A monocular vision-based vehicle driving environment dynamic risk map construction method

Abstract

The application discloses a kind of single vision-based vehicle driving environment dynamic risk map construction method, it is related to image recognition technical field, it is solved in the existing monocular vision perception technology, due to target jitter, depth overhang and repeated counting phenomenon, leading to the accuracy and reliability of the risk map finally output is lower;Based on reference support field analysis each target's ground state, contact point coordinates and foot point depth;Based on projection strategy and foot point depth, the contact point coordinates are projected, time sequence fusion and space deduplication, obtain obstacle candidate point set;And construct BEV occupancy grid map and BEV risk grid map;Based on pixel risk data and BEV occupancy grid map, generate image field risk map;And construct cognitive risk map;Output cognitive risk map, image field risk map and BEV risk grid map;The accuracy and reliability of final risk map output are improved.

Description

Technical Field

[0001] This application belongs to the field of image recognition technology, specifically a method for constructing a dynamic risk map of a vehicle driving environment based on monocular vision. Background Technology

[0002] Vehicle path planning and decision-making typically rely on computational representations of the environment, such as occupancy grid maps, accessibility maps, or potential cost maps. In autonomous driving systems, the visual perception module usually outputs results in an image coordinate system, such as 2D bounding boxes and segmentation masks. The planning module needs to convert this information into ground coordinates or BEV grids for path search and collision constraints. For cost-sensitive passenger vehicle driver assistance systems, low-speed unmanned delivery vehicles, or park logistics vehicles, monocular camera solutions offer significant deployment advantages.

[0003] However, mapping monocular vision perception results to BEV risk maps presents numerous challenges, such as unstable footpoint sampling and monocular ranging being highly dependent on the target's grounding point's position in the image. In real-world scenarios, the target's bottom is often affected by occlusion, shadow interference, and segmentation edge jitter, causing the selected footpoints to drift. The traditional "hard projection" method, which directly uses the midpoint of the detection box's bottom edge as the grounding point, easily leads to drastic jumps in the obstacle's position within the BEV space, causing frequent replanning or sudden braking by the planning module. Other challenges include depth noise and outliers. Existing technologies typically rely on monocular depth estimation networks to directly recover distance, but in dynamic scenarios, due to vehicle bumps, road reflections, or target occlusion, depth estimation often exhibits local outliers, resulting in "floating" or "ground-penetrating" phenomena in the BEV projection results. Furthermore, in dynamic scenarios, the same physical entity may be detected by different objects. The algorithm outputs multiple candidate boxes or mask fragments; if these are directly superimposed onto the BEV grid, the risk value of that area will be erroneously amplified, or even block feasible paths. These shortcomings result in a large number of duplicate or false targets in the risk map generated based on monocular vision perception, leading to a significant discrepancy between the output obstacle distribution based on the risk map and the actual obstacle distribution, resulting in low accuracy and reliability of the final risk map output. Therefore, a scheme is needed that can stably generate BEV risk maps and accessibility grids under monocular conditions through environmental semantic constraints, statistical filtering correction, and spatial deduplication strategies, providing a temporally continuous and usable cost field input for path planning. Summary of the Invention

[0004] This application provides a method for constructing a dynamic risk map of a vehicle driving environment based on monocular vision. It solves the technical problem that existing risk map construction methods based on monocular vision perception inherently result in a large number of duplicate or false targets in the risk map, causing a large discrepancy between the obstacle distribution in the output risk map and the actual obstacle distribution, leading to low accuracy and reliability of the final risk map output.

[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, a method for constructing a dynamic risk map of the vehicle driving environment based on monocular vision is provided, including: Acquire an image sequence of the vehicle, wherein the image sequence is an image sequence output by the vehicle's forward-facing monocular camera; perform target perception processing on the current frame image in the image sequence to obtain a target set; the target set includes each target, as well as the target's corresponding category, confidence level, historical motion state, and instance mask; Obtain environmental constraint information of the scene where the vehicle is located in the current frame, the environmental constraint information including a drivable area mask; construct a reference support domain based on the environmental constraint information; the reference support domain is the region corresponding to the region mask in the coordinate system; Based on the reference support domain, the landing constraints and depth corrections of each target are performed to obtain the ground-hugging state, contact point coordinates and landing depth of each target. The projection strategy is determined based on the ground contact state. Based on the projection strategy and the foot point depth, the corresponding contact point coordinates are projected onto the bird's-eye view, i.e., the BEV map, to obtain the obstacle candidate point set. The projection strategy includes precise projection and conservative projection. Based on the various targets in the BEV map, temporal fusion and spatial neighborhood deduplication are performed to obtain an updated obstacle candidate point set; and a BEV occupancy grid map and a BEV risk grid map are constructed based on the obstacle candidate point set. Pixel risk data is acquired, and an image domain risk map is generated based on the pixel risk data and the BEV occupancy raster map. The image domain risk value corresponding to each target is calculated. The pixel risk data includes semantic risk items, depth risk items, and future trajectory risk items corresponding to each pixel. Calculate the BEV risk value for each target; fuse the image domain risk value and the BEV risk value to obtain the fused risk of the target, and construct a cognitive risk map; Output cognitive risk map, image domain risk map, and BEV risk raster map.

[0006] Based on the above technical solution, in the method for constructing a dynamic risk map of a vehicle driving environment based on monocular vision provided in this application, the following steps are taken: First, the target set corresponding to the current frame image in the vehicle image sequence and the reference support domain corresponding to the vehicle scene are obtained. Second, foot-point landing constraints and depth corrections are applied to each target based on the reference support domain to obtain the ground-touching state, contact point coordinates, and foot-point depth corresponding to each target. Third, a projection strategy is determined based on the ground-touching state, and the corresponding contact point coordinates are projected onto the BEV map based on the projection strategy and foot-point depth to obtain an obstacle candidate point set. Fourth, temporal fusion and spatial neighborhood deduplication are performed on each target in the BEV map to obtain an updated obstacle candidate point set. Fifth, a BEV occupancy grid map and a BEV risk grid map are constructed based on the obstacle candidate point set. Sixth, pixel risk data is obtained, and an image domain risk map is generated based on the pixel risk data and the BEV occupancy grid map. Finally, the image domain risk value corresponding to each target is calculated. Calculate the BEV risk value for each target; fuse the image domain risk value and the BEV risk value to obtain the fused risk of the target, and construct a cognitive risk map; output the cognitive risk map, the image domain risk map, and the BEV risk grid map; by correcting the foot points of the identified targets, deduplicating spatial neighborhoods, and performing multi-risk fusion output, the number of duplicate or false targets in the risk map is reduced, significantly narrowing the gap between the output obstacle distribution based on the risk map and the actual obstacle distribution, and improving the accuracy and reliability of the final risk map output; fundamentally reducing coordinate drift caused by the instability of monocular depth estimation and projection errors; compared with the control scheme that does not use ground-level depth, the complete implementation scheme can significantly reduce the output of duplicate or false targets at the BEV level, making the perceived obstacle set closer to the real distribution of the physical world, thereby outputting a clearer and more interpretable BEV occupancy and risk grid.

[0007] In conjunction with the first aspect above, in one possible implementation, the step of performing foothold landing constraints and depth correction on each target based on the reference support domain to obtain the ground-hugging state, contact point coordinates, and foothold depth corresponding to each target includes: Construct sampling strips for each target; calculate the spatial consistency between the sampling strip and the environmental constraint information to generate the ground-hugging state of the target and determine the ground contact point coordinates for each target; perform robust correction on the foot point depth based on the ground contact point coordinates to obtain the final foot point depth of the target.

[0008] In conjunction with the first aspect above, in one possible implementation, the sampling strips are constructed in the following ways: Extract the mask corresponding to target i pixel set : ; in, For the goal Corresponding mask pixel set The pixel coordinates in the image, and The x-axis is... The vertical axis is used as the coordinate. Define the position of the bottom edge of target i. : ; In the mask The bottom edge region is constructed to a height of k The sampling strips correspond to the sampling domain : ; Where k is the set sampling strip height; The spatial consistency between the sampling strip and the environmental constraint information is calculated to generate the ground-hugging state of the target and determine the ground contact point coordinates of each target. Based on the ground contact point coordinates, the foot depth is robustly corrected to obtain the final foot depth of the target.

[0009] In conjunction with the first aspect above, in one possible implementation, the step of calculating the spatial consistency between the sampling strip and the environmental constraint information, generating the ground-hugging state of the target, and determining the ground contact point coordinates corresponding to each target includes: Several candidate footpoints were obtained based on the sampling strips. : ; ; in, It is a median function; For the target number The sampling domain of the corresponding sampling strip; Spatial consistency is measured based on the sampling strips and the reference support domain to obtain the overlap ratio. : ; That is, determining the overlap ratio between the region corresponding to the sampling strip and the region corresponding to the reference support domain; where, Let it be a very small positive number. For reference support domain; when the sampling strip corresponds to the sampling domain When empty, it avoids the denominator being zero, thus making the overlap ratio meaningful; at the same time, it ensures that targets that fail to reach the ground are not directly discarded, but are used in subsequent conservative projection strategies for the modeling domain unless the bottom layer is covered, thus avoiding the situation where targets are not identified due to discarding targets. When the overlap ratio is greater than or equal to the set ground contact threshold, the ground contact state corresponding to the target is set to ground contact; that is... When the overlap ratio is less than the set ground-touching threshold, the ground-touching state of the target is set to non-ground-touching; that is... When this happens, it indicates that the target's ground-hugging recognition failed, and the target's ground-hugging state is changed to non-ground-hugging; among them, The threshold value for determining ground contact is set by experts based on experience. Define the goal Corresponding ground contact point coordinates: Obtain the target Corresponding candidate footpoints ; When the candidate footpoint is not within the reference support domain, i.e. Then, the search proceeds vertically downwards along the image to find the first pixel that falls into the reference support domain, and the coordinates corresponding to that pixel are designated as the ground contact point coordinates; specifically, the ground contact point coordinates are... ;in ; ; Since the search is performed in the vertical direction, the horizontal coordinates of the candidate foot points are the same as the horizontal coordinates of their corresponding ground contact points. The vertical coordinates are obtained by searching within a set range. Specifically, the settings are... For the vertical region range corresponding to the candidate footpoints of target i, in The search is performed within the range, and the first point that falls into the reference support domain is recorded as the ground contact point; This is the maximum search distance set for vertical search; It is worth noting that when no corresponding ground contact point is found within the set range on the ground, the coordinates of the nearest reference support domain pixel point are searched within the candidate foot point's domain as the ground contact point coordinates; and the ground-attached state corresponding to the target is set to an untrusted state.

[0010] In conjunction with the first aspect above, in one possible implementation, the robust correction of the foot depth based on the ground contact point coordinates to obtain the final foot depth of the target includes: Perform local robust statistics on deep patching to determine ground contact points. Then, a depth image block is extracted centered on the ground contact point; Depth image patch centered on ground contact point The size s of the depth image patch is based on the vertical coordinate of the target in the image. Dynamic adjustment; Depth correction is performed by robust statistical filtering sampling within the reference support domain. The corrected footpoint depth is: ; in, The corrected foot point depth, It is a median function; Calculate the ground mass fraction domain ground-to-ground confidence coefficient corresponding to target i; Get the original depth of the foot point , ; Calculate depth deviation , ; Calculation of ground mass fraction based on depth deviation; ; in, Ground mass fraction; The overlap ratio; These are the weighting coefficients. This is the depth deviation scale parameter, used to control the rate of exponential decay; The ground contact determination threshold is obtained, and the ground contact confidence score is calculated based on the ground contact determination threshold and the ground quality score, specifically using a Sigmoid mapping. ; in, For ground-level confidence; For the set slope, The threshold for determining ground contact; The final footpoint depth is obtained by revising the corrected footpoint depth and the original footpoint depth based on the ground confidence level. ; The system integrates and outputs the ground contact point coordinates, corrected foot depth, overlap ratio, ground quality score, and ground contact confidence score.

[0011] In conjunction with the first aspect above, in one possible implementation, determining the projection strategy based on the ground-hugging state includes: When the ground-hugging state is ground-hugging, the projection strategy is set to precise projection; When the ground-hugging state is non-ground-hugging, the projection strategy is set to conservative projection.

[0012] In conjunction with the first aspect above, in one possible implementation, the step of projecting the corresponding contact point coordinates onto the BEV map based on the projection strategy and foot depth to obtain a set of obstacle candidate points includes: The coordinates of the ground contact point are back-projected onto the camera coordinate system to obtain the first ground contact point coordinate; Transform the ground contact point coordinates to the BEV coordinate system and construct the obstacle candidate points corresponding to the target; When the projection strategy corresponding to the target is conservative projection, the obstacle radius corresponding to the obstacle candidate point is extracted, and the obstacle radius is conservatively extended to obtain the final obstacle radius; the obstacle radius is used to replace the corresponding obstacle radius in the obstacle candidate point to obtain the final obstacle candidate point. The individual obstacle candidate points are integrated into an obstacle candidate point set.

[0013] In conjunction with the first aspect mentioned above, one possible implementation involves performing temporal fusion and spatial neighborhood deduplication on each target in the BEV graph to obtain an updated set of obstacle candidate points; including: Obtain the center coordinates and radius of obstacles corresponding to targets with the same ID number in each frame image; For each target, the obstacle center coordinates and obstacle radius are smoothed by exponential moving average to obtain the obstacle center coordinates and obstacle radius of each target in the current frame image after exponential moving average smoothing. Based on the obstacle center coordinates and obstacle radius, the obstacle center coordinates and obstacle radius of the obstacle candidate point set are updated. Obtain obstacle candidate points of the same category from the obstacle candidate point set; obtain the obstacle center coordinates and obstacle radius corresponding to each obstacle candidate point; calculate the BEV plane Euclidean distance between two obstacle candidate points based on the obstacle center coordinates; and calculate the scale similarity between two obstacle candidate points based on the obstacle radius. When the Euclidean distance in the BEV plane is less than a set distance threshold, and the scale similarity is also less than a set scale similarity threshold, the two obstacle candidate points are determined to be duplicate targets; thus, a set of duplicate candidate points is obtained; the set of candidate points includes several obstacle candidate points, and a representative candidate point of the candidate point set is generated based on each obstacle candidate point in the candidate point set; the corresponding obstacle candidate point in the obstacle candidate point set is replaced with the representative candidate node to obtain an updated set of obstacle candidate points; To eliminate duplicate observations of the same physical obstacle, the center coordinates of candidate point pairs with the same category or belonging to the same planning-related category set are used. and Calculate its Euclidean distance in the BEV plane: ; At the same time, combined with its physical scale radius , Calculate scale similarity: ; When both conditions are met and , it is determined that the pair of candidate points are duplicate targets, wherein the distance threshold can be configured according to the scenario and grid resolution.

[0014] In combination with the first aspect described above, in a possible implementation, the constructing a BEV occupancy grid map and a BEV risk grid map based on the obstacle candidate point set includes: acquiring the obstacle center coordinates and obstacle radius corresponding to each obstacle candidate point in the obstacle candidate point set, establishing a BEV grid, and constructing a circular kernel with the obstacle radius as the radius to perform filling and occupancy on the obstacle center coordinates: ; wherein, represents the occupancy state of the cell in the i-th row and j-th column in the BEV grid at time t, represents occupied when; represents unoccupied when; represents the coordinates of the center corresponding to the grid cell in the ground coordinate system; is the obstacle center coordinate of the obstacle candidate point corresponding to number i; is the obstacle radius; performing filling and occupancy on each obstacle center coordinate; to obtain the BEV occupancy grid map; acquiring the risk value corresponding to each grid cell and the basic risk level of the target to which the grid cell belongs; selecting and obtaining the current risk value of the grid cell through a formula; ; wherein, represents the current risk value of the grid cell, which is a scalar, usually a number normalized to the range of 0 to 1; represents the basic risk level of target i, which specifically depends on the type, speed, etc., for example: pedestrians have high risk, stationary objects have low risk; represents the distance from the center of the cell to the center of the target; represents the obstacle radius; the basic risk level is obtained in the following manner: ; wherein, represents a truncation / clipping function; is the inherent risk weight of the category; is the motion state correction factor; is the spatial position correction factor; is the perception confidence correction factor; constructing a BEV risk grid map based on the risk value of each grid cell.

[0015] In conjunction with the first aspect above, in one possible implementation, the step of generating an image domain risk map based on pixel risk data and a BEV occupancy raster map, and calculating the image domain risk value corresponding to each target, includes: Extracting semantic risk items from pixel risk data Deep risk items and future trajectory risk items ; The semantic risk item It is calculated using the following formula: ; in, For pixels The prior risk weights are set according to the category. For example, the prior risk weights for each pixel in the category of vehicle and person are set to be large, while the prior risk weights for each pixel in the category of road surface are set to 0. Semantic risk reflects the inherent danger level of different categories of objects. The deep risk item It is calculated using the following formula: ; in, For pixels The physical distance mapped to the depth value; The set proximity threshold, The threshold value is set at a certain distance. The risk item of future trajectory It is calculated using the following formula: ; in, Indicates the future number One predicted moment, K It is the set of predicted time steps. Indicates the first An obstacle in the future Predicted image coordinates, This represents the standard deviation of the Gaussian kernel; semantic risk items Deep risk items and future trajectory risk items Risks associated with obtaining the pixel position of the corresponding pixel through linear fusion: ; in, For pixels Corresponding pixel location risk; , and The weighting coefficients are set for the semantic risk item, the depth risk item, and the future trajectory risk item, respectively. This represents the truncation / limiting function; The pixel location risk corresponding to each pixel is calculated sequentially, and an image domain risk map is constructed based on the pixel location risk. Obtain the set of pixels in the filled area occupied by each target in the BEV-occupied raster image; calculate the mean of the pixel position risk of each pixel in the image domain risk map, and use it as the image domain risk value corresponding to the target.

[0016] In conjunction with the first aspect above, in one possible implementation, the fusion risk of the target obtained by fusing image domain risk values ​​and BEV risk values ​​includes: Calculate the BEV risk value for each objective: ; in, For the goal The corresponding BEV risk value; The category weight of the target category. Represents the vertical coordinate. The longitudinal distance attenuation threshold; Indicates the horizontal coordinate; The lateral coordinates of the lane centerline. The standard deviation of the lateral positional distribution; This is a reference threshold; For activation functions; Through formula Calculate the target The corresponding integration risks.

[0017] The beneficial effects of adopting the above technical solution are as follows: By correcting the footpoints of identified targets, deduplicating spatial neighborhoods, and fusing multiple risks into the output, this method avoids the possibility of a large number of duplicate or false targets in the risk map. This fundamentally reduces coordinate drift caused by unstable monocular depth estimation and projection errors, and significantly narrows the gap between the output obstacle distribution based on the risk map and the actual obstacle distribution. Compared with the control scheme that does not use ground-level depth, this method can significantly reduce the output of duplicate or false targets at the BEV level, making the perceived obstacle set closer to the real distribution in the physical world, and improving the accuracy and reliability of the final risk map output. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram illustrating the steps of the method for constructing a dynamic risk map of the vehicle driving environment in this application. Detailed Implementation

[0020] The technical solutions of this application will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application. Specific Implementation Example 1 Please see Figure 1 The first aspect of this application provides a method for constructing a dynamic risk map of a vehicle driving environment based on monocular vision, including: Acquire an image sequence of the vehicle, wherein the image sequence is an image sequence output by the vehicle's forward-facing monocular camera; perform target perception processing on the current frame image in the image sequence to obtain a target set; target perception is to identify each target in the frame image, and a trained image recognition model can be used for target perception. It is understood that the training process of the image recognition model is a relatively existing technology, which will not be elaborated on here; the target set includes each target, as well as the category, confidence level, historical motion state and instance mask of the target; The environmental constraint information of the scene where the vehicle is located in the current frame is obtained, and the environmental constraint information includes a drivable area mask; a reference support domain is constructed based on the environmental constraint information; the reference support domain is the region corresponding to the region mask in the coordinate system; specifically, the reference support domain includes one or more of the following combinations: a road area mask or a drivable area mask generated based on semantic segmentation, a reference plane projection region set based on geometric priors, and an effective region based on historical frame temporal propagation. In the image coordinate system, pixel coordinates are denoted as The x-axis The horizontal axis is positive to the right in this embodiment; the vertical axis is... The vertical coordinates are used, with downward being positive in this embodiment; the image width and height are respectively... Therefore , The current frame depth map is D. The environmental constraints are represented by a binary graph, i.e., the reference support domain. The two-dimensional region information of target i is the mask. ; Based on the reference support domain, the landing constraints and depth corrections of each target are performed to obtain the ground-hugging state, contact point coordinates and landing depth of each target. The projection strategy is determined based on the ground contact state. The corresponding contact point coordinates are projected onto the BEV map based on the projection strategy and foot depth to obtain a set of obstacle candidate points. The projection strategy includes precise projection and conservative projection. Based on the various targets in the BEV map, temporal fusion and spatial neighborhood deduplication are performed to obtain an updated obstacle candidate point set; and a BEV occupancy grid map and a BEV risk grid map are constructed based on the obstacle candidate point set. Pixel risk data is acquired, and an image domain risk map is generated based on the pixel risk data and the BEV occupancy raster map. The image domain risk value corresponding to each target is calculated. The pixel risk data includes semantic risk items, depth risk items, and future trajectory risk items corresponding to each pixel. Calculate the BEV risk value for each target; fuse the image domain risk value and the BEV risk value to obtain the fused risk of the target, and construct a cognitive risk map; Output a cognitive risk map, an image domain risk map, and a BEV risk raster map; the specific structured risk set represented by the cognitive risk map is as follows: ; in, For the goal Target ID, For the goal Category For the goal BEV coordinates For the goal obstacle radius, For the goal Integration risks For the goal Ground-level confidence; BEV risk grid; used for planning grid input.

[0022] Based on the above technical solution, in the method for constructing a dynamic risk map of a vehicle driving environment based on monocular vision provided in this application, the following steps are taken: First, the target set corresponding to the current frame image in the vehicle's image sequence and the reference support domain corresponding to the vehicle scene are obtained. Second, foot-point landing constraints and depth corrections are applied to each target based on the reference support domain to obtain the ground-touching state, contact point coordinates, and foot-point depth corresponding to each target. Third, a projection strategy is determined based on the ground-touching state, and the corresponding contact point coordinates are projected onto the BEV map based on the projection strategy and foot-point depth to obtain an obstacle candidate point set. Fourth, temporal fusion and spatial neighborhood deduplication are performed on each target in the BEV map to obtain an updated obstacle candidate point set. Finally, a BEV occupancy grid map is constructed based on the obstacle candidate point set. BEV risk raster map; acquire pixel risk data, generate image domain risk map based on pixel risk data and BEV occupancy raster map, and calculate image domain risk value corresponding to each target; calculate BEV risk value corresponding to each target; fuse image domain risk value and BEV risk value to obtain fused risk of the target, and construct cognitive risk map; output cognitive risk map, image domain risk map and BEV risk raster map; by correcting the foot points of the identified targets, deduplicating spatial neighborhoods, and fusion outputting multiple risks, the number of duplicate or false targets in the risk map is reduced, significantly narrowing the gap between the output obstacle distribution based on the risk map and the actual obstacle distribution, and improving the accuracy and reliability of the final risk map output.

[0023] In one possible implementation, foothold landing constraints and depth corrections are applied to each target based on the reference support domain to obtain the ground-hugging state, contact point coordinates, and foothold depth for each target, including: Construct sampling strips for each target; calculate the spatial consistency between the sampling strip and the environmental constraint information to generate the ground-hugging state of the target and determine the ground contact point coordinates for each target; perform robust correction on the foot point depth based on the ground contact point coordinates to obtain the final foot point depth of the target.

[0024] In one possible implementation, the sampling strips are constructed in the following ways: Extract target Corresponding mask pixel set : ; in, For the goal Corresponding mask pixel set The pixel coordinates in the image, and The x-axis is... The vertical axis is used as the coordinate. Define goals bottom edge position : ; Due to the definition in this embodiment The coordinates are vertical, with downward being positive; therefore, the target... Corresponding mask The bottom edge is The maximum value of the square; In the mask The bottom edge region is constructed to a height of The sampling strips correspond to the sampling domain : ; Where k is the set sampling strip height.

[0025] In one possible implementation, calculating the spatial consistency between the sampling strip and the environmental constraint information, generating the ground-hugging state of the target, and determining the ground contact point coordinates corresponding to each target includes: Several candidate footpoints were obtained based on the sampling strips. : ; ; in, It is a median function; For the target number The sampling domain of the corresponding sampling strip; It is worth noting that when the target When the corresponding mask is missing or the sampling strip constructed in the above manner is empty, the target can be... The center of the bottom edge of the corresponding two-dimensional detection box is used as a candidate foot point, and the target is... Recorded as a low-quality candidate; Spatial consistency is measured based on the sampling strips and the reference support domain to obtain the overlap ratio. : ; That is, determining the overlap ratio between the region corresponding to the sampling strip and the region corresponding to the reference support domain; where, Let it be a very small positive number. For reference support domain; when the sampling strip corresponds to the sampling domain When empty, it avoids the denominator being zero, thus making the overlap ratio meaningful; at the same time, it ensures that targets that fail to reach the ground are not directly discarded, but are used in subsequent conservative projection strategies for the modeling domain unless the bottom layer is covered, thus avoiding the situation where targets are not identified due to discarding targets. When the overlap ratio is greater than or equal to the set ground contact threshold, the ground contact state corresponding to the target is set to ground contact; that is... When the overlap ratio is less than the set ground-touching threshold, the ground-touching state of the target is set to non-ground-touching; that is... When this happens, it indicates that the target's ground-hugging recognition failed, and the target's ground-hugging state is changed to non-ground-hugging; among them, The threshold value for determining ground contact is set by experts based on experience. Define the goal Corresponding ground contact point coordinates: Obtain the target Corresponding candidate footpoints ; When the candidate footpoint is not within the reference support domain, i.e. Then, the search proceeds vertically downwards along the image to find the first pixel that falls into the reference support domain, and the coordinates corresponding to that pixel are designated as the ground contact point coordinates; specifically, the ground contact point coordinates are... ;in ; ; Since the search is performed in the vertical direction, the horizontal coordinates of the candidate foot points are the same as the horizontal coordinates of their corresponding ground contact points. The vertical coordinates are obtained by searching within a set range. Specifically, the settings are... For the goal The vertical region range corresponding to the candidate footpoints is in The search is performed within the range, and the first point that falls into the reference support domain is recorded as the ground contact point; This is the maximum search distance set for vertical search; It is worth noting that when no corresponding ground contact point is found within a set range on the ground, the coordinates of the nearest reference support pixel within the candidate foot point's neighborhood are searched as the ground contact point coordinates; and the ground-contact state corresponding to the target is set to an untrusted state; to avoid situations where the search is unsuccessful; specifically, The domain for setting candidate footpoints , ,in, The maximum search distance for the set horizontal coordinate; In the domain corresponding to the candidate footpoint Perform a search to find the coordinates of the nearest horizontally supporting pixel as the ground contact point coordinates. , ; If no corresponding ground contact point is found in the horizontal search, then let The target i is then designated as a ground-hugging untrusted state, and this type of target can be used for weight reduction processing in subsequent risk fusion.

[0026] In one possible implementation, robust correction is performed on the foothold depth based on the ground contact point coordinates to obtain the final foothold depth of the target, including: Perform local robust statistics on deep patching to determine ground contact points. Then, a depth image block is extracted centered on the ground contact point; Depth image patch centered on ground contact point The size s of the depth image patch is based on the vertical coordinate of the target in the image. Dynamic adjustment; for example, when the target is determined to be a nearby target by the vertical axis, s = 11 × 11 is taken; when the target is determined to be a distant target by the vertical axis, s = 5 × 5 is taken. Depth correction is performed by robust statistical filtering sampling within the reference support domain. The corrected footpoint depth is: ; in, The corrected foot point depth, It is a median function; Calculation target The corresponding ground mass fraction domain ground-level confidence coefficient; Get the original depth of the foot point , ; Calculate depth deviation , ; Calculation of ground mass fraction based on depth deviation; ; in, Ground mass fraction; The overlap ratio; These are the weighting coefficients. This is the depth deviation scale parameter, used to control the rate of exponential decay; The ground contact determination threshold is obtained, and the ground contact confidence score is calculated based on the ground contact determination threshold and the ground quality score, specifically using a Sigmoid mapping. ; in, For ground-level confidence; For the set slope, The threshold for determining ground contact; The final footpoint depth is obtained by revising the corrected footpoint depth and the original footpoint depth based on the ground confidence level. ; The system integrates and outputs the ground contact point coordinates, corrected foot depth, overlap ratio, ground quality score, and ground contact confidence score.

[0027] In one possible implementation, the projection strategy is determined based on the ground-hugging state, including: When the ground-hugging state is ground-hugging, the projection strategy is set to precise projection; When the ground-hugging state is non-ground-hugging, the projection strategy is set to conservative projection.

[0028] In one possible implementation, the coordinates of the corresponding contact points are projected onto the BEV map based on a projection strategy and foot depth to obtain a set of obstacle candidate points; including: The coordinates of the ground contact point are back-projected onto the camera coordinate system to obtain the first ground contact point coordinate; Transform the ground contact point coordinates to the BEV coordinate system and construct the obstacle candidate points corresponding to the target; When the projection strategy corresponding to the target is conservative projection, the obstacle radius corresponding to the obstacle candidate point is extracted, and the obstacle radius is conservatively extended to obtain the final obstacle radius; the obstacle radius is used to replace the corresponding obstacle radius in the obstacle candidate point to obtain the final obstacle candidate point. The individual obstacle candidate points are integrated into an obstacle candidate point set.

[0029] Specifically, the ground contact point coordinates are back-projected onto the camera coordinate system to obtain ground contact point coordinates one; specifically, the ground contact point coordinates are obtained. Set the second pixel coordinate vector corresponding to the ground contact point coordinates. ; By formula: ; in, Let the coordinates of the ground contact point in the camera coordinate system be the first coordinate. This represents the depth value corresponding to the ground contact point, and K is the camera intrinsic parameter matrix; The ground contact point coordinates are transformed into the BEV coordinate system to obtain two-dimensional coordinates. Specifically, the camera's extrinsic parameters, rotation matrix R and translation vector t, are obtained, and the ground contact point coordinates are transformed into a coordinate system centered on the vehicle to obtain the three-dimensional position coordinates. ; in, Let be the position coordinates, and This indicates the position of the target in the three-dimensional space around the vehicle. Building goals Mapped obstacle candidate points ; ;in, For the goal The corresponding ID number, For the goal The corresponding category; Map the position coordinates of target i in three-dimensional space to the coordinates of the obstacle center on the BEV plane, i.e. ; For the goal The corresponding obstacle radius, that is, the physical scale radius of the corresponding category, such as 2.5 meters for cars and 0.5 meters for pedestrians; Ground quality rating; In a preferred embodiment, directly take As its two-dimensional coordinates on the BEV plane, and the height component As a ground-contact consistency self-checking metric, it is used for subsequent identification of depth outliers. In another embodiment, it can be further... Project onto a preset ground plane to enhance the consistency of all target points in the height dimension.

[0030] Set the spatial extent of the BEV raster map With resolution res , continuously coordinate Discretize into a raster index: , ; Out-of-bounds index values ​​can be ignored or removed, which can prevent invalid data from consuming computing resources, ensure the stability of the raster data structure and the real-time processing, and adapt to the standardized input of downstream planning modules; it can also eliminate false risk interference caused by unreliable data outside the range, ensure the accuracy of risk assessment, and thus ensure the stability of the generated raster data structure.

[0031] A conservative projection and dilation strategy is implemented. If and only if the "ground-hugging" determination fails, the contact point calculated based on the mask is abandoned, and instead the geometric center of the bottom edge of the target's 2D detection frame is used. This serves as the projection source point. Geometrically, this point is generally more stable than the segmentation edge, although it may include the ground background, but it is safer under a conservative strategy. For conservative targets, a higher degree of positional uncertainty is assigned. The variance of the target's longitudinal distance is set to be a quadratic function of its depth value. , The variance of the longitudinal distance. For depth value, and These are the coefficients obtained after solving the problem.

[0032] The greater the distance, the greater the uncertainty in monocular ranging. When mapping the target to the BEV grid, its occupancy radius is dynamically adjusted based on the aforementioned uncertainties. :

[0033] in, This is the physical prior radius for that category, such as 2.5 meters for cars and 0.5 meters for pedestrians. The safety factor is positively correlated with the current vehicle speed; the faster the speed, the higher the safety factor. The larger the value, the more it compensates for the positional uncertainty caused by perception delay. This makes distant, non-grounded targets appear as a larger ellipse or circular area on the BEV map, thus increasing the probability of covering their true position and ensuring that the vehicle maintains a sufficient safe distance when passing through the area.

[0034] In one possible implementation, temporal fusion and spatial neighborhood deduplication are performed on each target in the BEV graph to obtain an updated set of obstacle candidate points; including: Obtain the center coordinates and radius of obstacles corresponding to targets with the same ID number in each frame image; For each target, the obstacle center coordinates and obstacle radius are smoothed by exponential moving average to obtain the obstacle center coordinates and obstacle radius of each target in the current frame image after exponential moving average smoothing. Based on the obstacle center coordinates and obstacle radius, the obstacle center coordinates and obstacle radius of the obstacle candidate point set are updated. Specifically, the obstacle center coordinates for the same tracking ID. Perform exponential moving average (EMA) smoothing: ; ; α is the coordinate smoothing coefficient, with a value range of (0,1), used to balance the weights of the current frame observation value and the historical smoothing value; the larger α is, the more reliable the current frame observation result is; the smaller α is, the stronger the smoothing effect is, and the better it can suppress inter-frame jitter at the target position. The radius is similar: ; This is the radius smoothing coefficient, with a value range of (0,1). It functions the same as the coordinate smoothing coefficient and can be the same as α, or it can be adjusted separately depending on the scenario. When consecutive frames are lost (not exceeding TTL), speed extrapolation is used for completion. ; , For the first The lateral and longitudinal motion velocities of each target in the BEV coordinate system are obtained by fitting the coordinate changes of the target in historical frames. k is the number of consecutive frames lost by the target, and its value does not exceed the preset maximum survival threshold TTL; This embodiment first corrects the obstacle center coordinates and obstacle radius corresponding to the target in different frames by analyzing the changes of the same target between different frames, eliminating sudden changes in target position caused by camera shake or other reasons; then, it merges different targets that actually belong to the same target through spatial reconstruction, which can be understood as removing ghosting; by performing spatiotemporal correction on the obstacle center coordinates and obstacle radius of the target in the above two steps, the number of duplicate or false targets in the current frame is reduced, which facilitates the subsequent generation of a high-accuracy risk map; Obtain obstacle candidate points of the same category from the obstacle candidate point set; obtain the obstacle center coordinates and obstacle radius corresponding to each obstacle candidate point; calculate the BEV plane Euclidean distance between two obstacle candidate points based on the obstacle center coordinates; and calculate the scale similarity between two obstacle candidate points based on the obstacle radius. When the Euclidean distance in the BEV plane is less than a set distance threshold, and the scale similarity is also less than a set scale similarity threshold, the two obstacle candidate points are determined to be duplicate targets; thus, a set of duplicate candidate points is obtained; the set of candidate points includes several obstacle candidate points, and a representative candidate point of the candidate point set is generated based on each obstacle candidate point in the candidate point set; the corresponding obstacle candidate point in the obstacle candidate point set is replaced with the representative candidate node to obtain an updated set of obstacle candidate points; To eliminate duplicate observations of the same physical obstacle, the center coordinates of candidate point pairs with the same category or belonging to the same planning-related category set are used. and Calculate its Euclidean distance in the BEV plane: ; At the same time, combined with its physical scale radius , Calculate scale similarity: ; When both conditions are met and When the candidate points are determined to be duplicate targets, the distance threshold is used. It can be configured according to the scene and raster resolution; There are multiple duplicate targets corresponding to the same entity of the same category in the candidate point set, for example, there are duplicate targets A, B, C and D in the candidate point set; the relationship between the duplicate targets A, B, C and D is that one obstacle candidate point is a duplicate target with the other three obstacle candidate points, for example, obstacle candidate point A and obstacle candidate point B are duplicate targets, obstacle candidate point A and obstacle candidate point C are duplicate targets, and obstacle candidate point A and obstacle candidate point D are duplicate targets; or the duplicate targets between the obstacle candidate points are in a serial relationship, for example, obstacle candidate point A and obstacle candidate point B are duplicate targets, obstacle candidate point B and obstacle candidate point C are duplicate targets, obstacle candidate point C and obstacle candidate point D are duplicate targets, etc.; Generating a representative candidate point of the candidate point set based on each obstacle candidate point in the candidate point set, including one of the following strategies: 1. Retain the one with a higher risk estimation value; 2. Retain the one with higher perception confidence; 3. Prioritize retaining the one with a successful ground-attached state; 4. Prioritize retaining the one with a higher trajectory quality score; 5. Merge the risk contributions of overlapping regions using a maximum value fusion method.

[0035] In a possible implementation, constructing a BEV occupancy grid map and a BEV risk grid map based on the obstacle candidate point set includes: Acquire the obstacle center coordinates and obstacle radius corresponding to each obstacle candidate point in the obstacle candidate point set, establish a BEV grid, and construct a circular kernel with the obstacle radius as the radius to perform filling and occupancy on the obstacle center coordinates:

[0036] wherein, represents the occupancy state of the cell in the i-th row and j-th column of the BEV grid at time t, indicates occupied; indicates free; represents the coordinates of the center corresponding to the grid cell in the ground coordinate system; is the obstacle center coordinate of the obstacle candidate point corresponding to number i; is the obstacle radius; Perform filling and occupancy on each obstacle center coordinate to obtain the BEV occupancy grid map; Acquire the risk value corresponding to each grid cell and the basic risk level of the target to which the grid cell belongs; select and obtain the current risk value of the grid cell through a formula; ; wherein, The current risk value of the grid cell is a scalar, usually normalized to a number in the range of 0 to 1; This indicates the basic risk level of target i, which depends on factors such as type and speed. For example, pedestrians pose a high risk, while stationary objects pose a low risk. This indicates the distance from the center of the cell to the center of the target cell; Indicates the radius of the obstacle; The basic risk level is obtained in the following way: ; in, This represents a truncation / limiting function, whose core function is to force the input value to be limited to a specified minimum and maximum value; It is the inherent risk weight of the category. Based on the risk level commonly used in the autonomous driving industry, it sets a fixed prior weight for different categories of targets. It is the benchmark value of basic risk and is normalized to the [0,1] range. It is a motion state correction factor, calculated based on the relative motion state of the target and the vehicle, and normalized to the [1, 1.2] interval. It can be understood that a moving target can amplify the basic risk. It is a spatial position correction factor, calculated based on the longitudinal distance and lateral lane position between the target and the vehicle, and normalized to the interval [0.8, 1.2]. It is the perception confidence correction factor, which is calculated by combining the target detection confidence and the ground-based confidence and normalized to the [0.5, 1.0] interval.

[0037] A BEV risk raster map is constructed based on the risk value of each raster cell.

[0038] The risk value decreases linearly with increasing distance, and the closer the distance, the higher the risk. This step outputs the BEV risk raster for the current frame. and occupy grid The grid map is updated after each frame is detected, supporting dynamic obstacle tracking.

[0039] In one possible implementation, an image domain risk map is generated based on pixel risk data and a BEV occupancy raster map, and the image domain risk value corresponding to each target is calculated, including: Extracting semantic risk items from pixel risk data Deep risk items and future trajectory risk items ; The semantic risk item It is calculated using the following formula: ; in, For pixels The prior risk weights are set according to the category. For example, the prior risk weights for each pixel in the category of vehicle and person are set to be large, while the prior risk weights for each pixel in the category of road surface are set to 0. Semantic risk reflects the inherent danger level of different categories of objects. The deep risk item It is calculated using the following formula: ; in, For pixels The depth values ​​are mapped to physical distances. The monocular depth estimation network directly outputs normalized relative depth values. These relative depth values ​​need to be converted to real-world physical distances using pre-calibrated camera intrinsics and a depth-distance mapping table. This depth-distance mapping table is a pre-calibrated table showing a one-to-one correspondence between the relative depth values ​​output by the monocular depth estimation network and real-world physical distances. The calibration method is as follows: with a fixed camera mounting pose, images of the calibration board at different known distances are acquired. The relative depth values ​​at the corresponding positions are output by the monocular depth estimation network, and the mapping relationship between the relative depth values ​​and the real physical distances is fitted to generate the mapping table. This mapping table allows the dimensionless relative depth values ​​output by the monocular depth estimation network to be converted into real physical distances in meters, which are then used for subsequent depth risk calculations and BEV coordinate projection. The set proximity threshold, The threshold value is set at a certain distance. The risk item of future trajectory It is calculated using the following formula: ; in, k Indicates the future number k One predicted moment, K It is the set of predicted time steps. Indicates the first i An obstacle in the future k Predicted image coordinates, This represents the standard deviation of the Gaussian kernel; semantic risk items Deep risk items and future trajectory risk items Risks associated with obtaining the pixel position of the corresponding pixel through linear fusion: ; in, For pixels Corresponding pixel location risk; , and The weighting coefficients are set for the semantic risk item, the depth risk item, and the future trajectory risk item, respectively. The pixel location risk corresponding to each pixel is calculated sequentially, and an image domain risk map is constructed based on the pixel location risk. Obtain the set of pixels in the filled area occupied by each target in the BEV occupied grid map; calculate the average risk of each pixel position in the image domain risk map, and use it as the image domain risk value corresponding to the target. Specifically, through the formula ;in, The image domain risk value corresponding to target i. For the goal The corresponding set of pixels.

[0040] In one possible implementation, the fusion risk of the target is obtained by fusing image domain risk values ​​and BEV risk values, including: Calculate the BEV risk value for each objective: ; in, For the goal The corresponding BEV risk value; The category weight of the target category. Represents the vertical coordinate. The longitudinal distance attenuation threshold is a preset fixed constant in meters, used to control the rate at which the risk value decreases with the longitudinal distance between the target and the vehicle. Indicates the horizontal coordinate; The lateral coordinates of the lane centerline. The standard deviation of the lateral position distribution is a preset fixed constant in meters, used to characterize the rate at which the risk of the target's lateral position deviating from the lane centerline decays. As a reference threshold, only when the risk impact radius of the target is greater than... Only when the activation function outputs a valid risk value will it output a low risk value close to 0; otherwise, it will output a low risk value close to 0. For activation functions; Through formula Calculate the target The corresponding fusion risk, the weight of this two-layer risk fusion is based on the ground-level confidence level. Dynamic adjustment: increase the BEV weight when the target ground confidence is high, and increase the image domain weight when the ground confidence is low. Specific Implementation Example 2 This embodiment provides a complete implementation of BEV obstacle perception. First, RGB image frames of a video sequence are acquired and preprocessed to construct an input tensor suitable for the XMem tracking network. Next, a monocular depth estimation network is used to generate a scene depth map, which is then aligned in size with the label map output by the tracking algorithm. When re-detection conditions such as first processing, arrival of an interval frame, or degradation of tracking quality are met, candidate target boxes are obtained using the GroundingDINO detector. After non-maximum suppression, a SAM segmentation model is used to generate a target mask, which is then used to initialize the XMem tracker. Simultaneously, a mapping relationship between tracking IDs and target categories is established.

[0042] Next, for each tracked target in each frame, a rigorous BEV coordinate extraction is performed: first, the median of the bottom 5 consecutive rows of pixels in the horizontal direction is extracted as the x-coordinate of the foot point. Using the bottom pixel of the mask as the ordinate Subsequently, a ground-based search is performed based on the road surface segmentation mask. Prioritizing vertical downward scanning up to 60 pixels to find road surface points, if this fails, a fallback search is performed within a horizontal radius of up to 80 pixels to determine the final ground-based point coordinates. Centered on the patch location, a 5×5 pixel neighborhood window is selected. Only pixels identified as road surfaces are filtered, and the median depth value is calculated as the target's depth sampling value. This process is supplemented by a series of gating filters, including the maximum allowable depth abrupt changes for various targets (e.g., vehicles 90, bicycles 70, pedestrians 60, traffic objects 80); the minimum proportion of road surface pixels within the patch (e.g., 0.45 to 0.40); and the lowest comprehensive ground quality score (0.30). The normalized depth values ​​are converted to metric distance using a calibration lookup table, and combined with camera parameters, a transformation from the image pixel coordinate system to the bird's-eye view coordinate system is performed to obtain the target's BEV center coordinates and occupancy radius. Simultaneously, the risk value of the target mask region in the image risk map is calculated. and BEV location-based calculation , by weight α A weighted fusion of values ​​equal to 0.4 is performed to obtain the final risk value. .

[0043] To further improve the continuity of the output, the target states with the same tracking ID are temporally fused. Specifically, their BEV coordinates and radius are smoothed using an exponential moving average (EMA) with a smoothing factor of 0.6. If a target is briefly lost, extrapolation based on historical velocity is allowed for completion within the next three consecutive frames, and the target is marked as "ghost," with its risk value decreasing exponentially with the number of lost frames. Finally, the fused obstacle set is mapped to a preset BEV grid, with the grid range set to [value missing]. X Direction [-20,20] meters, YWith a direction of [0,60] meters and a resolution of 0.2 meters per pixel, generate a binary occupying grid and a risk grid with values ​​of [0-1]. The risk raster uses maximum value fusion and overlays distance attenuation kernels. The results for each frame or a specified frame are exported in a structured form, including the record frame index, BEV metadata, obstacle list, a JSON file containing coordinates, radius, risk and category, and an NPZ file storing the two raster data. Specific Implementation Example 3 This embodiment illustrates the system implementation without using a temporal fusion mechanism. Its implementation steps are basically the same as in Embodiment 1, including video frame acquisition and preprocessing, depth estimation and alignment, detection-segmentation-redetection initialization, rigorous ground-level depth BEV coordinate extraction, BEV rasterization, and export. The difference from Embodiment 1 is that the temporal fusion step is omitted in this embodiment. Specifically, after obtaining the original obstacle state of each frame, EMA smoothing and extrapolation are not performed; instead, the result of the current frame is directly sent to the rasterization and export module. The ground-level depth sampling parameters, BEV grid parameters, deduplication threshold, and risk fusion strategy used in this embodiment are consistent with those in Embodiment 1.

[0045] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

[0046] How this application works: By acquiring the target set corresponding to the current frame image in the vehicle's image sequence, and the reference support domain corresponding to the vehicle scene; performing foothold landing constraints and depth correction on each target based on the reference support domain, obtaining the ground-hugging state, contact point coordinates, and foothold depth corresponding to each target; determining the projection strategy based on the ground-hugging state, and projecting the corresponding contact point coordinates onto the BEV map based on the projection strategy and foothold depth, obtaining the obstacle candidate point set; performing temporal fusion and spatial neighborhood deduplication on each target in the BEV map, obtaining an updated obstacle candidate point set; and constructing a BEV occupancy raster map and a BEV risk raster map based on the obstacle candidate point set; acquiring pixel risk data, generating an image domain risk map based on the pixel risk data and the BEV occupancy raster map, and calculating the image domain risk value corresponding to each target; calculating the BEV risk value corresponding to each target; and based on the image domain... The risk value and BEV risk value are fused to obtain the fused risk of the target, and a cognitive risk map is constructed. The cognitive risk map, image domain risk map, and BEV risk grid map are output. By correcting the foot points of the identified targets, deduplicating spatial neighborhoods, and fusion outputting multiple risks, the number of duplicate or false targets in the risk map is reduced, significantly narrowing the gap between the output obstacle distribution based on the risk map and the actual obstacle distribution, and improving the accuracy and reliability of the final risk map output. It fundamentally reduces coordinate drift caused by the instability of monocular depth estimation and projection errors. Compared with the control scheme that does not use ground-level depth, the complete implementation scheme can significantly reduce the output of duplicate or false targets at the BEV level, making the perceived obstacle set closer to the real distribution of the physical world, thereby outputting a clearer and more interpretable BEV occupancy and risk grid.

[0047] The above embodiments are only used to illustrate the technical methods of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of this application without departing from the spirit and scope of the technical methods of this application.

Claims

1. A method for constructing a dynamic risk map of a vehicle driving environment based on monocular vision, characterized in that, The target set corresponding to the current frame image in the vehicle image sequence and the reference support domain corresponding to the vehicle scene are obtained; the reference support domain is constructed based on environmental constraint information; the reference support domain is the region corresponding to the region mask in the coordinate system; the environmental constraint information includes the drivable region mask of the scene where the vehicle is located in the current frame. Based on the reference support domain, the landing constraints and depth corrections of each target are performed to obtain the ground-hugging state, contact point coordinates and landing depth of each target. The projection strategy is determined based on the ground contact state. Based on the projection strategy and the foot depth, the corresponding contact point coordinates are projected onto the BEV map to obtain the obstacle candidate point set. Based on the various targets in the BEV map, temporal fusion and spatial neighborhood deduplication are performed to obtain an updated obstacle candidate point set; and a BEV occupancy grid map and a BEV risk grid map are constructed based on the obstacle candidate point set. Acquire pixel risk data, generate image domain risk map based on pixel risk data and BEV occupancy raster map, and calculate image domain risk value for each target; Calculate the BEV risk value for each objective: ； in, The BEV risk value corresponding to target i; The category weight of the target category. Represents the vertical coordinate. The longitudinal distance attenuation threshold is a preset fixed constant in meters, used to control the rate at which the risk value decreases with the longitudinal distance between the target and the vehicle. Indicates the horizontal coordinate; The lateral coordinates of the lane centerline. The standard deviation of the lateral position distribution is a preset fixed constant in meters, used to characterize the rate at which the risk of the target's lateral position deviating from the lane centerline decays. As a reference threshold, only when the risk impact radius of the target is greater than... Only when the activation function outputs a valid risk value will it output a low risk value close to 0; otherwise, it will output a low risk value close to 0. For activation functions; Through formula The fusion risk corresponding to target i is calculated, where, The image domain risk value corresponding to target i; The fused risk of the target is obtained by fusing the image domain risk value and the BEV risk value, and a cognitive risk map is constructed. The specific structured risk set of the cognitive risk map is represented as follows: ； in, For target i, the target ID, For the category of target i, Let the BEV coordinates of target i be... Let the obstacle radius of target i be... The fusion risk for target i The ground-level confidence level for target i; the BEV risk grid is used for planning grid input; Output cognitive risk map, image domain risk map, and BEV risk raster map.

2. The method for constructing a dynamic risk map of a vehicle driving environment based on monocular vision according to claim 1, characterized in that, The process of applying foothold landing constraints and depth correction to each target based on the reference support domain yields the ground-hugging state, contact point coordinates, and foothold depth for each target, including: Construct sampling strips for each target, including: Extract the mask corresponding to target i pixel set : ； Define the position of the bottom edge of target i : ； In the mask The bottom edge region is constructed to a height of k The sampling strips correspond to the sampling domain : ； The spatial consistency between the sampling strip and the environmental constraint information is calculated to generate the ground-hugging state of the target and determine the ground contact point coordinates of each target. Based on the ground contact point coordinates, the foot depth is robustly corrected to obtain the final foot depth of the target.

3. The method for constructing a dynamic risk map of a vehicle driving environment based on monocular vision according to claim 2, characterized in that, The calculation of the spatial consistency between the sampling strip and the environmental constraint information, the generation of the target's ground-hugging state, and the determination of the ground contact point coordinates corresponding to each target include: Several candidate footpoints were obtained based on the sampling strips. : ； ； in, It is a median function; The sampling area of ​​the sampling strip corresponding to target i; Spatial consistency is measured based on the sampling strips and the reference support domain to obtain the overlap ratio. : ； in, It is a set, extremely small positive number; When the overlap ratio is greater than or equal to the set ground contact threshold, the ground contact state corresponding to the target is set to ground contact; when the overlap ratio is less than the set ground contact threshold, the ground contact state corresponding to the target is set to non-ground contact. When a candidate foot point is not within the reference support domain, the search is performed vertically downwards along the image to obtain the first pixel point that falls into the reference support domain, and the coordinates corresponding to the pixel point are set as the ground contact point coordinates. If no corresponding ground contact point is found within the set range on the ground, the coordinates of the nearest reference support domain pixel point within the candidate foot point's neighborhood are searched as the ground contact point coordinates.

4. The method for constructing a dynamic risk map of a vehicle driving environment based on monocular vision according to claim 2, characterized in that, The robust correction of the foot point depth based on the ground contact point coordinates to obtain the final foot point depth of the target includes: Perform local robust statistics on deep patching to determine ground contact points. Then, a depth image block is extracted centered on the ground contact point; Depth image patch centered on ground contact point The size s of the depth image patch is based on the vertical coordinate of the target in the image. v Dynamic adjustment; Depth correction is performed by robust statistical filtering sampling within the reference support domain. The corrected footpoint depth is: ； in, The corrected foot point depth, It is a median function; Calculate the ground mass fraction domain ground-to-ground confidence coefficient corresponding to target i; Get the original depth of the foot point , ; Calculate depth deviation , ; Calculation of ground mass fraction based on depth deviation; ； in, Ground mass fraction; The overlap ratio; These are the weighting coefficients. This is the depth deviation scale parameter, used to control the rate of exponential decay; The ground contact determination threshold is obtained, and the ground contact confidence score is calculated based on the ground contact determination threshold and the ground quality score, specifically using a Sigmoid mapping. ； in, is the ground confidence level; k is the set slope. The threshold for determining ground contact; The final footpoint depth is obtained by revising the corrected footpoint depth and the original footpoint depth based on the ground confidence level. 。 5. The method for constructing a dynamic risk map of a vehicle driving environment based on monocular vision according to claim 1, characterized in that, The projection strategy determined based on the ground-hugging state includes: When the ground-hugging state is ground-hugging, the projection strategy is set to precise projection; When the ground-hugging state is non-ground-hugging, the projection strategy is set to conservative projection.

6. The method for constructing a dynamic risk map of a vehicle driving environment based on monocular vision according to claim 5, characterized in that, The step of projecting the corresponding contact point coordinates onto the BEV map based on the projection strategy and foot depth to obtain a set of obstacle candidate points includes: The coordinates of the ground contact point are back-projected onto the camera coordinate system to obtain the first ground contact point coordinate; Transform the ground contact point coordinates to the BEV coordinate system and construct the obstacle candidate points corresponding to the target; When the projection strategy corresponding to the target is conservative projection, the obstacle radius corresponding to the obstacle candidate point is extracted, and the obstacle radius is conservatively extended to obtain the final obstacle radius; the obstacle radius is used to replace the corresponding obstacle radius in the obstacle candidate point to obtain the final obstacle candidate point. The individual obstacle candidate points are integrated into an obstacle candidate point set.

7. The method for constructing a dynamic risk map of a vehicle driving environment based on monocular vision according to claim 1, characterized in that, Based on the targets in the BEV map, temporal fusion and spatial neighborhood deduplication are performed to obtain an updated set of obstacle candidate points, including: Obtain the center coordinates and radius of obstacles corresponding to targets with the same ID number in each frame image; For each target, the obstacle center coordinates and obstacle radius are smoothed by exponential moving average to obtain the obstacle center coordinates and obstacle radius of each target in the current frame image after exponential moving average smoothing. Based on the obstacle center coordinates and obstacle radius, the obstacle center coordinates and obstacle radius of the obstacle candidate point set are updated. Obtain obstacle candidate points of the same category from the obstacle candidate point set; obtain the obstacle center coordinates and obstacle radius corresponding to each obstacle candidate point; calculate the BEV plane Euclidean distance between two obstacle candidate points based on the obstacle center coordinates; and calculate the scale similarity between two obstacle candidate points based on the obstacle radius. When the Euclidean distance in the BEV plane is less than a set distance threshold, and the scale similarity is also less than a set scale similarity threshold, the two obstacle candidate points are determined to be duplicate targets; thus, a set of duplicate candidate points is obtained; the set of candidate points includes several obstacle candidate points, and a representative candidate point of the candidate point set is generated based on each obstacle candidate point in the candidate point set; the corresponding obstacle candidate point in the obstacle candidate point set is replaced with the representative candidate node to obtain an updated set of obstacle candidate points.

8. The method for constructing a dynamic risk map of a vehicle driving environment based on monocular vision according to claim 1, characterized in that, The construction of the BEV occupancy grid map and the BEV risk grid map based on the obstacle candidate point set includes: Obtain the obstacle center coordinates and radius corresponding to each obstacle candidate point in the obstacle candidate point set, establish a BEV grid, and construct a circular core with the obstacle radius as the radius to fill and occupy the obstacle center coordinates: ； wherein, represents the occupancy state of the cell at row i and column j in the BEV grid at time t, represents occupied; represents free; represents the coordinates of the center corresponding to the grid cell in the ground coordinate system; is the obstacle center coordinate of the obstacle candidate point corresponding to number i; is the obstacle radius; Fill in the center coordinates of each obstacle to obtain the BEV occupancy grid map. Obtain the risk value corresponding to each grid cell, as well as the basic risk level of the target to which the grid cell belongs; select the current risk value of the grid cell using a formula; ； in, This indicates the current risk value of the raster cell; Indicates the basic risk level of target i; This indicates the distance from the center of the raster cell to the center of the target cell; The radius of the obstacle is indicated; the basic risk level is obtained in the following way: ； in, This represents the truncation / limiting function; It is the inherent risk weight of the category; It is a motion state correction factor; It is a spatial location correction factor; It is a perceived confidence correction factor; A BEV risk raster map is constructed based on the risk value of each raster cell.

9. The method for constructing a dynamic risk map of a vehicle driving environment based on monocular vision according to claim 1, characterized in that, The process of generating an image domain risk map based on pixel risk data and BEV occupancy raster map, and calculating the image domain risk value corresponding to each target, includes: Extracting semantic risk items from pixel risk data Deep risk items and future trajectory risk items ; semantic risk items Deep risk items and future trajectory risk items Risks associated with obtaining the pixel position of the corresponding pixel through linear fusion: ； in, For pixels Corresponding pixel location risk; , and The weighting coefficients are set for the semantic risk item, the depth risk item, and the future trajectory risk item, respectively. This represents the truncation / limiting function; The pixel location risk corresponding to each pixel is calculated sequentially, and an image domain risk map is constructed based on the pixel location risk. Obtain the set of pixels in the filled area occupied by each target in the BEV-occupied raster image; calculate the mean of the pixel position risk of each pixel in the image domain risk map, and use it as the image domain risk value corresponding to the target.

Images