Robot terrain recognition method based on depth feature dimension reduction and visual optimization

Abstract

The invention provides a robot terrain recognition method based on depth feature dimension reduction and visual optimization, and belongs to the technical field of robot visual perception and intelligent classification. Comprising the following steps: acquiring an input image of a terrain category by using a camera of a robot body, and preprocessing the input image to obtain a preprocessed input image; the method comprises the steps that a ResNet50 network structure is adjusted to obtain an improved ResNet50 network, ImageNet is used for pre-training the improved ResNet50 network, the improved ResNet50 network completing pre-training serves as a feature extractor, and a feature tensor is extracted from an input image; dimension reduction processing is carried out on the feature tensor to obtain a feature vector after dimension reduction, the feature vector after dimension reduction is mapped to a two-dimensional embedding space, and a cluster distribution structure of different terrain categories in the feature space is visually presented; and selecting a classifier to obtain a final classification score of the terrain category.

Description

A Robot Terrain Recognition Method Based on Deep Feature Dimensionality Reduction and Visualization Optimization Technical Field

[0001] This invention relates to the field of robot visual perception and intelligent classification technology, and in particular to a robot terrain recognition method based on depth feature dimensionality reduction and visualization optimization. Background Technology

[0002] Current methods for terrain perception in quadruped robots primarily rely on LiDAR, IMUs, or tactile sensors. While these methods provide some environmental feedback, they suffer from several drawbacks: high energy consumption and insufficient real-time performance, making them unsuitable for long-duration autonomous tasks; a lack of predictability, as foot sensors can only determine terrain after contact; and susceptibility to degradation in complex lighting, occlusion, and texture variations. In recent years, visual recognition methods based on Convolutional Neural Networks (CNNs) have achieved significant results in autonomous driving and remote sensing. However, their end-to-end structures often require a large number of labeled samples and suffer from feature redundancy, model overfitting, and poor interpretability. Furthermore, while traditional Principal Component Analysis (PCA) and t-distributed stochastic neighborhood embedding (t-SNE) can aid in understanding deep feature distributions, they lack adaptive mechanisms for robot terrain recognition tasks, failing to achieve an optimal balance between feature compression and classification performance. The lack of a unified optimization mechanism between PCA and Support Vector Machine (SVM) makes it difficult to achieve end-to-end feature compression and collaborative discrimination. In addition, the existing t-SNE method is only used for visualization and lacks quantitative guidance, which cannot effectively assist classification decisions. The overall model also lacks an adaptive tuning mechanism for task characteristics.

[0003] Therefore, it is essential to provide a robot terrain recognition method based on depth feature dimensionality reduction and visualization optimization to improve the motion stability and environmental adaptability of quadruped robots in complex underground environments. Summary of the Invention

[0004] In view of this, the present invention proposes a robot terrain recognition method based on deep feature dimensionality reduction and visualization optimization, which integrates deep feature extraction, feature dimensionality reduction, visualization analysis and classification decision-making.

[0005] This invention provides a robot terrain recognition method based on deep feature dimensionality reduction and visualization optimization, comprising the following steps: S1: Image acquisition and preprocessing: using the robot's built-in camera to acquire input images of terrain categories, preprocessing the input images to obtain preprocessed input images; S2: Feature extraction and optimization: adjusting the ResNet50 network structure to obtain an improved ResNet50 network, pre-training the improved ResNet50 network using ImageNet, using the pre-trained improved ResNet50 network as a feature extractor, and outputting the extracted feature tensors; S3: Feature dimensionality reduction and visualization discrimination: using the task-adaptive PCA dimensionality reduction method to process the feature tensors to obtain dimensionality-reduced feature vectors, mapping the dimensionality-reduced feature vectors to a two-dimensional embedding space, intuitively presenting the cluster distribution structure of different terrain categories in the feature space; selecting a classifier to obtain the final classification score of the terrain category.

[0006] Based on the above technical solutions, preferably, the preprocessing of the input image in step S1 includes size normalization, center cropping, brightness perturbation, and contrast enhancement preprocessing operations.

[0007] Based on the above technical solutions, preferably, the improved ResNet50 network obtained by adjusting the ResNet50 network structure in step S2 is such that the ResNet50 network structure includes, in sequence, a first convolutional module Conv_1x, a max pooling module Maxpool, a second convolutional module Conv_2x, a third convolutional module Conv_3x, a fourth convolutional module Conv_4x, a fifth convolutional module Conv_5x, an average pooling module Avgpool, and a fully connected layer FC. Each convolutional module includes several convolutional layers, and each convolutional layer is followed by a BatchNorm layer. Each module is used to process the input... The corresponding feature maps are obtained through line processing. The adjustments include: removing the last fully connected layer (FC) of the ResNet50 network structure, retaining the semantic feature outputs of the first convolutional module Conv_1x to the fifth convolutional module Conv_5x at the beginning of the ResNet50 network, adopting a freezing and fine-tuning strategy for the BatchNorm layer of each convolutional module, and only updating the parameters of the BatchNorm layer in the third convolutional module Conv_3x, the fourth convolutional module Conv_4x, and the fifth convolutional module Conv_5x; and introducing a custom channel attention fusion mechanism between the fifth convolutional module Conv_5x and the average pooling module Avgplool.

[0008] A preferred, custom channel attention fusion mechanism, with attention weights The calculation is based on the information entropy of the feature map and the average gradient magnitude of different channels.

[0009] In a further preferred embodiment, the improved ResNet50 network that has completed pre-training as described in step S2 is used as a feature extractor, and the extracted feature map is output by introducing an adaptive fusion coefficient α based on feature complexity. i Adaptive fusion coefficient α based on feature complexity i The fusion coefficient α is calculated based on the temperature adjustment coefficient, the normalized value of the feature map information entropy of the i-th channel, and the normalized value of the average gradient magnitude; the fusion coefficient α is calculated after the adaptive fusion mechanism based on feature complexity. i Substitute attention weights In the calculation formula, the enhancement weights of the texture complexity feature channels are calculated and applied to the corresponding channel feature map F. i The resulting output feature tensor is obtained by fusion and weighting. .

[0010] Further preferably, in step S3, the task-adaptive PCA dimensionality reduction method is used to process the feature tensor to obtain the dimensionality-reduced feature vector. This involves obtaining the variance contribution term corresponding to the number of candidate principal components, the inter-class separability term corresponding to the traces of the inter-class scatter matrix and the traces of the intra-class scatter matrix, and the output feature tensor. After projecting the numerical stability term of the covariance matrix onto the subspace spanned by the first k principal components, the dimensionality-reduced eigenvectors containing the first k principal components are calculated by weighted summation.

[0011] Furthermore, the weighting coefficients for the variance contribution term, inter-class separability term, and numerical stability term are respectively... , The weight coefficient λ2 is based on the F1 score F1(k) of the first k principal components on a pre-defined validation set for a classification task, divided by the score F1(k) and the average cross-entropy loss. The summation is obtained; the weighting coefficient λ3 is based on the average cross-entropy loss. Divide by the score F1(k) and the average cross-entropy loss The sum of the two is obtained.

[0012] More preferably, in step S3, the dimensionality-reduced feature vectors are mapped to a two-dimensional embedding space to intuitively present the clustering distribution structure of different terrain categories in the feature space. This is achieved by using t-distributed random neighborhood embedding (t-SNE) to map the dimensionality-reduced feature vectors to the two-dimensional embedding space. The perplexity parameter P of t-distributed random neighborhood embedding (t-SNE) is... t The global and local structure of the dimensionality-reduced feature vectors is determined. The clustering quality index Q is obtained by using the number of terrain categories C, the L2 norm of the distance between cluster centers of different categories, and different average cluster radii. t Through clustering quality index Q tThis determines the global and local structure of the eigenvectors after dimensionality reduction.

[0013] In a further preferred embodiment, the classifier selection in step S3 to obtain the final classification score of the terrain category is achieved by using a Support Vector Machine (SVM) with radial basis function kernel as the classifier to obtain the normalized confidence S of the output. SVM (x), combined with the normalized score S of the clustering distance of t-SNE with t-distributed random neighborhood embedding. t-SNE (x), and a weighted vote is performed to obtain the final classification score S(x): , where x is the position of the terrain to be classified in the two-dimensional embedding space; λ(x) is a weighting factor dynamically adjusted based on sample uncertainty. d, e, and f are empirical adjustment coefficients, and H(x) is the prediction entropy of the support vector machine (SVM) classification result. , It is the Platt calibration probability of the SVM classification result.

[0014] A further preferred approach is to use the normalized score S of the clustering distance of the t-distributed random neighborhood embedding t-SNE. t-SNE (x) is obtained by calculating the coordinates of the terrain to be classified in the two-dimensional embedding space and the cluster center vector of the predicted category of the terrain to be classified.

[0015] This invention provides a robot terrain recognition method based on deep feature dimensionality reduction and visualization optimization, which has the following advantages over existing technologies: 1. Compared with traditional end-to-end convolutional neural network classification methods, this invention has significant advantages in accuracy, computational efficiency, interpretability, and adaptability. Through multi-layer feature fusion and task adaptive optimization mechanisms, the model exhibits higher recognition accuracy and stability in complex terrain environments.

[0016] 2. By using a custom attention mechanism, the sensitivity of feature extraction to terrain texture is improved. The t-SNE algorithm is used to map the dimensionality-reduced features, which intuitively shows the clustering distribution of different terrains in the feature space. This provides a quantitative basis for model optimization and anomaly detection. Compared with traditional black-box CNN, the output results of this method are more understandable, and the classification basis and sources of misjudgment can be clearly analyzed, thus improving interpretability.

[0017] 3. The final classification score can realize the dual collaborative mechanism of vision and discrimination, which significantly improves the robustness of this method in scenarios with blurred boundaries. The t-SNE auxiliary discrimination mechanism can automatically correct the SVM classification results under blurred boundaries or noise interference. The t-SNE clustering quality index Q makes the visualized clustering more compact, and the misjudgment correction mechanism improves the classification accuracy of boundary samples. Attached Figure Description

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

[0019] Figure 1 is a classification flowchart of a robot terrain recognition method based on deep feature dimensionality reduction and visualization optimization according to the present invention; Figure 2 is a structural diagram of an improved ResNet50 network for a robot terrain recognition method based on deep feature dimensionality reduction and visualization optimization according to the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Existing methods have the following shortcomings in terrain recognition tasks: First, the high dimensionality and redundancy of deep features lead to high computational complexity and overfitting under small sample conditions, resulting in decreased classification performance. Second, the dimensionality reduction and classification stages are disconnected, and there is a lack of a unified optimization mechanism between PCA and SVM, making it difficult to achieve end-to-end feature compression and collaborative discrimination. In addition, the existing t-SNE method is only used for visualization and lacks quantitative guidance, failing to effectively assist in classification decisions. Finally, the model as a whole lacks an adaptive tuning mechanism tailored to task characteristics, and the ResNet structure is not specifically optimized for terrain texture and geometric features, resulting in limited feature representation capabilities.

[0022] In view of this, as shown in Figure 1, the present invention provides a robot terrain recognition method based on depth feature dimensionality reduction and visualization optimization, including the following steps: S1: Image acquisition and preprocessing: The robot body's built-in camera acquires the input image of the terrain category, and the input image is preprocessed to obtain the preprocessed input image.

[0023] The input image is preprocessed, including size normalization, center cropping, brightness perturbation, and contrast enhancement preprocessing operations.

[0024] Size normalization unifies the size of all input images; center cropping preserves the central main content of the image by cropping the edges within a preset rectangular area; brightness perturbation simulates changes in lighting to brighten or darken the pixel values ​​of the image; contrast enhancement expands the overall pixel value distribution range and enhances texture details in the image, such as gamma correction adjusting contrast. These are all common techniques in image processing and will not be elaborated upon further here.

[0025] S2: Feature Extraction and Optimization: The ResNet50 network structure is adjusted to obtain an improved ResNet50 network. The improved ResNet50 network is pre-trained using ImageNet. The pre-trained improved ResNet50 network is used as a feature extractor to output the extracted feature tensors.

[0026] Referring to Figure 2, the improved ResNet50 network, obtained by adjusting the ResNet50 network structure, consists of a first convolutional module (Conv_1x), a max-pooling module (Maxpool), a second convolutional module (Conv_2x), a third convolutional module (Conv_3x), a fourth convolutional module (Conv_4x), a fifth convolutional module (Conv_5x), an average pooling module (Avgpool), and a fully connected layer (FC). Each convolutional module includes several convolutional layers, and each convolutional layer is followed by a BatchNorm layer. Each module processes the input to obtain the corresponding... The feature maps were modified as follows: the last fully connected layer (FC) of the ResNet50 network structure was removed, the semantic feature outputs of the first convolutional module Conv_1x to the fifth convolutional module Conv_5x at the beginning of the ResNet50 network were retained, a freezing and fine-tuning strategy was adopted for the BatchNorm layer of each convolutional module, and the parameters of the BatchNorm layer in the third convolutional module Conv_3x, the fourth convolutional module Conv_4x, and the fifth convolutional module Conv_5x were updated only; a custom channel attention fusion mechanism was introduced between the fifth convolutional module Conv_5x and the Avgplool module.

[0027] A custom channel attention fusion mechanism with attention weights Calculate the information entropy and gradient magnitude of the feature maps based on different channels: H i Let G represent the entropy of the feature map information of the i-th channel. i α represents the average gradient magnitude. i It is the fusion coefficient, and T is the temperature coefficient.

[0028] To enhance the model's adaptability, an adaptive fusion mechanism based on feature complexity differences is further introduced: Let , , σ(•) is the Sigmoid function. The temperature regulation coefficient is calculated using the fusion coefficient α, where max(•) is the maximum value operation and min(•) is the minimum value operation. This fusion coefficient α is calculated using the adaptive fusion mechanism based on feature complexity. i Substitute attention weights In the calculation formula, the enhancement weights of the texture complexity feature channels are calculated and applied to the corresponding channel feature map F. i By combining and weighting, we obtain , This is the output feature tensor obtained after fusion.

[0029] The adaptive fusion mechanism based on feature complexity differences described above can adaptively enhance the feature responses of key texture regions, ultimately outputting a highly discriminative feature vector with 2048 dimensions. .

[0030] The content pre-trained using ImageNet involves using this dataset to construct training and validation sets for pre-training an improved ResNet50 network, thereby obtaining better model parameters. This will not be elaborated upon further here. ImageNet is a large visualization database used for research on visual object recognition software.

[0031] S3: Feature Dimensionality Reduction and Visual Discrimination: The task-adaptive PCA dimensionality reduction method is adopted to process the feature tensor and obtain the dimensionality-reduced feature vector. The dimensionality-reduced feature vector is mapped to a two-dimensional embedding space to intuitively present the cluster distribution structure of different terrain categories in the feature space; a classifier is selected to obtain the final classification score of the terrain category.

[0032] In the feature dimensionality reduction and redundancy suppression stage, step S3 describes using the task-adaptive PCA dimensionality reduction method to process the feature tensor and obtain the dimensionality-reduced feature vector. This involves defining the principal component selection criterion function J(k) to determine the optimal number of principal components. k is the number of candidate principal components, K is the total number of principal components, j∈K, It is the variance of the j-th principal component, obtained by applying the output feature tensor. The covariance matrix is ​​calculated to obtain S. b It is the inter-class scatter matrix, S w It is the within-class scatter matrix, and tr(•) represents the trace operation of the matrix. These are weighting coefficients. C k It is a k×k projection covariance matrix, representing the output feature tensor. The covariance matrix projected onto the subspace spanned by the first k principal components; the first k principal components constitute the dimensionality-reduced eigenvectors. The three terms on the right-hand side of the principal component selection criterion function J(k) are, respectively, the variance contribution term corresponding to the number of principal components, the inter-class separability term corresponding to the traces of the inter-class scatter matrix and the traces of the intra-class scatter matrix, and the output feature tensor. The numerical stability term corresponding to the covariance matrix after projection onto the subspace spanned by the first k principal components.

[0033] To ensure the task adaptability of the weights, this invention proposes a weight allocation strategy based on dynamic adjustment of the validation set performance: , F1(k) is the F1 score of a classification task based on the first k principal components on a predefined validation set. The average cross-entropy loss is used. Parameters are optimized using five-fold cross-validation, and the optimal balance between feature compression and performance is achieved by maximizing the principal component selection criterion function J(k), thus effectively reducing feature dimensionality and computational complexity while preserving key discriminative information. The dimensionality-reduced feature vectors are used not only for training Support Vector Machines (SVMs) but also provide foundational data for subsequent t-distributed randomized domain embedding (t-SNE).

[0034] The dimensionality-reduced feature vectors are mapped to a two-dimensional embedding space, intuitively presenting the cluster distribution structure of different terrain categories in the feature space. This is achieved using t-distributed random neighborhood embedding (t-SNE). The perplexity parameter P of t-distributed random neighborhood embedding (t-SNE) is... t The global and local structure of the dimensionality-reduced feature vectors is determined, and the clustering quality index is defined: , ||•||2 represents the L2 norm, C is the number of terrain categories, μ a and μ b The cluster centers of categories a and b are respectively, r a and r b These are the average clustering radii of categories a and b, respectively; they determine the global and local structure of the reduced feature vectors, and the update rule for the perplexity parameter is defined as follows. P t+1 For the updated perplexity parameter, P t The parameter representing the current perplexity, η, is the learning rate, and Q represents the perplexity parameter. * P is the target clustering quality threshold. min and P max These represent the minimum and maximum values ​​of the perplexity parameter; when the clustering quality index decreases, i.e. It will automatically increase the perplexity parameter P t This improves local discrimination accuracy and enables adaptive optimization of the perplexity parameter.

[0035] The step S3, which involves selecting a classifier to obtain the final classification score for the terrain category, uses a Support Vector Machine (SVM) with a radial basis function kernel as the classifier to obtain the normalized confidence S of the output. SVM (x), combined with the normalized score S of the clustering distance of t-SNE with t-distributed random neighborhood embedding. t-SNE (x), and a weighted vote is performed to obtain the final classification score S(x): , where x is the position of the terrain to be classified in the two-dimensional embedding space; λ(x) is a weighting factor dynamically adjusted based on sample uncertainty. d, e, and f are empirical adjustment coefficients, and H(x) is the prediction entropy of the support vector machine (SVM) classification result. , It is the Platt calibration probability of the SVM classification result.

[0036] Introducing the normalized score S of clustering distance for t-SNE with t-distributed random neighborhood embedding based on Mahalanobis distance. t-SNE (x) is calculated using the coordinates of the terrain to be classified in the two-dimensional embedding space and the cluster center vector of the predicted category of the terrain. Specifically: , Let be the coordinates of the dimensionality-reduced feature vector of the terrain to be classified in the two-dimensional embedding space. To predict the cluster center vector of the category, To predict the covariance matrix of the categories, a diagonal regularization term is added to ensure numerical stability.

[0037] In addition, this embodiment also provides a dynamic misjudgment tolerance radius based on cluster radius and inter-class distance. This is used to calculate the normalized confidence S of the output obtained by using a Support Vector Machine (SVM) with a radial basis function kernel as a classifier. SVM (x) When low, provide a decision correction mechanism: , , It is the global average cluster radius, d min γ is the minimum inter-class distance, β(x) is the adjustment coefficient, and β(x) is a dynamic weighting factor that quantifies the reliability or uncertainty of the SVM classification result and generates a dynamic weight used in the final classification scoring process. This mechanism ensures that low-confidence samples have a larger tolerance radius, thereby reducing misclassification under noisy conditions.

[0038] As shown in Figure 1, after pre-training the improved ResNet50 network, both the PCA projection matrix and SVM classification parameters are fitted using the training set. In the online deployment phase, only the improved ResNet50 forward computation and SVM inference are required to achieve real-time terrain type discrimination. This method structurally integrates deep feature extraction, dimensionality reduction optimization, and classification decision-making, significantly improving recognition accuracy and computational efficiency, while also enhancing the model's interpretability and scalability in complex environments.

[0039] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A robot terrain recognition method based on deep feature dimensionality reduction and visualization optimization, characterized in that, The steps include: S1: Image acquisition and preprocessing: The robot's built-in camera acquires the input image of the terrain category, and the input image is preprocessed to obtain the preprocessed input image; S2: Feature Extraction and Optimization: The ResNet50 network structure is adjusted to obtain an improved ResNet50 network. The improved ResNet50 network is pre-trained using ImageNet. The pre-trained improved ResNet50 network is used as a feature extractor to output the extracted feature tensors. S3: Feature Dimensionality Reduction and Visualization: The task-adaptive PCA dimensionality reduction method is used to process the feature tensors to obtain dimensionality-reduced feature vectors. The dimensionality-reduced feature vectors are mapped to a two-dimensional embedding space to intuitively present the cluster distribution structure of different terrain categories in the feature space. A classifier is selected to obtain the final classification score of the terrain category.

2. The robot terrain recognition method based on deep feature dimensionality reduction and visualization optimization according to claim 1, characterized in that, The preprocessing of the input image described in step S1 includes size normalization, center cropping, brightness perturbation, and contrast enhancement preprocessing operations.

3. The robot terrain recognition method based on deep feature dimensionality reduction and visualization optimization according to claim 1, characterized in that, The improved ResNet50 network obtained by adjusting the ResNet50 network structure in step S2 is defined as follows: the ResNet50 network structure includes, in sequence, a first convolutional module (Conv_1x), a max pooling module (Maxpool), a second convolutional module (Conv_2x), a third convolutional module (Conv_3x), a fourth convolutional module (Conv_4x), a fifth convolutional module (Conv_5x), an average pooling module (Avgpool), and a fully connected layer (FC). Each convolutional module includes several convolutional layers, and each convolutional layer is followed by a BatchNorm layer. Each module is used to process the input to obtain the corresponding... Feature maps; Adjustments include: removing the last fully connected layer (FC) of the ResNet50 network structure, retaining the semantic feature outputs of the first convolutional module Conv_1x to the fifth convolutional module Conv_5x at the beginning of the ResNet50 network, adopting a freezing and fine-tuning strategy for the BatchNorm layers of each convolutional module, and only updating the parameters of the BatchNorm layers in the third convolutional module Conv_3x, the fourth convolutional module Conv_4x, and the fifth convolutional module Conv_5x; and introducing a custom channel attention fusion mechanism between the fifth convolutional module Conv_5x and the average pooling module Avgplool.

4. The robot terrain recognition method based on deep feature dimensionality reduction and visualization optimization according to claim 3, characterized in that, A custom channel attention fusion mechanism with attention weights The calculation is based on the information entropy of the feature map and the average gradient magnitude of different channels.

5. The robot terrain recognition method based on deep feature dimensionality reduction and visualization optimization according to claim 4, characterized in that, The improved ResNet50 network, which has completed pre-training as described in step S2, serves as a feature extractor, outputting the extracted feature map. This is achieved by introducing an adaptive fusion coefficient α based on feature complexity. i Adaptive fusion coefficient α based on feature complexity i The fusion coefficient α is calculated based on the temperature adjustment coefficient, the normalized value of the feature map information entropy of the i-th channel, and the normalized value of the average gradient magnitude; the fusion coefficient α is calculated after the adaptive fusion mechanism based on feature complexity. i Substitute attention weights In the calculation formula, the enhancement weights of the texture complexity feature channels are calculated and applied to the corresponding channel feature map F. i The resulting output feature tensor is obtained by fusion and weighting. 。 6. The robot terrain recognition method based on deep feature dimensionality reduction and visualization optimization according to claim 5, characterized in that, Step S3 describes using a task-adaptive PCA dimensionality reduction method to process the feature tensor and obtain the dimensionality-reduced feature vector. This involves obtaining the variance contribution term corresponding to the number of candidate principal components, the inter-class separability term corresponding to the traces of the inter-class scatter matrix and the traces of the intra-class scatter matrix, and the output feature tensor. After projecting the numerical stability term of the covariance matrix onto the subspace spanned by the first k principal components, the dimensionality-reduced eigenvectors containing the first k principal components are calculated by weighted summation.

7. The robot terrain recognition method based on deep feature dimensionality reduction and visualization optimization according to claim 6, characterized in that, The weighting coefficients for the variance contribution term, the inter-class separability term, and the numerical stability term are respectively... ， The weight coefficient λ2 is based on the F1 score F1(k) of the first k principal components on a pre-defined validation set for a classification task, divided by the score F1(k) and the average cross-entropy loss. The summation is obtained; the weighting coefficient λ3 is based on the average cross-entropy loss. Divide by the score F1(k) and the average cross-entropy loss The sum of the two is obtained.

8. The robot terrain recognition method based on deep feature dimensionality reduction and visualization optimization according to claim 6, characterized in that, The dimensionality-reduced feature vectors described in step S3 are mapped to a two-dimensional embedding space, intuitively presenting the cluster distribution structure of different terrain categories in the feature space. This is achieved using t-distributed random neighborhood embedding (t-SNE) to map the dimensionality-reduced feature vectors to the two-dimensional embedding space. The perplexity parameter P of t-distributed random neighborhood embedding (t-SNE) is... t The global and local structure of the dimensionality-reduced feature vectors is determined. The clustering quality index Q is obtained by using the number of terrain categories C, the L2 norm of the distance between cluster centers of different categories, and different average cluster radii. t Through clustering quality index Q t This determines the global and local structure of the eigenvectors after dimensionality reduction.

9. The robot terrain recognition method based on deep feature dimensionality reduction and visualization optimization according to claim 8, characterized in that, The step S3, which involves selecting a classifier to obtain the final classification score for the terrain category, uses a Support Vector Machine (SVM) with a radial basis function kernel as the classifier to obtain the normalized confidence S of the output. SVM (x), combined with the normalized score S of the clustering distance of t-SNE with t-distributed random neighborhood embedding. t-SNE (x), and a weighted vote is performed to obtain the final classification score S(x): , where x is the position of the terrain to be classified in the two-dimensional embedding space; λ(x) is a weighting factor dynamically adjusted based on sample uncertainty. d, e, and f are empirical adjustment coefficients, and H(x) is the prediction entropy of the support vector machine (SVM) classification result. ， It is the Platt calibration probability of the SVM classification result.

10. A robot terrain recognition method based on deep feature dimensionality reduction and visualization optimization according to claim 9, characterized in that, Normalized score S of clustering distance for t-distributed random neighborhood embedding t-SNE t-SNE (x) is obtained by calculating the coordinates of the terrain to be classified in the two-dimensional embedding space and the cluster center vector of the predicted category of the terrain to be classified.

Images