Intelligent sensing method and device for traffic transfer data and medium
By combining the YoLov8 and DeepSort models, along with an image quality assessment and self-supervised learning pedestrian re-identification model, the lag and scenario adaptability issues of existing traffic transfer data perception methods are resolved, enabling real-time, accurate transfer passenger statistics and low-cost data acquisition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-10
- Publication Date
- 2026-03-10
AI Technical Summary
Existing methods for sensing traffic transfer data are characterized by strong lag, poor adaptability to different scenarios, and low matching accuracy. They cannot effectively cope with interference in complex scenarios and cannot meet the needs of refined operation and management of urban traffic.
By combining the YoLov8 model with the DeepSort tracking model, and integrating image quality assessment and self-supervised learning pedestrian re-identification models, we can identify transfer targets through monitoring video streams, extract feature data, and perform matching based on spatiotemporal constraints to achieve non-intrusive transfer passenger statistics.
It enables real-time and accurate passenger transfer statistics, reduces statistical latency, improves matching accuracy, enhances scenario adaptability and generalization ability, avoids privacy leaks, and reduces data acquisition costs.
Smart Images
Figure CN121640385A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent transportation sensing technology, and in particular to an intelligent sensing method, device and medium for traffic transfer data. Background Technology
[0002] With the acceleration of urbanization and the development of multi-dimensional public transportation networks, buses and subways, as the core carriers of urban transportation, directly impact residents' travel experience and the rationality of transportation resource allocation through their transfer efficiency. Accurate and real-time statistics on bus and subway transfer passengers are not only the core basis for transportation planning departments to optimize route layout and adjust station facilities, but also a key support for operators to achieve peak-hour capacity scheduling and emergency response, which is of great significance for alleviating traffic congestion and improving travel efficiency.
[0003] Currently, the statistics of passengers transferring between buses and subways mainly rely on two traditional methods: one is based on ticketing data, which obtains transfer information through passengers' active actions such as swiping cards and scanning codes. However, this method has significant drawbacks: First, the data is highly delayed, requiring passengers to complete their transfer payments before data can be aggregated, making it impossible to provide operators with real-time capacity scheduling information; second, cross-system data sharing is difficult, as bus and subway ticketing systems are often operated by different entities, resulting in complex data authorization processes and the risk of user privacy leaks; third, the coverage is limited, failing to count passengers who did not swipe their cards (such as those traveling for free or those who missed swiping their cards during transfers), leading to distorted statistical results.
[0004] Another category is existing statistical methods based on surveillance videos. Although these methods attempt to achieve non-intrusive statistics through target detection and tracking technologies, they still face several technical bottlenecks: First, they have poor scene adaptability. Fixed transfer area delineation cannot adapt to the layout differences and dynamic passenger flow density of different stations, easily leading to missed detections or false detections. Second, they have weak anti-interference capabilities. In scenarios with severe occlusion, sudden changes in lighting, or dense crowds, target tracking is prone to loss. Image quality evaluation indicators are singular and have fixed weights, making it difficult to extract effective features. Third, there is a contradiction between matching accuracy and real-time performance. Feature extraction relies solely on appearance information, lacking multimodal feature fusion. Furthermore, the spatiotemporal constraints are rigid and cannot be dynamically adjusted according to real-time traffic, resulting in low matching efficiency and a high false matching rate. Fourth, they lack generalization ability. Fixed model parameters cannot adapt to the differences in camera layout and shooting angles at different stations, requiring separate debugging for each station, leading to high deployment costs.
[0005] In summary, existing technologies struggle to balance the real-time nature, high accuracy, scenario adaptability, and privacy security of passenger transfer statistics. There is an urgent need for a passenger transfer perception and statistics method that requires no new hardware, relies on existing monitoring resources, and can effectively cope with interference in complex scenarios, in order to meet the actual needs of refined urban traffic operation and management. Summary of the Invention
[0006] This invention provides an intelligent sensing method, device, and medium for traffic transfer data to solve the following technical problems: existing methods for sensing traffic transfer data have strong lag, poor scene adaptability, and low matching accuracy, and cannot effectively cope with interference in complex scenarios.
[0007] The embodiments of the present invention adopt the following technical solutions: On one hand, embodiments of the present invention provide an intelligent sensing method for traffic transfer data, the method including: acquiring monitoring video streams of target traffic stations; The target tracking model identifies transfer targets in the surveillance video stream and generates a list of transfer targets. Extract the feature data of each transfer target in the transfer target list and store it in a vector database; In the vector database, based on spatiotemporal constraints, the same transfer target in different surveillance video streams is matched; For successfully matched transfer destinations, transfer data is processed and stored in the transportation transfer database.
[0008] In one feasible implementation, the monitoring video stream of the target transportation station is acquired, and the transfer targets in the monitoring video stream are identified through a target tracking model to generate a list of transfer targets, specifically including: The surveillance video streams are retrieved from each surveillance camera deployed at the target transportation station; wherein, the target transportation station is a bus stop or a subway station. The YoLov8 model is concatenated with the DeepSort tracking model to form the target tracking model, and the model is trained in advance. The current video frame is obtained from the monitoring video stream and input into the YoLov8 model to detect pedestrian targets. The detected pedestrian template is then input into the DeepSort tracking model for target tracking. Based on the trajectory of the center point of the target bounding box of the pedestrian target, it is determined whether the pedestrian target has entered the preset transfer area. If so, the pedestrian target is identified as a transfer target and added to the transfer target list.
[0009] In one feasible implementation, after inputting the detected pedestrian template into the DeepSort tracking model for target tracking, the method further includes: The target tracking result of the current video frame is input into the image quality estimation module to perform image quality evaluation on the target tracking result, and an image quality evaluation result is obtained; wherein, the image quality evaluation includes at least the following evaluation items: target occlusion degree, target blur degree, and target image size; The image quality assessment results are appended to the target tracking results.
[0010] In one feasible implementation, the target tracking result of the current video frame is input into the image quality estimation module, and the target tracking result is used to perform image quality evaluation to obtain the image quality evaluation result, specifically including: In the image quality estimation module, the target occlusion degree is calculated based on the target bounding box of the target tracking result. : ;when The smaller the value, the less the target is obscured; Among them, S i S is the intersection area of the target bounding box and other detection boxes. c The area of the target bounding box; The degree of blur of the target is obtained based on the pixel variance in the target bounding box. ; The larger the value, the clearer the image; Obtain the image size within the target bounding box. ; Degree of occlusion of the target The degree of ambiguity of the target and the image size The image quality assessment result is obtained by performing a weighted average calculation.
[0011] In one feasible implementation, feature data of each transfer target in the transfer target list is extracted and stored in a vector database, specifically including: A pedestrian re-identification model based on self-supervised learning is constructed; wherein, the backbone network of the pedestrian re-identification model adopts a swin-transformer network; The pedestrian re-identification model was pre-trained in a self-supervised manner on the LUPerson dataset, enabling the model to learn the semantic information of human images. The pre-trained pedestrian re-identification model is further trained using a pre-built dataset of pedestrian images from transportation stations. Based on the bounding box of each transfer target in the transfer target list, the original video frame image is cropped, and the cropped image is input into the trained pedestrian re-identification model to extract the feature data of the transfer target; wherein, the feature data includes at least shape features; The tracking ID, feature data, image quality assessment results, and shooting time of the transfer target are all stored in the vector database.
[0012] In one feasible implementation, after extracting the feature data of each transfer target in the transfer target list and storing it in a vector database, the method further includes: After the transfer target data of different transportation stations are aggregated into the vector database, it is determined whether the target has left the preset transfer area within a preset time based on the trajectory of the center point of the target box of each transfer target. If so, the data of the transfer target is removed. Image quality screening is performed on the remaining transfer targets in the vector database, and feature data of transfer targets whose image quality assessment results are less than a first preset threshold are removed; For transfer targets whose collected feature data volume exceeds a second preset threshold, interval sampling is performed, and the top five feature data ranked by image quality assessment results are averaged to obtain the final feature of the transfer target.
[0013] In one feasible implementation, in the vector database, based on spatiotemporal constraints, matching the same transfer target in different surveillance video streams specifically includes: In the vector database, the same transfer target is matched in different surveillance video streams based on the feature data of each transfer target; Once the same transfer destination is matched, several successfully matched surveillance video streams are obtained, and the acquisition location and acquisition time of each surveillance video stream are obtained. Based on the collection location and collection time, determine whether the transfer target meets the transfer time constraint and the transfer space constraint. If all conditions are met, then based on the agglomerative hierarchical clustering algorithm, the same transfer targets under different surveillance cameras are associated and matched, and included in the final transfer target to complete the statistics of the number of transfers.
[0014] In one feasible implementation, transfer data is processed for successfully matched transfer destinations with the same destination, and then stored in a transportation transfer database, specifically including: After the transfer target is confirmed, the transfer data of the transfer target is determined based on the location and distance of the transfer start station; wherein, the transfer data includes at least: transfer direction and average transfer time; The transfer data is aggregated into a transportation transfer database for data analysis. Update the transfer time constraint based on the average transfer time.
[0015] On the other hand, embodiments of the present invention also provide an intelligent sensing device for traffic transfer data, the device comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, enabling the at least one processor to execute the intelligent sensing method for traffic transfer data.
[0016] Finally, this embodiment of the invention also provides a storage medium, which is a non-volatile computer-readable storage medium storing at least one program, each program including instructions, which, when executed by a terminal, cause the terminal to execute the intelligent sensing method for traffic transfer data.
[0017] Compared with the prior art, the intelligent sensing method, device and medium for traffic transfer data provided in this embodiment of the invention have at least the following beneficial effects: 1. Significantly improved real-time performance: Through dynamic spatiotemporal constraint optimization, the amount of invalid calculations is greatly reduced, the delay in transfer data statistics is reduced, and the number of transfer passengers, directions and average transfer time can be output in real time, providing immediate decision support for peak-hour capacity scheduling.
[0018] 2. Significantly improved matching accuracy: The system integrates multi-dimensional dynamic image quality assessment based on occlusion level, blur level, and image size, effectively solving feature extraction challenges in scenarios such as occlusion, sudden changes in lighting, and dense crowds, thereby improving matching accuracy and reducing statistical errors.
[0019] 3. Enhanced scene adaptability and generalization ability: Dynamic transfer area delineation and adaptive image preprocessing adapt to different station layouts, passenger flow density, weather and lighting conditions. No need for individual debugging for a single station. It can be directly deployed in various bus and subway transfer scenarios, and the generalization ability is significantly improved.
[0020] 4. Outstanding advantages in privacy protection and cost: It adopts a non-intrusive statistical method, which does not require access to ticketing systems or third-party data platforms. It only uses existing surveillance videos, avoids access to passenger privacy information, has low data acquisition costs and no authorization threshold, and does not require new hardware equipment, thus fully exploring the potential value of existing surveillance resources.
[0021] In summary, this invention combines surveillance video with computer vision technology to achieve non-intrusive passenger counting during transfers, avoiding direct impact on passenger privacy. Traditional methods, on the other hand, require access to ticketing systems or third-party data platforms, and data authorization is difficult. Furthermore, this invention relies solely on surveillance video, exhibiting strong adaptability and scalability, fully exploring and utilizing the potential value of existing video data. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A flowchart of an intelligent sensing method for traffic transfer data provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an intelligent sensing device for traffic transfer data provided in an embodiment of the present invention. Detailed Implementation
[0023] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.
[0024] This invention provides an intelligent sensing method for traffic transfer data, such as... Figure 1 As shown, the intelligent sensing method for traffic transfer data specifically includes steps S101-S104: S101. Obtain the surveillance video stream of the target transportation station; identify the transfer targets in the surveillance video stream through the target tracking model and generate a list of transfer targets.
[0025] Specifically, the surveillance video streams are retrieved from each surveillance camera deployed at the target transportation station, which is either a bus stop or a subway station.
[0026] Furthermore, the YoLov8 model is concatenated with the DeepSort tracking model to form a target tracking model, and the model is trained in advance.
[0027] Furthermore, the current video frame is acquired from the surveillance video stream and input into the YoLov8 model to detect pedestrian targets. The detected pedestrian templates are then input into the DeepSort tracking model for target tracking. Based on the trajectory of the center point of the pedestrian target's bounding box, it is determined whether the pedestrian target has entered the preset transfer area. If so, the pedestrian target is identified as a transfer target and added to the transfer target list.
[0028] During target tracking, the target tracking result of the current video frame is input into the image quality estimation module to evaluate the image quality of the target tracking result, resulting in an image quality evaluation result. This image quality evaluation includes at least the following evaluation items: target occlusion degree, target blur degree, and target image size. The image quality evaluation result is then appended to the target tracking result.
[0029] As a feasible implementation method, the target tracking result of the current video frame is input into the image quality estimation module to perform image quality evaluation on the target tracking result, and the image quality evaluation result is obtained, specifically including: In the image quality estimation module, the degree of target occlusion is calculated based on the target bounding box from the target tracking results. : ;when The smaller the value, the less the target is occluded; where S i S is the area of intersection between the target bounding box and other detection boxes. c The area of the target bounding box; The degree of target blur is obtained based on the pixel variance within the target bounding box. ; The larger the value, the clearer the image; Get the image size within the target bounding box ; Target occlusion degree Degree of ambiguity of the target and image size A weighted average calculation is performed to obtain the image quality assessment result.
[0030] S102. Extract the feature data of each transfer target in the transfer target list and store it in the vector database.
[0031] Specifically, a pedestrian re-identification model based on self-supervised learning is constructed; the backbone network of the pedestrian re-identification model adopts the swin-transformer network.
[0032] Furthermore, the pedestrian re-identification model is pre-trained in a self-supervised manner on the LUPerson dataset, enabling the model to learn semantic information from human images. The pre-trained pedestrian re-identification model is then further trained using a pre-constructed dataset of pedestrian images from transportation stations.
[0033] Furthermore, based on the target bounding box of each transfer target in the transfer target list, the original video frame image is cropped, and the cropped image is input into the trained pedestrian re-identification model to extract the feature data of the transfer target; wherein, the feature data includes at least shape features.
[0034] Finally, the tracking ID, feature data, image quality assessment results, and shooting time of the transfer target are all stored in the vector database.
[0035] As a feasible implementation method, after aggregating transfer target data from different transportation stations into a vector database, the trajectory of the center point of the target bounding box for each transfer target is used to determine whether it has left the preset transfer area within a preset time. If so, the transfer target is removed from the data. Image quality screening is performed on the remaining transfer targets in the vector database, removing feature data of transfer targets whose image quality assessment results are lower than a first preset threshold. For transfer targets with a feature data volume greater than a second preset threshold, interval sampling is performed, and the average of the top five feature data with the highest image quality assessment results is calculated as the final feature of the transfer target.
[0036] S103. In the vector database, based on spatiotemporal constraints, match the same transfer target in different surveillance video streams.
[0037] Specifically, in the vector database, based on the feature data of each transfer target, the same transfer target is matched in different surveillance video streams. When the same transfer target is matched, several successfully matched surveillance video streams are obtained, and the acquisition location and acquisition time of each surveillance video stream are obtained.
[0038] Furthermore, based on the data collection location and time, it is determined whether the transfer target meets the transfer time constraint and the transfer space constraint. If both are met, then based on the agglomerative hierarchical clustering algorithm, the same transfer targets under different surveillance cameras are associated and matched, and included in the final transfer target to complete the statistics of the number of transfers.
[0039] S104. Perform transfer data calculations on the successfully matched same transfer destinations and store them in the transportation transfer database.
[0040] Specifically, after the transfer target is confirmed, the transfer data of the transfer target is determined based on the location and distance of the transfer starting station of the transfer target; the transfer data includes at least the transfer direction and the average transfer time.
[0041] Furthermore, the transfer data is aggregated into a transportation transfer database for data analysis; the aforementioned transfer time constraints are updated based on the average transfer time.
[0042] As a feasible implementation method, the present invention includes three main modules: a pedestrian tracking module, an image quality assessment module, and a feature extraction and matching module.
[0043] The pedestrian tracking module acquires video streams from the numerous cameras inevitably deployed in bus and subway stations, and uses YoLov8 and DeepSort to track pedestrians between stations. By analyzing the trajectory of the center point of the target bounding box, it determines whether the target has entered a preset transfer area; if so, it adds the target to a potential transfer target list.
[0044] The image quality assessment module is designed to extract high-quality features and reduce various interferences. After the pedestrian tracking module outputs the detection boxes for the current frame, it evaluates the image quality from the following three aspects: 1. Occlusion level: Determine the occlusion level of the target based on the degree of overlap of the detection boxes.
[0045] 2. Blur level: The variance of the target image. The larger the value, the richer the image texture and structure, and the more effective information it contains.
[0046] 3. Image size: Remove unreasonable narrow and elongated target boxes, as well as images that are too small in length and width (which cannot provide effective features and may even become noise, affecting the matching effect).
[0047] Finally, the results from the three aspects are weighted and averaged to obtain the image quality assessment result, which is then added to the tracking results for future reference.
[0048] The feature extraction and matching module extracts the appearance features of a target after it is detected in the image, using advanced pedestrian re-identification technology. It operates in parallel with the pedestrian tracking module, storing features in a vector database in real time and filtering matchable features based on spatiotemporal constraints and quality estimation conditions. Then, it uses agglomerative hierarchical clustering to aggregate transfer targets from different cameras based on a specified threshold. After confirming the transfer target, the module automatically calculates the transfer direction and average transfer time between stations and updates the time constraints, improving computational efficiency.
[0049] The following describes the solution process in more detail through specific implementation methods: First, in the pedestrian tracking module, video streams are retrieved from cameras deployed at bus and subway stations, and the current video frame is obtained and input into YoLov8 to detect pedestrian targets. The detection results are then sent to DeepSort for target association and tracking. The former has been fine-tuned in a private dataset to better suit images from a surveillance perspective. Simultaneously, the first occurrence time of each tracking ID is recorded, and the tracking results of the current frame are sent to the image quality estimation module and the feature extraction and matching module.
[0050] In the image quality estimation module, the latest detection results are obtained from the video frame detection results. (Target box BBOX) c The degree of occlusion is calculated using the following formula: ;where S i For BBOX cThe area of intersection with other detection boxes, S c For BBOX c The area, when The smaller the value, the less occlusion occurs. Blur level The variance within the detection box is used for calculation. Variance measures the dispersion of the data distribution; a larger variance indicates richer image texture information. When an image is blurry, it tends to lack clear boundaries and details disappear, resulting in a smaller variance. Based on prior knowledge, the detection box for a person in surveillance video tends to have H:W > 1. This means that extreme aspect ratios indicate extremely poor image quality; and when the height and width are less than 86 pixels, they are practically indistinguishable to the naked eye. Therefore, image size can measure image quality to some extent. This standard is denoted as [missing standard]. .therefore, Image quality is described from multiple perspectives, and three evaluation metrics are combined to make the quality estimation more robust. A weighted average is then applied. In addition, image quality will be added to the target tracking results.
[0051] The feature extraction and matching module first crops the original image based on the detection results and inputs it into a self-supervised learning-based person re-identification model to extract image features. It first performs self-supervised pre-training on the LUPerson dataset, enabling the model to learn the semantic information of human images. The backbone network of the model uses a Swin-transformer, which has extremely strong feature extraction capabilities. During the training phase, due to the limited number of samples from the monitoring perspective in the pre-training data, this invention uses approximately 20,000 bus-subway images (own data) for fine-tuning, enabling the model to achieve strong feature extraction capabilities in transfer scenarios. After feature extraction, the features, image quality, tracking ID, and time are stored in the database.
[0052] After aggregating data from different stations, the tracking IDs are screened for potential transfer targets based on whether the center point of the target trajectory has left the pre-defined transfer area within a short period of time. This step aims to remove interference from pedestrians, electric vehicles, and other objects appearing on the road. The next step involves filtering the features retained by the tracking IDs, eliminating low-quality features to prevent interference with the overall matching effect. Features with a quality score less than 0.6 are removed, and tracking IDs with a large number of features are sampled at intervals. Finally, the top 5 features with the highest quality scores are averaged to obtain the final features for that tracking ID.
[0053] Analysis of transfer data reveals that transferring passengers tend to seek the nearest subway exit corresponding to the bus stop, indicating a geographical correlation between buses and the subway. Therefore, the transfer spatial constraint is set as follows: ; in, Let this be the probability of a passenger transferring from subway station i to bus station j. Given the historical number of passengers transferring from subway station i to bus station j, the transfer probability matrix between stations can be derived. m represents the number of subway stations, and n represents the number of bus stops.
[0054] The relevant station transfer times fall within a certain range, meaning that passengers transferring within this range will inevitably appear in the surveillance videos of buses and subways. Using a Gaussian distribution to model the probability distribution of transfer times, the transfer time constraints can be obtained: ;in, Let be the probability of transfer time from subway station i to bus station j. Let be the average transfer time from subway station i to bus station j. The standard deviation of transfer times can be used to derive the probability distribution matrix of transfer times between stations. .
[0055] Therefore, by limiting the matching relationship and duration of bus-subway connections, spatiotemporal constraints can be implemented, reducing additional computation and significantly improving matching efficiency and accuracy, while ensuring the real-time nature of transfer statistics. For pedestrian features between different stations, a cosine similarity matrix S is calculated, and multiple weightings are applied to combine P and T. The final distance matrix is obtained, and agglomerative hierarchical clustering is performed to associate and match passengers transferring between different surveillance cameras. When a successful match is made between cameras at any bus stop and subway station, the target is included in the transfer target, and the number of transfers is finally counted.
[0056] In addition, embodiments of the present invention also provide an intelligent sensing device for traffic transfer data, such as... Figure 2 As shown, the intelligent sensing devices for traffic transfer data specifically include: At least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by at least one processor, so that at least one processor can perform the following: Acquire surveillance video streams from the target transportation station; The target tracking model identifies transfer targets in the surveillance video stream and generates a list of transfer targets. Extract the feature data of each transfer target in the transfer target list and store it in a vector database; In the vector database, based on spatiotemporal constraints, the same transfer target in different surveillance video streams is matched; For successfully matched transfer destinations, transfer data is processed and stored in the transportation transfer database.
[0057] Finally, this embodiment of the invention also provides a storage medium, which is a non-volatile computer-readable storage medium storing at least one program, each program including instructions, which, when executed by a terminal, cause the terminal to perform: Acquire surveillance video streams from the target transportation station; The target tracking model identifies transfer targets in the surveillance video stream and generates a list of transfer targets. Extract the feature data of each transfer target in the transfer target list and store it in a vector database; In the vector database, based on spatiotemporal constraints, the same transfer target in different surveillance video streams is matched; For successfully matched transfer destinations, transfer data is processed and stored in the transportation transfer database.
[0058] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.
[0059] The foregoing has described specific embodiments of the present invention. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.
[0060] The above description is merely an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the embodiments of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for intelligently sensing traffic transfer data, characterized in that, The method comprises: acquiring a monitoring video stream of a target traffic station; identifying transfer targets in the monitoring video stream through a target tracking model to generate a transfer target list; extracting feature data of each transfer target in the transfer target list and storing it in a vector database; matching the same transfer targets in different monitoring video streams based on spatiotemporal constraints in the vector database; performing transfer data operation on the successfully matched same transfer targets and storing them in a traffic transfer database. 2.The intelligent perception method for traffic transfer data according to claim 1, characterized in that, The method comprises: pulling monitoring video streams from each monitoring camera deployed at the target traffic station; wherein the target traffic station is a bus station or a subway station; connecting a YoLov8 model and a DeepSort tracking model in series to form the target tracking model and pre-training the model; acquiring a current video frame in the monitoring video stream, inputting it into the YoLov8 model to detect pedestrian targets, and inputting the detected pedestrian template into the DeepSort tracking model for target tracking; determining whether the pedestrian target has entered a preset transfer area based on the target box center point trajectory of the pedestrian target, and if so, determining the pedestrian target as a transfer target and adding it to the transfer target list. 3.The intelligent perception method of the traffic transfer data according to claim 2, characterized in that, After inputting the detected pedestrian template into the DeepSort tracking model for target tracking, the method further comprises: inputting the target tracking result of the current video frame into an image quality estimation module to perform image quality evaluation on the target tracking result and obtain an image quality evaluation result; wherein the image quality evaluation at least includes the following evaluation items: target occlusion degree, target blur degree and target image size; attaching the image quality evaluation result to the target tracking result. 4.The intelligent perception method of the traffic transfer data according to claim 3, characterized in that, Inputting the target tracking result of the current video frame into an image quality estimation module to perform image quality evaluation on the target tracking result and obtain an image quality evaluation result, specifically comprising: In the image quality estimation module, a target blocking degree is calculated according to a target frame of the target tracking result : ; the smaller the is, the smaller the degree of target blocking is Wherein, S i is the intersection area of the target frame and other detection frames, S c is the area of the target frame; obtain the target blur degree based on pixel variance in the target frame ; The greater the value, the clearer the image. acquiring an image size within the target frame ; weighting average calculation is performed on the target occlusion degree , the target blur degree , and the image size to obtain the image quality evaluation result. 5.The intelligent perception method for traffic transfer data according to claim 1, characterized in that, extracting feature data of each transfer target in the transfer target list and storing it in a vector database, specifically comprising: constructing a pedestrian re-identification model based on self-supervised learning; wherein the backbone network of the pedestrian re-identification model adopts a swin-transformer network; pre-training the pedestrian re-identification model on the LUPerson dataset in a self-supervised manner to enable the model to learn semantic information of human images; further training the pre-trained pedestrian re-identification model through a pre-constructed traffic station pedestrian image dataset; cropping the original video frame image according to the target box of each transfer target in the transfer target list and inputting the cropped image into the trained pedestrian re-identification model to extract feature data of the transfer target; wherein the feature data at least includes shape features; storing the tracking id, feature data, image quality evaluation result and shooting time of the transfer target in the vector database. 6.The intelligent perception method of traffic transfer data according to claim 1, characterized in that, After extracting the feature data of each transfer target in the transfer target list and storing it into the vector database, the method further comprises: After converging the transfer target data of different transportation stations into the vector database, it is judged whether each transfer target has walked out of the preset transfer area within the preset time according to the target box center point trajectory of each transfer target, and if so, the data of the transfer target is excluded; Image quality screening is performed on the remaining transfer targets in the vector database, and the feature data of the transfer targets with an image quality evaluation result less than a first preset threshold is excluded; Interval sampling is performed on the transfer targets with a feature data amount greater than a second preset threshold, and the feature data ranked in the top five in image quality evaluation result is selected for average calculation as the final feature of the transfer target. 7.The intelligent perception method of traffic transfer data according to claim 1, characterized in that, In the vector database, the same transfer targets in different monitoring video streams are matched based on the space-time constraint conditions, specifically comprising: In the vector database, the same transfer targets in different monitoring video streams are matched based on the feature data of each transfer target; After matching the same transfer targets, a plurality of monitoring video streams with matching success are obtained, and the collection position and collection time of each monitoring video stream are obtained; Based on the collection position and collection time, it is judged whether the transfer target satisfies the transfer time constraint condition and the transfer space constraint condition; If both conditions are satisfied, the same transfer targets under different monitoring cameras are associated and matched based on the condensed hierarchical clustering algorithm, and are counted into the final transfer target to complete the statistics of the number of transfer passengers. 8.The intelligent perception method of traffic transfer data according to claim 7, characterized in that, Transfer data operation is performed on the matched same transfer targets, and is stored into the traffic transfer database, specifically comprising: After the transfer target is confirmed, the transfer data of the transfer target is determined according to the position and distance of the transfer starting station of the transfer target; wherein the transfer data at least includes: transfer direction and average transfer time; The transfer data is summarized into the traffic transfer database for data analysis; The average transfer time is updated according to the average transfer time.
9. An intelligent sensing device for traffic transfer data, characterized in that, The device comprises: at least one processor; and a memory connected in communication with the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to execute the intelligent perception method for traffic transfer data according to any one of claims 1-8.
10. A storage medium, characterized by The storage medium is a non-volatile computer readable storage medium, and the non-volatile computer readable storage medium stores at least one program, each of which includes instructions that, when executed by a terminal, cause the terminal to execute the intelligent perception method for traffic transfer data according to any one of claims 1-8.