Dual-channel low-altitude violation warning method and system based on visual language model

Abstract

The application relates to the technical field of traffic monitoring, and provides a double-channel low-altitude illegal parking alarm method and system based on a visual language model, which comprises the following steps: performing sliding window cutting on an original image to obtain a slice image and record an offset; performing open-vocabulary target detection on the slice image to obtain a candidate target set, mapping the candidate target set to a global coordinate system of the original image to obtain a global candidate target set; performing redundancy suppression on the global candidate target set to obtain a merged candidate target set, cutting out a local subgraph in the original image, drawing corresponding visual prompt marks in the local subgraph, and obtaining a visual prompt graph; inputting the visual prompt graph and a text instruction into a pre-constructed visual language model to obtain a structured text; and matching and judging a verification type of the structured text and a coarse detection category of a candidate target in the merged candidate target set to obtain a structured alarm result; and the application can improve the efficiency of double-channel low-altitude illegal parking alarm.

Description

Technical Field

[0001] This invention relates to the field of traffic monitoring technology, and in particular to a dual-channel low-altitude illegal parking alarm method and system based on a visual language model. Background Technology

[0002] Current methods for detecting illegally parked non-motorized vehicles often employ a combination of object detection and geometric rules. However, in high-resolution traffic scenarios captured from a low-altitude perspective, directly performing object detection on the entire image suffers from insufficient feature extraction accuracy and a high false negative rate due to factors such as the small pixel proportion of non-motorized vehicles and mutual occlusion caused by densely parked vehicles. If slice detection is used to break down the image, overlapping slices can lead to the same physical target being detected multiple times, resulting in numerous duplicate alarms and distorting downstream statistical data. Furthermore, the illegal parking judgment logic based on region intersection-union ratio (IoU) heavily relies on fixed parking area calibration results, exhibiting poor adaptability to complex scenarios such as worn markings, overhead distortion, and vehicles obscuring markings. It also fails to recognize illegal parking behaviors with semantic attributes, resulting in significant limitations in judgment dimensions.

[0003] End-to-end illegal parking detection solutions based on visual language models typically directly input the entire scene image to generate detection results. This leads to significantly increased computational costs with increasing image resolution and scene complexity. Furthermore, the output lacks a unified structured format, with coordinate information often presented as unstructured natural language descriptions. This makes direct integration with urban governance work order systems and enforcement processes difficult, and in dense scenarios, it can easily result in missed detections and false positives. Existing combined applications of open vocabulary detectors and visual language models lack a traceable target numbering system and a unified coordinate mapping mechanism. The detection and verification results cannot be accurately aligned, making it difficult to control computational costs by reducing the scale of visual language model calls in the initial screening stage, and also failing to establish a one-to-one correspondence between semantic verification results and specific detection targets. The overall solution suffers from a high false alarm rate, making it difficult to form a stable business loop. Therefore, improving the accuracy of dual-channel low-altitude illegal parking alarms has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a dual-channel low-altitude illegal parking alarm method and system based on a visual language model to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a dual-channel low-altitude illegal parking alarm method based on a visual language model, comprising: A. Acquire original images of traffic scenes using a low-altitude platform, perform sliding window cropping on the original images to obtain slice images of the traffic scenes, and record the offset of the slice images in the original images; B. Using a preset non-motorized vehicle category text set as the query condition, perform open vocabulary target detection on the slice image to obtain the candidate target set of the slice image, and map the bounding box coordinates of the candidate target set to the global coordinate system of the original image according to the offset to obtain the global candidate target set of the traffic scene; C. Perform redundancy suppression based on confidence ranking on the global candidate target set to obtain the merged candidate target set of the traffic scenario. If the merged candidate target set is empty, output the result of no illegal parking alarm and end the process. D. Based on the global bounding box of the merged candidate target set, crop out a local sub-image containing contextual information from the original image, and draw corresponding visual cue markers in the local sub-image to obtain the visual cue map of the traffic scene; E. Input the visual cue image and the text instructions of the traffic scene into a pre-built visual language model to obtain the structured text of the traffic scene; F. Match and determine the verification type of the structured text with the coarse detection category of the candidate targets in the merged candidate target set to obtain the structured alarm result of the traffic scenario.

[0006] In a preferred embodiment, the process of acquiring original images of the traffic scene via a low-altitude platform, performing sliding window cropping on the original images to obtain slice images of the traffic scene, and recording the offset of the slice images within the original image includes: Raw images of traffic scenes were acquired using a low-altitude platform; Based on the preset sliding window size, determine the sliding step size of the original image in the width and height directions; Starting from the top left corner of the original image, the image is traversed sequentially along the width and height directions according to the sliding step size to obtain a slice image of the original image. Record the horizontal and vertical offsets of the top left corner of the cropping window corresponding to the slice image in the original image, and store the horizontal and vertical offsets in association with the slice image.

[0007] In a preferred embodiment, the step of performing open-vocabulary target detection on the slice image using a preset non-motorized vehicle category text set as the query condition to obtain a candidate target set for the slice image includes: The sliced ​​image is converted into an image feature representation of the traffic scene; The category text descriptions in the preset non-motorized vehicle category text set are converted into text feature representations of the traffic scene; The image feature representation and the text feature representation are correlated and matched to obtain the feature matching degree between the image feature representation and the text feature; Based on the feature matching degree, feature extraction is performed on the slice image to obtain candidate image regions of the slice image; The candidate image regions are labeled with corresponding category names and confidence scores, and the candidate image regions, their bounding box coordinates, category names, and confidence scores are summarized into a candidate target set for the sliced ​​image.

[0008] In a preferred embodiment, the step of mapping the bounding box coordinates of the candidate target set to the global coordinate system of the original image based on the offset to obtain the global candidate target set of the traffic scene includes: The horizontal coordinates of the bounding box coordinates in the candidate target set are superimposed with the horizontal offset of the slice image to obtain the global horizontal coordinates of the candidate target set in the global coordinate system of the original image. The ordinate value of the bounding box coordinates is added to the vertical offset of the slice image to obtain the global ordinate value of the candidate target set in the global coordinate system; Based on the global horizontal coordinate value and the global vertical coordinate value, determine the global bounding box of the candidate target set in the global coordinate system; By integrating the global bounding box, the category name, and the confidence score, and assigning a globally unique number to each candidate target in the candidate target set, a global candidate target set for the traffic scene is obtained.

[0009] In a preferred embodiment, the step of performing redundancy suppression based on confidence ranking on the global candidate target set to obtain the merged candidate target set for the traffic scene includes: Based on the confidence scores of the global candidate targets in the global candidate target set, the global candidate targets are ranked to obtain the sorted candidate target sequence for the traffic scenario; The candidate target with the highest confidence is selected from the ranked candidate target sequence as the benchmark target, and the benchmark target is removed from the ranked candidate target sequence to obtain the target sequence to be determined for the traffic scenario; Traverse the targets to be determined in the target sequence and obtain the intersection area and union area between the global bounding box of the target to be determined and the global bounding box of the reference target; The ratio between the intersection area and the union area is used as the first overlap metric of the traffic scene. The second overlap metric of the traffic scene is obtained by dividing the intersection area by the smaller of the global bounding box area of ​​the target to be determined and the global bounding box area of ​​the reference target. Calculate the comprehensive overlap metric value of the traffic scenario based on the first overlap metric value and the second overlap metric value; The comprehensive overlap metric is compared with a preset comprehensive overlap threshold. If the comprehensive overlap metric is greater than the comprehensive overlap threshold, the target to be determined is determined to be a redundant target and the target to be determined is removed. Otherwise, the target to be determined is retained to obtain the merged candidate target set of the traffic scene.

[0010] In a preferred embodiment, the formula for calculating the comprehensive overlap metric is as follows: ; in, This represents the comprehensive overlap metric value. This represents the first overlap metric value. This represents the second overlap metric. This represents the global bounding box area of ​​the reference target. This represents the area of ​​the global bounding box of the target to be determined. This represents the preset minimum positive number. This represents the preset overlap metric weight coefficient and .

[0011] In a preferred embodiment, the step of cropping a local sub-image containing contextual information from the original image based on the global bounding box of the merged candidate target set, and drawing corresponding visual cue markers in the local sub-image to obtain the visual cue map of the traffic scene includes: Centered on the global bounding box of the merged candidate target set, the cropping region is determined in the original image according to a preset expansion ratio to obtain a local sub-image containing environmental context information. The coordinate values ​​of the global bounding box are mapped to the coordinate system of the local subgraph to obtain the local bounding box of the local subgraph; On the local sub-graph, a marker box is drawn according to the local bounding box, and the reordered number of the merged candidate target set is superimposed and drawn at a preset adjacent position of the marker box to obtain the visual cue mark of the local sub-graph; By integrating the aforementioned visual cue markers and the local sub-images, a visual cue map of the traffic scene is obtained.

[0012] In a preferred embodiment, the step of inputting the visual cue image and the text instructions of the traffic scene into a pre-built visual language model to obtain the structured text of the traffic scene includes: Obtain the text instruction for the traffic scenario, the text instruction containing the verification number and the key points for determining illegal parking semantics; Extract the target number corresponding to the visual cue image, and verify the target number with the number to be verified to obtain the verified visual cue image sequence of the traffic scene; The verified visual cue sequence is associated and combined with the text instructions to obtain the joint data of the traffic scene; The joint data is input into a pre-built visual language model, and the original verification text returned by the visual language model is received. The original verification text includes a type description field and a scene description field corresponding to the target number. The original verification text is formatted to obtain the structured text of the traffic scenario.

[0013] In a preferred embodiment, the step of matching and determining the verification type of the structured text with the coarse detection category of the candidate targets in the merged candidate target set to obtain the structured alarm result for the traffic scenario includes: Extract the verification type corresponding to the target number from the structured text, and determine the verification type mapping relationship between the target number and the verification type; The reordering number, global bounding box, and corresponding coarse detection category of the candidate targets in the merged candidate target set are parsed to obtain the coarse detection information mapping relationship of the merged candidate target set; Using the target number as the associated key, the verification type mapping relationship and the coarse inspection information mapping relationship are aligned and merged to obtain a dual-channel information record of the target number; The consistency of the dual-channel information records is verified to obtain the alarm status of the target number; Based on the global bounding box, the coarse inspection category, the verification type, and the alarm status, a structured alarm result for the traffic scenario is generated.

[0014] To address the aforementioned problems, this invention also provides a dual-channel low-altitude illegal parking alarm system based on a visual language model, the system comprising: The image acquisition module is used to acquire original images of traffic scenes through a low-altitude platform, perform sliding window cropping on the original images to obtain slice images of the traffic scenes, and record the offset of the slice images in the original images; The image detection module is used to perform open-vocabulary target detection on the slice image using a preset non-motorized vehicle category text set as the query condition, to obtain a candidate target set of the slice image, and to map the bounding box coordinates of the candidate target set to the global coordinate system of the original image according to the offset, so as to obtain the global candidate target set of the traffic scene. The redundancy elimination module is used to perform redundancy suppression based on confidence ranking on the global candidate target set to obtain the merged candidate target set of the traffic scenario. If the merged candidate target set is empty, the no illegal parking alarm result is output and the process ends. The visual labeling module is used to crop out a local sub-image containing contextual information from the original image based on the global bounding box of the merged candidate target set, and draw corresponding visual cue labels in the local sub-image to obtain the visual cue image of the traffic scene. The semantic understanding module is used to input the visual cue image and the text instructions of the traffic scene into a pre-built visual language model to obtain the structured text of the traffic scene; The result alarm module is used to match and determine the verification type of the structured text with the coarse detection category of the candidate targets in the merged candidate target set, so as to obtain the structured alarm result of the traffic scenario.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention employs a dual-channel illegal parking recognition architecture to achieve hierarchical detection and accurate verification of illegally parked non-motorized vehicles. Coarse-grained localization of targets across the entire domain is achieved through sliding window cropping and open-vocabulary target detection, while fine-grained semantic determination of illegal parking is accomplished using visual cue tags and a visual language model, effectively improving the accuracy of illegal parking target recognition. The accompanying coordinate mapping and redundancy suppression mechanisms unify the global coordinate system of targets, eliminate redundant targets detected repeatedly, ensure consistency in target localization, and output standardized and reliable structured alarm results.

[0016] 2. This invention, through the image acquisition capabilities of a low-altitude platform and a sliding window cropping strategy, can cover the needs of illegal parking investigation in a wide range of traffic scenarios. The open vocabulary detection mode supports flexible configuration of non-motorized vehicle recognition categories, adapting to the illegal parking identification needs in different scenarios. The semantic parsing capability of the visual language model can directly output well-structured verification text, simplifying the alarm result generation process and comprehensively improving the processing efficiency and output quality of low-altitude illegal parking alarms, providing stable and efficient technical support for illegal parking management in traffic scenarios. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a dual-channel low-altitude illegal parking alarm method based on a visual language model provided in an embodiment of the present invention. Figure 2 This is a functional block diagram of a dual-channel low-altitude illegal parking alarm system based on a visual language model provided in an embodiment of the present invention. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides a dual-channel low-altitude illegal parking alarm method based on a visual language model. The executing entity of the dual-channel low-altitude illegal parking alarm method based on a visual language model includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the dual-channel low-altitude illegal parking alarm method based on a visual language model can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 The diagram shown is a flowchart illustrating a dual-channel low-altitude illegal parking alarm method based on a visual language model according to an embodiment of the present invention. In this embodiment, the dual-channel low-altitude illegal parking alarm method based on a visual language model includes: A. Acquire original images of traffic scenes using a low-altitude platform, perform sliding window cropping on the original images to obtain slice images of the traffic scenes, and record the offset of the slice images in the original images; In this embodiment of the invention, the step of acquiring original images of a traffic scene via a low-altitude platform, performing sliding window cropping on the original images to obtain slice images of the traffic scene, and recording the offset of the slice images within the original image includes: Raw images of traffic scenes were acquired using a low-altitude platform; Based on the preset sliding window size, determine the sliding step size of the original image in the width and height directions; Starting from the top left corner of the original image, the image is traversed sequentially along the width and height directions according to the sliding step size to obtain a slice image of the original image. Record the horizontal and vertical offsets of the top left corner of the cropping window corresponding to the slice image in the original image, and store the horizontal and vertical offsets in association with the slice image.

[0021] A low-altitude platform equipped with an image acquisition device flies to a preset acquisition position of the target traffic scene, adjusts its flight attitude so that the shooting direction of the image acquisition device is perpendicular to the traffic scene on the ground, and triggers the image acquisition operation after the platform's flight attitude is stable, acquiring original images covering the entire traffic scene.

[0022] The sliding window size is set to 640 pixels and the overlap ratio is set to 0.2. Based on the overall width and height of the original image, the distance the window moves in each step in the width direction and the distance the window moves in each step in the height direction are set respectively. This ensures that there is an overlapping area between adjacent windows after the window moves and that the window can eventually completely cover the entire screen area of ​​the original image. This gives the sliding step size of the original image in the width and height directions.

[0023] Starting from the top left corner of the original image, the cropping window moves sequentially along the width of the original image with a sliding step size in the width direction. Each time it moves to a new position, the image area covered by the current cropping window is cropped and saved. After completing the cropping operation of the entire row, it moves down one row along the height direction with a sliding step size in the height direction. The cropping operation in the width direction is repeated until the cropping window has traversed and covered the entire screen area of ​​the original image.

[0024] When the cropping window moves to the boundary of the original image, if the width or height of the remaining image is smaller than the size of the sliding window, the starting position of the cropping window is returned, so that the cropping window is close to the boundary of the original image and the window size remains unchanged, so as to avoid the target at the boundary being arbitrarily truncated, and finally all slice images corresponding to the original image are obtained.

[0025] For each cropped slice image, extract the horizontal and vertical position values ​​of the top-left corner vertex of the corresponding cropping window in the original image. Use the horizontal position value as the horizontal offset and the vertical position value as the vertical offset. Then, associate and store each slice image with its corresponding horizontal and vertical offsets, so that each slice image corresponds to a unique set of offset data.

[0026] The beneficial effects are that by splitting a large original image into multiple slice images through sliding window cropping, it can adapt to the input size requirements of the subsequent open vocabulary object detection stage. The overlapping areas retained by adjacent slices can avoid the problem of missed detection caused by the complete object being truncated by the window boundary. The offset and slice image associated storage can provide accurate positional basis for the global coordinate mapping of the candidate object set bounding box in the subsequent process, ensuring the accuracy of global coordinate restoration and data consistency.

[0027] B. Using a preset non-motorized vehicle category text set as the query condition, perform open vocabulary target detection on the slice image to obtain the candidate target set of the slice image, and map the bounding box coordinates of the candidate target set to the global coordinate system of the original image according to the offset to obtain the global candidate target set of the traffic scene; In this embodiment of the invention, the step of performing open-vocabulary target detection on the slice image using a preset non-motorized vehicle category text set as the query condition to obtain a candidate target set for the slice image includes: The sliced ​​image is converted into an image feature representation of the traffic scene; The category text descriptions in the preset non-motorized vehicle category text set are converted into text feature representations of the traffic scene; The image feature representation and the text feature representation are correlated and matched to obtain the feature matching degree between the image feature representation and the text feature; Based on the feature matching degree, feature extraction is performed on the slice image to obtain candidate image regions of the slice image; The candidate image regions are labeled with corresponding category names and confidence scores, and the candidate image regions, their bounding box coordinates, category names, and confidence scores are summarized into a candidate target set for the sliced ​​image.

[0028] The step of mapping the bounding box coordinates of the candidate target set to the global coordinate system of the original image based on the offset, to obtain the global candidate target set of the traffic scene, includes: The horizontal coordinates of the bounding box coordinates in the candidate target set are superimposed with the horizontal offset of the slice image to obtain the global horizontal coordinates of the candidate target set in the global coordinate system of the original image. The ordinate value of the bounding box coordinates is added to the vertical offset of the slice image to obtain the global ordinate value of the candidate target set in the global coordinate system; Based on the global horizontal coordinate value and the global vertical coordinate value, determine the global bounding box of the candidate target set in the global coordinate system; By integrating the global bounding box, the category name, and the confidence score, and assigning a globally unique number to each candidate target in the candidate target set, a global candidate target set for the traffic scene is obtained.

[0029] Multi-level feature encoding is performed on the sliced ​​image, extracting the low-level pixel features, mid-level texture features, and high-level semantic features in sequence and integrating them into a unified feature vector form to generate an image feature representation of the traffic scene.

[0030] Read all category text descriptions from the preset non-motorized vehicle category text set, which includes text descriptions of non-motorized vehicle types such as electric bicycles, shared bicycles, electric scooters, and tricycles; perform semantic encoding processing on each text description, extract the semantic connotation of the text and convert it into a feature vector form that is in the same feature space as the image features, and generate a text feature representation of the traffic scene.

[0031] Calculate the similarity value between the image feature representation and each set of text feature representations within a unified feature space, and use this similarity value as the feature matching degree between the image feature representation and the corresponding text feature.

[0032] Feature locations with matching scores higher than a preset threshold are selected based on the feature matching scores. The preset threshold is the coarse detection confidence threshold, typically set to 0.01. The complete image regions corresponding to these feature locations are located in the sliced ​​image to obtain the candidate image regions of the sliced ​​image.

[0033] For each candidate image region, a corresponding non-motorized vehicle category name is matched, and the feature matching degree value of the region is used as the confidence score for annotation. The bounding box coordinate information of each candidate image region is extracted, and the bounding box coordinate category name and confidence score of all candidate image regions are summarized and organized to generate a candidate target set of the slice image.

[0034] Extract the bounding box coordinates corresponding to each candidate target in the candidate target set, read all the horizontal coordinate values ​​of the bounding box coordinates one by one, and add each horizontal coordinate value to the horizontal offset corresponding to the current slice image to obtain the global horizontal coordinate value of the candidate target set in the global coordinate system of the original image.

[0035] Read all the ordinate values ​​in the bounding box coordinates one by one, and add each ordinate value to the vertical offset corresponding to the current slice image to obtain the global ordinate value of the candidate target set in the global coordinate system.

[0036] The global bounding box of the candidate target set is determined by combining the calculated global x-coordinate and global y-coordinate values ​​to form the coordinates of the four vertices of the bounding box.

[0037] The global bounding box category name and confidence information corresponding to each candidate target are bound one-to-one, and a unique identification number is assigned to each candidate target within the entire original image. All the bound candidate target information is then summarized to obtain the global candidate target set for the traffic scene.

[0038] The beneficial effects are that the text-driven open vocabulary detection mode can be used to complete the initial screening of non-motorized vehicle targets. It can directly adapt to the recognition needs of different non-motorized vehicle types through text category definition. At the same time, the slice-level detection results are unified to the global coordinate system of the original image through the coordinate offset superposition mapping method, which ensures the coordinate consistency of subsequent global redundancy suppression and illegal parking verification, and improves the positioning accuracy and processing continuity of non-motorized vehicle target detection in low-altitude and wide-field-of-view scenarios.

[0039] C. Perform redundancy suppression based on confidence ranking on the global candidate target set to obtain the merged candidate target set of the traffic scenario. If the merged candidate target set is empty, output the result of no illegal parking alarm and end the process. In this embodiment of the invention, the step of performing redundancy suppression based on confidence ranking on the global candidate target set to obtain the merged candidate target set for the traffic scene includes: Based on the confidence scores of the global candidate targets in the global candidate target set, the global candidate targets are ranked to obtain the sorted candidate target sequence for the traffic scenario; The candidate target with the highest confidence is selected from the ranked candidate target sequence as the benchmark target, and the benchmark target is removed from the ranked candidate target sequence to obtain the target sequence to be determined for the traffic scenario; Traverse the targets to be determined in the target sequence and obtain the intersection area and union area between the global bounding box of the target to be determined and the global bounding box of the reference target; The ratio between the intersection area and the union area is used as the first overlap metric of the traffic scene. The second overlap metric of the traffic scene is obtained by dividing the intersection area by the smaller of the global bounding box area of ​​the target to be determined and the global bounding box area of ​​the reference target. Calculate the comprehensive overlap metric value of the traffic scenario based on the first overlap metric value and the second overlap metric value; The comprehensive overlap metric is compared with a preset comprehensive overlap threshold. If the comprehensive overlap metric is greater than the comprehensive overlap threshold, the target to be determined is determined to be a redundant target and the target to be determined is removed. Otherwise, the target to be determined is retained to obtain the merged candidate target set of the traffic scene.

[0040] The formula for calculating the comprehensive overlap metric is as follows: ; in, This represents the comprehensive overlap metric value. This represents the first overlap metric value. This represents the second overlap metric. This represents the global bounding box area of ​​the reference target. This represents the area of ​​the global bounding box of the target to be determined. This represents the preset minimum positive number. This represents the preset overlap metric weight coefficient and .

[0041] Extract the confidence score value corresponding to each global candidate target in the global candidate target set, and arrange all global candidate targets in descending order of confidence score value to obtain the sorted candidate target sequence for the traffic scene.

[0042] The candidate target with the highest confidence is taken from the first position of the ranked candidate target sequence as the benchmark target. At the same time, the benchmark target is removed from the ranked candidate target sequence. All remaining candidate targets form the target sequence to be determined in the traffic scene.

[0043] Each target in the target sequence to be determined is selected in turn. The area of ​​the region covered by the global bounding box of the target to be determined and the global bounding box of the reference target overlaps is calculated as the intersection area. At the same time, the area of ​​the entire region covered by the two global bounding boxes is calculated as the union area.

[0044] Divide the intersection area by the union area and use the calculated ratio as the first overlap metric for the traffic scenario.

[0045] Calculate the total area of ​​the global bounding box of the target to be determined and the total area of ​​the global bounding box of the reference target respectively. Select the smaller area as the denominator and divide the value of the intersection area by the value of the smaller area to obtain the second overlap metric of the traffic scene.

[0046] The values ​​of the first overlap metric and the second overlap metric are combined for fusion calculation. During the calculation process, the proportion of the two metrics in the result is adjusted according to the area ratio of the baseline target and the target to be determined. At the same time, the calculation result is corrected according to the deviation of the two metrics to obtain the comprehensive overlap metric of the traffic scenario.

[0047] The first overlap metric is calculated by dividing the area of ​​the intersection of the global bounding box of the target to be determined and the global bounding box of the reference target by the area of ​​their union. The second overlap metric is calculated by dividing the area of ​​the intersection of the global bounding box of the target to be determined and the global bounding box of the reference target by the smaller of the two bounding box areas. The area of ​​the global bounding box of the reference target is obtained by multiplying the length and width of the global bounding box of the reference target with the highest confidence in the ranked candidate target sequence. The area of ​​the global bounding box of the target to be determined is obtained by multiplying the length and width of the global bounding box of each target to be determined in the target to be determined sequence.

[0048] The minimum positive value is a pre-set small positive value used to avoid division by zero anomalies, typically 10. -6 This is used to avoid division by zero during the calculation process; the overlap metric weight coefficient is a pre-set adjustment parameter with a value between zero and one, used to balance the proportion of the two overlap metrics in the calculation.

[0049] The comprehensive overlap metric is used to measure the degree of overlap between the global bounding boxes of the baseline target and the target to be judged. It is the core judgment indicator for determining whether the target to be judged is a redundant target in the redundancy suppression process.

[0050] This calculation method integrates two different dimensions of overlap measurement results, and adds an area ratio adjustment term and a difference penalty term, which can adapt to the overlap determination requirements of bounding boxes of different sizes. On the basis of conventional overlap determination logic, it further takes into account the overlap determination accuracy in scenarios where small targets are covered by large targets.

[0051] The overall overlap metric reaches its maximum value when two global bounding boxes completely overlap, and its minimum value when there is no overlap between the two global bounding boxes. As the area of ​​the overlapping region between the two global bounding boxes continues to increase, the overall overlap metric shows a continuous upward trend. When the area difference between the two global bounding boxes gradually increases, the overall overlap metric will be adjusted accordingly to avoid misjudgment when the small area box is completely inside the large area box. The value of the overlap metric weight coefficient will be biased towards a certain overlap metric value, and the change pattern of the overall overlap metric will be closer to the change characteristics of the corresponding metric value.

[0052] The value of the comprehensive overlap metric is compared with the preset comprehensive overlap threshold, which is typically 0. If the comprehensive overlap metric is greater than the comprehensive overlap threshold, the corresponding target to be judged is determined to be a redundant target and removed from the target sequence. If the comprehensive overlap metric is less than or equal to the comprehensive overlap threshold, the corresponding target to be judged is retained. After verifying all targets to be judged corresponding to the current benchmark target, the target with the highest confidence is selected again from the remaining target sequence as the new benchmark target. The overlap judgment and redundancy removal operations are repeated until all candidate targets are verified. Finally, all remaining candidate targets form the merged candidate target set of the traffic scene.

[0053] Count the total number of candidate targets in the merged candidate target set: if the total number of candidate targets is zero, the merged candidate target set is determined to be empty; after determining that the merged candidate target set is empty, generate the corresponding no illegal parking alarm result, output the generated no illegal parking alarm result, and at the same time end all current illegal parking alarm processing processes.

[0054] The beneficial effects are that the confidence-ranked round-by-round redundancy suppression mechanism, combined with the overlap measurement results of two different dimensions, calculates the comprehensive overlap degree, which can accurately identify and remove duplicate target boxes generated by sliding window clipping detection, avoid the same non-motorized vehicle target being counted multiple times, ensure the number of targets and the accuracy of positioning in the subsequent illegal parking verification process, and improve the reliability of non-motorized vehicle target detection results and subsequent processing efficiency in low-altitude, wide-field-of-view scenarios.

[0055] D. Based on the global bounding box of the merged candidate target set, crop out a local sub-image containing contextual information from the original image, and draw corresponding visual cue markers in the local sub-image to obtain the visual cue map of the traffic scene; In this embodiment of the invention, the step of cropping a local sub-image containing contextual information from the original image based on the global bounding box of the merged candidate target set, and drawing corresponding visual cue markers in the local sub-image to obtain the visual cue map of the traffic scene includes: Centered on the global bounding box of the merged candidate target set, the cropping region is determined in the original image according to a preset expansion ratio to obtain a local sub-image containing environmental context information. The coordinate values ​​of the global bounding box are mapped to the coordinate system of the local subgraph to obtain the local bounding box of the local subgraph; On the local sub-graph, a marker box is drawn according to the local bounding box, and the reordered number of the merged candidate target set is superimposed and drawn at a preset adjacent position of the marker box to obtain the visual cue mark of the local sub-graph; By integrating the aforementioned visual cue markers and the local sub-images, a visual cue map of the traffic scene is obtained.

[0056] Extract the global bounding box corresponding to each candidate target in the merged candidate target set one by one, locate the geometric center of each global bounding box, and expand the cropping range in the four directions of the image (up, down, left, and right) according to the preset expansion ratio. Determine the four boundary coordinates corresponding to the cropping area in the global coordinate system of the original image, and extract the corresponding range of image content from the original image along the determined boundaries to obtain a local sub-image containing environmental context information. The preset expansion ratio is typically 0.5.

[0057] Extract the global x-coordinate and global y-coordinate values ​​corresponding to the top left corner of the clipping region. Subtract the global x-coordinate value of the top left corner of the clipping region from the x-coordinate value of each vertex of the global bounding box to obtain the x-coordinate value of the corresponding vertex in the local subgraph coordinate system. Subtract the global y-coordinate value of the top left corner of the clipping region from the y-coordinate value of each vertex of the global bounding box to obtain the y-coordinate value of the corresponding vertex in the local subgraph coordinate system. Combine the calculated coordinate values ​​of the four vertices to form a closed rectangular frame, thus obtaining the local bounding box of the local subgraph.

[0058] Draw closed marker box lines along the four sides of the local bounding box at the corresponding position of the local sub-image. Extract the reorder number of the candidate target in the merged candidate target set. Overlay the number text at the preset adjacent position of the marker box. The marker box lines and the number text together constitute the visual cue mark of the local sub-image.

[0059] The completed visual cue markers are merged with the underlying local sub-images at the pixel level, so that the marker frame lines and numbered text are fixedly displayed in the corresponding positions of the local sub-images, forming a complete marked image and obtaining a visual cue map of the traffic scene.

[0060] The beneficial effect is that by expanding and pruning, the surrounding environmental context information of each candidate target is preserved. At the same time, by generating a bounding box that accurately matches the local sub-image through coordinate mapping and superimposing the corresponding number, the visual language model can be provided with clear target orientation and complete surrounding scene information. This helps the visual language model to accurately locate and verify the target and complete the semantic judgment of illegal parking in combination with the surrounding environment, thereby improving the positioning accuracy and semantic judgment reliability of the subsequent illegal parking verification process.

[0061] E. Input the visual cue image and the text instructions of the traffic scene into a pre-built visual language model to obtain the structured text of the traffic scene; In this embodiment of the invention, the step of inputting the visual cue image and the text instructions of the traffic scene into a pre-built visual language model to obtain the structured text of the traffic scene includes: Obtain the text instruction for the traffic scenario, the text instruction containing the verification number and the key points for determining illegal parking semantics; Extract the target number corresponding to the visual cue image, and verify the target number with the number to be verified to obtain the verified visual cue image sequence of the traffic scene; The verified visual cue sequence is associated and combined with the text instructions to obtain the joint data of the traffic scene; The joint data is input into a pre-built visual language model, and the original verification text returned by the visual language model is received. The original verification text includes a type description field and a scene description field corresponding to the target number. The original verification text is formatted to obtain the structured text of the traffic scenario.

[0062] The text instructions for the traffic scenario are read from the pre-stored verification rule configuration file. The text instructions contain the verification number of the target to be verified and the key points of illegal parking semantic judgment that need to be followed.

[0063] Extract the target number drawn overlaid on each visual cue image one by one, compare each extracted target number with the number to be verified in the text instruction, remove visual cue images that do not have a corresponding number to be verified, and arrange the visual cue images with matching numbers in numerical order to obtain the verified visual cue image sequence of the traffic scene.

[0064] Each visual cue image in the verified visual cue image sequence is bound one-to-one with the corresponding verification number and the key points for determining illegal parking in the text instructions, so that each visual cue image matches the corresponding verification rule content. All the bound image and text corresponding data are integrated to obtain the joint data of the traffic scene.

[0065] After the integrated joint data is adjusted to the input format supported by the pre-built visual language model, it is fed into the model. The visual language model performs inference analysis by combining the image content of the visual cue map with the verification requirements of the text instructions. After the analysis is completed, the corresponding results are output. The original verification text returned by the visual language model is received. The original verification text contains the type description field and scene description field corresponding to each target number.

[0066] Extract the type description field and scenario description field content corresponding to each target number from the original verification text. Organize and arrange the two types of content according to a unified field order and text format, remove redundant descriptions, and ensure that all verification results follow the same text structure to obtain structured text of the traffic scenario.

[0067] The beneficial effects are that invalid visual cue images are eliminated by comparing the target number with the number to be verified in advance, and then the visual cue images are bound and combined with the corresponding text instructions and input into the visual language model. This can guide the visual language model to accurately focus on the specified target and complete the semantic analysis of illegal parking according to the established rules. Finally, the structured text with a unified structure is obtained through format normalization, which provides verification data with a standard format for subsequent dual-channel matching and judgment, ensuring the accuracy of illegal parking semantic verification and the standardization of the results.

[0068] F. Match and determine the verification type of the structured text with the coarse detection category of the candidate targets in the merged candidate target set to obtain the structured alarm result of the traffic scenario.

[0069] In this embodiment of the invention, the step of matching and determining the verification type of the structured text with the coarse detection category of the candidate targets in the merged candidate target set to obtain the structured alarm result of the traffic scenario includes: Extract the verification type corresponding to the target number from the structured text, and determine the verification type mapping relationship between the target number and the verification type; The reordering number, global bounding box, and corresponding coarse detection category of the candidate targets in the merged candidate target set are parsed to obtain the coarse detection information mapping relationship of the merged candidate target set; Using the target number as the associated key, the verification type mapping relationship and the coarse inspection information mapping relationship are aligned and merged to obtain a dual-channel information record of the target number; The consistency of the dual-channel information records is verified to obtain the alarm status of the target number; Based on the global bounding box, the coarse detection category, the verification type, and the alarm status, a structured alarm result for the traffic scenario is generated.

[0070] The entire content of the structured text is parsed line by line to identify each target number field and its corresponding verification type field. Each target number and its matching verification type are stored in pairs to form a one-to-one correspondence between the target number and the verification type, thus determining the verification type mapping relationship between the target number and the verification type.

[0071] Read the complete data of all candidate targets in the merged candidate target set, parse the global bounding box coordinates and corresponding coarse detection category information of each candidate target, and associate and bind each reordered number with its corresponding global bounding box and coarse detection category to form a one-to-one correspondence between the reordered number and the target attribute, thus obtaining the coarse detection information mapping relationship of the merged candidate target set.

[0072] Using the target number as the common association key value for the two sets of mapping relationships, each record in the verification type mapping relationship is matched and aligned one by one with the record with the same number in the coarse inspection information mapping relationship. The verification type information and coarse inspection attribute information under the same number are merged into the same complete data record, resulting in the dual-channel information record corresponding to each target number.

[0073] Extract the coarse inspection category information and verification type information from each dual-channel information record, compare whether the category attributes of the two types of information match, and mark the corresponding alarm status for the corresponding target number according to the comparison result to obtain the alarm status of each target number.

[0074] Integrate the global bounding box coarse check category verification type and alarm status information corresponding to each target record, organize and arrange them according to a unified field order and data format, form a complete alarm data set with standardized information, and generate structured alarm results for traffic scenarios.

[0075] The beneficial effects are that the target information of the open vocabulary coarse detection is accurately aligned and merged with the verification information of the visual language model through number association, and the final alarm status of the target is determined through category consistency verification. Finally, multi-dimensional information is integrated to generate a structured alarm result. The accuracy of illegal parking alarms can be improved by combining the judgment results of dual channels. At the same time, the unified structured output facilitates the storage management and business calls of subsequent alarm data.

[0076] like Figure 2 The diagram shown is a functional block diagram of a dual-channel low-altitude illegal parking alarm system based on a visual language model provided in an embodiment of the present invention.

[0077] The dual-channel low-altitude illegal parking alarm system based on a visual language model described in this invention can be installed in electronic devices. Depending on the functions implemented, the dual-channel low-altitude illegal parking alarm system based on a visual language model may include an image acquisition module, an image detection module, a redundancy removal module, a visual tagging module, a semantic understanding module, and a result alarm module. The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, stored in the memory of the electronic device.

[0078] In this embodiment, the functions of each module / unit are as follows: The image acquisition module is used to acquire original images of traffic scenes through a low-altitude platform, perform sliding window cropping on the original images to obtain slice images of the traffic scenes, and record the offset of the slice images in the original images. The image detection module is used to perform open-vocabulary target detection on the slice image using a preset non-motorized vehicle category text set as the query condition, to obtain a candidate target set of the slice image, and to map the bounding box coordinates of the candidate target set to the global coordinate system of the original image according to the offset, so as to obtain the global candidate target set of the traffic scene. The redundancy elimination module is used to perform redundancy suppression based on confidence ranking on the global candidate target set to obtain the merged candidate target set of the traffic scenario. If the merged candidate target set is empty, the no illegal parking alarm result is output and the process ends. The visual labeling module is used to crop out a local sub-image containing contextual information from the original image based on the global bounding box of the merged candidate target set, and draw corresponding visual cue labels in the local sub-image to obtain the visual cue image of the traffic scene. The semantic understanding module is used to input the visual cue image and the text instructions of the traffic scene into a pre-built visual language model to obtain the structured text of the traffic scene; The result alarm module is used to match and determine the verification type of the structured text with the coarse detection category of the candidate targets in the merged candidate target set, so as to obtain the structured alarm result of the traffic scenario.

[0079] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0080] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0081] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0082] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0083] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A dual-channel low-altitude illegal parking alarm method based on a visual language model, characterized in that, The method includes: A. Acquire original images of traffic scenes using a low-altitude platform, perform sliding window cropping on the original images to obtain slice images of the traffic scenes, and record the offset of the slice images in the original images; B. Using a preset non-motorized vehicle category text set as the query condition, perform open vocabulary target detection on the slice image to obtain the candidate target set of the slice image, and map the bounding box coordinates of the candidate target set to the global coordinate system of the original image according to the offset to obtain the global candidate target set of the traffic scene; C. Perform redundancy suppression based on confidence ranking on the global candidate target set to obtain the merged candidate target set of the traffic scenario. If the merged candidate target set is empty, output the result of no illegal parking alarm and end the process. D. Based on the global bounding box of the merged candidate target set, crop out a local sub-image containing contextual information from the original image, and draw corresponding visual cue markers in the local sub-image to obtain the visual cue map of the traffic scene; E. Input the visual cue image and the text instructions of the traffic scene into a pre-built visual language model to obtain the structured text of the traffic scene; F. Match and determine the verification type of the structured text with the coarse detection category of the candidate targets in the merged candidate target set to obtain the structured alarm result of the traffic scenario.

2. The dual-channel low-altitude illegal parking alarm method based on a visual language model as described in claim 1, characterized in that, The process of acquiring original images of the traffic scene via a low-altitude platform, performing sliding window cropping on the original images to obtain slice images of the traffic scene, and recording the offset of the slice images within the original images includes: Raw images of traffic scenes were acquired using a low-altitude platform; Based on the preset sliding window size, determine the sliding step size of the original image in the width and height directions; Starting from the top left corner of the original image, the image is traversed sequentially along the width and height directions according to the sliding step size to obtain a slice image of the original image. Record the horizontal and vertical offsets of the top left corner of the cropping window corresponding to the slice image in the original image, and store the horizontal and vertical offsets in association with the slice image.

3. The dual-channel low-altitude illegal parking alarm method based on a visual language model as described in claim 1, characterized in that, The step of performing open-vocabulary target detection on the sliced ​​image using a preset non-motorized vehicle category text set as the query condition to obtain a candidate target set for the sliced ​​image includes: The sliced ​​image is converted into an image feature representation of the traffic scene; The category text descriptions in the preset non-motorized vehicle category text set are converted into text feature representations of the traffic scene; The image feature representation and the text feature representation are correlated and matched to obtain the feature matching degree between the image feature representation and the text feature; Based on the feature matching degree, feature extraction is performed on the slice image to obtain candidate image regions of the slice image; The candidate image regions are labeled with corresponding category names and confidence scores, and the candidate image regions, their bounding box coordinates, category names, and confidence scores are summarized into a candidate target set for the sliced ​​image.

4. The dual-channel low-altitude illegal parking alarm method based on a visual language model as described in claim 3, characterized in that, The step of mapping the bounding box coordinates of the candidate target set to the global coordinate system of the original image based on the offset, to obtain the global candidate target set of the traffic scene, includes: The horizontal coordinates of the bounding box coordinates in the candidate target set are superimposed with the horizontal offset of the slice image to obtain the global horizontal coordinates of the candidate target set in the global coordinate system of the original image. The ordinate value of the bounding box coordinates is added to the vertical offset of the slice image to obtain the global ordinate value of the candidate target set in the global coordinate system; Based on the global horizontal coordinate value and the global vertical coordinate value, determine the global bounding box of the candidate target set in the global coordinate system; By integrating the global bounding box, the category name, and the confidence score, and assigning a globally unique number to each candidate target in the candidate target set, a global candidate target set for the traffic scene is obtained.

5. The dual-channel low-altitude illegal parking alarm method based on a visual language model as described in claim 1, characterized in that, The process of performing redundancy suppression based on confidence ranking on the global candidate target set to obtain the merged candidate target set for the traffic scene includes: Based on the confidence scores of the global candidate targets in the global candidate target set, the global candidate targets are ranked to obtain the sorted candidate target sequence for the traffic scenario; The candidate target with the highest confidence is selected from the ranked candidate target sequence as the benchmark target, and the benchmark target is removed from the ranked candidate target sequence to obtain the target sequence to be determined for the traffic scenario; Traverse the targets to be determined in the target sequence and obtain the intersection area and union area between the global bounding box of the target to be determined and the global bounding box of the reference target; The ratio between the intersection area and the union area is used as the first overlap metric of the traffic scene. The second overlap metric of the traffic scene is obtained by dividing the intersection area by the smaller of the global bounding box area of ​​the target to be determined and the global bounding box area of ​​the reference target. Calculate the comprehensive overlap metric value of the traffic scenario based on the first overlap metric value and the second overlap metric value; The comprehensive overlap metric is compared with a preset comprehensive overlap threshold. If the comprehensive overlap metric is greater than the comprehensive overlap threshold, the target to be determined is determined to be a redundant target and the target to be determined is removed. Otherwise, the target to be determined is retained to obtain the merged candidate target set of the traffic scene.

6. The dual-channel low-altitude illegal parking alarm method based on a visual language model as described in claim 5, characterized in that, The formula for calculating the comprehensive overlap metric is as follows: ； in, This represents the comprehensive overlap metric value. This represents the first overlap metric value. This represents the second overlap metric. This represents the global bounding box area of ​​the reference target. This represents the area of ​​the global bounding box of the target to be determined. This represents the preset minimum positive number. This represents the preset overlap metric weight coefficient and .

7. The dual-channel low-altitude illegal parking alarm method based on a visual language model as described in claim 1, characterized in that, The step involves cropping a local sub-image containing contextual information from the original image based on the global bounding box of the merged candidate target set, and drawing corresponding visual cue markers in the local sub-image to obtain the visual cue map of the traffic scene, including: Centered on the global bounding box of the merged candidate target set, the cropping region is determined in the original image according to a preset expansion ratio to obtain a local sub-image containing environmental context information. The coordinate values ​​of the global bounding box are mapped to the coordinate system of the local subgraph to obtain the local bounding box of the local subgraph; On the local sub-graph, a marker box is drawn according to the local bounding box, and the reordered number of the merged candidate target set is superimposed and drawn at a preset adjacent position of the marker box to obtain the visual cue mark of the local sub-graph; By integrating the aforementioned visual cue markers and the local sub-images, a visual cue map of the traffic scene is obtained.

8. The dual-channel low-altitude illegal parking alarm method based on a visual language model as described in claim 1, characterized in that, The step of inputting the visual cue image and the text instructions of the traffic scene into a pre-built visual language model to obtain the structured text of the traffic scene includes: Obtain the text instruction for the traffic scenario, the text instruction containing the verification number and the key points for determining illegal parking semantics; Extract the target number corresponding to the visual cue image, and verify the target number with the number to be verified to obtain the verified visual cue image sequence of the traffic scene; The verified visual cue sequence is associated and combined with the text instructions to obtain the joint data of the traffic scene; The joint data is input into a pre-built visual language model, and the original verification text returned by the visual language model is received. The original verification text includes a type description field and a scene description field corresponding to the target number. The original verification text is formatted to obtain the structured text of the traffic scenario.

9. The dual-channel low-altitude illegal parking alarm method based on a visual language model as described in claim 1, characterized in that, The step of matching and determining the verification type of the structured text with the coarse detection category of the candidate targets in the merged candidate target set to obtain the structured alarm result for the traffic scenario includes: Extract the verification type corresponding to the target number from the structured text, and determine the verification type mapping relationship between the target number and the verification type; The reordering number, global bounding box, and corresponding coarse detection category of the candidate targets in the merged candidate target set are parsed to obtain the coarse detection information mapping relationship of the merged candidate target set; Using the target number as the associated key, the verification type mapping relationship and the coarse inspection information mapping relationship are aligned and merged to obtain a dual-channel information record of the target number; The consistency of the dual-channel information records is verified to obtain the alarm status of the target number; Based on the global bounding box, the coarse detection category, the verification type, and the alarm status, a structured alarm result for the traffic scenario is generated.

10. A dual-channel low-altitude illegal parking alarm system based on a visual language model, characterized in that, The system for implementing the dual-channel low-altitude illegal parking alarm method based on a visual language model as described in claim 1 includes: The image acquisition module is used to acquire original images of traffic scenes through a low-altitude platform, perform sliding window cropping on the original images to obtain slice images of the traffic scenes, and record the offset of the slice images in the original images; The image detection module is used to perform open-vocabulary target detection on the slice image using a preset non-motorized vehicle category text set as the query condition, to obtain a candidate target set of the slice image, and to map the bounding box coordinates of the candidate target set to the global coordinate system of the original image according to the offset, so as to obtain the global candidate target set of the traffic scene. The redundancy elimination module is used to perform redundancy suppression based on confidence ranking on the global candidate target set to obtain the merged candidate target set of the traffic scenario. If the merged candidate target set is empty, the no illegal parking alarm result is output and the process ends. The visual labeling module is used to crop out a local sub-image containing contextual information from the original image based on the global bounding box of the merged candidate target set, and draw corresponding visual cue labels in the local sub-image to obtain the visual cue image of the traffic scene. The semantic understanding module is used to input the visual cue image and the text instructions of the traffic scene into a pre-built visual language model to obtain the structured text of the traffic scene; The result alarm module is used to match and determine the verification type of the structured text with the coarse detection category of the candidate targets in the merged candidate target set, so as to obtain the structured alarm result of the traffic scenario.

Images