi=0).

[0059] The weights Bi are strictly inversely proportional to the selected region level: the first layer, because it contains personalized details (such as stickers, scratches, bracket deformation, etc.), has a high degree of differentiation and is given a high weight (e.g., B1=0.6); the second layer has a lower weight due to increased commonality (e.g., B2=0.3); and the third layer, because it is mainly composed of regions with greater commonality, has the lowest weight (e.g., B3=0.1).

[0060] Furthermore, since the weight Bi is inversely proportional to the level number of the locally effective comparison area, the first level area, which is closest to the license plate edge and has the smallest coverage, mainly includes highly sensitive details such as the license plate frame, mounting bracket, and adjacent paint surface. It is highly susceptible to significant visual differences due to personalized modifications (such as stickers, scratches, stains, bent license plate edges, or differences in brackets), exhibiting strong distinguishability and high specificity, and is therefore assigned a higher weight (e.g., B1=0.6). The second and third level areas expand outwards, gradually incorporating large, common body parts such as bumpers and hoods. These areas are highly consistent across vehicles of the same model, diluting personalized information and significantly reducing distinguishability; their weights gradually decrease (e.g., B2=0.3, B3=0.1). Therefore, smaller areas are more likely to carry discriminative features similar to "individual fingerprints" and should be assigned a higher weight, while larger areas, due to their greater commonality and diluted individual characteristics, have correspondingly lower weights. This design prioritizes the most discriminative and detailed local information, effectively suppressing common interference while improving matching accuracy, and significantly enhancing the accuracy and robustness of vehicle re-identification in highly similar scenarios.

[0061] The two dynamic processes of region selection and weight assignment are interdependent: the selection of a region determines the value of Bi, while the setting of Bi reflects the discriminative value of that region. The resulting comprehensive similarity S can fully utilize global appearance consistency (through SimReID) and intelligently focus on the most individual-specific local visual cues (through SimPlatei and Bi), thereby accurately identifying individuals among a large number of vehicles of the same type. This mechanism not only avoids reliance on OCR but also significantly improves the system's robustness, accuracy, and adaptability in scenarios with damaged license plates, modifications, or high similarity.

[0062] As another possible embodiment of the present invention, multiple cameras are typically deployed at different locations within the park, with each camera having a different installation height and shooting angle. In this embodiment, each camera is pre-configured with a list of observable vehicle parts and additional object types that it supports for recognition.

[0063] This embodiment first performs prior configuration for each camera: based on its physical location and imaging characteristics, it predefines its "observable vehicle parts" (such as the windshield area, door side windows, trunk, roof, etc.) and the "list of additional item types that can be identified for that part" (for example: the windshield area can identify "company logo stickers", "safety stickers", "water cups", the side windows can identify "contact information stickers", and the visible areas inside the vehicle can identify "ornaments", "hanging ornaments", etc.).

[0064] S400: When any vehicle enters the optimal shooting area of ​​a camera, detect whether there are any items in the additional item type list in the observable parts of the vehicle from the corresponding video frame.

[0065] S500: If it exists, extract the type identifier and its local feature vector of the additional item, and bind and store them with the vehicle to form the additional item corresponding to the vehicle.

[0066] The additional item feature library is dynamically updated during vehicle operation:

[0067] S501: When a subsequent camera detects a new type of additional item on a vehicle with the same identity ID, the type identifier and local feature vector of that item are added to the vehicle's additional item feature library.

[0068] S502: In subsequent vehicle comparisons, SimExtraItem calculations are performed based on the updated additional item feature library to improve the accuracy of vehicle identity matching.

[0069] Specifically, additional items include company logos, contact stickers, safety labels, or car interior decorations. Local feature vectors are extracted using a convolutional neural network model.

[0070] Specifically, the technical solutions defined by S400 and S500 are illustrated with the following examples:

[0071] When any vehicle enters the optimal shooting area of ​​a camera (i.e., an area with high image clarity, complete target, and no severe obstruction), S400 is executed:

[0072] Using object detection models (such as YOLO or Faster R-CNN), the preset observable parts of the relevant image frames are located in the corresponding video frames; within these parts, it is further detected whether there are objects belonging to the "additional item type list".

[0073] If a valid item is detected (S500), then:

[0074] Extract its type identifier (such as "Company A logo" or "red safety sticker"); and extract its local feature vector through a lightweight convolutional neural network (such as MoBileNet or ResNet-18); bind the item information with the current vehicle's identity ID and store it in the vehicle's exclusive additional item feature library.

[0075] In addition, this feature library supports a dynamic update mechanism (S501–S502):

[0076] When other cameras subsequently observe the same vehicle ID and discover new types of additional items (such as ornaments not seen at the entrance but captured by an internal overhead camera), the type and feature vector of the new items are automatically added to the vehicle's feature library. This allows for subsequent cross-camera comparisons where, if the main ReID and features around the license plate are insufficient to distinguish highly similar vehicles, this expanded feature library can be used to calculate the local feature similarity between additional items of the same type, serving as a supplementary criterion for identity verification.

[0077] S600: When generating similarity S, if Bi=0, then the local feature similarity of additional items is introduced, and S satisfies the following condition:

[0078] S=A×SimReID+C×SimExtraItem.

[0079] Where A+C=1, SimExtraItem is the local feature similarity of the same type of extra item in the vehicle image to be compared and the target tracking vehicle image, and C is the local feature similarity weight of the extra item. Specifically, C can be set to a weight value less than 0.5, so that the size of SimExtraItem will affect S to a small extent, thus making S different due to SimExtraItem. The specific value of C can be adaptively adjusted by those skilled in the art according to actual usage needs.

[0080] If multiple additional items of the same type are detected in the vehicle image to be compared and the target tracking vehicle image, the local feature similarity of each additional item is weighted and averaged or the maximum value is taken to generate the final SimExtraItem.

[0081] Step S600 is the key step in realizing multi-source feature fusion matching in this embodiment. Its core is that when the area around the license plate cannot provide effective discrimination information (i.e., Bi=0), the local feature similarity of additional items is automatically introduced as an alternative criterion to maintain high-precision identity matching capability.

[0082] The triggering conditions and switching logic of S600 are as follows: First, S100–S203 are executed to determine whether there is a valid local comparison area. If the surrounding areas of all Level 3 license plates are highly consistent with the standard template (i.e., there is no individual difference), then Bi=0 is determined, indicating that the area around the license plate does not have the ability to distinguish. At this time, the system dynamically switches to the additional item-assisted matching mode, and the calculation logic of S600 is enabled to avoid misjudgment caused by relying solely on ReID features due to the lack of local features.

[0083] In this embodiment, the generation mechanism for the additional item similarity SimExtraItem is as follows:

[0084] The first step is type alignment matching, which extracts all registered item records from the additional item feature libraries of both the target tracking vehicle and the vehicle to be compared. Only item pairs of the same type are retained; items of different types are not included in the comparison to ensure semantic consistency.

[0085] Then there is the aggregation strategy in the case of multiple items. If multiple instances of both vehicles are detected under a certain type (for example, there are two "safety stickers" in different locations inside the vehicle), then the local feature similarity (usually cosine similarity) is calculated for each pair of items of the same type.

[0086] The final SimExtraItem is aggregated in one of the following ways:

[0087] Weighted average: Assign different weights to items in an image based on their sharpness, area, or confidence level, and then calculate a weighted average.

[0088] Take the maximum value: Select the pair of items with the most discriminative similarity as the representative. This method is suitable for key identifiers (such as unique company logos). If this method is used, the specific category of the identifier needs to be set in advance so that it can be directly called.

[0089] In this embodiment, the formula for the overall similarity S is updated to: S = A × SimReID + C × SimExtraItem.

[0090] Where A+C=1, the specific C can be dynamically adjusted according to the reliability of the additional items (for example: for high-confidence additional item types, such as a unique company logo or contact number identifier, C=0.4; for low-confidence additional item types, such as a water cup placed in the driver's cab, C=0.2).

[0091] This embodiment focuses on improving matching accuracy when, during vehicle re-identification, the area surrounding the license plate cannot provide effective discrimination information, i.e., when the technical solution (S100-S300) in the previous embodiment cannot be effectively implemented. It involves calculating the total similarity S by introducing the local feature similarity of additional items, thereby enhancing the overall similarity S. Steps S400 to S600 describe this process in detail, and its technical effect is a significant improvement in adaptability to complex environments and special situations. Specifically, when the traditionally relied-upon ReID features or license plate personalized features are unavailable (i.e., Bi=0), this method can automatically switch to using additional items carried on the vehicle for auxiliary judgment. These additional items, such as interior decorations and vehicle stickers, provide additional "fingerprints" for the vehicle due to their strong subjectivity and uniqueness. By performing similarity analysis on the local features of these additional items and combining them with the original ReID features, a multi-level, adaptive discrimination system is formed.

[0092] Compared to the technical solution in the previous embodiment, the two complement each other, constructing a more comprehensive vehicle re-identification solution. The first example focuses on extracting personalized local features from the license plate and its surrounding area to enhance the distinguishability between individual vehicles. The second embodiment is a supplement and extension based on this, especially suitable for scenarios where license plate information is limited or unclear. By incorporating the local features of additional objects into the consideration, not only are the potential discrimination gaps in the first invention point filled, but the robustness and accuracy of the entire system are further improved. In addition, this multi-source feature fusion approach enables the system to maintain high recognition performance even in extreme situations (such as no OCR information, license plate obscured, etc.), realizing the transformation from single feature dependence to multi-feature collaborative work, greatly improving the reliability and efficiency of the vehicle re-identification system in practical applications.

[0093] As another possible embodiment of the present invention, such as Figure 2 As shown, a vehicle traffic control method based on a queue pool is also provided. By dynamically maintaining the queue pool and using a hierarchical comparison mechanism, efficient vehicle re-identification and trajectory generation are achieved. This not only solves the problems of high computational complexity and high resource consumption in existing technologies, but also significantly improves the robustness and accuracy of the system in high-similarity environments, providing solid technical support for intelligent traffic management in parks.

[0094] Specifically, the method includes the following steps:

[0095] A100: Vehicle image information acquired through cameras at the park entrances and exits is used to construct and dynamically maintain a queue pool. The queue pool contains multiple queues, each aggregating vehicles with similar appearances within the park and storing the corresponding vehicle's identity ID, ReID features, and license plate area image. Specifically, the vehicle's identity ID is its license plate number.

[0096] Specifically, the process of building a queue pool includes:

[0097] A101: When any vehicle enters through the entrance camera, extract its ReID features f. ReID and identity ID.

[0098] A102: Traverse the average ReID characteristics of all existing queues in the current queue pool.

[0099] If there exists a queue that satisfies:

[0100] Sim(f ReID f ReID，AVG )≥T cluster Then, the vehicle's identity ID, ReID features, and license plate area image are added to the queue, and the average ReID features of the queue are updated. Where f ReID，AVG This represents the average ReID feature of any existing queue in the current queue pool. The average ReID feature of the queues is updated using either a moving average or incremental updates.

[0101] Specifically, when a vehicle enters the park, its ReID features are compared with the average ReID features of each queue to determine whether it belongs to a known appearance group. If the similarity reaches or exceeds a preset clustering threshold T, the clustering process is initiated. cluster This means that if a vehicle has a sufficiently consistent overall appearance with the members in the queue, it should be included in that queue.

[0102] In this embodiment, each queue aggregates several vehicles with similar appearances within the park, and an average ReID feature is used to represent the overall appearance of the entire queue. When a new vehicle joins or an old vehicle leaves, the queue members change. Typically, in some large parks, the number of vehicles in a queue may be dozens or even hundreds. If the arithmetic mean of all members is recalculated each time, the computational load will be large, with a time complexity of O(N), which is significant in high-concurrency or large fleet scenarios. Therefore, this embodiment uses a low-latency, low-memory, and approximately accurate update strategy of moving average and incremental update.

[0103] Incremental update

[0104] The system maintains two auxiliary variables for each queue: the current number of vehicles n and the sum of ReID features Sn.

[0105] When a new vehicle is added, the number of vehicles is updated synchronously from n to n+1, and the sum of ReID features changes from Sn to Sn+1 = Sn+f. new .

[0106] Updated average ReID feature f n+1 Calculate using the following formula: fReID，AVG =(Sn+f new f / (n+1). new ReID features for newly added vehicles.

[0107] When a vehicle leaves, its features are subtracted from the current Sn, and the current n is updated to n−1, and the average value is recalculated.

[0108] Moving average

[0109] The number of queue members is not recorded; only the current average feature f is maintained. ReID，AVG .

[0110] Whenever a new car enters (regardless of whether an old car is removed), update directly using the exponential smoothing formula:

[0111] f ReID，AVG =α×f new +(1﹣α)×f old , α∈(0,0.1). f old The average ReID feature of the current queue when a new vehicle is added (i.e., the average feature of previous time steps).

[0112] In this method, the influence of old vehicles in the queue decays naturally over time without the need for explicit deletion, making it particularly suitable for scenarios where vehicles frequently enter and exit and it is difficult to accurately track their departure status.

[0113] Both of the above-mentioned methods for updating the average ReID features of the queue (incremental update and moving average) have a time complexity of O(1). Among them, incremental update is mathematically strictly equivalent to full recalculation, which can accurately reflect the true mean of the features of all vehicles in the current queue; although the moving average is an approximate method, it effectively tracks the dynamic change trend of the queue appearance through the exponential smoothing mechanism. Together, they ensure that the queue representation always closely follows the actual distribution of the current vehicle set.

[0114] The update strategy employed in this embodiment requires only constant time to complete feature maintenance, achieving a leap in computational efficiency from O(N) to O(1), significantly reducing CPU load, and supporting real-time response capabilities under high concurrency. Crucially, the average ReID feature of the queue is the sole basis for selecting candidate queues in subsequent steps A300–A400. If this feature deviates from its true distribution due to update lag or distortion, it can easily lead to missed screenings (recall rate decrease) or incorrect screenings (accuracy loss). Through the aforementioned dynamic update mechanism, the timeliness and representativeness of the queue representation can be continuously maintained, thereby effectively ensuring high recall and high accuracy in the primary coarse screening stage.

[0115] A103: If f ReID The similarity with all queues is less than T.cluster If so, a new queue is created, and the vehicle is made the first member of the new queue.

[0116] This situation typically occurs when a newly arriving vehicle is dissimilar to any vehicle in the existing queue, indicating that it may be a new model or combination of appearance features, requiring a separate new queue for management.

[0117] The dynamic update process of the queue pool includes:

[0118] A104: When any vehicle leaves the exit camera, remove the vehicle's relevant information from its queue.

[0119] A105: If the number of vehicles in the queue reaches zero, then delete the queue.

[0120] When removing vehicle information, the average ReID characteristics of the corresponding queues need to be updated synchronously to maintain an accurate representation of queue characteristics. As vehicles enter and exit, some queues may gradually become empty. Timely clearing of these empty queues not only saves storage space but also improves the efficiency of subsequent comparisons.

[0121] A200: When any camera inside the park captures an image of a vehicle to be compared, extract the ReID features, license plate area image, image acquisition time, and image acquisition location of the vehicle to be compared.

[0122] ReID features: Global appearance embedding vectors are extracted through pre-trained vehicle re-identification models (such as OSNet or ResNet-50) to characterize the overall visual features of the vehicle.

[0123] License plate area image: The license plate detection module (such as YOLOv5+CRNN) is used to accurately crop the area where the license plate is located, providing input for subsequent OCR or local image comparison;

[0124] Image acquisition time: Records frame timestamps for trajectory time-series sorting;

[0125] Image acquisition location: Associated with the geographic coordinates or logical number of the camera, used to construct a spatial trajectory.

[0126] This step provides structured input for subsequent matching and trajectory generation, and is a key bridge connecting the perception layer and the decision-making layer.

[0127] A300: Calculate the similarity between the ReID features of the current vehicle to be compared and the average ReID features of each queue.

[0128] A400: If the similarity between the ReID features of the current vehicle to be compared and the average ReID features of a certain queue is greater than a preset threshold, then the queue will be used as a candidate queue.

[0129] The similarity (usually cosine similarity) between the ReID features of the current vehicle and the average ReID features of all queues is calculated one by one to determine the candidate queues. This scheme is a first-level coarse screening based on a queue pool. By utilizing the dynamically maintained queue pool in A100, the original O(N) problem of comparing hundreds of vehicles in the park is simplified to comparing with K queues (K «N), significantly improving efficiency. In addition, the average features of the queues are maintained with high representativeness through moving average or incremental updates, ensuring that the coarse screening results have both high recall (not missing true matches) and low redundancy (avoiding the introduction of irrelevant queues).

[0130] A500: Calculate the similarity between the license plate area image of the current vehicle to be compared and the license plate area image of each individual vehicle in the candidate queue to generate a unique identity ID for the current vehicle to be compared.

[0131] Within the candidate queue, vehicles in the same queue exhibit high similarity in appearance features. Therefore, when determining subsequent queues, the method no longer relies on global ReID features but instead focuses on fine-grained comparison of license plate region images, resulting in higher comparison accuracy. The specific operation involves two complementary implementation paths:

[0132] One approach is to identify vehicle license plate numbers using OCR and then compare them to determine the vehicle's unique ID. This method can be used when OCR recognition of the license plate area image yields a clear and reliable license plate number.

[0133] A511: Perform OCR recognition on the license plate area image of the vehicle to be compared to obtain the first license plate number.

[0134] A512: Iterate through each vehicle in the candidate queue and obtain its second license plate number recorded by the entrance camera when it enters the park.

[0135] A513: Perform string comparison between the first license plate number and each of the second license plate numbers.

[0136] A514: If a vehicle's second license plate number is exactly the same as its first license plate number, then assign the vehicle's unique ID to the vehicle being compared.

[0137] The second method is a non-OCR license plate recognition method. Specifically, it determines the unique ID of a vehicle by comparing the similarity of image features of local areas of the license plates (excluding the license plate number area) between two vehicles. This method is typically used when OCR cannot accurately identify the license plate number, or when its identification results are unreliable, or when a match cannot be found through license plate number comparison. Specifically, the non-OCR license plate recognition method in this part is the same as steps S100 to S300 disclosed in the above embodiment (a vehicle trajectory generation method for traffic management), and can be implemented with reference to the disclosed content in the above steps. In this embodiment, a vehicle in the candidate queue is the target tracking vehicle.

[0138] A521: If there is no completely matching license plate number, then obtain the overall ReID feature similarity SimReID between each individual vehicle in the candidate queue and the current vehicle to be compared, as well as the similarity SimPlatei of the local effective comparison regions.

[0139] A531: Based on SimReID and SimPlatei, generate the similarity S between two vehicles, where S satisfies the following condition:

[0140] S = A × SimReID + Bi × SimPlatei.

[0141] Where A is the overall ReID feature weight, Bi is the weight of the local effective comparison region, and i is the level number of the local effective comparison region, i=1, 2 or 3, and satisfies A+Bi=1.

[0142] The method for determining Bi and the local effective comparison region is as follows:

[0143] Based on the license plate area, three progressively expanding areas are defined around the license plate, with the license plate characters removed.

[0144] The image features of three regions in the license plate area image of each vehicle in the candidate queue are extracted and compared with the similarity features of the corresponding standard template image. The standard template image features are the image features of the outer region around the license plate of the corresponding vehicle model in a standard state.

[0145] A541: If the image features of the area surrounding the license plate of a vehicle in the candidate queue have a similarity to the standard template features of the corresponding vehicle model that is lower than a preset difference threshold, then this area is identified as a locally effective comparison area, and a corresponding weight Bi is determined for that vehicle in the candidate queue accordingly. Bi is inversely proportional to the level number of the locally effective comparison area.

[0146] A551: Based on S, determine whether the vehicle in the current vehicle image to be compared is the same vehicle as a certain vehicle in the candidate queue, and assign the corresponding identity ID to the vehicle in the current vehicle image to be compared.

[0147] A600: Based on the chronological order of image acquisition time, multiple image acquisition locations corresponding to vehicles with the same identity ID are arranged to generate the vehicle's running trajectory for traffic control operations such as park traffic flow monitoring, parking management, or emergency response.

[0148] This embodiment acquires vehicle image information through cameras at the park entrances and exits, constructs and dynamically maintains a queue pool based on appearance similarity, and each queue stores the vehicle's identity ID, ReID features, and license plate area image. Compared to traditional solutions that require comparing each vehicle one by one, leading to a linear increase in computational complexity, this embodiment only needs to compare the ReID features of the vehicle to be compared with the average ReID features of each queue, significantly reducing computational complexity and improving recognition accuracy by utilizing high-quality prior information (such as license plate numbers).

[0149] Furthermore, this invention employs a method of calculating the similarity between the ReID features of the vehicle to be compared and the average ReID features of each queue, further optimizing matching efficiency and accuracy. When the similarity exceeds a preset threshold, the corresponding queue is considered a candidate queue. This approach not only quickly locates potential matching objects but also effectively avoids false matches, especially when dealing with vehicles that are highly similar in appearance. For example, in logistics bases or closed management areas, there are often a large number of transport vehicles of the same brand and model, which traditional methods can easily confuse in such situations. However, this invention, through a hierarchical comparison mechanism, first filters out queues with similar appearances and then performs detailed comparisons on the vehicles within these queues, ensuring accurate identification of each vehicle even in environments with high similarity.

[0150] Next, the similarity between the license plate area image of the vehicle to be compared and the license plate area image of each individual vehicle in the candidate queue is calculated to generate a unique identity ID. This step greatly improves the accuracy of identity verification. In practical applications, the license plate is one of the most direct and unique identifiers of a vehicle, and its image comparison results have extremely high reference value. By combining ReID features with license plate image verification, the system's adaptability to images captured from different angles and under different lighting conditions is enhanced, and it can also resist attempts to tamper with license plate information to a certain extent, thereby ensuring the security and reliability of the entire traffic control system.

[0151] Finally, multiple image acquisition locations corresponding to the same vehicle ID are arranged according to the order of image acquisition time to generate the vehicle's trajectory. This method provides solid data support for park traffic flow monitoring, parking management, and emergency response. Through effective analysis of vehicle trajectories, managers can grasp the traffic situation within the park in real time, rationally allocate resources, optimize parking arrangements, and quickly respond to emergencies. This not only improves the overall operational efficiency of the park but also brings a more convenient service experience to users. At the same time, this technical solution solves the problems of high computational complexity, high resource consumption, and the difficulty in balancing efficiency and accuracy mentioned in the background technology, achieving efficient and robust intelligent traffic management.

[0152] Furthermore, although the steps of the method in this disclosure are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps may be omitted, multiple steps may be combined into one step, and / or a step may be broken down into multiple steps.

[0153] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, mobile terminal, or network device, etc.) to execute the methods according to the embodiments of this disclosure.

[0154] In an exemplary embodiment of this disclosure, an electronic device capable of implementing the above-described method is also provided.

[0155] Those skilled in the art will understand that various aspects of the present invention can be implemented as systems, methods, or program products. Therefore, various aspects of the present invention can be specifically implemented in the following forms: entirely in hardware, entirely in software (including firmware, microcode, etc.), or in a combination of hardware and software, collectively referred to herein as “circuit,” “module,” or “system.”

[0156] An electronic device according to this embodiment of the invention. The electronic device is merely an example and should not be construed as limiting the functionality or scope of the embodiments of the invention.

[0157] Electronic devices are manifested in the form of general-purpose computing devices. The components of an electronic device may include, but are not limited to: at least one processor, at least one memory, and a bus connecting different system components (including memory and processor).

[0158] The memory stores program code that can be executed by a processor, causing the processor to perform the steps described in the "Exemplary Methods" section above, according to various exemplary embodiments of the present invention.

[0159] The storage may include readable media in the form of volatile storage, such as random access memory (RAM) and / or cache memory, and may further include read-only memory (ROM).

[0160] The storage may also include programs / utilities having a set (at least one) of program modules, including but not limited to: an operating system, one or more applications, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0161] A bus can represent one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus that uses any of the various bus architectures.

[0162] The electronic device can also communicate with one or more external devices (e.g., keyboards, pointing devices, Bluetooth devices, etc.), one or more devices that enable a user to interact with the electronic device, and / or any device that enables the electronic device to communicate with one or more other computing devices (e.g., routers, modems, etc.). This communication can be performed via input / output (I / O) interfaces. Furthermore, the electronic device can communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via a network adapter. The network adapter communicates with other modules of the electronic device via a bus. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with the electronic device, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0163] In exemplary embodiments of this disclosure, a computer-readable storage medium is also provided, on which a program product capable of implementing the methods described above is stored. In some possible embodiments, various aspects of the present invention may also be implemented as a program product comprising program code that, when the program product is run on a terminal device, causes the terminal device to perform the steps of the various exemplary embodiments of the present invention described in the "Exemplary Methods" section above.

[0164] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0165] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting programs for use by or in conjunction with an instruction execution system, apparatus, or device.

[0166] The program code contained on the readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0167] Program code for performing the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0168] Furthermore, the accompanying drawings are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention and are not intended to be limiting. It is readily understood that the processes shown in the above drawings do not indicate or limit the temporal order of these processes. Additionally, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0169] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0170] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for generating vehicle trajectories for traffic management, characterized in that, Includes the following steps: Based on the target vehicle image and the comparison vehicle image, obtain the overall ReID feature similarity SimReID and the local effective comparison region similarity SimPlatei between the target vehicle and the comparison vehicle; Based on SimReID and SimPlatei, generate the similarity S between the two vehicles, where S satisfies the following condition: S = A × SimReID + Bi × SimPlatei; Where A is the overall ReID feature weight, Bi is the weight of the local effective comparison region, and i is the level number of the local effective comparison region, i=1, 2 or 3, and satisfies A+Bi=1; The method for determining Bi and the local effective comparison region is as follows: Based on the license plate area, define three progressively expanding areas around the license plate area, after removing the license plate characters; The image features of three regions in the target vehicle image are extracted and compared with the corresponding standard template image features. The standard template image features are the image features of the license plate surrounding area of ​​the corresponding vehicle model in the standard state. The standard template image features are the pre-stored image features of the three license plate surrounding areas of the target vehicle model in the unmodified state. If the similarity between the image features of the outer region surrounding a license plate in the target vehicle image and the features of the standard template is lower than a preset difference threshold, then the region is determined as a locally effective comparison region, and Bi is determined; wherein Bi is inversely proportional to the level number of the locally effective comparison region. Based on S, it is determined whether the vehicle in the vehicle image to be compared is the same vehicle as the target tracking vehicle, and an identity ID corresponding to the vehicle in the vehicle image to be compared is assigned. The vehicle record points of the same identity ID that are continuously identified are connected to generate the vehicle driving trajectory, which is used for traffic flow monitoring, abnormal behavior early warning and traffic dispatch control in parks, roads or specific areas.

2. The vehicle trajectory generation method as described in claim 1, characterized in that, The three progressively expanding areas around the license plate, after removing the license plate characters, include: The first outer expansion area is the smallest area that is right next to the outer edge of the license plate and after the license plate characters have been removed by masking. The second extended region is the region formed by extending the first extended region outward by a first preset pixel value; The third extended region is the region formed by extending the second extended region outward by a second preset pixel value.

3. The vehicle trajectory generation method as described in claim 2, characterized in that, When determining the effective local comparison area, starting from the first outward expansion area, the image features of the corresponding area in the target tracking vehicle image are compared with the features of the standard template image. If the similarity of a certain extended region is lower than the preset difference threshold, then the region is determined as a locally valid comparison region, and the comparison of larger extended regions is stopped; If the similarity of the three outer regions is higher than the preset difference threshold, it is determined that there is no personalized difference around the license plate, and Bi is set to zero.

4. The vehicle trajectory generation method as described in claim 3, characterized in that, Bi is inversely proportional to the level number i of the locally effective comparison region, specifically satisfying the following: When i=1, Bi=0.6; when i=2, Bi=0.3; when i=3, Bi=0.

1.

5. The vehicle trajectory generation method according to any one of claims 1 to 4, characterized in that, Multiple cameras are deployed in different locations within the park, each with a different installation height and shooting angle; Pre-configure each camera with a list of observable vehicle parts and additional object types it supports for recognition; When any vehicle enters the optimal shooting area of ​​a certain camera, the presence of an item from the list of additional item types is detected in the observable parts of the vehicle from the corresponding video frame. If it exists, extract the type identifier and its local feature vector of the additional item, and bind and store them with the vehicle.

6. The vehicle trajectory generation method as described in claim 5, characterized in that, When generating the similarity S, if Bi=0, then the local feature similarity of the additional items is introduced, and S satisfies the following condition: S=A×SimReID+C×SimExtraItem; Where A+C=1, SimExtraItem is the local feature similarity of the same type of extra item in the vehicle image to be compared and the target tracking vehicle image, and C is the local feature similarity weight of the extra item.

7. The vehicle trajectory generation method as described in claim 6, characterized in that, The additional items include company logos, contact information stickers, safety stickers, or vehicle interior decorations; The local feature vectors are extracted using a convolutional neural network model; If multiple additional items of the same type are detected in the vehicle image to be compared and the target tracking vehicle image, the local feature similarity of each additional item is weighted and averaged or the maximum value is taken to generate the final SimExtraItem.

8. The vehicle trajectory generation method as described in claim 7, characterized in that, The additional item feature library is dynamically updated during vehicle operation: When subsequent cameras detect new additional item types on vehicles with the same identity ID, the type identifier and local feature vector of that item are added to the vehicle's additional item feature database. In subsequent vehicle comparisons, SimExtraItem calculations are performed based on the updated additional item feature library to improve the accuracy of vehicle identity matching.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores a computer program that, when executed by a processor, implements a vehicle trajectory generation method for traffic management as described in any one of claims 1 to 8.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a vehicle trajectory generation method for traffic management as described in any one of claims 1 to 8.

Images