Edge extraction method and system for target segmentation

Abstract

The invention discloses an edge extraction method and system for target segmentation, and belongs to the field of target segmentation, and the method comprises the steps: obtaining a vehicle collection image set, and constructing an edge detection model based on the vehicle collection image set; acquiring a real-time perception image based on a vehicle visual perception system; inputting the real-time perception image into an edge detection model to obtain a semantic edge probability graph; performing binarization processing on the semantic edge probability graph to obtain a binarized edge graph, and performing pixel clustering on the binarized edge graph to obtain a plurality of continuous edge communication domains; and carrying out edge distribution intensity calculation to obtain a plurality of edge intensities corresponding to the plurality of continuous edge connected domains, and obtaining an edge extraction result based on the plurality of edge intensities. The technical problem of low target segmentation precision caused by poor edge extraction accuracy for target segmentation in the prior art is solved, and the technical effects of improving the edge extraction accuracy for target segmentation and improving the target segmentation precision are achieved.

Description

Technical Field

[0001] The present invention relates to the field of target segmentation, and in particular to an edge extraction method and system for target segmentation. Background Art

[0002] In the image processing of intelligent vehicles, it is necessary to segment the road scene and separate the vehicles, pedestrians, road signs and other targets from the background. The first step of target segmentation is to extract the edge and obtain the contour information of the target. However, traditional edge extraction methods such as Sobel and Canny have problems such as broken and adhered edges, which have an adverse effect on the subsequent segmentation results. Summary of the invention

[0003] The present application provides an edge extraction method and system for target segmentation, aiming to solve the technical problem in the prior art that edge extraction used for target segmentation is poor in accuracy, resulting in low target segmentation precision.

[0004] In view of the above problems, the present application provides an edge extraction method and system for target segmentation.

[0005] The first aspect disclosed in the present application provides an edge extraction method for target segmentation, the method comprising: collecting vehicle collected images, obtaining a vehicle collected image set, and constructing an edge detection model based on the vehicle collected image set; a vehicle visual perception system of an interactive target vehicle, based on a real-time perception image obtained by the vehicle visual perception system; inputting the real-time perception image into the edge detection model, performing semantic edge extraction on the real-time perception image, and obtaining a semantic edge probability map; binarizing the semantic edge probability map to obtain a binarized edge map, and performing pixel clustering on the binarized edge map to obtain multiple continuous edge connection domains based on the pixel clustering results; performing edge distribution strength calculation on the multiple continuous edge connection domains to obtain multiple edge intensities corresponding to the multiple continuous edge connection domains, and obtaining an edge extraction result based on the multiple edge intensities.

[0006] Another aspect disclosed in the present application provides an edge extraction system for target segmentation, the system comprising: a monitoring model construction unit, used to collect vehicle-collected images, obtain a vehicle-collected image set, and construct an edge detection model based on the vehicle-collected image set; a real-time perception image unit, used to interact with the vehicle visual perception system of the target vehicle, based on the real-time perception image obtained by the vehicle visual perception system; a semantic edge extraction unit, used to input the real-time perception image into the edge detection model, perform semantic edge extraction on the real-time perception image, and obtain a semantic edge probability map; an edge connection domain unit, used to binarize the semantic edge probability map, obtain a binarized edge map, and perform pixel clustering on the binarized edge map, and obtain multiple continuous edge connection domains based on the pixel clustering results; an edge extraction result unit, used to calculate the edge distribution strength of the multiple continuous edge connection domains, obtain multiple edge strengths corresponding to the multiple continuous edge connection domains, and obtain an edge extraction result based on the multiple edge strengths.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0008] The invention adopts the following technical solutions: collecting vehicle-collected images, constructing an edge detection model for detecting edges from images; obtaining real-time perception images of the vehicle visual perception system to provide basic information for edge detection; passing the real-time perception images into the edge detection model to extract semantic edges, reflecting the probability information that each position in the image may be an edge; binarizing the semantic edge probability map, and then performing pixel clustering to divide adjacent edge pixels into multiple continuous edge connection domains; for each continuous edge connection domain, calculating the edge distribution strength of the continuous edge connection domain to obtain the edge extraction result, which solves the technical problem of poor edge extraction accuracy for target segmentation in the prior art, resulting in low target segmentation accuracy, and achieves the technical effect of improving the edge extraction accuracy for target segmentation and improving the target segmentation accuracy.

[0009] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A schematic flow chart of an edge extraction method for target segmentation is provided for an embodiment of the present application;

[0011] Figure 2 A schematic flow chart of constructing an edge detection model in an edge extraction method for target segmentation is provided for an embodiment of the present application;

[0012] Figure 3 A structural schematic diagram of an edge extraction system for target segmentation is provided for an embodiment of the present application.

[0013] Explanation of the accompanying drawings: monitoring model construction unit 11, real-time perception image unit 12, semantic edge extraction unit 13, edge connection domain unit 14, edge extraction result unit 15. DETAILED DESCRIPTION

[0014] The overall idea of ​​the technical solution provided by this application is as follows:

[0015] The embodiment of the present application provides an edge extraction method and system for target segmentation. First, a large number of vehicle collection images are collected and a special edge detection model is constructed to ensure efficient edge detection. Then, a real-time perception image is obtained through the vehicle visual perception system to provide input for subsequent processing. By inputting the real-time perception image into the edge detection model, semantic edge extraction is performed to generate a semantic edge probability map. Next, the probability map is binarized to obtain a binary edge map, and pixel clustering is performed to form multiple continuous edge connection domains. Finally, the edge distribution strength of these edge connection domains is calculated to obtain multiple edge strengths, and finally an effective edge extraction result is obtained, which realizes edge extraction in an efficient and accurate manner, providing important support for target segmentation.

[0016] After introducing the basic principles of the present application, various non-limiting implementation methods of the present application will be specifically described below in conjunction with the drawings in the specification.

[0017] Embodiment 1

[0018] like Figure 1 As shown, an embodiment of the present application provides an edge extraction method for target segmentation, the method comprising:

[0019] Collecting vehicle collected images to obtain a vehicle collected image set, and building an edge detection model based on the vehicle collected image set;

[0020] Further, such as Figure 2 As shown, the embodiment of the present application also includes:

[0021] Set the diversity dimension of vehicle image collection;

[0022] Collect vehicle captured images based on the diversity dimension, pre-process the vehicle captured images, and obtain a vehicle captured image set;

[0023] Performing semantic annotation on the vehicle collected image set to obtain a collected image annotation set;

[0024] A semantic edge extraction network is built with an encoding-decoding fully convolutional network structure, and the semantic edge extraction network is trained based on the collected image annotation set to obtain an edge detection model.

[0025] In a preferred embodiment, first, the diversity dimension of vehicle image acquisition is set, and the diversity dimension includes but is not limited to factors such as the position, angle, type, and lighting conditions of the vehicle, which helps to ensure that the acquired images are diverse and representative. Then, according to the set diversity dimension, vehicle acquisition images are acquired, and these images come from different angles, positions, and lighting conditions. Next, the acquired vehicle acquisition images are preprocessed, such as cutting, scaling, data enhancement, etc., to ensure that the size and quality of the acquired vehicle acquisition images are suitable for subsequent processing, so as to form a vehicle acquisition image set.

[0026] Then, all vehicle-collected images in the vehicle-collected image set are semantically annotated to provide labels and descriptions for the edges and other semantic information in the images, and to provide training data for training edge detection models. Subsequently, a semantic edge extraction network is built using an encoder-decoder fully convolutional network structure, which includes an encoder and a decoder. The encoder part uses an existing classification network (such as VGG16) as the backbone network to extract the semantic features of the image in a stacked convolutional layer manner; the connection structure uses dilated and reduced convolutions to process the encoder output to simultaneously improve the semantic information and retain edge details; the decoder part gradually upsamples and connects the feature maps, and finally generates a semantic edge probability map, i.e., a semantic edge prediction result, through convolution classification. In addition, the training process uses the collection image annotation set, and uses optimization methods such as cross entropy loss function and weight decay to train the semantic edge extraction network to ensure that the semantic edge extraction network can accurately detect the edges in the vehicle-collected images, thereby obtaining a reliable edge detection model for real-time perception image processing.

[0027] A vehicle visual perception system of an interactive target vehicle, based on a real-time perception image acquired by the vehicle visual perception system;

[0028] In the embodiment of the present application, the vehicle visual perception system of the target vehicle is first interacted with, for example, communicating with sensors (such as cameras, radars, or lidars) installed on the target vehicle to obtain real-time image data. Once the connection with the vehicle's visual perception system is established, real-time perception images are obtained from these systems, representing the current environment around the target vehicle, for tasks such as edge detection and object segmentation.

[0029] Inputting the real-time perception image into the edge detection model, performing semantic edge extraction on the real-time perception image, and obtaining a semantic edge probability map;

[0030] Furthermore, this step includes:

[0031] Preprocessing the real-time perception image to obtain an input perception image;

[0032] Loading an edge detection model, performing forward calculation on the input perception image, outputting the probability that each pixel in the input perception image belongs to an edge, and generating the semantic edge probability map, wherein the semantic edge probability map is a two-dimensional probability matrix of the same size as the input perception image.

[0033] Furthermore, this step also includes:

[0034] When the image pixels of the input perception image are larger than the input pixels of the edge detection model, setting a sliding window according to the input pixels;

[0035] Sliding the sliding window on the input perception image to obtain a window perception image;

[0036] The window perception images are sequentially input into the edge detection model to obtain the semantic edge probability map.

[0037] In a preferred embodiment, after acquiring the real-time perception image, first, the acquired real-time perception image is preprocessed to prepare for use as an input of the edge detection model. Specifically, the size of the real-time perception image is adjusted to ensure that they meet the input requirements of the edge detection model; the resolution of the real-time perception image is adjusted and scaled to match the size expected by the model, thereby obtaining an input perception image suitable for use as an input of the edge detection model. Then, the edge detection model that has been constructed and trained is loaded, the preprocessed input perception image is input into the edge detection model, and a forward calculation is performed on the input perception image to calculate the probability that each pixel belongs to the edge, and a semantic edge probability map is obtained. The semantic edge probability map is a two-dimensional probability matrix of the same size as the input perception image. In this probability map, the output value corresponding to each pixel represents the probability that the pixel belongs to the edge. These predicted probability values ​​vary between 0 and 1. The higher the predicted probability value, the more likely it is to be the target edge. An obvious semantic edge may have a predicted probability of 0.8 or higher, while a position with a low predicted probability represents a more likely position to belong to the smooth target interior or background area. Therefore, compared with the binary edge map output by the general edge detection algorithm, the probability map predicted by the semantic edge network contains richer information. It can not only obtain a clear edge distribution, but also retain important uncertainty information, which is more conducive to subsequent processing.

[0038] When the actual input perceived image has more pixels than the expected input pixels of the edge detection model, the edge detection model cannot directly process the actual input perceived image. In order to deal with this situation, a sliding window technique is used. First, when the number of pixels of the input perceived image exceeds the number of pixels required by the edge detection model, a sliding window is set according to the size of the actual input pixels. The sliding window is a fixed-size rectangle that is consistent with the expected input size of the model, such as 224x224 pixels. This sliding window slides on the input perceived image to cover different areas of the entire input perceived image.

[0039] Then, use the set sliding window to slide on the input perception image with a certain step size (for example, 32 pixels to the right or down each time). Each time the window moves to a new position, the image in the sliding window is captured as the window perception image. Next, for each window perception image, it is input into the loaded edge detection model in turn. The edge detection model will perform forward calculations on each window perception image and generate a 224x224 pixel semantic edge probability map for each window. The prediction results of each window will be recorded. When the sliding window covers the entire input perception image, the prediction results of all window positions are spliced ​​together to generate the final large image edge prediction result, that is, the semantic edge probability map. Through the sliding window, input perception images of different sizes can be effectively processed, so that the model can adapt to various image sizes and provide accurate edge detection results.

[0040] Binarizing the semantic edge probability map to obtain a binary edge map, performing pixel clustering on the binary edge map, and obtaining a plurality of continuous edge connection domains based on the pixel clustering result;

[0041] Furthermore, this step includes:

[0042] Set the probability binarization threshold;

[0043] Traversing the semantic edge probability map to obtain a first unit pixel;

[0044] Retrieving the unit edge probability of the first unit pixel, subtracting the unit edge probability from the probability binarization threshold, and obtaining a first probability judgment result;

[0045] When the first probability judgment result is positive or 0, setting the value of the first unit pixel to 1;

[0046] When the first probability judgment result is negative, setting the value of the first unit pixel to 0;

[0047] All probability judgment results are counted to obtain a binary edge map.

[0048] Furthermore, this step also includes:

[0049] Traversing each pixel point of the binary edge map to obtain a first binary pixel;

[0050] Retrieving a first adjacent edge pixel set based on the first binary pixel;

[0051] If there is a pixel point with a value of 1 in the first adjacent edge pixel set, the first binary pixel and the adjacent edge pixels are clustered into the same cluster to obtain a plurality of edge pixel clusters;

[0052] A plurality of continuous edge connection domains are obtained according to the plurality of edge pixel clusters.

[0053] In a preferred embodiment, first, a probability binarization threshold is set to convert the semantic edge probability map into a binary edge map. In the target segmentation of vehicle acquisition images, the image size is not fixed, and the lighting conditions of the image vary greatly. In order to adapt to different images and lighting conditions, an adaptive threshold selection method is adopted, that is, the threshold is determined according to the local information around each pixel. For example, using the local adaptive threshold method, the image is divided into small blocks, and different thresholds are used in each small block to adapt to local lighting changes, reduce the need for manual adjustment, and is suitable for images of different sizes and lighting conditions.

[0054] Then, the entire semantic edge probability map is traversed to obtain each pixel therein as the first unit pixel to process each pixel one by one. Next, for the first unit pixel, the corresponding unit edge probability value is retrieved to indicate the probability that the pixel belongs to the edge, and this unit edge probability value is subtracted from the previously set probability binarization threshold to obtain the first probability judgment result. If the first probability judgment result is positive (greater than or equal to zero) or zero, it means that the pixel is judged to be an edge or has a high edge probability. In this case, the value of the pixel is set to 1, indicating that the pixel is an edge. If the first probability judgment result is negative, it means that the pixel is judged to be a non-edge or has a low edge probability. In this case, the value of the pixel is set to 0, indicating that the pixel is not an edge. Finally, all the processed probability judgment results are counted to obtain the final binary edge map, in which each pixel is clearly marked as an edge (1) or a non-edge (0).

[0055] Next, load the binary edge image, and use nested loops to traverse each pixel of the binary edge map. For example, use Python language to traverse the rows of the image through an external loop and traverse the columns of the image through an internal loop to obtain the first binary pixel. Then, obtain the position of the first binary pixel, and check the first adjacent edge pixel set around the first binary pixel to determine which pixels are edge pixels. The first adjacent edge pixel set includes 8 adjacent edge pixels adjacent to the first binary pixel, which are the pixels above, below, left, right and four diagonal directions of the first binary pixel. For each pixel in the first adjacent edge pixel set, check whether their values ​​are 1. If the pixel value is 1, it means that it is an edge pixel, which means that there is an edge connected to the first binary pixel, and these pixels will be aggregated into the same edge pixel cluster. Repeat this process until no new edge pixels can be added, thereby obtaining multiple edge pixel clusters, wherein the pixels in each cluster are interconnected.

[0056] After that, for each edge pixel cluster, a depth-first search algorithm is used to find connected pixels. Starting from a pixel in the edge pixel cluster, its adjacent pixels are recursively visited and marked as visited until it is impossible to continue. This series of visited pixels constitutes a connected domain, which represents a group of interconnected pixels that belong to the same target area. After determining the connected domain, its contour range is calculated by determining the minimum and maximum row and column coordinates of the pixels in the connected domain to obtain the contour representing the connected domain. For each edge pixel cluster, the above process is repeated to obtain the contour range of all connected domains, and the contour range of each connected domain is recorded to obtain multiple continuous edge connected domains.

[0057] The edge distribution strength of the multiple continuous edge connection domains is calculated to obtain multiple edge strengths corresponding to the multiple continuous edge connection domains, and the edge extraction result is obtained based on the multiple edge strengths.

[0058] In the embodiment of the present application, for each continuous edge connection domain, first, the intensity distribution of the edge pixels therein is determined based on the brightness gradient, color difference and other characteristics of the pixels. In the input perception image, for each pixel, its edge intensity value is compared with the adjacent pixels to consider the association between pixels, which helps to identify the continuity and consistency of the edge. Based on the obtained edge intensity value, an intensity distribution map of each continuous edge connection domain can be obtained, where the value of each pixel represents its edge intensity, thereby obtaining multiple edge intensities.

[0059] Then, based on multiple edge intensities, a threshold judgment method is used to determine which edge responses are strong enough to be regarded as target edges. For example, a threshold is set to divide the edges into strong edges and weak edges. Among them, the weak edge response represents some interference or insignificant edges in the image, which will be removed to reduce noise or irrelevant edge information, thereby obtaining an edge extraction result. The edge extraction result is a binary image that only contains the screened strong edges, while the weak edges are removed, making the edge information in the image more accurate and targeted to improve the accuracy of target segmentation.

[0060] In summary, the edge extraction method for target segmentation provided by the embodiment of the present application has the following technical effects:

[0061] Collect vehicle acquisition images, obtain a vehicle acquisition image set, and build an edge detection model based on the vehicle acquisition image set for edge detection in subsequent steps, so as to identify the contour and edge of the target object. The vehicle visual perception system of the interactive target vehicle provides input data for edge detection based on the real-time perception image obtained by the vehicle visual perception system. Input the real-time perception image into the edge detection model, perform semantic edge extraction on the real-time perception image, obtain a semantic edge probability map, and provide probability information that the image may contain an edge. Binarize the semantic edge probability map to obtain a binary edge map, perform pixel clustering on the binary edge map, and obtain multiple continuous edge connection domains based on the pixel clustering results to provide more continuous edge information. Calculate the edge distribution strength of multiple continuous edge connection domains to obtain multiple edge strengths corresponding to the multiple continuous edge connection domains, obtain edge extraction results based on the multiple edge strengths, distinguish each edge connection domain, and provide available edge information for target segmentation.

[0062] Embodiment 2

[0063] Based on the same inventive concept as the edge extraction method for target segmentation in the aforementioned embodiment, Figure 3 As shown, an embodiment of the present application provides an edge extraction system for target segmentation, the system comprising:

[0064] A monitoring model building unit 11 is used to collect vehicle collected images, obtain a vehicle collected image set, and build an edge detection model based on the vehicle collected image set;

[0065] A real-time perception image unit 12, used for interacting with a vehicle visual perception system of a target vehicle, based on a real-time perception image acquired by the vehicle visual perception system;

[0066] A semantic edge extraction unit 13, used for inputting the real-time perception image into the edge detection model, performing semantic edge extraction on the real-time perception image, and obtaining a semantic edge probability map;

[0067] The edge connection domain unit 14 is used to perform binarization processing on the semantic edge probability map to obtain a binarized edge map, perform pixel clustering on the binarized edge map, and obtain a plurality of continuous edge connection domains based on the pixel clustering result;

[0068] The edge extraction result unit 15 is used to calculate the edge distribution strength of the multiple continuous edge connection domains to obtain multiple edge strengths corresponding to the multiple continuous edge connection domains, and obtain edge extraction results based on the multiple edge strengths.

[0069] Furthermore, the monitoring model building unit 11 includes the following execution steps:

[0070] Set the diversity dimension of vehicle image collection;

[0071] Collect vehicle captured images based on the diversity dimension, pre-process the vehicle captured images, and obtain a vehicle captured image set;

[0072] Performing semantic annotation on the vehicle collected image set to obtain a collected image annotation set;

[0073] A semantic edge extraction network is built with an encoding-decoding fully convolutional network structure, and the semantic edge extraction network is trained based on the collected image annotation set to obtain an edge detection model.

[0074] Furthermore, the semantic edge extraction unit 13 includes the following execution steps:

[0075] Preprocessing the real-time perception image to obtain an input perception image;

[0076] Loading an edge detection model, performing forward calculation on the input perception image, outputting the probability that each pixel in the input perception image belongs to an edge, and generating the semantic edge probability map, wherein the semantic edge probability map is a two-dimensional probability matrix of the same size as the input perception image.

[0077] Furthermore, the semantic edge extraction unit 13 further includes the following execution steps:

[0078] When the image pixels of the input perception image are larger than the input pixels of the edge detection model, setting a sliding window according to the input pixels;

[0079] Sliding the sliding window on the input perception image to obtain a window perception image;

[0080] The window perception images are sequentially input into the edge detection model to obtain the semantic edge probability map.

[0081] Furthermore, the edge communication domain unit 14 includes the following execution steps:

[0082] Set the probability binarization threshold;

[0083] Traversing the semantic edge probability map to obtain a first unit pixel;

[0084] Retrieving the unit edge probability of the first unit pixel, subtracting the unit edge probability from the probability binarization threshold, and obtaining a first probability judgment result;

[0085] When the first probability judgment result is positive or 0, setting the value of the first unit pixel to 1;

[0086] When the first probability judgment result is negative, setting the value of the first unit pixel to 0;

[0087] All probability judgment results are counted to obtain a binary edge map.

[0088] Furthermore, the edge communication domain unit 14 further includes the following execution steps:

[0089] Traversing each pixel point of the binary edge map to obtain a first binary pixel;

[0090] Retrieving a first adjacent edge pixel set based on the first binary pixel;

[0091] If there is a pixel point with a value of 1 in the first adjacent edge pixel set, the first binary pixel and the adjacent edge pixels are clustered into the same cluster to obtain a plurality of edge pixel clusters;

[0092] A plurality of continuous edge connection domains are obtained according to the plurality of edge pixel clusters.

[0093] Any step of the method described above can be stored as a computer instruction or program in an unlimited computer memory, and can be called and recognized by an unlimited computer processor to implement any method in the embodiments of the present application, without any unnecessary restrictions.

[0094] Furthermore, the first or second mentioned above may not only represent an order relationship, but may also represent a specific concept, and / or refer to multiple elements that can be selected individually or in full. Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application intends to include these modifications and variations.

Claims

1. An edge extraction method for target segmentation, characterized in that: The method comprises: Collecting vehicle collected images to obtain a vehicle collected image set, and building an edge detection model based on the vehicle collected image set; A vehicle visual perception system of an interactive target vehicle, based on a real-time perception image acquired by the vehicle visual perception system; Inputting the real-time perception image into the edge detection model, performing semantic edge extraction on the real-time perception image, and obtaining a semantic edge probability map; Binarizing the semantic edge probability map to obtain a binary edge map, performing pixel clustering on the binary edge map, and obtaining a plurality of continuous edge connection domains based on the pixel clustering result; The edge distribution strength of the multiple continuous edge connection domains is calculated to obtain multiple edge strengths corresponding to the multiple continuous edge connection domains, and the edge extraction result is obtained based on the multiple edge strengths.

2. The method according to claim 1, characterized in that Collecting vehicle collected images to obtain a vehicle collected image set, and building an edge detection model based on the vehicle collected image set, including: Set the diversity dimension of vehicle image collection; Collect vehicle captured images based on the diversity dimension, pre-process the vehicle captured images, and obtain a vehicle captured image set; Performing semantic annotation on the vehicle collected image set to obtain a collected image annotation set; A semantic edge extraction network is built with an encoding-decoding fully convolutional network structure, and the semantic edge extraction network is trained based on the collected image annotation set to obtain an edge detection model.

3. The method according to claim 1, characterized in that Inputting the real-time perception image into the edge detection model, performing semantic edge extraction on the real-time perception image, and obtaining a semantic edge probability map, including: Preprocessing the real-time perception image to obtain an input perception image; Loading an edge detection model, performing forward calculation on the input perception image, outputting the probability that each pixel in the input perception image belongs to an edge, and generating the semantic edge probability map, wherein the semantic edge probability map is a two-dimensional probability matrix of the same size as the input perception image.

4. The method according to claim 3, characterized in that The method further comprises: When the image pixels of the input perception image are larger than the input pixels of the edge detection model, setting a sliding window according to the input pixels; Sliding the sliding window on the input perception image to obtain a window perception image; The window perception images are sequentially input into the edge detection model to obtain the semantic edge probability map.

5. The method according to claim 1, characterized in that Binarizing the semantic edge probability map to obtain a binary edge map includes: Set the probability binarization threshold; Traversing the semantic edge probability map to obtain a first unit pixel; Retrieving the unit edge probability of the first unit pixel, subtracting the unit edge probability from the probability binarization threshold, and obtaining a first probability judgment result; When the first probability judgment result is positive or 0, setting the value of the first unit pixel to 1; When the first probability judgment result is negative, setting the value of the first unit pixel to 0; All probability judgment results are counted to obtain a binary edge map.

6. The method according to claim 1, characterized in that Pixel clustering is performed on the binary edge map, and a plurality of continuous edge connection domains are obtained based on the pixel clustering result, including: Traversing each pixel point of the binary edge map to obtain a first binary pixel; Retrieving a first adjacent edge pixel set based on the first binary pixel; If there is a pixel point with a value of 1 in the first adjacent edge pixel set, the first binary pixel and the adjacent edge pixels are clustered into the same cluster to obtain a plurality of edge pixel clusters; A plurality of continuous edge connection domains are obtained according to the plurality of edge pixel clusters.

7. An edge extraction system for target segmentation, characterized in that: The system comprises: A monitoring model building unit, the monitoring model building unit is used to collect vehicle collected images, obtain a vehicle collected image set, and build an edge detection model based on the vehicle collected image set; A real-time perception image unit, the real-time perception image unit is used for interacting with a vehicle visual perception system of a target vehicle, based on a real-time perception image acquired by the vehicle visual perception system; A semantic edge extraction unit, wherein the semantic edge extraction unit is used to input the real-time perception image into the edge detection model, perform semantic edge extraction on the real-time perception image, and obtain a semantic edge probability map; An edge connection domain unit, the edge connection domain unit is used to perform binarization processing on the semantic edge probability map, obtain a binarized edge map, perform pixel clustering on the binarized edge map, and obtain a plurality of continuous edge connection domains based on the pixel clustering result; The edge extraction result unit is used to calculate the edge distribution strength of the multiple continuous edge connection domains to obtain multiple edge strengths corresponding to the multiple continuous edge connection domains, and obtain edge extraction results based on the multiple edge strengths.

Images