A terrain recognition method based on a DEM terrain recognition model

Abstract

This application relates to the field of terrain recognition, and particularly to a terrain recognition method based on a DEM terrain recognition model. The method includes: determining an implementable gait and DEM terrain information; obtaining an initial data set determined after a hexapod robot walks on a simulated terrain through the implementable gait by means of a tactile system of a foot-end pressure sensor; preprocessing the initial data set to obtain a training data set, and training a basic PNN with the training data set to obtain a DEM terrain recognition model; obtaining an actual data set including a plurality of pressure signals determined after the hexapod robot walks on a target terrain through the implementable gait, inputting the actual data set into the DEM terrain recognition model, and the DEM terrain recognition model outputs a terrain type. This application trains the DEM terrain recognition model with the pressure signals received by the tactile system designed at the foot end of the hexapod robot, replacing the image recognition mode in the traditional method, and solving the problem that a large amount of effort is required for traditional feature extraction methods.

Description

Technical Field

[0001] This application relates to the field of terrain recognition, and particularly to a terrain recognition method based on a DEM terrain recognition model. Background Art

[0002] In the research of terrain recognition, terrain recognition methods include unimodal terrain recognition methods and multimodal terrain recognition methods. Unimodal terrain recognition methods include vision-based terrain recognition methods, audio-based terrain recognition methods, vibration-based terrain recognition methods, lidar-based terrain recognition methods, and touch-based terrain recognition methods, etc. Multimodal terrain recognition methods include various sensor fusion-based terrain recognition methods. These methods can realize the recognition of local features (such as obstacle size, distance, etc.) of the terrain in a specific environment.

[0003] For different modal terrain recognition methods, the extracted terrain features are different, so the adopted terrain recognition algorithms are also different. Currently, the classification algorithms used for terrain recognition mainly include tensor voting algorithm, decision tree model, support vector machine (SVM), random forest (RF), k-nearest neighbor algorithm (KNN), naive Bayes algorithm (NB), extreme learning machine (ELM), and convolutional neural network. Convolutional neural network is used to extract the features required for terrain recognition in visual images. Using traditional feature extraction methods to extract picture features requires manual selection and optimization of algorithms, with a long design cycle and a large amount of effort.

[0004] In addition, although both unimodal and multimodal terrain recognition methods can complete terrain recognition, the terrain parameters recognized by these methods do not include geomorphic factors such as slope, aspect, and slope change rate, so they cannot be fused with DEM data. As a digital expression of the terrain surface, DEM data has advantages that traditional topographic maps cannot match and is applicable to all industries and application fields. Summary of the Invention

[0005] This application provides a terrain recognition method based on a DEM terrain recognition model, which can solve the problems existing in the prior art that DEM data cannot be applied, and the problems that traditional feature extraction methods are used to extract picture features, require manual selection and optimization of algorithms, have a long design cycle, and require a large amount of effort.

[0006] The technical solution of this application is a terrain recognition method based on a DEM terrain recognition model, which is implemented based on a hexapod robot and a foot-end pressure sensor tactile system used in cooperation with the hexapod robot. The method includes:

[0007] S1: Determine the implementable gait of the hexapod robot and the DEM terrain information;

[0008] S2: Through the tactile system of the foot-end pressure sensors, obtain the initial data set including the pressure signals corresponding to different implemented gaits and the DEM terrain information determined after the hexapod robot walks on the simulated terrain corresponding to the DEM terrain information by implementing the gait.

[0009] S3: Preprocess the initial data set to correspondingly obtain the training data set, and train the basic PNN with the training data set to obtain a DEM terrain recognition model with the pressure signal corresponding to the implemented gait as the input value and the terrain type corresponding to the DEM terrain information as the output value.

[0010] S4: Obtain the actual data set including a number of pressure signals determined after the hexapod robot walks on the target terrain by implementing the gait and input the actual data set into the DEM terrain recognition model, and the DEM terrain recognition model outputs the terrain type corresponding to the target terrain.

[0011] Optionally, the DEM terrain information includes: slope, slope change rate, and slope aspect.

[0012] And, the terrain types corresponding to the DEM terrain information include: flat ground, convex, concave, uphill slope, steep slope, and gentle slope.

[0013] Optionally, the tactile system of the foot-end pressure sensors includes: a central control module, and a servo control module and an environment perception module respectively connected to the central control module.

[0014] And, the step S2 includes:

[0015] S21: Send a control command from the central control module to the servo control module, and the servo control module controls the hexapod robot to walk on the simulated terrain corresponding to the DEM terrain information by implementing the gait.

[0016] S22: Through the environment perception module arranged at the foot-end, obtain a number of pressure values corresponding to different implemented gaits and transmit the pressure values to the central control module.

[0017] S23: The central control module converts the pressure values into pressure signals and outputs an initial data set including a number of pressure signals corresponding to different implemented gaits.

[0018] Optionally, the step S3 includes:

[0019] S31: Perform normalization processing on the initial data set to obtain a preprocessed data set.

[0020] S32: Extract the DEM features from the preprocessed data set to obtain a feature extraction set including the pressure signals corresponding to different implemented gaits and the DEM terrain information.

[0021] S33: Divide the feature extraction set in a ratio of 7:3 to obtain a training data set and a test data set accordingly;

[0022] S34: Train the basic PNN with the training data set to obtain an initial DEM model with the pressure signal corresponding to the implemented gait as the input value and the terrain type corresponding to the DEM terrain information as the output value;

[0023] S35: Based on the expansion coefficient of the radial basis function, test the initial DEM models corresponding to the expansion coefficients of different radial basis functions with the test data set to obtain the terrain recognition accuracy corresponding to the expansion coefficient of the radial basis function;

[0024] S36: Select the initial DEM model corresponding to the expansion coefficient of the radial basis function with the highest terrain recognition accuracy as the DEM terrain recognition model.

[0025] Optionally, the implemented gait includes: tripod gait, quadruped gait, and pentapod gait;

[0026] And, the training data set includes: tripod gait training set, quadruped gait training set, and pentapod gait training set;

[0027] And, the types of the DEM terrain recognition models include: tripod gait recognition PNN, quadruped gait recognition PNN, and pentapod gait recognition PNN.

[0028] And, the step S34 includes:

[0029] S341: Train different basic PNNs with the tripod gait training set, quadruped gait training set, and pentapod gait training set respectively to obtain an initial tripod gait PNN, an initial quadruped gait PNN, and an initial pentapod gait PNN with the pressure signal corresponding to the tripod gait as the input value and the terrain type corresponding to the DEM terrain information as the output value;

[0030] And, the test data set includes: tripod gait test set, quadruped gait test set, and pentapod gait test set;

[0031] And, the step S35 includes:

[0032] S351: Determine the expansion coefficients of several radial basis functions for the initial DEM model;

[0033] S352: Based on the expansion coefficients of the radial basis function, the initial PNNs for three-legged gait, four-legged gait, and five-legged gait are respectively tested through the three-legged gait test set, four-legged gait test set, and five-legged gait test set, and the terrain recognition accuracy corresponding to the expansion coefficients of the radial basis function is obtained;

[0034] And, the step S36 includes:

[0035] S361: For the initial PNNs of three-legged gait, four-legged gait, and five-legged gait, the initial PNNs of three-legged gait, four-legged gait, and five-legged gait corresponding to the expansion coefficients of the radial basis function corresponding to the highest terrain recognition accuracy are respectively selected as the recognition PNNs of three-legged gait, four-legged gait, and five-legged gait.

[0036] Optionally, the training data set includes a number of training samples;

[0037] The training samples include: pressure signals and DEM terrain information corresponding to different implementation gaits;

[0038] The structure of the DEM terrain recognition model includes: an input layer, a pattern layer, a summation layer, and an output layer;

[0039] And, the step S4 includes:

[0040] S41: Obtain the actual data set including a number of pressure signals determined after the hexapod robot walks on the target terrain through the implementation gait and input the actual data set into the DEM terrain recognition model;

[0041] S42: Through the input layer, it is used to receive the actual data set corresponding to the target terrain;

[0042] S43: Through the pattern layer, calculate a number of initial distance amounts in the actual data set corresponding to the training samples based on the training samples;

[0043] S44: Through the summation layer, calculate the distance weighted amounts corresponding to the classification of the training samples of a number of initial distance amounts in the actual data set based on the classification of the training samples;

[0044] S45: Through the output layer, perform normalization and probability calculation operations on the distance weighted amounts in sequence to obtain the terrain type corresponding to the target terrain.

[0045] Beneficial effects:

[0046] This application takes a hexapod robot as the implementation object. Based on the characteristics that the legs of the hexapod robot are flexible and can adapt to the surrounding environment faster and better, a tactile system is designed at the foot end of the hexapod robot. During the movement of the hexapod robot, the values of the foot-end pressure sensors contain rich terrain information. The terrain feature values and terrain classification are determined through on-site DEM data. The tactile force feedback value is used as the input value of the DEM terrain recognition model, and the DEM terrain recognition model is used for terrain classification recognition. Through experiments, the accuracy of the hexapod robot's terrain recognition is verified and evaluated. After repeated optimization, the terrain can be accurately recognized.

[0047] In summary, this application uses DEM data to improve the accuracy of terrain recognition, and trains the DEM terrain recognition model through the pressure signals received by the tactile system designed at the foot end of the hexapod robot, replacing the image recognition mode in the traditional method. It can effectively solve the problems of the traditional feature extraction method for extracting picture features, which requires manual selection and optimization of algorithms, has a long design cycle, and requires a large amount of effort. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of this application, the accompanying drawings required for the embodiments will be briefly introduced below. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0049] Figure 1 It is a schematic flowchart of the terrain recognition method based on the DEM terrain recognition model in the embodiment of this application;

[0050] Figure 2 It is a schematic structural diagram of the foot-end pressure sensor tactile system in the embodiment of this application;

[0051] Figure 3 It is a data acquisition flowchart of the foot-end pressure sensor tactile system in the embodiment of this application;

[0052] Figure 4 is a data graph of the pressure values obtained when the hexapod robot implements different implementation gaits in the embodiment of this application;

[0053] Figure 5 is a schematic structural diagram of the probabilistic neural network corresponding to different implementation gaits in the embodiment of this application;

[0054] Figure 6 is a data graph of the pressure values after the hexapod robot implements three implementation gaits in the first terrain in the embodiment of this application;

[0055] Figure 7 is a data graph of the pressure values after the hexapod robot implements three implementation gaits in the second terrain in the embodiment of this application;

[0056] Figure 8 is a graph of pressure value data after the six-legged robot implements three implementation gaits in the third terrain in the embodiment of the present application;

[0057] Figure 9 It is a data distribution diagram of the training data set in the embodiment of the present application;

[0058] Figure 10 It is a data distribution diagram of the test data set in the embodiment of the present application;

[0059] Figure 11 It is the running time of the DEM terrain recognition model in the embodiment of the present application;

[0060] Figure 12 is a topographic schematic diagram in the comparative experiment of the DEM terrain recognition model and other terrain recognition algorithms in the embodiment of the present application;

[0061] Figure 13 It is a schematic flow diagram of the DEM terrain recognition model in the embodiment of the present application;

[0062] In the figure, 1 - Central control module; 2 - Servo control module; 3 - Environment perception module. Detailed implementation mode

[0063] The embodiments will be described in detail below, and the examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementation manners described in the following embodiments do not represent all implementation manners consistent with the present application. They are only examples of systems and methods consistent with some aspects of the present application described in detail in the claims.

[0064] The embodiment of the present application studies the terrain recognition method of the robot, and proposes a terrain recognition method based on the DEM terrain recognition model. This method can recognize terrain parameters such as slope, aspect, and slope change rate, and then realize terrain classification. The obtained terrain parameters such as slope, aspect, and slope change rate meet the DEM data requirements.

[0065] Specifically, this method is implemented based on a six-legged robot and a tactile system of foot-end pressure sensors used in conjunction with the six-legged robot, as Figure 1 shown, Figure 1 It is a schematic flow diagram of the terrain recognition method based on the DEM terrain recognition model in the embodiment of the present application. The method includes:

[0066] S1: Determine the implementable implementation gaits of the six-legged robot and the DEM terrain information.

[0067] Specifically, as the basic data of national geographic information, DEM is widely used in the fields related to terrain feature detection. The geomorphic factors of DEM include slope change rate, aspect, and slope, etc. The slope (I) represents the degree of steepness of the surface unit, and is usually defined as the ratio of the vertical height h of the slope surface to the horizontal distance l, that is, the tangent value of the slope angle. The slope change rate (S) represents the reflection of the terrain's concave and convex changes, reflecting the terrain structure and morphology. The aspect (A) is the direction of the projection of the slope normal on the horizontal plane.

[0068] The terrain data comes from the tactile system of the foot-end pressure sensors of the hexapod robot. Due to the limitations of the body posture and walking structure of the hexapod robot, the hexapod robot cannot move on terrains with high slopes or high slope change rates. Therefore, based on the DEM data, the slope grading table that can be recognized in the embodiments of the present application is shown in Table 1, the slope change rate grading is shown in Table 2, and the aspect grading table that can be recognized is shown in Table 3.

[0069] Table 4 is the correspondence table between the terrain feature values and the terrain classifications in the embodiments of the present application.

[0070] Table 1 Slope Grading Table

[0071]

[0072] Table 2 Slope Change Rate Grading Table

[0073]

[0074] Table 3 Aspect Grading Table

[0075]

[0076] Table 4 Correspondence Table between Terrain Feature Values and Terrain Classifications

[0077]

[0078]

[0079] S2: Through the tactile system of the foot-end pressure sensors, obtain the initial data set including the pressure signals corresponding to different implemented gaits and the DEM terrain information determined after the hexapod robot walks through the implemented gait on the simulated terrain corresponding to the DEM terrain information;

[0080] And, step S2 includes:

[0081] S21: Send a control command from the central control module to the servo control module, and the servo control module controls the hexapod robot to walk through the implemented gait on the simulated terrain corresponding to the DEM terrain information;

[0082] S22: Obtain several pressure values corresponding to different implemented gaits through the environmental perception module set at the foot end, and transmit the pressure values to the central control module;

[0083] S23: The central control module converts the pressure values into pressure signals and outputs an initial data set including several pressure signals corresponding to different implemented gaits.

[0084] Specifically, the implemented gaits include: tripod gait, quadruped gait, and pentapod gait.

[0085] In the embodiment of the present application, the hexapod robot traverses and walks to sense the terrain of the target area. The hexapod robot can adopt various gaits when walking on the ground. Under the same terrain conditions, the characteristics of the sampling data obtained by the tactile system of the hexapod robot are inconsistent when it walks with different gaits.

[0086] The hexapod robot senses the terrain information by traversing the target area and requires an appropriate tactile information feedback system. In this paper, a tactile system of foot-end pressure sensors is designed and installed on the soles of the feet of the hexapod robot. The leg structure of the overall structure of the hexapod robot includes arm servo motors, leg servo motors, foot servo motors, and foot ends.

[0087] As Figure 2 shown, Figure 2 is a schematic structural diagram of the tactile system of foot-end pressure sensors in the embodiment of the present application. The tactile system of foot-end pressure sensors includes: a central control module 1, and a servo control module 2 and an environmental perception module 3 respectively connected to the central control module 1.

[0088] The environmental perception module 3 senses the pressure value between the foot end and the ground through the foot-end pressure sensors F i (i = 1, …, 6).

[0089] The central control module 1 collects the analog signals output by the environmental perception module 3, and converts them into digital signals by the AD module inside the STM32F103VET6 to complete the acquisition of environmental information.

[0090] The servo control module 2 includes a servo control board and servo motors.

[0091] The servo control board communicates with the central control module 1 through a serial port, obtains the commands of the central control module 1, and then controls the servo motors to complete the corresponding gaits.

[0092] As Figure 3 shown, Figure 3 is the data acquisition flowchart of the tactile system of foot-end pressure sensors in the embodiment of the present application. The central control module 1 controls the AD module to sample the output values of the foot-end pressure acquisition circuits C i (i = 1, …, 6) periodically, and the sampling frequency is 10 Hz.

[0093] In the process of obtaining the initial data set, it is necessary to collect relevant tactile data of the hexapod robot under different conditions through different implemented gaits.

[0094] The tests are as follows:

[0095] The test terrain conditions are: slope I = 0, slope change rate S = 0, and slope length 1.5 m.

[0096] As shown in Figure 4, Figure 4 is a graph of pressure value data obtained when the hexapod robot implements different implemented gaits in an embodiment of the present application.

[0097] When the hexapod robot travels in a tripod gait, the pressure values of the sole pressure sensors F i (i = 1,..., 6) are as Figure 4a shown.

[0098] When the hexapod robot travels in a quadruped gait, the pressure values of the sole pressure sensors F i (i = 1,..., 6) are as Figure 4b shown.

[0099] When the hexapod robot travels in a pentapod gait, the pressure values of the sole pressure sensors F i (i = 1,..., 6) are as Figure 4c shown.

[0100] From the test results in Figure 4, the following conclusions can be drawn:

[0101] (1) Under the same terrain conditions, when traveling in different gaits, the feedback values of the sole pressure sensors F i (i = 1,..., 6) in the tactile system of the foot-end pressure sensors are different.

[0102] (2) Under the same terrain conditions, when using one gait, the feedback values of the foot-end pressure sensors F i (i = 1,..., 6) have an obvious change pattern.

[0103] According to conclusion (2), in an embodiment of the present application, the change pattern of the feedback values of the foot-end pressure sensors F i (i = 1,..., 6) can be used to design a probabilistic neural network algorithm to identify terrain features. According to conclusion (1), the present application needs to design the structure and parameters of the probabilistic neural network algorithm for different gaits to identify terrain features.

[0104] S3: Preprocess the initial data set, and correspondingly obtain a training data set, and train the basic PNN with the training data set to obtain a DEM terrain recognition model with the pressure signal corresponding to the implemented gait as the input value and the terrain type corresponding to the DEM terrain information as the output value;

[0105] Step S3 includes:

[0106] S31: Normalize the initial data set to obtain a preprocessed data set;

[0107] S32: Extract DEM features from the preprocessed data set to obtain a feature extraction set including pressure signals and DEM terrain information corresponding to different implemented gaits;

[0108] S33: Divide the feature extraction set in a ratio of 7:3 to correspondingly obtain a training data set and a test data set;

[0109] S34: Train the basic PNN with the training data set to obtain a DEM initial model with the pressure signal corresponding to the implemented gait as the input value and the terrain type corresponding to the DEM terrain information as the output value;

[0110] And, step S34 includes:

[0111] S341: Train different basic PNNs with the tripod gait training set, quadruped gait training set, and pentapod gait training set respectively to correspondingly obtain a tripod gait initial PNN, a quadruped gait initial PNN, and a pentapod gait initial PNN with the pressure signal corresponding to the tripod gait as the input value and the terrain type corresponding to the DEM terrain information as the output value;

[0112] Specifically, the training data set includes: a tripod gait training set, a quadruped gait training set, and a pentapod gait training set;

[0113] The types of DEM terrain recognition models include: a tripod gait recognition PNN, a quadruped gait recognition PNN, and a pentapod gait recognition PNN.

[0114] The test data set includes: a tripod gait test set, a quadruped gait test set, and a pentapod gait test set;

[0115] The structure of the DEM terrain recognition model includes: an input layer, a pattern layer, a summation layer, and an output layer;

[0116] Let the input layer be x = [x 1 , …, x m T , and let m = 6, then x i (i = 1, …, 6) respectively represent the sampling values of the foot-end pressure sensors F i (i = 1, …, 6).

[0117] The output layer is Y = [y 1 , y 2 , …, y p T , where y​​i (i = 1, …, p) represents the category of training samples. In this paper, p = 6, where y i (i = 1, 2, 3, 4) represents the slope grading, y i (i = 5, 6) represents the slope change rate grading. The slope direction is calculated based on the Beidou navigation and positioning information during the walking process of the hexapod robot. Therefore, this parameter does not require the PNN network to participate in the operation. The pattern layer is D = [D 1 , D 2 , …, D q T , where q represents the number of neurons in the pattern layer, and q is the same as the number of training samples.

[0118] The main function of the pattern layer is to perform a weighted summation operation on the input signal and send it to the next layer after passing through an activation function operation.

[0119] In the embodiment of this application, the activation function is taken as the Gaussian function. The summation layer is S = [S 1 , S 2 , …, S l T , where l represents the number of neurons in the summation layer, and l is the same as the category value of the training samples, that is, l = 6, then S = [S 1 , S 2 , S 3 , S 4 , S 5 , S 6 T .

[0120] The main function of the summation layer is to perform a weighted average on the outputs of the neurons of the same category in the pattern layer. Finally, the output layer normalizes the output value of the summation layer, calculates the probabilities of the test sample corresponding to different training sample categories, and outputs the category of the sample according to the maximum probability criterion. Under different gaits, the input layer data of the probabilistic neural network all come from the sampling values of F i (i = 1, …, 6), that is, the number of neuron nodes in the input layer is 6; the output sample categories are all 6 categories.

[0121] Therefore, under different gaits, the input layer, summation layer, and output structure of the probabilistic neural network are the same.

[0122] Specifically, since the value of q in the pattern layer D = [D 1 , D 2 , …, D q T is related to the number of training samples, and the number of training samples for terrain recognition is different under different gaits. Therefore, under different gaits, the structure of the probabilistic neural network for terrain recognition is different.

[0123] ​​​​The specific processing process of the training samples of the probabilistic neural network under different gaits is as follows:

[0124] (1) Input layer

[0125] The dimension of each training sample is m = 6, that is, x = [x 1 ,…,x 6 T , and the number of samples of x i (i = 1,…,6) is n for each. The sample matrix can be obtained as follows:

[0126]

[0127] (2) Pattern layer

[0128] Each neuron in the pattern layer has a center. The distance from the data input from the input layer to each center is calculated in the pattern layer. That is, the output of the j-th neuron corresponding to x i (i = 1,…,6) in the pattern layer is as follows:

[0129]

[0130] where n is the number of training samples of x i (i = 1,…,6); σ is the smoothing factor; x ij is the center vector of x i (i = 1,…,6) in the j-th neuron in the pattern layer, and it is also the connection weight between the input layer and the pattern layer.

[0131] (3) Summation layer

[0132] The number of neurons l in the summation layer is the same as the class value of the training samples, that is, l = 6. Each neuron in this layer is responsible for weighted averaging the outputs of the neurons of the same class in the pattern layer to obtain as follows:

[0133]

[0134] where i = 1,…,6, and k i is the number of training samples of the i-th class.

[0135] (4) Output layer

[0136] The output layer normalizes the output value of the summation layer to obtain the probability y i of the test sample corresponding to different training sample classes, and takes the class corresponding to the maximum value as the expected class η(x) of the test sample.

[0137] ​

[0138] η(x) = argmax[y i ;

[0139] where p i = k i / q is the prior probability of the class. q is the number of neurons in the pattern layer, which is the total number of training samples.

[0140] In summary, different training samples n will affect the parameters of the pattern layer, summation layer, and output layer.

[0141] The training dataset for training the PNN for terrain recognition in the embodiments of this application comes from the tactile system of the foot-end pressure sensors of a hexapod robot.

[0142] The training data comes from the feedback data of the tactile system of the hexapod robot at the maximum traveling speeds of the three-legged gait, four-legged gait, and five-legged gait respectively. We design the foot-end pressure sensor tactile system to obtain feedback data at a fixed sampling frequency.

[0143] When the hexapod robot passes through a terrain of a fixed length, due to the different maximum traveling speeds of the three-legged gait, four-legged gait, and five-legged gait, the training sample n is different, so the PNN parameters are also different.

[0144] In the embodiments of this application, the designed maximum traveling speed of the three-legged gait of the hexapod robot is 0.08 m / s; the maximum traveling speed of the four-legged gait is 0.06 m / s; the maximum traveling speed of the five-legged gait is 0.03 m / s.

[0145] Set the length of the terrain that can be recognized by each PNN operation to 10 cm.

[0146] Then, n = 13 needs to be set for the three-legged gait; n = 17 needs to be set for the four-legged gait; n = 33 needs to be set for the five-legged gait.

[0147] As shown in Figure 5, Figure 5 is a schematic structural diagram of the probability neural network corresponding to different implementation gaits in the embodiments of this application.

[0148] The initial PNN of the three-legged gait is as Figure 5a shown, where the training samples are divided into 6 categories, and each category has 13 training samples.

[0149] The initial PNN of the four-legged gait is as Figure 5b shown, where the training samples are divided into 6 categories, and each category has 17 training samples.

[0150] The initial PNN of the five-legged gait is as Figure 5c shown, where the training samples are divided into 6 categories, and each category has 33 training samples.

[0151] S35: Based on the expansion coefficients of the radial basis functions, use the test data set to test the DEM initial models corresponding to the expansion coefficients of different radial basis functions, and obtain the terrain recognition accuracy corresponding to the expansion coefficients of the radial basis functions;

[0152] And, step S35 includes:

[0153] S351: Determine the expansion coefficients of several radial basis functions for the DEM initial model;

[0154] S352: Based on the expansion coefficients of the radial basis functions, use the tripod gait test set, quadruped gait test set, and pentapod gait test set to test the tripod gait initial PNN, quadruped gait initial PNN, and pentapod gait initial PNN respectively, and obtain the terrain recognition accuracy corresponding to the expansion coefficients of the radial basis functions;

[0155] S36: Select the DEM initial model corresponding to the expansion coefficient of the radial basis function with the highest terrain recognition accuracy as the DEM terrain recognition model.

[0156] And, step S36 includes:

[0157] S361: For the tripod gait initial PNN, quadruped gait initial PNN, and pentapod gait initial PNN, respectively select the tripod gait initial PNN, quadruped gait initial PNN, and pentapod gait initial PNN corresponding to the expansion coefficient of the radial basis function with the highest terrain recognition accuracy as the tripod gait recognition PNN, quadruped gait recognition PNN, and pentapod gait recognition PNN.

[0158] Specifically, to test the performance of the probabilistic neural network system for terrain recognition, (1) we first designed the gait control plan for the hexapod robot; (2) designed the experimental environment, and sampled the data sets for terrain recognition according to the tripod gait, quadruped gait, and pentapod gait plans; (3) trained and tested the probabilistic neural network for terrain recognition; (4) compared the terrain recognition algorithm proposed in this paper with other algorithms.

[0159] To identify the slope change rate and the grading of the slope, the embodiments of the present application set the test terrain as three terrains, which are: the first terrain is a flat terrain with a certain inclination angle α; the second terrain is a concave terrain; the third terrain is a convex terrain.

[0160] For the first terrain, under the condition of keeping the α angle unchanged during each test, drive the hexapod robot through this flat terrain according to the gait control plan, and sample and record the measurement data of the tactile system of the foot-end pressure sensors under the tripod gait, quadruped gait, and pentapod gait respectively.

[0161] For the second type of terrain, the hexapod robot is driven according to the gait control plan to pass through the sunken terrain with a slope change rate in the range of -4 to -0.5, and the measurement data of the tactile system of the foot-end pressure sensors under the three-legged gait, four-legged gait, and five-legged gait are sampled and recorded respectively.

[0162] For the third type of terrain, the hexapod robot is driven according to the gait control plan to pass through the convex terrain with a slope change rate in the range of 0.5 to 4, and the measurement data of the tactile system of the foot-end pressure sensors under the three-legged gait, four-legged gait, and five-legged gait are sampled and recorded respectively.

[0163] During the experiment, the hexapod robot is driven to walk on the three types of terrain according to different gaits, and each walk is repeated 10 times. The measurement data is recorded to form a data set for terrain recognition.

[0164] As shown in Figure 6, Figure 6 is a graph of the pressure value data of the hexapod robot after implementing three implementation gaits on the first type of terrain in the embodiment of the present application. For the first type of terrain, in order to observe the relationship between the pressure values of the foot-end pressure sensors F i (i = 1,..., 6) and the tilt angle α, Figure 6(a) represents the three-legged gait, Figure 6(b) represents the four-legged gait, and Figure 6(c) represents the five-legged gait.

[0165] As shown in Figure 7, Figure 7 is a graph of the pressure value data of the hexapod robot after implementing three implementation gaits on the second type of terrain in the embodiment of the present application. For the second type of terrain, in order to observe the relationship between the pressure values of the foot-end pressure sensors F i (i = 1,..., 6) and the sunken terrain, Figure 7(a) represents the three-legged gait, Figure 7(b) represents the four-legged gait, and Figure 7(c) represents the five-legged gait.

[0166] As shown in Figure 8, Figure 8 is a graph of the pressure value data of the hexapod robot after implementing three implementation gaits on the third type of terrain in the embodiment of the present application. For the third type of terrain, in order to observe the relationship between the pressure values of the foot-end pressure sensors F i (i = 1,..., 6) and the convex terrain, Figure 8(a) represents the three-legged gait, Figure 8(b) represents the four-legged gait, and Figure 8(c) represents the five-legged gait.

[0167] 100 groups of data are extracted from the feature extraction set, of which 70 groups are used as the training data set, as Figure 9 shown, Figure 9 is the data distribution diagram of the training data set in the embodiment of the present application, and 30 groups are used as the test data set, as Figure 10 shown, Figure 10 is the data distribution diagram of the test data set in the embodiment of the present application.

[0168] The creation function of the PNN network is net = newpnn(P, T, SPREAD), where P is the input matrix, T is the target output, and SPREAD is the spread coefficient of the radial basis function. To analyze the influence of SPREAD on the network performance, the SPREAD value of the PNN is continuously adjusted during the training process. During the simulation process, SPREAD is taken as 2, 1.5, 1, 0.5, and 0.1 respectively for training and classification. The network model is executed 10 times respectively, and the average recognition accuracies are 91%, 96%, 94%, 82%, and 85% respectively, as shown in Table 5.

[0169] For the probabilistic neural network algorithm, if SPREAD is too large or too small, the accuracy of the network model is relatively poor. It is advisable to take the data between 1 and 1.5, especially around 1.5 is better. Therefore, in the embodiment of the present application, the Spread value is taken as 1.5, and 30 test samples are input into the trained PNN network. At this time, the average recognition accuracy of the PNN network is 96%.

[0170] Table 5 Influence of the spread coefficient of the radial basis function on the network performance in 10 models

[0171] Spread setting value spread = 2 spread = 1.5 spread = 1 spread = 0.5 spread = 0.1 Average accuracy rate 94％ 96.30％ 90％ 82.60％ 83.60％

[0172] As Figure 11 shown, Figure 11 This is the running time of the DEM terrain recognition model in the embodiment of the present application. When spread = 1.5, from the running time of 10 models, since the first time Matlab reads memory and other resource occupations are more, the time is longer, and it tends to be stable after the third time. Thus, it can be seen that the DEM terrain recognition model has a fast processing speed and a short running time.

[0173] In addition, to show the superiority of the DEM terrain recognition model, the recognition accuracies of two terrain recognition algorithms, namely the error backpropagation neural network and the ELM algorithm, are compared.

[0174] After establishing the network and conducting tests, it is necessary to verify the recognition results of the network in the actual terrain to ensure the accuracy of its terrain recognition. To verify the practicality and accuracy of the terrain recognition algorithm proposed in the embodiment of the present application, actual environment experiments are carried out. As shown in Figure 12, Figure 12 is a schematic diagram of the terrain in the comparison experiment of the DEM terrain recognition model and other terrain recognition algorithms in the embodiment of the present application. Figure 12(a) represents a gentle slope terrain, Figure 12(b) represents a steep slope terrain, and Figure 12(c) represents an inclined uphill terrain, which respectively pass through the flat terrain, convex terrain, concave terrain, and sandy complex terrain with a certain slope angle (including gentle slope terrain, steep slope terrain, and inclined uphill terrain) in the actual environment.

[0175] Using the dataset of the actual terrain as the terrain recognition dataset, the proposed algorithm model is trained and tested, and the performance of the evaluation model is verified. In order to evaluate the proposed model and verify the effectiveness of feature selection for terrain recognition based on haptic force feedback in this model, a series of field experiments are conducted in this paper, and the recognition of different terrains is implemented using a probabilistic neural network (PNN), a backpropagation (BP) neural network, and an extreme learning machine (ELM). 50 groups are randomly selected from the datasets of 6 terrains as the test sets respectively. The accuracy results of each category of the three algorithm test sets are compared as shown in Table 6. Among them, the average recognition accuracy of the probabilistic neural network (PNN) for various terrains is 96%; the average recognition accuracy of the backpropagation (BP) neural network for various terrains is 82%; the average recognition accuracy of the extreme learning machine (ELM) for various terrains is 63%. In addition, the calculation times of the probabilistic neural network (PNN), the backpropagation (BP) neural network, and the extreme learning machine (ELM) are 0.06 s, 1 s, and 0.08 s respectively. Therefore, the probabilistic neural network (PNN) has superiority in both recognition accuracy and calculation speed.

[0176] Table 6 Recognition accuracy of each category for terrain recognition based on different algorithms

[0177] Terrain category Flat ground Gentle convex Gentle concave Uphill slope Steep slope Gentle slope PNN algorithm accuracy 100％ 96％ 90％ 100％ 92％ 98％ BP algorithm accuracy 96％ 86％ 70％ 88％ 64％ 90％ ELM algorithm accuracy 78％ 70％ 60％ 52％ 64％ 58％

[0178] S4: Obtain the actual dataset including a number of pressure signals determined after the hexapod robot walks on the target terrain by implementing a gait, and input the actual dataset into the DEM terrain recognition model. The DEM terrain recognition model outputs the terrain type corresponding to the target terrain.

[0179] And, step S4 includes:

[0180] S41: Obtain the actual dataset including a number of pressure signals determined after the hexapod robot walks on the target terrain by implementing a gait, and input the actual dataset into the DEM terrain recognition model;

[0181] S42: Through the input layer, receive the actual dataset corresponding to the target terrain;

[0182] S43: Through the pattern layer, calculate a number of distance initial values corresponding to the neurons in the actual dataset based on the neurons;

[0183] S44: Through the summation layer, calculate the distance weighted values corresponding to the neuron classification of a number of distance initial values in the actual dataset based on the classification of the neurons;

[0184] S45: Through the output layer, perform normalization and probability calculation operations on the distance weighted values in sequence to obtain the terrain type corresponding to the target terrain.

[0185] Specifically, as Figure 13 shown Figure 13 is a schematic flow diagram of the DEM terrain recognition model in the embodiments of the present application.

[0186] The above has described the embodiments of the present application in detail, but the content is only the preferred embodiments of the present application and cannot be considered as limiting the scope of implementation of the present application. Any equivalent changes and improvements made within the scope of the present application should still fall within the scope covered by the patent of the present application.

Claims

1. A terrain recognition method based on a DEM terrain recognition model, characterized in that, it is implemented based on a hexapod robot and a tactile system of foot-end pressure sensors used in cooperation with the hexapod robot, and the method includes: S1: Determine the implementable gait of the hexapod robot and the DEM terrain information; S2: Through the tactile system of foot-end pressure sensors, obtain an initial data set including pressure signals corresponding to different implementable gaits and DEM terrain information determined after the hexapod robot walks on a simulated terrain corresponding to the DEM terrain information by means of the implementable gait; S3: Preprocess the initial data set to obtain a training data set accordingly, and train a basic PNN with the training data set to obtain a DEM terrain recognition model with the pressure signal corresponding to the implementable gait as the input value and the terrain type corresponding to the DEM terrain information as the output value; S4: Obtain an actual data set including a number of pressure signals determined after the hexapod robot walks on the target terrain by means of the implementable gait and input the actual data set into the DEM terrain recognition model, and the DEM terrain recognition model outputs the terrain type corresponding to the target terrain; The DEM terrain information includes: slope, slope change rate, and slope direction; And, the terrain types corresponding to the DEM terrain information include: flat ground, convex, concave, uphill slope, steep slope, and gentle slope; The S3 includes: S31: Normalize the initial data set to obtain a preprocessed data set; S32: Extract DEM features from the preprocessed data set to obtain a feature extraction set including pressure signals corresponding to different implementable gaits and DEM terrain information; S33: Divide the feature extraction set in a ratio of 7:3 to obtain a training data set and a test data set accordingly; S34: Train a basic PNN with the training data set to obtain a DEM initial model with the pressure signal corresponding to the implementable gait as the input value and the terrain type corresponding to the DEM terrain information as the output value; S35: Based on the expansion coefficient of the radial basis function, test the DEM initial models corresponding to different expansion coefficients of the radial basis function with the test data set to obtain the terrain recognition accuracy corresponding to the expansion coefficient of the radial basis function; S36: Select the DEM initial model corresponding to the expansion coefficient of the radial basis function with the highest terrain recognition accuracy as the DEM terrain recognition model; The implementable gaits include: tripod gait, quadruped gait, and pentapod gait.

2. The terrain recognition method based on the DEM terrain recognition model according to claim 1, characterized in that, the tactile system of foot-end pressure sensors includes: a central control module, and a servo control module and an environment perception module respectively connected to the central control module; And, the S2 includes: S21: Send a control command from the central control module to the servo control module, and the servo control module controls the hexapod robot to walk on a simulated terrain corresponding to the DEM terrain information by means of the implementable gait; S22: Obtain a number of pressure values corresponding to different implemented gaits through the environmental perception module set at the foot end and transmit the pressure values to the central control module; S23: The central control module converts the pressure values into pressure signals and outputs an initial data set including a number of pressure signals corresponding to different implemented gaits.

3. The terrain recognition method based on the DEM terrain recognition model according to claim 1, characterized in that, the training data set includes: a three-legged gait training set, a four-legged gait training set, and a five-legged gait training set; the types of the DEM terrain recognition models include: a three-legged gait recognition PNN, a four-legged gait recognition PNN, and a five-legged gait recognition PNN; the S34 includes: S341: Respectively train different basic PNNs through the three-legged gait training set, the four-legged gait training set, and the five-legged gait training set, and correspondingly obtain a three-legged gait initial PNN, a four-legged gait initial PNN, and a five-legged gait initial PNN with the pressure signals corresponding to the three-legged gait, the four-legged gait, and the five-legged gait as input values and the terrain types corresponding to the DEM terrain information as output values; and, the test data set includes: a three-legged gait test set, a four-legged gait test set, and a five-legged gait test set; and, the S35 includes: S351: Determine the expansion coefficients of a number of radial basis functions for the DEM initial model; S352: Based on the expansion coefficients of the radial basis functions, respectively test the three-legged gait initial PNN, the four-legged gait initial PNN, and the five-legged gait initial PNN through the three-legged gait test set, the four-legged gait test set, and the five-legged gait test set, and obtain the terrain recognition accuracy corresponding to the expansion coefficients of the radial basis functions; and, the S36 includes: S361: For the three-legged gait initial PNN, the four-legged gait initial PNN, and the five-legged gait initial PNN, respectively select the three-legged gait initial PNN, the four-legged gait initial PNN, and the five-legged gait initial PNN corresponding to the expansion coefficients of the radial basis functions corresponding to the highest terrain recognition accuracy as the three-legged gait recognition PNN, the four-legged gait recognition PNN, and the five-legged gait recognition PNN.

4. The terrain recognition method based on the DEM terrain recognition model according to claim 1, characterized in that, the training data set includes a number of training samples; the training samples include: pressure signals corresponding to different implemented gaits and DEM terrain information; the structure of the DEM terrain recognition model includes: an input layer, a pattern layer, a summation layer, and an output layer; and, the S4 includes: S41: Obtain an actual data set including a number of pressure signals determined after the six-legged robot walks through the implemented gait on the target terrain and input the actual data set into the DEM terrain recognition model; S42: Through the input layer, be used to receive the actual data set corresponding to the target terrain; S43: Through the pattern layer, calculate a number of distance initial amounts corresponding to the training samples in the actual data set based on the training samples; S44: Through the summation layer, calculate the distance weighting quantities corresponding to the classification of the training samples for several distance initial quantities in the actual dataset based on the classification of the training samples; S45: Through the output layer, perform normalization and probability calculation operations on the distance weighting quantities in sequence to obtain the terrain type corresponding to the target terrain.

Images