Risk road section identification and evaluation method and system

Abstract

The invention discloses a risk road section identification and evaluation method and system, and belongs to the technical field of image processing, and the method specifically comprises the steps: obtaining a target object image generated by an image sensor; feature differences between a target object and a background in the target object image are preliminarily evaluated, a first coordinate transformation model is constructed, the feature differences comprise differences of texture, color and edge features, and the first coordinate transformation model is constructed through feature difference preliminary evaluation and geometric transformation; performing coordinate deformation mapping conversion on the first coordinate conversion model parameters to obtain a second coordinate conversion model, and performing risk road section identification and evaluation based on the second coordinate conversion model; the target object is accurately positioned and classified by fusing the image saliency, the coordinate response and the time sequence change information, and the method is particularly suitable for a complex background, a weak light environment or a dynamic scene.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and in particular relates to a method and system for identifying and evaluating risky road sections. Background Art

[0002] Existing target object recognition methods mostly rely on static feature extraction and fixed coordinate models for recognition. When faced with complex backgrounds, lighting changes, or motion and deformation of target objects, problems such as decreased recognition accuracy, target positioning deviation, or missed detection and misjudgment are prone to occur.

[0003] Existing technologies usually use image processing algorithms to identify target objects, such as edge detection, feature extraction, and region segmentation. However, these image processing methods often find it difficult to completely eliminate background interference, especially for obstacle identification in dynamic environments, which still poses great challenges, resulting in large errors in the identification and assessment of risky road sections.

[0004] Therefore, a new method is needed to solve these problems, especially in the identification of vehicles and roadside obstacles in complex road environments. Summary of the Invention

[0005] In view of the deficiencies of the existing technology, the present invention proposes a method and system for identifying and evaluating risky road sections.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for identifying and evaluating risky road sections, comprising:

[0008] Acquire an image of a target object generated by an image sensor;

[0009] Performing a preliminary evaluation of feature differences between a target object and a background in the target object image to construct a first coordinate transformation model, wherein the feature differences include differences in texture, color, and edge features, and the first coordinate transformation model is constructed through the preliminary feature difference evaluation and geometric transformation;

[0010] Performing deformation mapping on the first coordinate transformation model parameters to obtain a second coordinate transformation model;

[0011] Based on the second coordinate transformation model, risky road sections are identified and evaluated.

[0012] Specifically, a preliminary evaluation is performed on the feature difference between the target object and the background in the target object image to construct a first coordinate transformation model, including:

[0013] Divide the target object image area and extract the features of different areas in the target object image, including texture, color and edge features;

[0014] Perform a preliminary evaluation of the feature difference between the target object and the background to obtain the preliminary feature difference between the target object and the background;

[0015] A first coordinate transformation model is constructed using geometric transformation according to preliminary feature differences between the target object and the background, wherein the preliminary feature differences between the target object and the background are used to determine initial parameters of the coordinate transformation model.

[0016] Specifically, dividing the target object image area and extracting features of different areas in the target object image includes:

[0017] Use image saliency detection to preliminarily divide the target object image and generate preliminary candidate regions;

[0018] Perform superpixel segmentation on the preliminary candidate area and divide it into superpixel areas;

[0019] Use multi-scale to extract texture, color and edge features of superpixel regions.

[0020] Specifically, performing a preliminary evaluation on the feature difference between the target object and the background to obtain the preliminary feature difference between the target object and the background includes:

[0021] The superpixel region is divided into a target object region candidate set R1 and a background region candidate set R2, and the difference score between each region in R1 and R2 and the target object image is calculated;

[0022] The difference scores between each region and the target object image are sorted, and the regions corresponding to the top n difference scores are retained as the preliminary feature differences between the target object and the background.

[0023] Specifically, the method of constructing a preliminary coordinate transformation model using geometric transformation based on the preliminary feature difference between the target object and the background includes:

[0024] Based on the preliminary feature differences between the target object and the background, a preliminary coordinate transformation model is constructed and preliminary geometric transformation parameters are calculated;

[0025] Extract feature points from the retained target object area and background area, determine the mapping relationship between the target object area and the background area, and calculate the preliminary transformation matrix;

[0026] Adjusting the preliminary transformation matrix according to the local feature differences within the retained target object area to obtain a first transformation matrix;

[0027] The preliminary coordinate transformation model is adjusted according to the first transformation matrix to obtain a first coordinate transformation model.

[0028] Specifically, performing coordinate deformation mapping conversion on the first coordinate transformation model parameters to obtain a second coordinate transformation model includes:

[0029] In each frame of the target object image, the target object image is divided into scale levels and a coordinate response map is constructed;

[0030] Analyze the differences in displacement direction, amplitude, and angle between the target object area and the background area, and extract the difference gradient map;

[0031] The coordinate response map is fused with the difference gradient map to obtain the candidate region of the deformation map;

[0032] A coordinate deformation mapping conversion function is generated for each deformation mapping candidate region and uniformly projected into the first coordinate transformation model to obtain a second coordinate transformation model.

[0033] Specifically, the identification and assessment of risky road sections based on the second coordinate transformation model includes:

[0034] According to the second coordinate transformation model, the coordinate change difference of the target object area is calculated;

[0035] According to the coordinate change difference of the target object area, the coordinate change amplitude of the target object area is calculated to generate a response intensity map;

[0036] Select high-value areas in the response intensity map as high-confidence object candidates, and determine the image of the object based on the aggregation characteristics of the high-confidence object candidates;

[0037] Input the object image into the classification neural network to obtain the category of the target object;

[0038] Based on the identified target objects and target object categories, risky road sections are identified and evaluated.

[0039] Specifically, the target objects include vehicles and roadside obstacles.

[0040] A risky road section identification and assessment system, used to implement the risky road section identification and assessment method, comprising: an image acquisition module, a first model building module, a second model building module and a target object recognition module;

[0041] The image acquisition module is used to acquire the target object image generated by the image sensor;

[0042] The first model building module is used to perform a preliminary evaluation of feature differences between the target object and the background in the target object image and build a first coordinate transformation model;

[0043] The second model building module is used to perform deformation mapping on the first coordinate transformation model parameters to obtain a second coordinate transformation model;

[0044] The target object recognition module is used to identify and evaluate risky road sections based on the second coordinate transformation model.

[0045] Specifically, the first model building module includes: a regional feature extraction unit, a preliminary feature difference calculation unit and a first coordinate transformation model building unit;

[0046] The region feature extraction unit is used to divide the target object image region and extract features of different regions in the target object image;

[0047] The preliminary feature difference calculation unit is used to perform a preliminary evaluation on the feature difference between the target object and the background to obtain the preliminary feature difference between the target object and the background;

[0048] The first coordinate transformation model construction unit is configured to construct a first coordinate transformation model by using geometric transformation according to preliminary feature differences between the target object and the background.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] This paper proposes a method and system for identifying and assessing risky road sections. This method dynamically constructs and optimizes a coordinate transformation model based on the characteristic differences between the target object and its background. By fusing a coordinate response map, a difference gradient map, and the coordinate change differences between image frames, a response intensity map is dynamically generated. This method effectively identifies target objects in complex backgrounds, at night, or under occlusion, achieving precise location and classification of the target objects. Compared to existing image processing technologies, this method achieves higher recognition accuracy in complex backgrounds, low-light environments, or dynamic scenes, making it particularly suitable for practical applications such as nighttime, occlusion, and multiple targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A flow chart of a risky road section identification and assessment method provided by the present invention;

[0052] Figure 2 Schematic diagram of region division and feature extraction provided by the present invention;

[0053] Figure 3 A schematic diagram of the second coordinate transformation model provided by the present invention;

[0054] Figure 4 A coordinate change amplitude response intensity diagram provided by the present invention;

[0055] Figure 5 This is an architecture diagram of a risky road section identification and assessment system provided by the present invention. DETAILED DESCRIPTION

[0056] The present application is described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present application, but are not intended to limit the present application in any form. It should be noted that those skilled in the art may make several variations and improvements without departing from the scope of the present application. These all fall within the scope of protection of the present application.

[0057] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0058] It should be noted that, if there is no conflict, the various features in the embodiments of the present application can be combined with each other and are all within the scope of protection of the present application. In addition, although the functional modules are divided in the device schematic and the logical order is shown in the flow chart, in some cases, the steps shown or described can be performed in a different order than the module division in the device or the order in the flow chart. In addition, the words "first", "second", "third", etc. used in this application do not limit the data and execution order, but only distinguish between the same items or similar items with basically the same functions and effects.

[0059] Unless otherwise defined, all technical and scientific terms used in this specification have the same meanings as those commonly understood by those skilled in the art to which this application belongs. The terms used in this specification and in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application. The term "and / or" as used in this specification includes any and all combinations of one or more of the relevant listed items.

[0060] Example 1

[0061] See also Figure 1-Figure 4 The present invention provides an embodiment of a method for identifying and assessing risky road sections. The method can be applied to dynamic detection of vehicles and obstacles on highways, can predict collisions in advance, and automatically identify and assess traffic risky road sections. The method includes the following specific steps:

[0062] Step S1: Acquire a target object image generated by an image sensor.

[0063] It should be noted that the image sensor is a surveillance camera with low-light imaging capabilities and a wide dynamic range. It is suitable for image acquisition at night or under complex lighting conditions. This image acquisition is not only suitable for static image acquisition, but also for video stream acquisition.

[0064] Step S2: Preliminarily evaluate the feature differences between the target object and the background in the target object image, and construct a first coordinate transformation model. The feature differences include differences in texture, color, and edge features. The first coordinate transformation model is constructed through preliminary feature difference evaluation and geometric transformation.

[0065] It should be noted that the target objects include but are not limited to vehicles, roadside obstacles, pedestrians and other obstacles. In a highway environment, this application is particularly suitable for identifying and detecting vehicles and roadside obstacles, and thereby identifying risks.

[0066] The specific steps of step S2 are:

[0067] Step S201: Divide the target object image area and extract features of different areas in the target object image, including texture, color and edge features.

[0068] like Figure 2 As shown, a highway scene with a complex background at night is provided. Step S201 is used to perform region segmentation and feature extraction on the target object image. The specific steps are as follows:

[0069] Step S2011: Use image saliency detection to preliminarily divide the target object image and generate preliminary candidate regions.

[0070] On the one hand, first obtain an image of the target object, such as Figure 2 As shown in the upper left corner, the image contains multiple vehicle targets and a complex night background, including street lights, trees, and car light spots.

[0071] On the other hand, based on image saliency detection, saliency analysis is performed on the target object image to extract the focus area in the image. Through this saliency analysis, multiple preliminary candidate regions are generated, such as Figure 2 As shown in the upper right corner, these areas are considered to contain potential target objects or their edge contours, or are background areas. The preliminary candidate areas have strong information density and contrast.

[0072] Step S2012: Perform superpixel segmentation on the preliminary candidate area to divide it into superpixel areas.

[0073] Each preliminary candidate region is then further refined. In this embodiment, superpixel segmentation is used to generate multiple image blocks with high structural consistency. Each superpixel region has high color consistency and edge continuity.

[0074] like Figure 2 As shown in the lower right area, the preliminary candidate region can be subdivided into several irregular, locally consistent image regions. Figure 2A more detailed segmentation is not shown in the figure. This superpixel division method is more accurate than the traditional rectangular segmentation method and is suitable for extracting the boundaries of complex obstacle shapes.

[0075] Step S2013: Use multi-scale extraction to extract texture, color and edge features of the superpixel area.

[0076] In this step, multi-scale feature extraction operations are performed on all superpixel areas. It should be noted that the existing technology is used to extract texture features in the superpixel area using local binary patterns and gray-level co-occurrence matrix methods, including texture distribution, texture contrast, and texture directionality; the pixels in the superpixel area are converted to HSV space, the color histogram is calculated and normalized, and the color features are extracted; the Canny operator is used to detect boundary information, and the Sobel gradient is combined to calculate the regional edge density to extract the edge features of the superpixel area for modeling the boundary between the obstacle outline and the background.

[0077] Through step S201 described in this embodiment, the distinction between the background area and the target object area can be effectively improved while ensuring regional consistency.

[0078] Step S202: Preliminary evaluation is performed on the feature difference between the target object and the background to obtain a preliminary feature difference between the target object and the background.

[0079] The specific steps of step S202 are:

[0080] Step S2021: Divide the superpixel region into a target object region candidate set R1 and a background region candidate set R2, and calculate the difference score between each region in R1 and R2 and the target object image.

[0081] In this embodiment, after completing the superpixel region feature extraction, all superpixel regions are divided into two sets based on their saliency response, color distribution, position deviation and other indicators: the target object region candidate set R1, including superpixel regions with color and texture similar to the obstacle and dense edge features; the background region candidate set R2, including image edge regions, low saliency response regions, or regions with texture distribution significantly different from obstacle characteristics.

[0082] For each region in R1 and R2, the difference score with the target object image region preliminarily located in the entire image is calculated. The difference score is scored based on texture difference, color difference and edge structure difference. Each difference score comprehensively reflects the overall difference between the current region and the target object in texture, color and edge.

[0083] Step S2022: Sort the difference scores between each region and the target object image, and retain the regions corresponding to the top n difference scores as the preliminary feature differences between the target object and the background.

[0084] In this embodiment, n=5 is set as an example, and the five selected areas include the headlights, the rear contour, the high-contrast edge, and some background interference areas; it is possible to screen out areas with significant feature differences from multiple candidate areas, and further distinguish the target object area and the background area.

[0085] Step S203: determining initial parameters of a coordinate transformation model based on preliminary feature differences between the target object and the background, and constructing a first coordinate transformation model using geometric transformation.

[0086] The specific steps of step S203 are:

[0087] Step S2031: constructing a preliminary coordinate transformation model based on preliminary feature differences between the target object and the background, and calculating preliminary geometric transformation parameters;

[0088] In this embodiment, affine transformation or perspective transformation is introduced as the formal framework of the preliminary coordinate transformation model. The preliminary feature differences between the target object and the background, such as the center point of the headlights, the edge line of the rear of the vehicle, the background street light line, etc., are used to derive the initial control point pairs required for the transformation model, and based on this, a set of geometric transformation parameters are calculated.

[0089] Step S2032: extracting feature points from the retained target object region and background region, determining a mapping relationship between the target object region and the background region, and calculating a preliminary transformation matrix;

[0090] In this step, feature points are extracted from the candidate target object area and the background area respectively to establish a spatial mapping relationship. The scale-invariant feature transform (SIFT) algorithm is used to detect key points, and the image matching algorithm is combined to perform point pair matching. It should be noted that the selection of feature point pairs prioritizes ensuring that there are obvious differences between the point pairs, such as the taillights of a car and the road light poles. Through the feature point pairs, an affine or perspective relationship between the point sets is constructed, and a preliminary transformation matrix is ​​calculated to achieve alignment of the target object area and the background area in the geometric space.

[0091] Step S2033: adjusting the preliminary transformation matrix according to the local feature differences within the retained target object area to obtain a first transformation matrix;

[0092] In this embodiment, the transformation matrix obtained by preliminary calculation has errors caused by background disturbance or target deformation. A local optimization strategy is introduced to perform fine-grained feature difference correction on the retained target object area. Specifically, the transformation error of the feature points in each area is calculated, and a local affine perturbation factor is introduced. The parameters in the preliminary transformation matrix are fine-tuned with the goal of minimizing the coordinate residuals in the area. The transformation matrix is ​​iteratively optimized to finally obtain the first transformation matrix.

[0093] Step S2034: adjusting the preliminary coordinate transformation model according to the first transformation matrix to obtain a first coordinate transformation model.

[0094] In this embodiment, the first transformation matrix is ​​set to T, and the first coordinate transformation model M(x, y) is constructed. Then The three-dimensional coordinates representing the target object are obtained by expanding the two-dimensional coordinates (x, y) in order to include translation transformations in matrix operations. If only two-dimensional coordinates are used, only linear transformations such as rotation, scaling, and symmetry can be expressed, but translation cannot be expressed. It should be noted that the image content of the target object is dynamically updated.

[0095] Step S3: performing deformation mapping on the first coordinate transformation model parameters to obtain a second coordinate transformation model.

[0096] like Figure 3 As shown, in response to slight deformations, posture changes or background interference of the target object in the time series image, on the basis of the first coordinate transformation model, a second coordinate transformation model is further constructed by modeling the differences between the response changes and background interference within the target object area, thereby enhancing the adaptability to target positioning in complex environments.

[0097] The specific steps of step S3 are:

[0098] Step S301: In each frame of the target object image, the target object image is divided into scale levels, and a coordinate response map is constructed.

[0099] On the one hand, to adapt to the deformation of vehicles of different sizes in the image, an image pyramid strategy is adopted in this embodiment to divide each frame of the image into multiple scale levels, for example, the original size, 0.75 times, 0.5 times, etc., and coordinate response modeling is performed separately at each scale.

[0100] On the other hand, in each scale layer, a coordinate response map is constructed based on the position changes after coordinate transformation between the target object area and the background area. Each pixel value in the response map represents the response intensity of the position under the first coordinate transformation model, such as the coordinate offset and residual size. The larger the value, the more sensitive the area is to the coordinate transformation.

[0101] For example, in a nighttime target object image, the headlight area exhibits consistent and high coordinate response values ​​at multiple scale levels due to its high brightness and clear edges, while the background tree shadows exhibit low or no response.

[0102] Step S302: Analyze the differences in displacement direction, amplitude, and angle between the target object area and the background area, and extract a difference gradient map.

[0103] Specifically, the displacement direction, amplitude, and angle changes of the target object area and the background area in adjacent image frames are analyzed to construct a difference gradient map, including: displacement direction, which represents its transformation direction vector; displacement amplitude, which represents the intensity difference of the coordinate response of each area; and angle gradient, which reflects the angular torsion trend after transformation.

[0104] In this way, the sensitivity of the target object region to deformation can be structurally extracted and random changes caused by background noise can be distinguished.

[0105] Step S303: Fusing the coordinate response map with the difference gradient map to obtain a candidate region for deformation mapping;

[0106] The fusion method uses weighted superposition or pixel-level product enhancement to improve the response intensity of the deformation area and finally obtains the deformation mapping candidate area.

[0107] In this embodiment, regions where both the response value and the gradient difference exceed the set threshold are exemplarily selected as candidate regions. These regions are mostly concentrated in: the rear edge of the vehicle, the light area, and the contour area.

[0108] Step S304: generating a coordinate deformation mapping conversion function for each deformation mapping candidate region, and uniformly projecting it into the first coordinate transformation model to obtain a second coordinate transformation model.

[0109] Specifically, for each deformation mapping candidate region, a local coordinate transformation function is constructed based on the local coordinate response of the feature points in the region. The function can be a local affine, local perspective or spline-based nonlinear function f i (x,y), all local transformation functions f i The unified projection is put into the first coordinate transformation model M for fusion. The specific formula is: M1(x,y) represents the second coordinate transformation model, ω i It represents the weight of the i-th local coordinate transformation function, which is dynamically assigned according to the regional response strength and recognition confidence. I represents the number of local coordinate transformation functions.

[0110] Finally, the second coordinate transformation model M1 is obtained, as Figure 3As shown in the figure, the coordinate transformation model not only contains the original global coordinate mapping relationship, but also integrates the response characteristics of local micro-deformation, which can more accurately reflect the real position and posture of the target object in complex scenes.

[0111] Step S4: Based on the second coordinate transformation model, identify and evaluate risky road sections.

[0112] The specific steps of step S4 are:

[0113] Step S401: Calculating the coordinate change difference of the target object area according to the second coordinate transformation model;

[0114] Specifically, in the time series, each frame of the target object image uses the second coordinate transformation model, and the target object area is mapped to the standard coordinate system. In the t-th frame and the t+1-th frame, the coordinates of the key points in the target object area are tracked and compared to obtain the coordinate change difference between the frames.

[0115] Step S402: Calculating the coordinate change amplitude of the target object region based on the coordinate change difference of the target object region, and generating a response intensity map;

[0116] Perform a modulus operation on the coordinate change difference obtained in the previous step to obtain the coordinate change amplitude of the target object area, such as Figure 4 As shown, the bright white closed-loop area in the lower left corner represents the rear of the vehicle, which shows continuous displacement and morphological changes in multiple frames, with a large response amplitude. The small bright block in the middle is a vehicle in the distance, and the dot-shaped bright spot in the upper right corner is the background street light. Its position changes slightly at multiple time points, but the response amplitude is small, and it is sparsely distributed in a dot-like manner. Because street lights are generally evenly spaced, they are relatively static, resulting in a small response amplitude.

[0117] Step S403: selecting high-value areas in the response intensity map as high-confidence object candidate areas, and determining the image of the object based on the aggregation characteristics of the high-confidence object candidate areas;

[0118] In the embodiment, a response intensity threshold is set, high response areas are extracted, and a cluster analysis method is used to identify concentrated and closed high response blocks, such as Figure 4 As shown in the figure, the brightest large area blocks are selected as candidate areas of the main object, such as vehicles, while point-like areas are excluded or treated with low weights, such as street lights or trees with uniform spacing. It should be noted that Figure 4 The examples provided are relatively uniform and obvious.

[0119] It should be noted that the target object image is an image containing the target object, background, obstacles, etc., while the object image only contains the target object.

[0120] Step S404: inputting the object image into the classification neural network to obtain the category of the target object;

[0121] In this embodiment, a lightweight convolutional neural network can be used. After the lightweight convolutional neural network is trained, an object image is input to obtain the category of the target object.

[0122] Step S405: Based on the identified target objects and target object categories, risky road sections are identified and evaluated.

[0123] It should be noted that, in this embodiment, the identified target object categories can not only be used for basic vehicle recognition tasks, but can also be further combined with traffic scene information to achieve automatic recognition and assessment of traffic risk sections.

[0124] Specifically, in continuous video frames, the coordinates of the identified target objects are tracked and the response intensity is monitored, and an information sequence is constructed, including: the target motion trajectory, which records the position changes of the target object in each frame under the second coordinate transformation model; the response intensity evolution diagram, which records the response amplitude changes of the target object area in each frame; and the dynamic aggregation index, which calculates the regional dynamic aggregation characteristics based on the motion trends, spacing convergence and direction consistency of multiple vehicle targets.

[0125] On the other hand, high-risk behavior features are extracted; when any one or more of the following features appear in the same spatial area, the area is marked as a high-risk candidate section. The features include: the vehicle area always shows violent coordinate response in continuous frames, such as frequent sudden braking, deflection, and pause; the target trajectory is distorted or crossed, and the trajectories of multiple vehicles overlap, intersect, or suddenly approach each other in space; the regional response diffusion is enhanced, and the response graph shows a trend of spreading from a single point highlight to the surrounding area, which is suspected to be an accident or sudden aggregation; the inter-frame response suddenly changes, and the response distribution between frames jumps significantly, which is determined to be caused by sudden behavior, such as lane changing and collision; it is determined that there is a traffic risk in the high-risk candidate section.

[0126] For example, in an image sequence, if the system continuously detects: the response value in the rear area of ​​the vehicle remains high; the following vehicles approach frequently; multiple targets suddenly slow down and gather in the intersection area; then the corresponding position of the image is marked as a road section with traffic risks.

[0127] Example 2

[0128] See also Figure 2 , another embodiment provided by the present invention: a risky road section identification and assessment system, comprising: an image acquisition module, a first model construction module, a second model construction module and a target object recognition module;

[0129] The image acquisition module is used to acquire the target object image generated by the image sensor;

[0130] The first model building module is used to perform a preliminary evaluation of feature differences between the target object and the background in the target object image and build a first coordinate transformation model;

[0131] The second model building module is used to perform deformation mapping on the first coordinate transformation model parameters to obtain a second coordinate transformation model;

[0132] The target object recognition module is used to identify and evaluate risky road sections based on the second coordinate transformation model.

[0133] The first model building module includes: a regional feature extraction unit, a preliminary feature difference calculation unit and a first coordinate transformation model building unit;

[0134] The region feature extraction unit is used to divide the target object image region and extract features of different regions in the target object image;

[0135] The preliminary feature difference calculation unit is used to perform a preliminary evaluation on the feature difference between the target object and the background to obtain the preliminary feature difference between the target object and the background;

[0136] The first coordinate transformation model construction unit is configured to construct a first coordinate transformation model by using geometric transformation according to preliminary feature differences between the target object and the background.

[0137] In addition, the parts of the above technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0138] The above-described specific embodiments further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above description is merely a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for identifying and evaluating risky road sections, characterized in that: include: Acquire an image of a target object generated by an image sensor; Performing a preliminary evaluation of feature differences between a target object and a background in the target object image to construct a first coordinate transformation model, wherein the feature differences include differences in texture, color, and edge features, and the first coordinate transformation model is constructed through the preliminary feature difference evaluation and geometric transformation; Performing coordinate deformation mapping conversion on the first coordinate transformation model parameters to obtain a second coordinate transformation model, wherein the coordinate deformation mapping conversion obtains a deformation mapping candidate region by fusing a coordinate response map and a difference gradient map, and generating a coordinate deformation mapping conversion function based on the deformation mapping candidate region for conversion; The coordinate response map is obtained by dividing the target object image into scale levels; The difference gradient map is obtained by analyzing the change difference between the target object area and the background area; Based on the second coordinate transformation model, risky road sections are identified and evaluated.

2. A risky road section identification and assessment method according to claim 1, characterized in that: Preliminarily evaluating the feature difference between the target object and the background in the target object image and constructing a first coordinate transformation model includes: Divide the target object image area and extract the features of different areas in the target object image, including texture, color and edge features; Perform a preliminary evaluation of the feature difference between the target object and the background to obtain the preliminary feature difference between the target object and the background; A first coordinate transformation model is constructed using geometric transformation according to preliminary feature differences between the target object and the background, wherein the preliminary feature differences between the target object and the background are used to determine initial parameters of the coordinate transformation model.

3. A risky road section identification and assessment method according to claim 2, characterized in that: The step of dividing the target object image region and extracting features of different regions in the target object image includes: Use image saliency detection to preliminarily divide the target object image and generate preliminary candidate regions; Perform superpixel segmentation on the preliminary candidate area and divide it into superpixel areas; Use multi-scale to extract texture, color and edge features of superpixel regions.

4. A risky road section identification and assessment method according to claim 3, characterized in that: The preliminary evaluation of the feature difference between the target object and the background to obtain the preliminary feature difference between the target object and the background includes: The superpixel region is divided into a target object region candidate set R1 and a background region candidate set R2, and the difference score between each region in R1 and R2 and the target object image is calculated; The difference scores between each region and the target object image are sorted, and the regions corresponding to the top n difference scores are retained as the preliminary feature differences between the target object and the background.

5. A risky road section identification and assessment method according to claim 4, characterized in that: The method of constructing a preliminary coordinate transformation model using geometric transformation based on the preliminary feature differences between the target object and the background includes: Based on the preliminary feature differences between the target object and the background, a preliminary coordinate transformation model is constructed and preliminary geometric transformation parameters are calculated; Extract feature points from the retained target object area and background area, determine the mapping relationship between the target object area and the background area, and calculate the preliminary transformation matrix; Adjusting the preliminary transformation matrix according to the local feature differences within the retained target object area to obtain a first transformation matrix; The preliminary coordinate transformation model is adjusted according to the first transformation matrix to obtain a first coordinate transformation model.

6. A risky road section identification and assessment method according to claim 5, characterized in that: Performing coordinate deformation mapping conversion on the first coordinate transformation model parameters to obtain a second coordinate transformation model, including: In each frame of the target object image, the target object image is divided into scale levels and a coordinate response map is constructed; Analyze the differences in displacement direction, amplitude, and angle between the target object area and the background area, and extract the difference gradient map; The coordinate response map is fused with the difference gradient map to obtain the candidate region of the deformation map; A coordinate deformation mapping conversion function is generated for each deformation mapping candidate region and uniformly projected into the first coordinate transformation model to obtain a second coordinate transformation model.

7. The method for identifying and evaluating risky road sections according to claim 1, wherein: The identification and assessment of risky road sections based on the second coordinate transformation model includes: According to the second coordinate transformation model, the coordinate change difference of the target object area is calculated; According to the coordinate change difference of the target object area, the coordinate change amplitude of the target object area is calculated to generate a response intensity map; Select high-value areas in the response intensity map as high-confidence object candidates, and determine the image of the object based on the aggregation characteristics of the high-confidence object candidates; Input the object image into the classification neural network to obtain the category of the target object; Based on the identified target objects and target object categories, risky road sections are identified and evaluated.

8. The method for identifying and evaluating risky road sections according to claim 1, wherein: The target objects include vehicles and roadside obstacles.

9. A risky road section identification and assessment system, used to implement a risky road section identification and assessment method according to any one of claims 1 to 6, characterized in that: include: An image acquisition module, a first model building module, a second model building module, and a target object recognition module; The image acquisition module is used to acquire the target object image generated by the image sensor; The first model building module is used to perform a preliminary evaluation of feature differences between the target object and the background in the target object image and build a first coordinate transformation model; The second model building module is used to perform deformation mapping on the first coordinate transformation model parameters to obtain a second coordinate transformation model; The target object recognition module is used to identify and evaluate risky road sections based on the second coordinate transformation model.

10. A risky road section identification and assessment system according to claim 9, characterized in that: The first model building module includes: a regional feature extraction unit, a preliminary feature difference calculation unit and a first coordinate transformation model building unit; The region feature extraction unit is used to divide the target object image region and extract features of different regions in the target object image; The preliminary feature difference calculation unit is used to perform a preliminary evaluation on the feature difference between the target object and the background to obtain the preliminary feature difference between the target object and the background; The first coordinate transformation model construction unit is configured to construct a first coordinate transformation model by using geometric transformation according to preliminary feature differences between the target object and the background.

Images