Dangerous goods vehicle detection method and system under high-position monitoring view angle

Abstract

The invention discloses a dangerous goods vehicle detection method and system under a high-position monitoring view angle, and the method comprises the steps: obtaining vehicle recognition original data, carrying out the data labeling of the vehicle recognition original data, obtaining vehicle recognition labeled data, carrying out the data enhancement processing of the vehicle recognition labeled data, and obtaining a target detection training set; and training a target detection model based on the target detection training set to obtain a dangerous goods vehicle detection model, identifying the to-be-detected data based on the dangerous goods vehicle detection model, and outputting an identification result. According to the method, the problem of identifying the dangerous goods vehicle under the high point view angle is solved, the recall rate of the algorithm on the target is improved, the early warning capability of the system on the dangerous goods vehicle is enhanced, the manual marking cost is reduced through data enhancement, a more robust target detection model can be rapidly trained, and the model iteration period and times are reduced.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision technology, and in particular relates to a method and system for detecting vehicles carrying dangerous goods under a high-point monitoring viewing angle. Background Art

[0002] Object detection is a technology in the field of computer vision that is used to identify multiple objects in an image and locate them. The object detection model can identify multiple objects in a picture and can locate different objects. This technology is widely used in fields such as unmanned driving and security systems. Dangerous goods vehicle object detection uses object detection technology to identify dangerous goods vehicles. The requirements for dangerous goods vehicles are as follows: During the road transportation of dangerous goods, the transport vehicles shall install and hang warning signs that meet the standards for road transportation of dangerous goods. Vehicles transporting explosives and highly toxic chemicals shall also install and affix safety signs that meet the safety standards for transporting dangerous goods. Based on the above description, since there are some special signs on the body of dangerous goods, some current methods will also use the method of identifying special signs to distinguish whether it is a dangerous goods vehicle. In the current hazardous goods vehicle identification system, the height of the surveillance camera is usually less than 10 meters. In the tower visual scene, the surveillance camera is usually hung at a height of 15-50 meters. At this height, it is difficult for the camera to obtain clear texture information (i.e. warning signs) on the target vehicle body. As the monitoring point increases, the recognition problem of small targets and blurred targets will become more obvious. It is difficult to ensure the recognition effect by training a target detection model for warning sign recognition based on such data. Therefore, the current recognition method of hazardous chemical vehicles in the low-point monitoring perspective cannot be directly reused in the high-point monitoring scene. In addition, in terms of data collection, the frequency of hazardous goods vehicles on the actual road is much lower than that of other normal models. Therefore, in the data collection stage, it is necessary to conduct long-term screening in the real high-point monitoring data, which requires a lot of time and cost to find a small number of pictures containing hazardous goods vehicles. In addition, in some sections where large trucks are prohibited from driving, or in some special places, there are basically no hazardous goods vehicles, which reduces the richness of the data scene in the data preparation stage, resulting in weak generalization ability of the final detection model and inability to effectively cope with complex and changing scenes. Summary of the invention

[0003] To solve the above problems, the present invention provides a method and system for detecting hazardous goods vehicles under a high-point monitoring perspective, so as to solve the problem that the current method for identifying hazardous chemicals vehicles is not suitable for high-point monitoring scenarios.

[0004] A method for detecting dangerous goods vehicles under a high-point monitoring perspective, comprising: Obtaining vehicle identification raw data and annotating it to obtain vehicle identification annotated data; Perform data enhancement processing on the vehicle identification and annotation data to obtain the target detection training set; The target detection model is trained based on the target detection training set to obtain a dangerous goods vehicle detection model; Based on the dangerous goods vehicle detection model, the data to be detected is identified and the identification results are output.

[0005] According to a specific embodiment of the present invention, the original vehicle identification data is obtained and data is annotated to obtain the vehicle identification annotated data, including: Select multiple vehicle identification images from high-point surveillance videos as vehicle identification raw data; The original vehicle identification data is labeled according to three categories: dangerous goods vehicles, suspected dangerous goods vehicles and non-dangerous goods vehicles to obtain vehicle identification labeling data.

[0006] According to a specific embodiment of the present invention, data enhancement processing is performed on vehicle identification and labeling data to obtain a target detection training set including: Create a dangerous goods vehicle material library based on vehicle identification and annotation data; Randomly select a vehicle identification image from the vehicle identification annotation data as the background image, and determine the area to be generated; Randomly select a foreground image of a dangerous goods vehicle from the dangerous goods vehicle library and scale the image; The scaled foreground image is placed in the to-be-generated area of ​​the background image to generate a target detection image; Generate an object detection training set based on multiple object detection images.

[0007] According to a specific embodiment of the present invention, creating a dangerous goods vehicle material library based on vehicle identification and annotation data includes: Select dangerous goods vehicle images and corresponding target boxes from vehicle identification and annotation data; Using the target box as the prompt information, the semantic segmentation model is used to segment the dangerous goods vehicle image and extract the foreground pixel information; Create a dangerous goods vehicle library based on foreground pixel information.

[0008] According to a specific embodiment of the present invention, a method for determining a region to be generated includes: Get all target boxes from the background image, select the bottom center point of each target box, and move the bottom center point downward along the vertical axis of the image to one tenth of the height of the corresponding target box. If the moved bottom center point falls into any of the target boxes, delete the bottom center point and use the remaining bottom center point as prompt information. Use the semantic segmentation model to segment the image and obtain the vehicle driving area, which is determined as the area to be generated.

[0009] According to a specific embodiment of the present invention, randomly selecting a foreground image of a dangerous goods vehicle from a dangerous goods vehicle material library and performing image scaling includes: Select any pixel point from the area to be generated and determine the n nearest target boxes of the pixel point; Calculate the average size of the n nearest target boxes and scale the foreground image to the same size as the average size.

[0010] According to a specific embodiment of the present invention, placing the scaled foreground image in the to-be-generated area of ​​the background image, and generating the target detection image includes: Select any pixel point from the area to be generated as the target center point, place the scaled foreground image at the target center point, and generate a target detection image if the overlap rate between the mask of the foreground image and the mask of the area to be generated is greater than 0.5.

[0011] According to a specific embodiment of the present invention, identifying the data to be detected based on the dangerous goods vehicle detection model and outputting the identification result includes: Input the data to be detected into the dangerous goods vehicle detection model for the first identification, and output the first identification result, including dangerous goods vehicles and suspected dangerous goods vehicles; Based on the first recognition result, determine whether it is a suspected dangerous goods vehicle. If so, perform a second recognition and output the second recognition result, including the dangerous goods vehicle.

[0012] According to a specific embodiment of the present invention, judging whether the vehicle is suspected of being a dangerous goods vehicle based on the first recognition result, and if so, performing a second recognition, and outputting the second recognition result includes: It is determined whether the target image of the suspected dangerous goods vehicle is smaller than a preset size threshold. If so, the output result is a dangerous goods vehicle. Otherwise, it is determined whether the confidence of the target image of the suspected dangerous goods vehicle meets the preset confidence threshold. If so, the output result is a dangerous goods vehicle.

[0013] According to a specific embodiment of the present invention, the preset confidence threshold is 0.5-0.7.

[0014] A dangerous goods vehicle detection system under high-point monitoring viewing angle, comprising: The data collection and data annotation module is used to obtain the original vehicle identification data and annotate it to obtain the vehicle identification annotation data; The data enhancement module is used to perform data enhancement processing on the vehicle identification and annotation data to obtain the target detection training set; A model creation module is used to train the target detection model based on the target detection training set to obtain a dangerous goods vehicle detection model; The vehicle detection module is used to identify the data to be detected based on the dangerous goods vehicle detection model and output the recognition result.

[0015] According to a specific embodiment of the present invention, the data collection and data annotation module further includes: A data acquisition module is used to select multiple vehicle identification images from high-point monitoring videos as vehicle identification raw data; The data labeling module is used to label the original vehicle identification data according to three categories: dangerous goods vehicles, suspected dangerous goods vehicles and non-dangerous goods vehicles, to obtain vehicle identification labeling data.

[0016] According to a specific embodiment of the present invention, the data enhancement module further includes: A material library creation module is used to create a dangerous goods vehicle material library based on vehicle identification and annotation data; The module for determining the area to be generated is used to randomly select a vehicle identification image from the vehicle identification annotation data as a background image and determine the area to be generated; The image scaling module is used to randomly select a foreground image of a dangerous goods vehicle from the dangerous goods vehicle material library and scale the image; The target detection image generation module is used to place the scaled foreground image in the to-be-generated area of ​​the background image to generate a target detection image; The target detection training set generation module is used to generate a target detection training set based on multiple target detection images.

[0017] According to a specific embodiment of the present invention, the vehicle detection module further includes: A first identification module is used to input the data to be detected into the dangerous goods vehicle detection model for first identification, and output a first identification result, including dangerous goods vehicles and suspected dangerous goods vehicles; The second identification module is used to determine whether it is a suspected dangerous goods vehicle based on the first identification result. If so, a second identification is performed and a second identification result is output, including the dangerous goods vehicle.

[0018] According to a specific embodiment of the present invention, the second identification module further includes: The first judgment module is used to judge whether the target image of the suspected dangerous goods vehicle is smaller than a preset size threshold, and if it is smaller than the preset size threshold, the output result is a dangerous goods vehicle; The second judgment module is used to judge whether the confidence of the target image of the suspected dangerous goods vehicle meets a preset confidence threshold. If so, the output result is a dangerous goods vehicle.

[0019] An electronic device includes: a processor and a memory, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement the above-mentioned dangerous goods vehicle detection method under the high-point monitoring perspective.

[0020] A computer-readable storage medium stores a computer program, which is loaded and executed by a processor to implement the above-mentioned dangerous goods vehicle detection method under the high-point monitoring perspective.

[0021] Compared with the prior art, the method and system for detecting dangerous goods vehicles under a high-point monitoring perspective provided by the present invention have the following advantages: 1. The dangerous goods vehicle detection method under high-point monitoring viewing angle proposed in the present invention can effectively solve the problem of unclear target identification and difficulty in analyzing target detail texture information under high-point viewing angle.

[0022] 2. The present invention performs secondary differentiation through the intermediate state of "suspected dangerous goods vehicles", effectively improving the recall rate of dangerous goods vehicle targets and improving the early warning capability of the system.

[0023] 3. The data enhancement method proposed in the present invention can efficiently generate targets including dangerous goods vehicles in different scene background images, while reducing the workload of data collection and effectively improving the generalization ability of the model.

[0024] 4. The dangerous goods vehicle detection system under the high-point monitoring perspective proposed in the present invention can realize automated data collection, data enhancement, and model training after the model is trained, thereby accelerating the optimization and iteration process of the model.

[0025] 5. The hazardous vehicle detection system proposed in the present invention can be quickly applied to target recognition tasks in other high-point traffic scenes. The system can cover a larger monitoring range and effectively reduce the investment cost of monitoring equipment deployment. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0027] Figure 1 The present invention provides a flowchart of a method for detecting dangerous goods vehicles under a high-point monitoring perspective according to an embodiment of the present invention.

[0028] Figure 2It is a flow chart of a method for acquiring data and performing data annotation according to an embodiment of the present invention.

[0029] Figure 3 It is a flow chart of a method for performing data enhancement processing on vehicle identification and labeling data provided according to an embodiment of the present invention.

[0030] Figure 4 is a flow chart of a material library creation method provided according to an embodiment of the present invention.

[0031] Figure 5 is a flow chart of a picture scaling method provided according to an embodiment of the present invention.

[0032] Figure 6 It is a flow chart of a method for identifying data to be detected provided according to an embodiment of the present invention.

[0033] Figure 7 It is a schematic diagram of a data enhancement method provided according to an embodiment of the present invention.

[0034] Figure 8 The figure is a flowchart of dangerous goods vehicle detection according to an embodiment of the present invention.

[0035] Fig. 9 It is a structural diagram of a dangerous goods vehicle detection system under a high-point monitoring perspective provided according to an embodiment of the present invention.

[0036] Fig.10 It is a structural diagram of a data collection and data annotation module provided according to an embodiment of the present invention.

[0037] Fig.11 It is a structural diagram of a data enhancement module provided according to an embodiment of the present invention.

[0038] Fig.12 is a structural diagram of a vehicle detection module provided according to an embodiment of the present invention.

[0039] Fig.13 is a structural diagram of a second identification module provided according to an embodiment of the present invention.

[0040] Fig.14 It is a schematic diagram of the structure of a computer device provided according to an embodiment of the present invention.

[0041] Reference numerals: 01-Data collection and data annotation module; 02-Data enhancement module; 03-Model creation module; 04-Vehicle detection module; 011-data acquisition module; 012-data annotation module; 021-Material library creation module; 022-Area determination module to be generated; 023-Image scaling module; 024-Target detection graphics generation module; 025-Target detection training set generation module; 041-first identification module; 042-second identification module; 0421-first judgment module; 0422-second judgment module. DETAILED DESCRIPTION

[0042] In order to make those skilled in the art understand the concept and thought of the present invention more clearly, the present invention is described in detail below in conjunction with specific embodiment.It should be understood that the embodiment provided herein is only a part of all embodiments that the present invention may have.Those skilled in the art, after reading the specification of the application, have the ability to make improvements, transformations, or replacements to part or the whole of the following embodiments, and these improvements, transformations, or replacements are also included in the scope of the present invention.

[0043] In this article, the terms "preview", "entry" and other similar words are not intended to imply any order, quantity and importance, but are only used to distinguish different elements. In this article, the terms "one", "an" and other similar words are not intended to indicate that there is only one thing, but to indicate that the relevant description is only for one of the things, and the thing may have one or more. In this article, the terms "comprise", "include" and other similar words are intended to indicate logical relationships, and cannot be regarded as indicating spatial structural relationships. For example, "A includes B" is intended to indicate that B belongs to A logically, but does not mean that B is located inside A in space. In addition, the meanings of the terms "comprise", "include" and other similar words should be regarded as open, not closed. For example, "A includes B" is intended to indicate that B belongs to A, but B does not necessarily constitute the whole of A, and A may also include other elements such as C, D, and E.

[0044] In this document, the terms "embodiment", "this embodiment", "one embodiment", and "an embodiment" do not mean that the relevant description is only applicable to a specific embodiment, but rather that the description may also be applicable to one or more other embodiments. Those skilled in the art should understand that in this document, any description of a certain embodiment can be replaced, combined, or combined in other ways with the relevant descriptions in one or more other embodiments, and the new embodiments generated by the replacement, combination, or combination in other ways are easily conceivable by those skilled in the art and fall within the scope of protection of the present invention.

[0045] Example 1 Additional aspects and advantages of embodiments of the present invention will be described in part in the following description, and will become apparent from the following description, or will be learned through practice of embodiments of the present invention. Figure 1-Figure 8 The embodiment of the present invention provides a method for detecting dangerous goods vehicles under a high-point monitoring perspective, comprising: S1: Obtain vehicle identification raw data and perform data annotation on it to obtain vehicle identification annotation data.

[0046] S2: Perform data enhancement processing on the vehicle identification and annotation data to obtain the target detection training set.

[0047] S3: Train the target detection model based on the target detection training set to obtain a dangerous goods vehicle detection model.

[0048] S4: Identify the data to be detected based on the dangerous goods vehicle detection model and output the identification result.

[0049] The dangerous goods vehicle detection method under the high-point monitoring perspective proposed by the present invention can effectively detect targets at the high-point perspective of the tower, which is helpful to improve the management level of various types of road dangerous goods transportation, and can be extended to the identification of other traffic targets. In machine learning and deep learning, data is a key component of the training model. More and more diverse data usually help the model to better generalize and adapt to new data sets. Therefore, the present invention uses a data enhancement scheme to increase the number and diversity of training data to improve the performance and robustness of the machine learning model, reduce the cost of manual annotation, and can quickly train a more robust target detection model to reduce the model iteration cycle and number. Due to the large loss of texture features of image targets collected under the high-point perspective, the present invention uses a machine learning algorithm to train a dangerous goods vehicle detection model to identify "dangerous goods vehicles" and "suspected dangerous goods vehicles" targets, and then uses a low-pixel small target recall strategy for "suspected dangerous goods vehicles" to increase the recall of suspected dangerous goods vehicles and improve the early warning capability of the system.

[0050] Specifically, step S1 obtains the original vehicle identification data and performs data annotation on it, and the obtained vehicle identification annotation data includes: S11: Select multiple vehicle recognition images from the high-point monitoring video as vehicle recognition raw data; S12: Labeling the original vehicle identification data according to three categories: dangerous goods vehicles, suspected dangerous goods vehicles, and non-dangerous goods vehicles, to obtain vehicle identification labeling data.

[0051] During the road transportation of dangerous goods, since there are some special signs installed on the body of dangerous goods, such as warning signs or safety signs, these special signs are identified to distinguish whether it is a dangerous goods vehicle. However, in the visual connection scene of the iron tower, the height of the monitoring camera is generally 15-50 meters. As the monitoring point increases, the recognition problem of small targets and blurred targets will become more and more obvious. The present invention identifies dangerous goods vehicles for the application scenario of high-point monitoring. First, it is necessary to screen out multiple vehicle recognition images in the high-point monitoring video as the original data for training the dangerous goods vehicle detection model. In addition to the pictures of dangerous goods vehicles, the screened data also includes pictures of non-dangerous goods vehicles and suspected dangerous goods vehicles, so as to expand the richness of the scene.

[0052] After filtering out the raw vehicle identification data, the data needs to be labeled. Since the present application is based on data obtained in a high-point monitoring scenario, the present invention defines the target labels as hazardous goods vehicles, suspected hazardous goods vehicles, and non-hazardous goods vehicles, and labels the raw data according to the above three categories, where the "hazardous goods vehicle" label represents a vehicle that can be clearly distinguished as a hazardous goods vehicle, the "suspected hazardous goods vehicle" label represents a vehicle that is difficult to identify directly with the naked eye, and the "non-hazardous goods vehicle" label represents an ordinary vehicle that can be clearly distinguished.

[0053] Specifically, step S2 performs data enhancement processing on the vehicle identification and annotation data to obtain a target detection training set including: S21: Create a dangerous goods vehicle material library based on vehicle identification and annotation data.

[0054] S22: Randomly select a vehicle recognition image from the vehicle recognition annotation data as a background image, and determine the area to be generated.

[0055] S23: randomly selecting a foreground image of a dangerous goods vehicle from the dangerous goods vehicle library and scaling the image.

[0056] S24: placing the scaled foreground image in the to-be-generated area of ​​the background image to generate a target detection image.

[0057] S25: Generate a target detection training set based on multiple target detection images.

[0058] Since during the data collection stage, the frequency of dangerous goods vehicles appearing on actual roads is much lower than that of other ordinary vehicles, a lot of time and cost is required to collect data, and the collected data is not rich enough, which will lead to weak generalization ability of the final detection model and cannot effectively cope with complex and changeable scenarios. At the same time, the trained detection model may have the problem of insufficient robustness. Therefore, the present invention proposes an improved data enhancement method to amplify the original data to increase the diversity of the data.

[0059] Specifically, step S21 of creating a dangerous goods vehicle material library based on vehicle identification and annotation data includes: S211: Select a hazardous goods vehicle image and a corresponding target frame from the vehicle identification and annotation data.

[0060] S212: Using the target frame as prompt information, a semantic segmentation model is used to segment the dangerous goods vehicle image and extract foreground pixel information.

[0061] S213: Creating a dangerous goods vehicle material library based on the foreground pixel information.

[0062] In a specific embodiment of the present invention, first, all vehicle identification images containing the label of "hazardous goods vehicle" are screened out from the annotated vehicle identification raw data, and the target box marked on the image is used as prompt information. The SAM segmentation model is used to perform image segmentation on the hazardous goods vehicle image, and the foreground pixel information of all hazardous goods vehicles is extracted and used as a material library of hazardous goods vehicles.

[0063] Specifically, step S22 randomly selects a vehicle recognition image from the vehicle recognition label data as a background image, and determines the area to be generated. The method for determining the area to be generated includes: Get all target boxes from the background image, select the bottom center point of each target box, and move the bottom center point downward along the vertical axis of the image to one tenth of the height of the corresponding target box. If the moved bottom center point falls into any of the target boxes, delete the bottom center point and use the remaining bottom center point as prompt information. Use the semantic segmentation model to segment the image and obtain the vehicle driving area, which is determined as the area to be generated.

[0064] In a specific embodiment of the present invention, first, a picture is randomly selected from the annotated vehicle recognition image as the background picture, then all the target frames annotated on the background picture are obtained, the bottom edge center point of the target frame is selected, and the bottom edge center point is moved downward along the longitudinal axis of the image, and the moving distance is one tenth of the height of the corresponding target frame. If the moved center point falls into any of the target frames, the bottom edge center point is deleted, and the remaining bottom edge center point is the pixel point on the vehicle driving area. Finally, the remaining bottom edge center point is used as prompt information, and the SAM segmentation model is used to segment the vehicle driving area of ​​the background picture, and the vehicle driving area is used as the area to be generated.

[0065] Specifically, step S23 randomly selects a foreground image of a dangerous goods vehicle from the dangerous goods vehicle material library and performs image scaling, including: S231: Select any pixel point from the area to be generated and determine n nearest target boxes of the pixel point.

[0066] S232: Calculate the average size of the n nearest target frames, and scale the foreground image to the same size as the average size.

[0067] In a specific embodiment of the present invention, first, a foreground image of a hazardous goods vehicle is randomly selected from a hazardous goods vehicle material library, and then any pixel point is selected from the to-be-generated area of ​​the background image as the center point of the generated target, and n nearest target frames around the center point of the generated target are selected, and then the size average S of the n target frames is calculated, and finally the selected foreground image is scaled to the same size as the size average S and placed at the position of the target center point, wherein n is given according to an empirical value, and in the embodiment of the present invention, n=3.

[0068] Specifically, step S24 places the scaled foreground image in the to-be-generated area of ​​the background image, and generating the target detection image includes: Select any pixel point from the area to be generated as the center point of the generated target, place the scaled foreground image at the center point of the generated target, and generate a target detection image if the overlap rate between the mask of the foreground image and the mask of the area to be generated is greater than 0.5.

[0069] In a specific embodiment of the present invention, first, any pixel point is selected from the area to be generated as the center point of the generated target, and then the scaled foreground image is placed at the generated target center point. If the overlap rate of the foreground image mask after placement and the vehicle driving area mask in the background image is greater than 0.5, that is, the overlapping pixel area of ​​the foreground image mask and the vehicle driving area mask divided by the area of ​​the foreground image mask is greater than 0.5, then it is considered that the target detection image is successfully generated, otherwise return to step S22.

[0070] Through the above steps, multiple target detection images are generated, and finally a target detection model training set is formed.

[0071] Specifically, in step S3, the target detection model is trained based on the target detection training set to obtain a dangerous goods vehicle detection model. In a specific embodiment of the present invention, the target detection model used is the yolov8 model, which can be trained for two types of dangerous goods vehicles and suspected dangerous goods vehicles. In another specific embodiment of the present invention, any other target detection model can also be used for training.

[0072] Specifically, step S4 identifies the data to be detected based on the dangerous goods vehicle detection model, and outputs the identification results including: S41: Input the data to be detected into the dangerous goods vehicle detection model for first recognition, and output a first recognition result, including dangerous goods vehicles and suspected dangerous goods vehicles.

[0073] S42: Determine whether the vehicle is suspected of being a dangerous goods vehicle based on the first recognition result. If so, perform a second recognition and output a second recognition result, including the vehicle being a dangerous goods vehicle, specifically including: It is determined whether the target image of the suspected dangerous goods vehicle is smaller than the preset size threshold. If it is smaller than the preset size threshold, the output result is a dangerous goods vehicle. Otherwise, it is determined whether the confidence of the target image of the suspected dangerous goods vehicle meets the preset confidence threshold. If it does, the output result is a dangerous goods vehicle. The preset confidence threshold is 0.5~0.7.

[0074] In a specific embodiment of the present invention, the image to be detected is first input into a trained dangerous goods vehicle detection model for a first recognition, and a first recognition result is output. The output first recognition result is a dangerous goods vehicle or a suspected dangerous goods vehicle. If the output result is a suspected dangerous goods vehicle, it is determined whether the target image of the suspected dangerous goods vehicle is smaller than a preset size threshold. The preset size threshold given in the embodiment of the present invention is 50x50 pixels. If it is smaller than the preset size threshold, the output result is a dangerous goods vehicle. If it is greater than or equal to the preset size threshold, it is determined whether the confidence of the target image of the suspected dangerous goods vehicle meets the preset confidence threshold, that is, 0.5<confidence<0.7. If it meets, the output result is a dangerous goods vehicle. Otherwise, it is discarded. In the present invention, the confidence is given by model prediction. Since the labelers have slightly different judgment scales for "suspected hazardous vehicle" during the actual data labeling process, the inconsistent judgment scales will be reflected in the confidence of the model output target during the experiment, that is, the target with low confidence is more likely to be a hazardous vehicle, and the target with high confidence is more likely to be a non-hazardous vehicle. Therefore, the present invention selects the target with lower confidence for recall, and uses a low confidence threshold recall strategy for the suspected hazardous target, which can increase the recall of suspected hazardous vehicles and improve the early warning capability of the system. The model outputs the target identified as "hazardous vehicle" as the final result of the recognition.

[0075] Example 2 Based on the above method, the embodiment of the present invention also provides a dangerous goods vehicle detection system under a high-point monitoring perspective, such as Figure 9-13 As shown, including: The data collection and data annotation module 01 is used to obtain the original vehicle identification data and annotate it to obtain the vehicle identification annotation data.

[0076] The data enhancement module 02 is used to perform data enhancement processing on the vehicle identification and annotation data to obtain a target detection training set.

[0077] The model creation module 03 is used to train the target detection model based on the target detection training set to obtain a dangerous goods vehicle detection model.

[0078] The vehicle detection module 04 is used to identify the data to be detected based on the dangerous goods vehicle detection model and output the identification result.

[0079] The dangerous goods vehicle detection system under the high-point monitoring perspective proposed by the present invention can effectively detect targets at the high-point perspective of the tower, which is helpful to improve the management level of various types of road dangerous goods transportation, and can be extended to the identification of other traffic targets. In machine learning and deep learning, data is a key component of the training model. More and more diverse data usually help the model to better generalize and adapt to new data sets. Therefore, the present invention uses a data enhancement module to increase the number and diversity of training data to improve the performance and robustness of the machine learning model, reduce the cost of manual annotation, and can quickly train a more robust target detection model to reduce the model iteration cycle and number. Due to the large loss of texture features of image targets collected under the high-point perspective, the present invention uses a machine learning algorithm to train a dangerous goods vehicle detection model to identify "dangerous goods vehicles" and "suspected dangerous goods vehicles" targets, and then uses a low-pixel small target recall strategy for "suspected dangerous goods vehicles" to increase the recall of suspected dangerous goods vehicles and improve the system's early warning capability.

[0080] Specifically, the data collection and data annotation module 01 also includes: The data acquisition module 011 is used to select a plurality of vehicle identification images from the high-point monitoring video as vehicle identification raw data.

[0081] The data labeling module 012 is used to label the original vehicle identification data according to three categories: dangerous goods vehicles, suspected dangerous goods vehicles and non-dangerous goods vehicles, to obtain vehicle identification labeling data.

[0082] The present invention performs hazardous vehicle identification for application scenarios of high-point monitoring. First, a plurality of vehicle identification images are screened out through the data acquisition module 011 as raw data for training a hazardous vehicle detection model. In addition to images of hazardous vehicles, the screened data also includes images of non-hazardous vehicles and suspected hazardous vehicles, so as to expand the richness of the scene.

[0083] After filtering out the original vehicle identification data, the collected data is annotated through the data annotation module 012. Since the present application is data obtained in a high-point monitoring scenario, the present invention defines the target labels as dangerous goods vehicles, suspected dangerous goods vehicles and non-dangerous goods vehicles, and annotates the original data according to the above three categories, where the "dangerous goods vehicle" label represents a vehicle that can be clearly distinguished as a dangerous goods vehicle, the "suspected dangerous goods vehicle" label represents a vehicle that is difficult to directly identify with the naked eye, and the "non-dangerous goods vehicle" label represents an ordinary vehicle that can be clearly distinguished.

[0084] Specifically, the data enhancement module 02 further includes: The material library creation module 021 is used to create a dangerous goods vehicle material library based on vehicle identification and annotation data.

[0085] The to-be-generated region determination module 022 is used to randomly select a vehicle recognition image from the vehicle recognition label data as a background image and determine the to-be-generated region.

[0086] The image scaling module 023 is used to randomly select a foreground image of a dangerous goods vehicle from the dangerous goods vehicle material library and scale the image.

[0087] The target detection image generation module 024 is used to place the scaled foreground image in the to-be-generated area of ​​the background image to generate a target detection image.

[0088] The target detection training set generation module 025 is used to generate a target detection training set based on multiple target detection images.

[0089] In a specific embodiment of the present invention, firstly, the material library creation module 021 is used to filter out all vehicle identification images containing the label "dangerous goods vehicle" from the annotated vehicle identification raw data, and the target frame annotated on the image is used as the prompt information. The SAM segmentation model is used to perform image segmentation on the dangerous goods vehicle image, and the foreground pixel information of all dangerous goods vehicles is extracted, and it is used as the material library of dangerous goods vehicles. Then, the module for determining the area to be generated 022 is used to obtain all target frames from the background image, select the bottom edge center point of each target frame, and move the bottom edge center point downward along the vertical axis of the image to one tenth of the height of the corresponding target frame. If the bottom edge center point after the movement falls into any of the target frames, the bottom edge center point is deleted, and the remaining bottom edge center point is used as the prompt information. The semantic segmentation model is used to segment the image to obtain the vehicle driving area, and the vehicle driving area is determined as the area to be generated. Then, the image scaling module 023 is used to select any pixel point from the area to be generated and determine the n nearest target frames of the pixel point. The size average of the n nearest target frames is calculated, and the foreground image is scaled to the same size as the size average. Finally, the target detection image generation module 024 is used to select any pixel point from the area to be generated as the center point of the generated target, and the scaled foreground image is placed at the center point of the generated target. If the overlap rate between the mask of the foreground image and the mask of the area to be generated is greater than 0.5, the target detection image is generated. The target detection training set generation module 025 is used to generate a target detection model training set from multiple target detection images.

[0090] Specifically, the vehicle detection module 04 also includes: The first identification module 041 is used to input the data to be detected into the dangerous goods vehicle detection model for the first identification, and output a first identification result, including dangerous goods vehicles and suspected dangerous goods vehicles.

[0091] The second identification module 042 is used to determine whether it is a suspected dangerous goods vehicle based on the first identification result. If so, a second identification is performed and a second identification result is output, including the dangerous goods vehicle.

[0092] The second identification module 042 further includes: The first judgment module 0421 is used to judge whether the target image of the suspected dangerous goods vehicle is smaller than a preset size threshold. If it is smaller than the preset size threshold, the output result is a dangerous goods vehicle.

[0093] The second judgment module 0422 is used to judge whether the confidence of the target image of the suspected dangerous goods vehicle meets a preset confidence threshold. If so, the output result is a dangerous goods vehicle.

[0094] In a specific embodiment of the present invention, the first recognition module 041 is first used to input the image to be detected into the trained dangerous goods vehicle detection model for the first recognition, and the first recognition result is output. The output first recognition result is a dangerous goods vehicle or a suspected dangerous goods vehicle. If the output result is a suspected dangerous goods vehicle, the second recognition module 042 is used to determine whether the target image of the suspected dangerous goods vehicle is smaller than a preset size threshold. The preset size threshold given in the embodiment of the present invention is 50x50 pixels. If it is smaller than the preset size threshold, the output result is a dangerous goods vehicle. If it is greater than or equal to the preset size threshold, it is determined whether the confidence of the target image of the suspected dangerous goods vehicle meets the preset confidence threshold, that is, 0.5<confidence<0.7. If it meets, the output result is a dangerous goods vehicle, otherwise, it is discarded. The present invention selects targets with lower confidence for recall and uses a low confidence threshold recall strategy for suspected dangerous goods targets, which can increase the recall of suspected dangerous goods vehicles and improve the early warning capability of the system.

[0095] Example 3 like Fig.14 As shown, an embodiment of the present invention further provides an electronic device, including: a processor and a memory, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement the above-mentioned dangerous goods vehicle detection method under the high-point monitoring perspective. The device in the present invention can be a server, a PC, a PAD, a mobile phone, etc.

[0096] Furthermore, an embodiment of the present invention also provides a computer-readable storage medium, in which a computer program is stored. The computer program is loaded and executed by a processor to implement the above-mentioned method for detecting dangerous goods vehicles under the high-point monitoring perspective.

[0097] In summary, the method and system for detecting dangerous goods vehicles under high-point monitoring viewing angle described in the present invention have the following advantages: 1. The dangerous goods vehicle detection method under high-point monitoring viewing angle proposed in the present invention can effectively solve the problem of unclear target identification and difficulty in analyzing target detail texture information under high-point viewing angle.

[0098] 2. The present invention performs secondary differentiation through the intermediate state of "suspected dangerous goods vehicles", effectively improving the recall rate of dangerous goods vehicle targets and improving the early warning capability of the system.

[0099] 3. The data enhancement method proposed in the present invention can efficiently generate targets including dangerous goods vehicles in different scene background images, while reducing the workload of data collection and effectively improving the generalization ability of the model.

[0100] 4. The dangerous goods vehicle detection system under the high-point monitoring perspective proposed in the present invention can realize automated data collection, data enhancement, and model training after the model is trained, thereby accelerating the optimization and iteration process of the model.

[0101] 5. The hazardous vehicle detection system proposed in the present invention can be quickly applied to target recognition tasks in other high-point traffic scenes. The system can cover a larger monitoring range and effectively reduce the investment cost of monitoring equipment deployment.

[0102] The above describes in detail the concepts, principles and ideas of the present invention in conjunction with specific implementation methods (including embodiments and examples). Those skilled in the art should understand that the implementation methods of the present invention are not limited to the forms given above. After reading this application document, those skilled in the art can make any possible improvements, substitutions and equivalent forms to the steps, methods, systems and components in the above implementation methods. These improvements, substitutions and equivalent forms should be deemed to fall within the scope of the present invention, and the protection scope of the present invention shall be subject only to the claims.

Claims

1. A method for detecting dangerous goods vehicles under high-point monitoring viewing angle, characterized in that: include: Obtaining vehicle identification raw data and annotating it to obtain vehicle identification annotated data; Performing data enhancement processing on the vehicle identification and annotation data to obtain a target detection training set; Training the target detection model based on the target detection training set to obtain a dangerous goods vehicle detection model; The data to be detected is identified based on the dangerous goods vehicle detection model, and an identification result is output.

2. The method for detecting dangerous goods vehicles under high-point monitoring viewing angle according to claim 1 is characterized in that: The obtaining of the original vehicle identification data and the data annotation thereof to obtain the vehicle identification annotation data comprises: Select multiple vehicle identification images from high-point surveillance videos as vehicle identification raw data; The original vehicle identification data is labeled according to three categories: dangerous goods vehicles, suspected dangerous goods vehicles and non-dangerous goods vehicles to obtain vehicle identification labeling data.

3. The method for detecting dangerous goods vehicles under high-point monitoring viewing angle according to claim 2 is characterized in that: The data enhancement processing is performed on the vehicle identification and labeling data to obtain a target detection training set, which includes: Creating a dangerous goods vehicle material library based on the vehicle identification and annotation data; Randomly select a vehicle identification image from the vehicle identification label data as a background image, and determine the area to be generated; Randomly select a foreground image of a dangerous goods vehicle from the dangerous goods vehicle material library and scale the image; Placing the scaled foreground image in the to-be-generated area of ​​the background image to generate a target detection image; Generate an object detection training set based on multiple object detection images.

4. The method for detecting dangerous goods vehicles under high-point monitoring viewing angle according to claim 3 is characterized in that: The creating of a dangerous goods vehicle material library based on the vehicle identification and labeling data comprises: Selecting a dangerous goods vehicle image and a corresponding target frame from the vehicle identification and annotation data; Using the target frame as prompt information, a semantic segmentation model is used to segment the dangerous goods vehicle image and extract foreground pixel information; A dangerous goods vehicle material library is created based on the foreground pixel information.

5. The method for detecting dangerous goods vehicles under high-point monitoring viewing angle according to claim 3 is characterized in that: The method for determining the area to be generated comprises: Obtain all target frames from the background image, select the bottom center point of each target frame, and move the bottom center point downward along the vertical axis of the image to one tenth of the height of the corresponding target frame. If the moved bottom center point falls into any of the target frames, delete the bottom center point, and use the remaining bottom center point as prompt information. Use the semantic segmentation model to segment the image to obtain the vehicle driving area, and determine the vehicle driving area as the area to be generated.

6. The method for detecting dangerous goods vehicles under high-point monitoring viewing angle according to claim 5 is characterized in that: The randomly selecting a foreground image of a dangerous goods vehicle from the dangerous goods vehicle material library and performing image scaling comprises: Select any pixel point from the area to be generated and determine n nearest target frames of the pixel point; An average size of the n nearest target frames is calculated, and the foreground image is scaled to a size that is the same as the average size.

7. The method for detecting dangerous goods vehicles under high-point monitoring viewing angle according to claim 6 is characterized in that: Placing the scaled foreground image in the to-be-generated area of ​​the background image to generate the target detection image includes: Select any pixel point from the area to be generated as the target center point, place the scaled foreground image at the target center point, and generate a target detection image if the mask overlap rate between the foreground image and the mask of the area to be generated is greater than 0.

5.

8. The method for detecting dangerous goods vehicles under high-point monitoring viewing angle according to claim 1 is characterized in that: The identifying the data to be detected based on the dangerous goods vehicle detection model and outputting the identification result includes: Inputting the data to be detected into the dangerous goods vehicle detection model for first identification, and outputting a first identification result, including dangerous goods vehicles and suspected dangerous goods vehicles; Based on the first recognition result, determine whether it is a suspected dangerous goods vehicle. If so, perform a second recognition and output a second recognition result, including the dangerous goods vehicle.

9. The method for detecting dangerous goods vehicles under high-point monitoring viewing angle according to claim 8 is characterized in that: The determining whether the vehicle is suspected of being a dangerous goods vehicle based on the first recognition result, and if so, performing a second recognition, and outputting a second recognition result includes: Determine whether the target image of the suspected dangerous goods vehicle is smaller than a preset size threshold. If so, output the result as a dangerous goods vehicle. Otherwise, determine whether the confidence of the target image of the suspected dangerous goods vehicle meets a preset confidence threshold. If so, output the result as a dangerous goods vehicle.

10. The method for detecting dangerous goods vehicles under high-point monitoring viewing angle according to claim 9 is characterized in that: The preset confidence threshold is 0.5-0.

7.

11. A dangerous goods vehicle detection system under high-point monitoring viewing angle, characterized in that: include: The data collection and data annotation module is used to obtain the original vehicle identification data and annotate it to obtain the vehicle identification annotation data; A data enhancement module is used to perform data enhancement processing on the vehicle identification and annotation data to obtain a target detection training set; A model creation module, used to train the target detection model based on the target detection training set to obtain a dangerous goods vehicle detection model; The vehicle detection module is used to identify the data to be detected based on the dangerous goods vehicle detection model and output the identification result.

12. The dangerous goods vehicle detection system under high-point monitoring viewing angle according to claim 11 is characterized in that: The data collection and data annotation module also includes: A data acquisition module is used to select multiple vehicle identification images from high-point monitoring videos as vehicle identification raw data; The data labeling module is used to label the original vehicle identification data according to three categories: dangerous goods vehicles, suspected dangerous goods vehicles and non-dangerous goods vehicles, to obtain vehicle identification labeling data.

13. The dangerous goods vehicle detection system under high-point monitoring viewing angle according to claim 12 is characterized in that: The data enhancement module also includes: A material library creation module, used to create a dangerous goods vehicle material library based on the vehicle identification and annotation data; A module for determining an area to be generated, used to randomly select a vehicle identification image from the vehicle identification label data as a background image, and determine an area to be generated; An image zooming module is used to randomly select a foreground image of a dangerous goods vehicle from the dangerous goods vehicle material library and zoom the image; A target detection image generation module, used for placing the scaled foreground image in the to-be-generated area of ​​the background image to generate a target detection image; The target detection training set generation module is used to generate a target detection training set based on multiple target detection images.

14. The dangerous goods vehicle detection system under high-point monitoring viewing angle according to claim 11, characterized in that: The vehicle detection module also includes: A first identification module, used for inputting the data to be detected into the dangerous goods vehicle detection model for first identification, and outputting a first identification result, including dangerous goods vehicles and suspected dangerous goods vehicles; The second identification module is used to determine whether it is a suspected dangerous goods vehicle based on the first identification result. If so, a second identification is performed and a second identification result is output, including the dangerous goods vehicle.

15. The dangerous goods vehicle detection system under high-point monitoring viewing angle according to claim 14 is characterized in that: The second identification module also includes: A first judgment module is used to judge whether the target image of the suspected dangerous goods vehicle is smaller than a preset size threshold, and if it is smaller than the preset size threshold, output a result that the target image is a dangerous goods vehicle; The second judgment module is used to judge whether the confidence of the target image of the suspected dangerous goods vehicle meets a preset confidence threshold. If so, the output result is a dangerous goods vehicle.

16. An electronic device, characterized in that: include: A processor and a memory, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement the dangerous goods vehicle detection method under the high-point monitoring perspective described in any one of claims 1 to 10 above.

17. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which is loaded and executed by a processor to implement the method for detecting dangerous goods vehicles under a high-point monitoring perspective as described in any one of claims 1 to 10 above.

Images