A method for off-road terrain recognition based on feature membership fusion and temporal arbitration
By using a feature membership fusion and temporal arbitration method based on the vehicle's own sensor signals, the problems of external sensor dependence and unstable recognition in off-road vehicle terrain recognition are solved, realizing a stable and continuous terrain recognition and control strategy, and reducing system complexity and hardware cost.
Patent Information
- Application Number
- CN202611615345.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-19
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies for terrain recognition in off-road vehicles rely heavily on external sensing sensors, are easily affected by environmental interference, and lack stable and easily calibrated recognition methods, resulting in discontinuous recognition results and unstable control strategies.
A feature membership fusion and temporal arbitration method based on the vehicle's own sensor signals is adopted. The vehicle signals are processed by filtering, denoising, synchronization and normalization to calculate basic and secondary features. The terrain category confidence is calculated using offline statistical parameters and feature weight table. Jitter is suppressed by maximum value logic, priority processing and sliding window voting consistency logic to output the final terrain category.
It reduces reliance on external sensors, minimizes the impact of environmental factors, improves the stability and continuity of identification results, facilitates engineering calibration and the implementation of control strategies, and enhances the controllability and security of the system.
Smart Images

Figure CN123116178A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of vehicle terrain environment perception and intelligent chassis control technology, specifically involving an off-road terrain recognition method based on feature membership fusion and temporal arbitration. Background Technology
[0002] When off-road vehicles travel on terrains such as mud, sand, gravel, and rocks, the interaction between the tires and the ground changes significantly. Signals such as the vehicle's slip ratio, longitudinal drive force demand, and vertical vibration response also vary with the terrain. Terrain recognition results can provide prior information for drive distribution, braking control, suspension control, and traction control strategies. Therefore, terrain recognition for off-road conditions is an important research direction in the field of intelligent vehicle control.
[0003] In existing technologies, road surface or terrain recognition solutions mainly include visual / point cloud methods, vehicle-mounted dynamic signal methods, and multi-source fusion methods. Visual or point cloud methods rely on cameras or LiDAR to acquire ground appearance or geometric information, providing a wealth of external environmental information, but they are sensitive to lighting conditions, occlusion, mud and water coverage, and sensor costs. Multi-source fusion methods jointly identify road surface types using multi-source data such as images, inertial navigation, and vehicle signals. These methods typically require rigorous sensor calibration and spatiotemporal alignment, resulting in high system complexity during engineering deployment. For example, existing solutions that use image and vehicle response feature fusion to identify road surface types are dependent on image input. When the vehicle is in low-light, occluded, or mud-covered scenarios, the image-based output may be significantly affected, placing higher demands on the fusion strategy. Vehicle-mounted dynamic signal methods are mainly based on extracting features from the vehicle's own sensor signals and classifying or discriminating them. They have a weak dependence on external sensors, strong engineering feasibility, and are almost unaffected by external environmental factors such as lighting and occlusion. However, they lack support for the continuity and stable output of off-road multi-type terrain recognition and do not provide an interpretable, calibrable terrain recognition structure with arbitration post-processing capabilities.
[0004] In summary, existing technologies have the following shortcomings: some solutions rely heavily on external sensing sensors and are easily affected by environmental interference under typical off-road conditions; some vehicle-mounted signal solutions are biased towards road event or roughness assessment and lack sufficient support for multi-category off-road terrain classification; some solutions lack a unified confidence output mechanism and timing arbitration mechanism, which can easily lead to jitter in recognition results and affect the continuity of control strategies; and some solutions lack an engineering parameter organization method, increasing the difficulty of calibration and deployment. Therefore, there is a need for an off-road terrain recognition method that does not rely on expensive external sensors, provides stable recognition results, and is easy to calibrate in engineering. Summary of the Invention
[0005] To address the aforementioned technical problems, this embodiment provides an off-road terrain recognition method based on feature membership fusion and temporal arbitration, comprising the following steps:
[0006] S1. Acquire vehicle-owned sensor signals, wherein the vehicle-owned sensor signals include at least one of wheel speed, vehicle speed, longitudinal acceleration, lateral acceleration, vertical acceleration, drive torque, braking pressure, and suspension response;
[0007] S2. Preprocess the vehicle's own sensor signals, the preprocessing including at least one of filtering, noise reduction, synchronization and normalization;
[0008] S3. Calculate the basic characteristics within a preset time window, wherein the basic characteristics include at least one of rolling resistance and slip ratio;
[0009] S4. The basic features are segmented into windows, and statistical features and / or time-domain features are calculated within the windows as secondary features to form a feature vector of fixed dimensions.
[0010] S5, for each terrain category The confidence level of the terrain category is calculated based on the input feature vector using offline statistical parameters.
[0011] The offline statistical parameters include the effective interval, equal-width binning threshold, interval membership table, and feature weight table; for the i-th dimension of the feature vector... ,like Not in category effective range Inside, then this feature Category membership degree If it is within the valid range, then the threshold for equal-width binning is used to determine the value. Determine the interval to which it belongs, and then look up the interval membership table to obtain the membership degree. The category score is obtained by fusing the membership degrees of each dimension according to the feature weight table. :
[0012]
[0013] And obtain the confidence level ,in, Indicates the feature weights.
[0014] Further, step S5 includes the following steps:
[0015] S5.1 Valid interval determination;
[0016] The effective intervals for each category and each feature were obtained by statistical analysis using offline training data.
[0017] Category With feature dimension Define the lower limit of the interval and the upper limit of the interval They are respectively:
[0018]
[0019] During online inference, if the input features Not in the valid range Within, the input feature is related to the category. Set the membership degree to 0:
[0020]
[0021] in, Indicates the terrain category index. , Indicates the total number of terrain categories; Indicates the feature dimension index. ; Indicates the categories in the training set The sample in the first The set of values in a dimension; This represents the quantile operator, which returns a set of... quantiles; Indicate category Features The membership function, with an output range of ;
[0022] S5.2 equal-width binning threshold;
[0023] For each feature dimension Preset number of boxes Within the effective range The upper part is divided into equal widths. Each segment has the same width;
[0024] The width of equal-width sub-boxes is defined as:
[0025]
[0026] The threshold sequence is defined as:
[0027]
[0028] The container interval is defined as:
[0029]
[0030] in, Indicates the first The number of bins for the dimensional feature is a preset parameter; Indicate category With features Equal width of the box; The equal-width binning threshold representing the lower limit of the interval. Indicates the first Each equal-width binning threshold The equal-width binning threshold representing the upper limit of the interval; Indicates the first Each container compartment;
[0031] S5.3 Initialization of interval probability proportion and interval membership;
[0032] After binning to equal width, the percentage of training samples falling into each bin interval is statistically analyzed, and the interval percentage is defined. for:
[0033]
[0034] in, This indicates that the training samples fall within the binning interval. The number of samples; Indicate category In feature dimension The total number of samples;
[0035] The initialization rule is to set the membership degree of the largest interval proportion to 1.
[0036]
[0037] The initial membership degrees of the remaining intervals are normalized according to the proportion of the interval with the largest relative membership degree:
[0038]
[0039] in, Indicates the index of the maximum range percentage; This indicates the index operator that maximizes the objective. Indicates the initial value of the interval membership degree;
[0040] Subsequently, the interval membership degrees were recalibrated during offline training, while maintaining the interval membership degree constraints. Online membership output is a piecewise constant:
[0041]
[0042] in, This represents the interval membership calibration value, which is a calibrable parameter; Representation of features Category The membership degree, the output range is ;
[0043] S5.4 Feature weight fusion and category confidence calculation;
[0044] For each category Set feature weights The feature weights satisfy:
[0045]
[0046] Category score is defined as:
[0047]
[0048] Confidence level is defined as:
[0049]
[0050] in, Indicate category For feature dimensions The weights; Indicate category At any moment Category scores, ranging from ; Indicate category At any moment The confidence level, ranging from .
[0051] S6. Execute the maximum value logic and take the terrain category with the highest category score as the instantaneous terrain category; when there are ties for the maximum value, perform a tie-breaking process according to the preset terrain priority sequence to obtain the instantaneous terrain category;
[0052] Furthermore, the instantaneous terrain category is represented as:
[0053]
[0054] in, Indicates time Instantaneous terrain category; This indicates the index operator that maximizes the objective. Indicate category At any moment Category scores;
[0055] When there are ties for the highest value, a tie is processed based on priority; the priority is preset by engineering requirements; terrain with higher difficulty is assigned higher priority, and the instantaneous terrain category output for tie processing is:
[0056]
[0057] in, Represents the set of the largest parallel categories; Indicate category The priority sequence parameter.
[0058] S7. Employ a sliding window voting consensus logic, with a length of... Within a sliding window, the percentage of each instantaneous terrain type is counted, and the type with the highest percentage and reaching a consistency threshold is selected. The category is output as the final terrain category;
[0059] Furthermore, a sliding window voting consensus decision is adopted to suppress jitter, and the length of the sliding window is denoted as... The consistency threshold is denoted as The percentage of instantaneous terrain types appearing within the window is statistically analyzed. for:
[0060]
[0061] in, This represents the first sampling point within the sliding window voting filter window; This indicates an indicator function that takes the value 1 if the condition is true and 0 otherwise.
[0062] According to the category with the largest proportion Make a consistency judgment:
[0063]
[0064] in, This indicates the index operator that maximizes the objective.
[0065] The final output category is defined as:
[0066]
[0067] in, This indicates the final output terrain category at time t. This represents the maximum percentage of each instantaneous terrain type appearing within the sliding window at time t.
[0068] S8. Output the final terrain category and its corresponding confidence level.
[0069] Compared with the prior art, the present invention has at least the following beneficial effects:
[0070] First, the present invention mainly relies on the vehicle's own signals to achieve terrain recognition, which can reduce the dependence on external environmental sensors such as cameras and lidar. Therefore, it reduces the impact of off-road environmental factors such as changes in lighting, occlusion, mud and water coverage and dust on the recognition effect, and also reduces hardware costs and system integration complexity.
[0071] Second, the present invention adopts a confidence calculation mechanism that integrates feature membership degree and weight, which can output confidence vectors for each terrain. The confidence has a uniform range and interpretable meaning, which makes it convenient for the controller to switch strategies, perform progressive control or protection control based on the confidence, and also facilitates diagnosis and calibration during the experimental stage.
[0072] Third, this invention uses effective intervals, equal-width binning thresholds, interval membership tables, and weight tables obtained through offline statistics to organize complex distribution information into parameters that can be looked up in tables. This facilitates offline calibration, parameter version management, and rapid iteration in MATLAB / Simulink or other environments, and also makes it easy to deploy in embedded controllers to meet real-time requirements.
[0073] Fourth, this invention introduces maximum value arbitration, priority tie processing and sliding window voting consistency logic, which can suppress category jitter caused by instantaneous misjudgment, improve the continuity and stability of output categories, and better adapt to the requirements of off-road control strategies for mode stability.
[0074] Fifth, this invention supports special requirement constraint logic, which can set entry and exit conditions for specific terrain patterns, thus enhancing the controllability and safety of the project. Attached Figure Description
[0075] Figure 1 This is a schematic diagram of the off-road terrain recognition system architecture according to an embodiment of the present invention.
[0076] Figure 2 This is a schematic diagram of the overall process of the off-road terrain recognition method described in this invention;
[0077] Figure 3 This is a schematic diagram of the terrain estimator used by the present invention to calculate the confidence level of various terrain features;
[0078] Figure 4 This is a schematic diagram of the terrain arbitration and post-processing process of the present invention; Detailed Implementation
[0079] The embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the present invention is not limited to the following embodiments. Those skilled in the art can make various modifications or substitutions without departing from the concept of the present invention, and all such modifications or substitutions should fall within the protection scope of the present invention.
[0080] like Figure 1As shown, the off-road terrain recognition system based on this embodiment includes: an input signal interface unit, a discriminant feature calculation module A, a confidence calculation module B, a terrain arbitration module C, and an output interface unit. The discriminant feature calculation module A includes a preprocessing unit, a basic feature extraction unit, and a secondary feature extraction unit; the terrain arbitration module includes maximum value logic, priority logic, sliding window voting consistency logic, and special requirement constraint logic.
[0081] Example 1
[0082] like Figure 2 As shown in the figure, this embodiment provides an off-road terrain recognition method based on feature membership fusion and temporal arbitration, which includes the following steps:
[0083] S1. Acquire signals from the vehicle's own sensors;
[0084] The input signal interface unit acquires the vehicle's own sensor signals, including wheel speed, vehicle speed, longitudinal acceleration, lateral acceleration, vertical acceleration, drive torque, braking pressure, and suspension response; the system updates the identification results at a fixed period or a fixed window.
[0085] S2. Preprocess the signals from the vehicle's own sensors;
[0086] The preprocessing unit preprocesses the input vehicle's own sensor signals, including filtering, noise reduction, synchronization, and normalization.
[0087] S3. Calculate the basic features within the preset time window;
[0088] The basic feature extraction unit is used to calculate basic features, which include vehicle dynamic response features such as rolling resistance and slip ratio that can indirectly and interpretably characterize road surface characteristics.
[0089] S4. Calculate secondary features to form and output high-dimensional feature vectors;
[0090] The secondary feature extraction unit segments the basic features into windows and calculates statistical features and / or temporal features within the window as secondary features to form a fixed-dimensional feature vector, thereby achieving high-dimensional derivation of the basic features.
[0091] The final output feature vector is:
[0092]
[0093] in, Indicates the window index or the index for identifying the time; Indicates time eigenvectors; Indicates the first dimensional features at time The value of , ; Indicates the number of feature dimensions.
[0094] S5. Calculate the confidence level for each terrain category;
[0095] like Figure 3 As shown, the confidence calculation module is designed for each terrain category. Construct a terrain estimator that calculates the confidence level of the terrain category based on the input feature vector using offline statistical parameters;
[0096] The offline statistical parameters include the effective interval, equal-width binning threshold, interval membership table, and feature weight table; for the i-th dimension of the feature vector... ,like Not in category effective range Inside, then this feature Category membership degree If it is within the valid range, then the threshold for equal-width binning is used to determine the value. Determine the interval to which it belongs, and then look up the interval membership table to obtain the membership degree. The category score is obtained by fusing the membership degrees of each dimension according to the feature weight table. :
[0097]
[0098] And obtain the confidence level ,in, Indicates the feature weights.
[0099] Specifically, in this embodiment, step S5 includes the following steps:
[0100] S5.1 Valid interval determination;
[0101] This embodiment uses offline training data statistics to obtain the effective interval for each category and each feature;
[0102] Category With feature dimension Define the lower limit of the interval and the upper limit of the interval They are respectively:
[0103]
[0104] During online inference, if the input features Not in the valid range Within, the input feature is related to the category. Set the membership degree to 0:
[0105]
[0106] in, Indicates the terrain category index. , Indicates the total number of terrain categories; Indicates the feature dimension index. ; Indicates the categories in the training set The sample in the first The set of values in a dimension; This represents the quantile operator, which returns a set of... quantiles; Indicate category Features The membership function, with an output range of ;
[0107] S5.2 equal-width binning threshold;
[0108] This embodiment addresses each feature dimension. Preset number of boxes Within the effective range The upper part is divided into equal widths. Each segment has the same width;
[0109] The width of equal-width sub-boxes is defined as:
[0110]
[0111] The threshold sequence is defined as:
[0112]
[0113] The container interval is defined as:
[0114]
[0115] in, Indicates the first The number of bins for the dimensional feature is a preset parameter; Indicate category With features Equal width of the box; The equal-width binning threshold representing the lower limit of the interval. Indicates the first Each equal-width binning threshold The equal-width binning threshold representing the upper limit of the interval; Indicates the first Each container compartment;
[0116] S5.3 Initialization of interval probability proportion and interval membership;
[0117] In this embodiment, after equal-width binning, the percentage of training samples falling into each bin interval is statistically analyzed, and the interval percentage is defined. for:
[0118]
[0119] in, This indicates that the training samples fall within the binning interval. The number of samples; Indicate category In feature dimension The total number of samples;
[0120] In this embodiment, the membership degree of the maximum interval proportion is 1 as the initialization rule:
[0121]
[0122] The initial membership degrees of the remaining intervals are normalized according to the proportion of the interval with the largest relative membership degree:
[0123]
[0124] in, Indicates the index of the maximum range percentage; This indicates the index operator that maximizes the objective. Indicates the initial value of the interval membership degree;
[0125] Subsequently, the interval membership degrees were recalibrated during offline training, while maintaining the interval membership degree constraints. The online membership output is a piecewise constant, resulting in an interval membership table:
[0126]
[0127] in, This represents the interval membership calibration value, which is a calibrable parameter; Representation of features Category The membership degree, the output range is ;
[0128] S5.4 Feature weight fusion and category confidence calculation;
[0129] This embodiment describes each category Set feature weights The feature weights satisfy:
[0130]
[0131] Category score is defined as:
[0132]
[0133] Confidence level is defined as:
[0134]
[0135] in, Indicate category For feature dimensions The weights; Indicate category At any moment Category scores, ranging from ; Indicate category At any moment The confidence level, ranging from .
[0136] S6. Take the terrain category with the highest category score as the instantaneous terrain category;
[0137] Specifically, the terrain arbitration module executes the maximum value logic, taking the terrain category with the highest category score as the instantaneous terrain category; when there are ties for the maximum value, a tie is processed according to a preset terrain priority sequence to obtain the instantaneous terrain category;
[0138] like Figure 4 As shown, the terrain arbitration module includes maximum value logic, priority logic, sliding window voting consensus logic, and special requirement constraint logic; the instantaneous terrain category is represented as:
[0139]
[0140] in, Indicates time Instantaneous terrain category; This indicates the index operator that maximizes the objective. Indicate category At any moment Category scores;
[0141] When there are ties for the highest value, the terrain arbitration module performs a tie-breaking process based on priority. Priority is preset by engineering requirements. In this embodiment, terrain with higher difficulty in escaping is assigned a higher priority, namely, rock category > mud category > sand category > snow category > road category. The instantaneous terrain category output for the tie-breaking process is:
[0142]
[0143] in, Represents the set of the largest parallel categories; Indicate category The priority sequence parameter.
[0144] S7. Employ a sliding window voting consensus logic, with a length of... Within a sliding window, the percentage of each instantaneous terrain type is counted, and the type with the highest percentage and reaching a consistency threshold is selected. The category is output as the final terrain category;
[0145] Specifically, a sliding window voting consensus decision is used to suppress jitter, and the length of the sliding window is denoted as... This is a preset parameter; a reasonable value can be selected based on the operating cycle of the vehicle system. For example, if the operating cycle of the vehicle system is 10 milliseconds, then... A value of 200 can be used to avoid frequent jumps; the consistency threshold is denoted as... , which is a preset parameter, such as 0.95; Statistical analysis of the percentage of terrain types appearing instantaneously within the window. for:
[0146]
[0147] in, This represents the first sampling point within the sliding window voting filter window; This indicates an indicator function that takes the value 1 if the condition is true and 0 otherwise.
[0148] According to the category with the largest proportion Make a consistency judgment:
[0149]
[0150] in, This indicates the index operator that maximizes the objective.
[0151] The final output category is defined as:
[0152]
[0153] in, This indicates the final output terrain category at time t. This represents the maximum percentage of each instantaneous terrain type appearing within the sliding window at time t.
[0154] When the consistency threshold With a value of 0.95, this strategy can output a new instantaneous terrain category when the 95% consistency requirement is not met.
[0155] S8, Output final terrain category And the corresponding confidence level information.
[0156] Example 2
[0157] This embodiment provides an off-road terrain recognition method based on feature membership fusion and temporal arbitration. Based on steps S1-S7 of embodiment 1, step S8 is added: special requirement constraint logic.
[0158] This embodiment allows for the inclusion of special requirement constraint logic steps. The logic in step S8 is used to restrict the entry or exit conditions for certain terrain modes. For example, when the final terrain is rock and the vehicle speed is below a threshold, the vehicle is likely still near the rock terrain, or temporarily in an atypical rock scenario with relatively gentle excitation, but this does not mean that the vehicle has already left the rock terrain. Therefore, even if the confidence level of other terrains is high enough, the vehicle should not exit the rock terrain. Only when the vehicle begins to gradually accelerate to a speed or acceleration threshold that is difficult to reach on the rock terrain should it be allowed to exit the rock terrain. This logic is used in engineering to enhance the safety and controllability of the strategy.
[0159] Step S9 is the same as step S8 in Example 1.
Claims
1. A method for off-road terrain recognition based on feature membership fusion and temporal arbitration, characterized in that: Includes the following steps: S1. Acquire signals from the vehicle's own sensors; S2. Preprocess the signals from the vehicle's own sensors; S3. Calculate the basic characteristics within a preset time window, wherein the basic characteristics include at least one of rolling resistance and slip ratio; S4. The basic features are segmented into windows, and statistical features and / or time-domain features are calculated within the windows as secondary features to form a feature vector of fixed dimensions. S5, for each terrain category The confidence level of the terrain category is calculated based on the input feature vector using offline statistical parameters; the offline statistical parameters include the effective interval, equal-width binning threshold, interval membership table, and feature weight table. For the i-th feature in the feature vector ,like Not in category effective range Inside, then this feature Category membership degree ; If it is within the valid range, then it is determined according to the equal-width binning threshold. Determine the interval to which it belongs, and then look up the interval membership table to obtain the membership degree. ; The category score is obtained by fusing the membership degrees of each dimension according to the feature weight table. : And obtain the confidence level ,in, Indicates feature weights; S6. Execute the maximum value logic and take the terrain category with the highest category score as the instantaneous terrain category; when there are ties for the maximum value, perform a tie-breaking process according to the preset terrain priority sequence to obtain the instantaneous terrain category; S7. Employ a sliding window voting consensus logic, with a length of... Within a sliding window, the percentage of each instantaneous terrain type is counted, and the type with the highest percentage and reaching a consistency threshold is selected. The category is output as the final terrain category; S8. Output the final terrain category and its corresponding confidence level.
2. The off-road terrain recognition method based on feature membership fusion and temporal arbitration according to claim 1, characterized in that: In step S5, the determination of the valid interval is as follows: The effective intervals for each category and each feature were obtained by statistical analysis using offline training data. Category With feature dimension Define the lower limit of the interval and the upper limit of the interval They are respectively: During online inference, if the input features Not in the valid range Within, the input feature is related to the category. Set the membership degree to 0: in, Indicates the terrain category index. , Indicates the total number of terrain categories; Indicates the feature dimension index. ; Indicates the categories in the training set The sample in the first The set of values in a dimension; This represents the quantile operator, which returns a set of... quantiles; Indicate category Features The membership function, with an output range of .
3. The off-road terrain recognition method based on feature membership fusion and temporal arbitration according to claim 1, characterized in that: In step S5, the equal-width binning threshold is defined as follows: For each feature dimension Preset number of boxes Within the effective range The upper part is divided into equal widths. Each segment has the same width; The width of equal-width sub-boxes is defined as: The threshold sequence is defined as: The container interval is defined as: in, Indicates the first The number of bins for the dimensional feature is a preset parameter; Indicate category With features Equal width of the box; The equal-width binning threshold representing the lower limit of the interval. Indicates the first Each equal-width binning threshold The equal-width binning threshold representing the upper limit of the interval; Indicates the first Each container compartment.
4. The off-road terrain recognition method based on feature membership fusion and temporal arbitration according to claim 1, characterized in that: In step S5, the interval membership degree is calculated as follows: After binning to equal width, the percentage of training samples falling into each bin interval is statistically analyzed, and the interval percentage is defined. for: in, This indicates that the training samples fall within the binning interval. The number of samples; Indicate category In feature dimension The total number of samples; The initialization rule is to set the membership degree of the largest interval proportion to 1. The initial membership degrees of the remaining intervals are normalized according to the proportion of the interval with the largest relative membership degree: in, Indicates the index of the maximum range percentage; This indicates the index operator that maximizes the objective. Indicates the initial value of the interval membership degree; Subsequently, the interval membership degrees were recalibrated during offline training, while maintaining the interval membership degree constraints. Online membership output is a piecewise constant: in, This represents the interval membership calibration value, which is a calibrable parameter; Representation of features Category The membership degree, the output range is .
5. The off-road terrain recognition method based on feature membership fusion and temporal arbitration according to claim 1, characterized in that: In step S5, the feature weight fusion and class confidence calculation are as follows: For each category Set feature weights The feature weights satisfy: Category score is defined as: Confidence level is defined as: in, Indicate category For feature dimensions The weights; Indicate category At any moment Category scores, ranging from ; Indicate category At any moment The confidence level, ranging from .
6. The off-road terrain recognition method based on feature membership fusion and temporal arbitration according to claim 1, characterized in that: The instantaneous terrain category mentioned in step S6 is represented as follows: in, Indicates time Instantaneous terrain category; This indicates the index operator that maximizes the objective. Indicate category At any moment Category scores; When there are ties for the highest value, a tie is processed based on priority; the priority is preset by engineering requirements; terrain with higher difficulty is assigned higher priority, and the instantaneous terrain category output for tie processing is: in, Represents the set of the largest parallel categories; Indicate category The priority sequence parameter.
7. The off-road terrain recognition method based on feature membership fusion and temporal arbitration according to claim 1, characterized in that: In step S7, a sliding window voting consensus decision is used to suppress jitter. The length of the sliding window is denoted as... The consistency threshold is denoted as The percentage of instantaneous terrain types appearing within the window is statistically analyzed. for: in, This represents the first sampling point within the sliding window voting filter window; This indicates an indicator function that takes the value 1 if the condition is true and 0 otherwise. According to the category with the largest proportion Make a consistency judgment: in, This indicates the index operator that maximizes the objective. The final output category is defined as: in, This indicates the final output terrain category at time t. This represents the maximum percentage of each instantaneous terrain type appearing within the sliding window at time t.
