A method for predicting traversability in unstructured terrain environments

Abstract

The application discloses a passability prediction method for unstructured terrain environment, and is applied to the technical field of vehicle traffic capacity prediction. The application constructs a three-dimensional grid map based on a multi-source sensor, combines semantic terrain classification and elevation structure feature extraction, and forms a geometric and semantic joint representation of the environment. By comparing the control instruction with the motion response of the vehicle in the actual terrain, a traffic capacity coefficient is constructed as a supervision signal, and a conditional variational autoencoder is trained to model the traffic capacity coefficient distribution and its uncertainty under different terrain conditions. Further, intermediate layer features are extracted from the semantic recognition network to establish a category statistical model for identifying terrain categories that do not appear in the training set. Finally, the traffic capacity confidence and terrain reliability are fused to output a grid-level traffic probability, realizing the passability prediction in the unstructured environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle passability prediction, in particular to a passability prediction method for unstructured terrain environment. BACKGROUND

[0002] In structured road environment, the road boundary is clear, the pavement material is uniform, and the geometric properties are relatively stable, so the judgment of the vehicle on the passability state has high certainty and predictability. However, in tasks such as field operation, disaster response and complex terrain survey, special vehicles need to operate in unstructured environment. The terrain in such scenes is diverse, the geometric structure is complex, and there is a lack of available prior map information, which has high uncertainty, resulting in significant challenges in passability judgment. The existing mainstream methods mostly rely on semantic segmentation and rule mapping strategy, which directly converts the identified terrain category into a judgment label of whether it is passable or not. Although this method has a certain practicality, it ignores the physical response differences of different terrains under the same semantic category in the actual driving process, and it is difficult to fully reflect the real passability of the ground. In addition, the existing models generally lack the ability to distinguish their own cognitive boundaries, and when faced with new terrain categories not seen in the training set, they are prone to misjudgment with high confidence, resulting in failure of path planning or the vehicle being trapped in a dangerous area. Therefore, how to provide a passability prediction method for unstructured terrain environment is a problem that those skilled in the art need to solve. SUMMARY

[0003] Therefore, the present application provides a passability prediction method for unstructured terrain environment, which can represent the physical accessibility of the terrain and the cognitive boundaries of the model, and improve the reliability of the passability decision of the special vehicle in the complex and highly uncertain environment and the adaptability to unknown terrain.

[0004] In order to achieve the above purpose, the present application provides the following technical scheme:

[0005] A passability prediction method for unstructured terrain environment, comprising the following steps:

[0006] S1, collecting terrain information by using a multi-source sensor, fusing the terrain information to estimate a pose transformation matrix at each time, constructing a grid map, and predicting the terrain category of each grid in the grid map based on an image point cloud collaborative semantic segmentation network;

[0007] S2, establishing a three-dimensional semantic map, calculating an elevation feature vector of each two-dimensional grid based on the point cloud height information of each grid in the grid map;

[0008] S3, recording the actual speed vector of the vehicle in each grid during the collection process of the multi-source sensor, defining and calculating a passability coefficient to represent the ability of the vehicle to execute control instructions in complex terrain;

[0009] S4, modeling the vehicle passability coefficient as a conditional random variable under terrain conditions, using a conditional variational autoencoder to model the conditional distribution;

[0010] S5, in the actual operation of the vehicle, introducing Mahalanobis distance to construct a recognition mechanism to judge whether the currently observed terrain belongs to the learned distribution support domain;

[0011] S6, jointly evaluating the passability of the terrain at the grid based on the passability distribution of the vehicle to the terrain in S4 and the cognitive credibility of the terrain in S5.

[0012] Optionally, the multi-source sensor collects terrain information, including: a laser radar scan point cloud sequence , represents the point cloud of the i-th frame; a camera image sequence t , represents the i-th frame image; an IMU measurement sequence , t represents the acceleration and angular velocity at the i-th frame. t Optionally, S1 is specifically:

[0013] by a factor graph-based nonlinear optimization method to fuse the information of laser radar, camera vision and IMU, to estimate the pose transformation matrix :

[0014]

[0015] ;

[0016] wherein, represents a rotation matrix, represents a three-dimensional Euclidean group, represents a three-dimensional orthogonal group, represents a translation vector, and the registered point cloud is projected into the world coordinate system to construct a dense map; a voxelized three-dimensional occupancy grid mapping method is used to generate a grid map OctoMap , a semantic segmentation network is introduced to cooperate with the image point cloud, the image and the corresponding projected point cloud are input into the semantic segmentation network, and the terrain category of each grid is predicted:

[0017] ;

[0018] wherein, is the semantic prediction label of the grid , ​​​representing a set of predefined terrain semantic categories, representing a semantic segmentation network.

[0019] Optionally, S2 is specifically:

[0020] According to the intrinsic matrix of the camera and the pose transformation matrix , the semantic prediction label is back-projected to the three-dimensional map coordinate system to give the voxel grid semantic attributes, and a three-dimensional semantic map is established :

[0021] ;

[0022] In the formula, x represents a coordinate vector, represents a depth value, and a two-dimensional grid is calculated based on the point cloud height information of each grid in the grid map OctoMap :

[0023] ;

[0024] Where is the elevation feature vector of the grid , the geometric feature of the ground surface is represented, is the point cloud set falling into the grid , represents the maximum height of all point clouds at the corresponding position, is the mean of all heights, is the standard deviation of all heights, n is the number of point clouds.

[0025] Optionally, S3 is specifically:

[0026] During the acquisition process of the multi-source sensor, the actual speed vector of the vehicle in the grid :

[0027] ;

[0028] In the formula, represents the linear velocity, represents the angular velocity, and records the expected control command speed :

[0029] ;

[0030] In the formula, is the expected control command linear velocity, is the expected control command angular velocity, and defines the traffic capacity coefficient ​To characterize the ability of a vehicle to execute control instructions in complex terrain:

[0031] ;

[0032] passability coefficient represents the strength of the vehicle's response to the expected control under certain terrain conditions, the greater the value, the less the terrain hinders movement, and the smaller the value, the more limited the movement.

[0033] Optionally, S4 is specifically:

[0034] The passability coefficient of the vehicle is modeled as a conditional random variable under terrain conditions to characterize this uncertainty, and the conditional probability distribution is learned; a conditional variational autoencoder is used to model the conditional distribution, which includes a conditional encoding module, an encoder network, and a decoder network. The collected data is used to construct a training process sample triad dataset:

[0035] ;

[0036] The conditional encoding module encodes the conditional input into a conditional representation vector , and the encoder network is defined as the posterior distribution:

[0037] ;

[0038] wherein is a latent variable, denotes the encoder network parameters, and the encoder network maps the observation and terrain conditions to Gaussian distribution parameters to construct an approximate posterior, in the generation process, the latent variable is first sampled from a standard normal distribution , and is input into the decoder network together with the conditional vector to predict the conditional probability distribution of the passability , which is modeled as a Gaussian distribution, and the decoding form is:

[0039] ;

[0040] wherein and are the outputs of the decoder network, and the decoder network parameters.

[0041] Optionally, S5 is specifically:

[0042] Deep feature representation is extracted from the last level of the encoder of the semantic segmentation network , for describing the high-dimensional semantic feature embedding of the current image region:

[0043] ;

[0044] wherein, is a feature extraction function, is a semantic segmentation network, is an input grid parameter vector; in the training phase, for each known semantic terrain class , a corresponding deep feature set is collected from the training set , the mean vector of the distribution is calculated and the covariance matrix :

[0045] ;

[0046] wherein, is the number of semantic terrain classes, and a feature Gaussian model is constructed for each terrain class:

[0047] :

[0048] In the inference phase, for the grid , the intermediate layer feature vector is extracted, and the Mahalanobis distance with each semantic class Gaussian model is calculated:

[0049] ;

[0050] The class with the smallest distance among all classes is taken as the discrimination result:

[0051] ;

[0052] The value of the discrimination result is used to evaluate whether the current terrain falls within the known distribution range in the training set.

[0053] Optionally, S6 is specifically:

[0054] The Gaussian distribution modeling the conditional distribution is mapped to the probability through , and a safety threshold is set to distinguish whether the vehicle can respond to the control instruction, and the possibility of successful passage of the vehicle is evaluated by calculating the right tail cumulative distribution function of the passage capacity distribution:

[0055] ;

[0056] wherein, is the passage capacity coefficient ; the probability of exceeding the safety threshold ; define a credibility decay factor the Mahalanobis distance fused into a unified scoring system:

[0057] ;

[0058] where, is the adjustment term, controlling the influence of the degree of cognitive bias on the final score, is the Mahalanobis distance between the current terrain feature and the mean of its closest semantic category, reflecting its typicality in the training distribution; finally, the probability of terrain passability is defined as the combination of the probability of the traffic capacity distribution and the cognitive credibility of the terrain :

[0059] ;

[0060] The probability of terrain passability simultaneously considers the actual impact of the terrain on the response capability of the vehicle and the credibility of the perception model.

[0061] Through the above technical solution, compared with the prior art, the present application provides a passability prediction method for unstructured terrain environment, which has the following beneficial effects:

[0062] 1. The present application introduces the actual motion response information of the vehicle under different terrain conditions, constructs the traffic capacity coefficient as a supervision signal, combines the elevation feature and the semantic label to train the traffic capacity prediction model, introduces the probability modeling mechanism to express the uncertainty of the prediction result, and improves the accuracy of the trafficability evaluation;

[0063] 2. The present application extracts the intermediate layer features in the semantic recognition network, combines the feature distribution modeling of the terrain category, discriminates whether the current input belongs to the known terrain in the training set, reduces the misjudgment risk of the model to the unfamiliar terrain, and improves the robustness and safety of the system in unknown environment;

[0064] 3. The present application constructs a joint evaluation mechanism of traffic capacity confidence and terrain recognition credibility, integrates the motion response information of the vehicle and the consistency of the model's cognition of the current terrain into the trafficability judgment, outputs the grid-level passability probability, and improves the risk perception and path evaluation ability of the system in complex environment. BRIEF DESCRIPTION OF DRAWINGS

[0065] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.

[0066] Figure 1A flowchart of the unstructured terrain environment accessibility prediction method of the present application;

[0067] Figure 2 A flowchart of the multi-source sensing information driven traffic capacity dataset construction of the present application;

[0068] Figure 3 A flowchart of the conditional variational autoencoder based traffic capacity probability modeling of the present application;

[0069] Figure 4 A flowchart of the Mahalanobis distance based terrain cognition matching judgment of the present application;

[0070] Figure 5 A flowchart of the traffic probability calculation considering terrain uncertainty of the present application. DETAILED DESCRIPTION

[0071] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0072] The embodiments of the present application disclose an unstructured terrain environment accessibility prediction method, as shown in Figure 1 The method comprises the following steps:

[0073] S1, collecting terrain information by using a multi-source sensor, fusing the terrain information to estimate a pose transformation matrix at each time, constructing a grid map, and predicting a terrain category of each grid in the grid map based on an image point cloud collaborative semantic segmentation network;

[0074] S2, establishing a three-dimensional semantic map, calculating an elevation feature vector of each two-dimensional grid based on point cloud height information of each grid in the grid map;

[0075] S3, recording an actual speed vector of the vehicle in each grid during the multi-source sensor collection process, defining and calculating a traffic capacity coefficient to represent the ability of the vehicle to execute a control instruction in a complex terrain;

[0076] S4, modeling the vehicle traffic capacity coefficient as a conditional random variable under terrain conditions, and modeling the conditional distribution by using a conditional variational autoencoder;

[0077] S5, in the actual operation process of the vehicle, introducing a Mahalanobis distance to construct a recognition mechanism to judge whether the currently observed terrain belongs to a learned distribution support domain;

[0078] S6, combined with the terrain accessibility distribution of vehicles S4 and the terrain perception credibility of vehicle S5, comprehensively assesses the accessibility of the terrain at the grid.

[0079] Furthermore, the terrain information acquired by the multi-source sensors includes: lidar scan point cloud sequences. , Indicates the first t Frame point cloud; camera image sequence , For the first t Frame images; IMU measurement sequences , Indicates the first t Acceleration and angular velocity at frame time.

[0080] Furthermore, such as Figure 2 As shown, S1 is specifically:

[0081] The pose transformation matrix at each time step is estimated by fusing information from LiDAR, camera vision, and IMU using a factor graph-based nonlinear optimization method. :

[0082] ;

[0083] In the formula, Represents the rotation matrix. Representing the three-dimensional Euclidean group, Represents the three-dimensional orthogonal group. Represents the translation vector, which will be used to register the point cloud. Projecting the map onto a world coordinate system, a dense map is constructed; a voxel-based 3D occupancy raster mapping method is used to generate a raster map, OctoMap. A semantic segmentation network based on image point cloud collaboration is introduced to segment the image... With the corresponding projected point cloud The input is fed into a semantic segmentation network to predict the terrain category for each raster:

[0084] ;

[0085] in, It is a grid semantic prediction labels, This represents a predefined set of terrain semantic categories. This represents a semantic segmentation network.

[0086] Furthermore, such as Figure 2 As shown, S2 is specifically:

[0087] Based on the camera's intrinsic parameter matrix and pose transformation matrix The semantic prediction label is back-projected to a three-dimensional map coordinate system to give the voxel grid semantic attributes, and a three-dimensional semantic map is established :

[0088] ;

[0089] wherein x represents a coordinate vector, represents a depth value, and a two-dimensional grid is calculated based on point cloud height information of each grid in an OctoMap

[0090] ;

[0091] wherein h is an elevation feature vector of the grid , which represents the geometric features of the ground surface, is a point cloud set falling into the grid , represents the maximum height of all point clouds at the corresponding position, is the mean of all heights, is the standard deviation of all heights, n is the number of point clouds.

[0092] Further, as shown in Figure 2 , S3 is specifically:

[0093] During the acquisition process of the multi-source sensor, the actual speed vector of the vehicle in the grid is recorded:

[0094] ;

[0095] wherein represents the linear speed, represents the angular speed, which can be obtained from the pose transformation of two frames, and the speed recorded by the control system represents the expected control command speed:

[0096] ;

[0097] wherein is the expected control command linear speed, is the expected control command angular speed, and a traffic capacity coefficient is defined to represent the ability of the vehicle to execute the control instruction in a complex terrain:

[0098] ;

[0099] The traffic capacity coefficient ​​It represents the strength of the vehicle's response to the expected control under certain terrain conditions, the greater the value, the less the terrain hinders the movement, and the smaller the value, the more the movement is limited.

[0100] In the embodiments of the present application, considering the instability of small speed division in the actual system, a robust form is introduced:

[0101] ;

[0102] In the formula, The stable denominator parameter is a small integer.

[0103] Further, as shown in Figure 3 , S4 is specifically:

[0104] In a complex terrain environment, even if the terrain conditions are similar, the actual movement effect of the vehicle also has certain differences, and this difference is manifested as the unstable fluctuation of the traffic capacity coefficient , the traffic capacity coefficient of the vehicle is modeled as a conditional random variable under terrain conditions to characterize this uncertainty, and the conditional probability distribution is learned; a conditional variational autoencoder is used to model the conditional distribution, and the conditional variational autoencoder includes a conditional encoding module, an encoder network and a decoder network three parts, and a training process sample three tuple data set is constructed by using the collected data:

[0105] ;

[0106] The conditional encoding module jointly encodes the condition input into a conditional representation vector , and the encoder network is defined as the posterior distribution:

[0107] ;

[0108] In the formula, is a latent variable, , the encoder network maps the observation and the terrain condition into Gaussian distribution parameters to construct the approximate posterior, in the generation process, the latent variable is first sampled from the standard normal distribution, and then input into the decoder network together with the condition vector to predict the conditional probability distribution of the traffic capacity , and the Gaussian distribution is used to model , and the decoding form is:

[0109] ;

[0110] , wherein and is the output of the decoder network, is the decoder network parameter.

[0111] In embodiments of the present application, the training objective of the conditional variational autoencoder is to maximize the variational lower bound (ELBO) of the conditional marginal likelihood, which is formulated as:

[0112] ;

[0113] where, is the KL divergence, measuring the distance between the latent distribution predicted by the encoder network and the standard Gaussian prior distribution , is the reconstruction loss, measuring the fitting degree of the Gaussian distribution generated by the decoder to the observation , which is specifically expressed as:

[0114] .

[0115] Further, as shown in Figure 4 , S5 is specifically:

[0116] extracting the deep feature representation from the last stage of the encoder of the semantic segmentation network, for describing the high-dimensional semantic feature embedding of the current image region:

[0117] ;

[0118] where, is the feature extraction function, is the semantic segmentation network, is the input grid parameter vector; in the training stage, for each known semantic terrain class , the corresponding deep feature set is collected from the training set, and the mean vector and the covariance matrix of the distribution are calculated:

[0119] ;

[0120] where, is the number of semantic terrain classes, and a feature Gaussian model of each terrain class is constructed:

[0121] ;

[0122] In the inference stage, for the grid , the intermediate layer feature vector , calculate Mahalanobis distance with each semantic class Gaussian model:

[0123] ;

[0124] Take the class with the smallest distance in all classes as the discrimination result:

[0125] ;

[0126] The value of the discrimination result is used to evaluate whether the current terrain falls within the known distribution range in the training set. When is higher, it indicates that the terrain feature has deviated from the training distribution, and there is an extrapolation risk for the traffic capacity model.

[0127] Further, as shown in Figure 5 , S6 specifically includes:

[0128] The conditional variational autoencoder outputs a Gaussian distribution representing the range of traffic capacity that may be possessed under the current grid terrain conditions, and the Gaussian distribution modeling the conditional distribution is mapped to the probability to distinguish whether the vehicle can respond to the control instruction. The possibility of successful traffic of the vehicle is evaluated by calculating the right tail cumulative distribution function of the traffic capacity distribution:

[0129] ;

[0130] wherein is the traffic capacity coefficient ; the probability of exceeding the safety threshold ; when is larger or is smaller, tends to 1, indicating high traffic stability; and when is smaller or is larger, the right tail probability will decrease significantly, indicating that the terrain has high traffic uncertainty; define the confidence decay factor , and fuse the Mahalanobis distance into a unified scoring system:

[0131] ;

[0132] wherein is an adjustment term that controls the influence of cognitive deviation on the final score, is the Mahalanobis distance between the current terrain feature and the mean of its closest semantic class, reflecting its typicality in the training distribution; when the feature distribution deviates significantly, the confidence decreases exponentially; when the terrain is more consistent with the training experience The closer to 1; the probability of final combination of the passability distribution and the cognitive credibility of the terrain, defines the terrain passability probability :

[0133] ;

[0134] Terrain passability probability At the same time, the actual impact of the terrain on the response ability of the vehicle and the credibility of the perception model are considered.

[0135] In the embodiments of the present application, the passability judgment is binarized based on the passability probability The probability threshold value is set, when , the terrain is considered passable, otherwise it is considered as a high-risk area.

[0136] Each of the embodiments in the specification is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts between the embodiments can be referred to each other.

[0137] Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A accessibility prediction method for unstructured terrain environments, characterized in that, Includes the following steps: S1. Collect terrain information using multi-source sensors, fuse the terrain information to estimate the pose transformation matrix at each time step, construct a grid map, and predict the terrain category of each grid in the grid map based on a semantic segmentation network that is based on image point cloud collaboration. S2. Establish a three-dimensional semantic map and calculate the elevation feature vector of each two-dimensional grid based on the point cloud height information of each grid in the grid map; S3. During the multi-source sensor acquisition process, record the actual speed vector of the vehicle in each grid, define and calculate the traffic capacity coefficient to characterize the vehicle's ability to execute control commands in complex terrain; S4. Model the vehicle capacity coefficient as a conditional random variable under terrain conditions, and use a conditional variational autoencoder to model the conditional distribution. S5. During the actual operation of the vehicle, a Mahalanobis distance recognition mechanism is introduced to determine whether the currently observed terrain belongs to the learned distribution support domain. S6, combined with the terrain accessibility distribution of S4 vehicles and the terrain perception credibility of S5, comprehensively evaluates the terrain accessibility at the grid. S3 specifically refers to: Record vehicle movement within the grid during multi-source sensor acquisition. actual velocity vector : ； In the formula, Indicates linear velocity. Represents angular velocity, and records the speed at which the control system issues the desired control command. : ； In the formula, To achieve the desired control command linear velocity, Define the traffic capacity coefficient to represent the desired control command angular velocity. To characterize the vehicle's ability to execute control commands in complex terrain: ； Traffic capacity coefficient This indicates the strength of a vehicle's response to desired control under certain terrain conditions. A larger value indicates that the terrain does not significantly hinder movement, while a smaller value indicates that movement is restricted.

2. The accessibility prediction method for unstructured terrain environments according to claim 1, characterized in that, Terrain information acquired by multiple sensors includes: lidar scan point cloud sequences. , Indicates the first t Frame point cloud; camera image sequence , For the first t Frame images; IMU measurement sequences , Indicates the first t Acceleration and angular velocity at frame time.

3. The accessibility prediction method for unstructured terrain environments according to claim 2, characterized in that, S1 specifically refers to: The pose transformation matrix at each time step is estimated by fusing information from LiDAR, camera vision, and IMU using a factor graph-based nonlinear optimization method. : ； In the formula, Represents the rotation matrix. Representing the three-dimensional Euclidean group, Represents the three-dimensional orthogonal group. Represents the translation vector, which will be used to register the point cloud. Projecting the map onto a world coordinate system, a dense map is constructed; a voxel-based 3D occupancy raster mapping method is used to generate a raster map, OctoMap. A semantic segmentation network based on image point cloud collaboration is introduced to segment the image... With the corresponding projected point cloud The input is fed into a semantic segmentation network to predict the terrain category for each raster: ； in, It is a grid semantic prediction labels, This represents a predefined set of terrain semantic categories. This represents a semantic segmentation network.

4. The accessibility prediction method for unstructured terrain environments according to claim 3, characterized in that, S2 specifically refers to: Based on the camera's intrinsic parameter matrix and pose transformation matrix The semantically predicted labels are back-projected onto the 3D map coordinate system to assign semantic attributes to the voxel raster, thereby establishing a 3D semantic map. : ； In the formula, x represents the coordinate vector. Represents depth values, based on the OctoMap raster map. The point cloud height information of each cell in the two-dimensional grid is used to calculate the grid. Elevation eigenvectors: ； Among them, is the grid. The elevation eigenvector represents the geometric features of the Earth's surface. For all falling into the grid The collection of point clouds, This represents the maximum height of all point clouds at the corresponding location. The average of all heights. For the standard deviation of all heights, n This represents the number of point clouds.

5. The accessibility prediction method for unstructured terrain environments according to claim 1, characterized in that, S4 specifically refers to: Vehicle capacity coefficient Modeling for terrain conditions To characterize this uncertainty, conditional random variables are used to learn conditional probability distributions. ; A conditional variational autoencoder (CVA) is used to model the conditional distribution. The CVA consists of three parts: a conditional encoder module, an encoder network, and a decoder network. A training process sample tripartite dataset is constructed using the collected data. ； The conditional encoding module will input the condition. Joint encoding as a conditional representation vector The encoder network is defined as a posterior distribution: ； In the formula, As latent variables, This represents the encoder network parameters, which map observations to terrain conditions using Gaussian distribution parameters. To construct an approximate posterior, during the generation process, latent variables are first sampled from the standard normal distribution. Then with the condition vector The data is input together into the decoder network to predict throughput. The conditional probability distribution is obtained by using a Gaussian distribution as the basis. The modeling and decoding methods are as follows: ； in, and For the output of the decoder network, These are the decoder network parameters.

6. The accessibility prediction method for unstructured terrain environments according to claim 1, characterized in that, S5 specifically refers to: Extracting deep feature representations from the last stage of the encoder in a semantic segmentation network. High-dimensional semantic feature embeddings used to describe the current image region: ； In the formula, For feature extraction function, For semantic segmentation networks, The input is a raster parameter vector; during the training phase, for each known semantic terrain category... Collect the corresponding deep feature set from the training set. The mean vector of the statistical distribution With covariance matrix : ； In the formula, To determine the number of semantic terrain categories, construct a Gaussian model for the features of each terrain category: ； During the reasoning phase, for the grid Extracting intermediate layer feature vectors Calculate the Mahalanobis distance with each semantic class Gaussian model: ； The category with the smallest distance among all categories is taken as the discrimination result: ； The value of the discrimination result is used to assess whether the current terrain falls within the known distribution range of the training set.

7. The accessibility prediction method for unstructured terrain environments according to claim 1, characterized in that, S6 specifically refers to: Gaussian distribution modeled as a conditional distribution Mapping to pass probability, setting a safety threshold To distinguish whether a vehicle can respond to control commands, the probability of a vehicle successfully passing through is assessed by calculating the right-tail cumulative distribution function of the capacity distribution: ； In the formula, Traffic capacity coefficient Exceeding the safety threshold The probability; defining the credibility decay factor. Markov distance Integrate into a unified scoring system: ； in, As a moderating term, it controls for the impact of the degree of cognitive deviation on the final score. It is the Mahalanobis distance between the current terrain feature and the mean of its nearest semantic category, reflecting its typicality in the training distribution; finally, combining the probability of the accessibility distribution with the cognitive credibility of the terrain, the terrain accessibility probability is defined. : ； Terrain accessibility probability Simultaneously, the actual impact of terrain on vehicle responsiveness and the credibility of the perception model are considered.

Images