Intelligent road surface recognition method based on force perception and image analysis

Abstract

The invention discloses an intelligent road surface recognition method based on force perception and image analysis. The method comprises the following steps: S1, acquiring image data and contact force data in the advancing process of a mobile robot; s2, inputting the image into a two-channel feature network, and extracting structural features and fuzzy region information; s3, processing through an unwrapping model to form an image semantic vector; s4, constructing a contact time sequence, extracting behavior characteristics, and generating a contact event code; s5, performing alignment and weighted fusion on the image semantic vector and the contact code, and outputting a combined pavement state vector; s6, inputting a trafficability scoring model, identifying a ground type and generating a trafficability score; and S7, transmitting the result to a navigation module for path planning and motion control. Visual and force sense information is combined, and the complex terrain recognition accuracy and the robot traffic capacity are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of mobile robot environment perception and intelligent recognition, and in particular to an intelligent road surface recognition method based on force perception and image analysis. Background Art

[0002] As the application of mobile robots in scenarios such as urban delivery, agricultural operations, and forest inspections continues to expand, their adaptability to the ground environment has become a crucial factor influencing operational stability and mission efficiency. In particular, in unstructured or unknown complex terrain, such as soft sand, slippery turf, and gravel potholes, rapidly changing road conditions pose challenges to robots' path planning and motion control. If robots cannot accurately identify road surface types, they are prone to sinking, slipping, and path deviation, potentially leading to mission failure or equipment damage.

[0003] In existing technologies, road surface recognition methods can be primarily categorized as those based on visual perception and those based on force perception. The former typically relies on monocular or binocular cameras to capture images and extract terrain features for classification using deep learning models; the latter uses force sensors deployed on wheelbases or support structures to sense the robot's contact status with the ground. However, both approaches still face several key challenges in complex real-world environments. These include: Visual methods are sensitive to environmental changes: Complex terrain is often accompanied by uneven lighting, image blur, and mixed terrain, leading to misjudgments in visual models and difficulty distinguishing fine-grained road surface differences, such as wet mud and dry soil. Force perception methods suffer from limited information dimensionality: Single-frame force sensor data struggles to capture dynamic contact behavior and cannot reflect potential risk features such as slippage and sinking. The ability to identify areas not visible to the naked eye is particularly limited. Multimodal information fusion mechanisms are weak: Existing methods often employ simple splicing or weighting methods to fuse images and force perception, lacking unified alignment and confidence adjustment mechanisms. This makes it difficult to effectively leverage the advantages of multiple sources when information conflicts arise. Lack of quantitative output of passability: Although some systems can achieve classification and identification, it is difficult to assess the degree of passability of the ground and cannot provide effective path risk parameter support for the navigation system.

[0004] Therefore, how to provide an intelligent road surface recognition method based on force perception and image analysis is an urgent problem that those skilled in the art need to solve. Summary of the Invention

[0005] One purpose of the present invention is to propose an intelligent road surface recognition method based on force perception and image analysis, which fully integrates the disentanglement complementary learning model and the contact event mapping algorithm, and describes in detail the process of mobile robots realizing road surface type recognition and trafficability assessment in complex terrain through joint modeling of image semantic features and contact behavior features. It has the advantages of high recognition accuracy, strong adaptability and low computational overhead.

[0006] According to an embodiment of the present invention, an intelligent road surface recognition method based on force sensing and image analysis includes the following steps:

[0007] S1. Obtain environmental image data and contact force perception data collected by the mobile robot during its movement as visual perception input and force perception input respectively;

[0008] S2. The visual perception input is processed through the constructed dual-channel image feature extraction network. The main channel extracts the landform structure features, and the complementary channel extracts the difference features of the semantically fuzzy area to form an image feature representation;

[0009] S3. Input the image feature representation into the disentanglement complementary learning model, perform feature disentanglement and semantic enhancement operations, and generate an image semantic representation vector;

[0010] S4. Construct a contact time series based on the force perception input, extract the dynamic unstable behavior characteristics during the contact process through the contact event mapping algorithm, and generate a contact event encoding vector;

[0011] S5. Align the image semantic representation vector with the contact event encoding vector, fuse them through a dynamic confidence weighting mechanism, and output a joint road surface state representation vector.

[0012] S6. Input the joint road surface state representation vector into the ground accessibility scoring model to generate a current ground type recognition result and a accessibility scoring result;

[0013] S7. Transmit the type recognition result and the pass score result to the mobile robot navigation module for path planning and motion decision execution.

[0014] The present invention constructs a dual-channel image feature extraction network and a disentanglement complementary learning model to extract landform structure and semantic fuzzy features, and combines it with a contact event mapping algorithm based on contact time series to extract dynamic behavior trends such as sinking and slipping, thereby realizing joint modeling of image and force perception. It fuses multimodal information through a dynamic confidence weighting mechanism to generate a highly robust joint road state representation, which is further input into a passability scoring model to generate terrain types and passability scores that can be used for navigation control, significantly improving the recognition accuracy, autonomous decision-making ability and path efficiency of mobile robots in complex terrain.

[0015] Optionally, the S1 specifically includes: visual perception input through a binocular camera system installed on the front of the mobile robot to collect RGB images and depth images in front of the travel path; force perception input through a three-axis force sensor arranged at the wheel or foot contact part to collect the normal force, tangential force and friction force at each frame timestamp; the collected image data and force data are format standardized and unit normalized respectively, and linear interpolation and alignment are performed according to a unified time base, and finally an image input tensor and a force input sequence are constructed for subsequent network input processing.

[0016] The present invention installs a binocular camera system on the front of the mobile robot to collect RGB and depth images, and deploys three-axis force sensors at the wheeled or footed contact points to collect normal force, tangential force and friction force. The image and force data are standardized in format, normalized in units and aligned by linear interpolation in combination with timestamps, and an image input tensor and force input sequence under a unified time base are constructed, providing high-quality, time-consistent data input for subsequent multimodal feature extraction and fusion modeling, effectively improving perception accuracy and network modeling stability, and providing a solid data foundation for recognition and decision-making in complex terrain environments.

[0017] Optionally, the S2 specifically includes:

[0018] S21, constructing a dual-channel image feature extraction network, wherein the network includes a main feature channel and a complementary feature channel, and the image data in the visual perception input is input into the main feature channel and the complementary feature channel respectively for parallel processing;

[0019] S22. In the main feature channel, a landform structure encoding unit is set. The landform structure encoding unit includes multiple serially connected two-dimensional convolutional layers and an integrated feature compression module. The two-dimensional convolutional layer extracts terrain boundaries, texture directions, and geometric structure features in the image. The integrated feature compression module includes batch normalization and maximum pooling structures to regularize the convolution output and reduce the spatial dimension. All extracted features are integrated along the channel dimension and then flattened to form a main channel feature representation.

[0020] S23. In the complementary feature channel, a semantic fuzzy detection unit is set. The semantic fuzzy detection unit includes an attention enhancement module and a multi-scale feature extraction module. The attention enhancement module includes a spatial weighting path and a channel importance path. The spatial weighting path generates a saliency map based on the regional response. The channel importance path estimates the semantic dimension weight based on the global pooling and activation function. The multi-scale feature extraction module includes multiple two-dimensional convolution paths with different receptive fields to identify edge and texture changes in the fuzzy area. All path outputs are spliced ​​and weightedly fused in the channel dimension to generate a complementary channel feature representation.

[0021] S24. Concatenate the main channel feature representation and the complementary channel feature representation in the channel dimension to generate a fused feature tensor, and perform channel compression and spatial consistency processing to form an image feature representation.

[0022] The present invention constructs a dual-channel image feature extraction network, and sends visual perception input into the main channel and the complementary channel for parallel processing. The main channel extracts structural features such as terrain boundaries and texture directions through convolution and pooling operations, while the complementary channel combines attention enhancement and multi-scale convolution to identify detailed differences in semantically ambiguous areas. Through channel splicing and consistency processing, a fused feature tensor is generated, which significantly improves the image feature representation's ability to express fine-grained terrain differences.

[0023] Optionally, the S3 specifically includes:

[0024] S31, inputting the image feature representation into the structural semantic extraction channel and the pseudo semantic difference extraction channel at the same time, and performing parallel processing in the disentanglement complementary learning model;

[0025] S32, the structural semantic extraction channel receives the image feature representation, extracts the dominant landform structural features based on the texture consistency, geometric boundary continuity and local area statistics of the image in the spatial domain, and forms the main semantic representation tensor through two-dimensional convolution, pooling and channel compression operations in sequence;

[0026] S33, the pseudo semantic difference extraction channel receives the image feature representation and the main semantic representation tensor, and constructs a residual signal by calculating the pixel-level residual map between them. The residual signal is sent to the convolution processing under the multi-scale receptive field, and combined with the spatial attention mechanism to extract the texture difference information of the semantically ambiguous area to form a complementary semantic representation tensor;

[0027] S34. The fusion decision channel receives the main semantic representation tensor and the complementary semantic representation tensor, calculates the mean difference and coefficient of variation of the two tensors in the channel dimension, constructs the channel fusion weight, performs channel weighting and summation on the two tensors, and generates a fused image semantic representation vector.

[0028] The present invention constructs a disentangled complementary learning model and inputs the image feature representation into the structural semantic channel and the pseudo-semantic difference channel in parallel. The former extracts dominant structural features such as spatial texture consistency and landform boundaries, while the latter extracts texture difference information in semantically fuzzy areas through main semantic residual analysis combined with multi-scale convolution and attention mechanism. In the fusion decision channel, fusion weights are generated based on channel mean difference and coefficient of variation to achieve weighted fusion of main semantics and complementary semantics, and output image semantic representation vectors with difference discrimination capabilities, which effectively enhances the recognition ability and expression robustness of fine-grained semantic differences in complex terrain.

[0029] Optionally, the disentangled complementary learning model includes:

[0030] The structural semantic extraction channel consists of a series of two-dimensional convolutional layers, pooling layers, channel attention structures, and channel compression structures. The two-dimensional convolutional layer contains two standard convolution operations with a kernel size of 3×3. The pooling layer uses the maximum pooling method. The channel attention structure consists of global average pooling, a fully connected layer, and a sigmoid activation function. The channel compression structure is a convolution operation with a kernel size of 1×1, which is used to output the main semantic representation tensor.

[0031] The pseudo semantic difference extraction channel includes a residual signal construction module, a multi-scale convolution module, and a spatial attention module. The residual signal construction module contains an element-by-element difference unit and an upsampling unit to construct a residual signal between the image and the main semantic representation. The multi-scale convolution module consists of three convolution kernels of sizes 1×1, 3×3, and 5×5. The spatial attention module consists of an average pooling branch, a maximum pooling branch, a convolution fusion layer, and a sigmoid activation structure, and outputs a complementary semantic representation tensor.

[0032] The fusion decision channel includes a fusion weight generation unit and a weighted fusion unit. The fusion weight generation unit constructs a channel fusion weight vector based on the channel variance and channel mean difference of the main semantic representation tensor and the complementary semantic representation tensor. The weighted fusion unit multiplies the channels of the two tensors by the fusion weights respectively and performs a sum operation to output the image semantic representation vector.

[0033] The present invention designs a structured disentanglement complementary learning model to collaboratively construct a structural semantic extraction channel and a pseudo-semantic difference extraction channel. The former extracts main semantic features through standard convolution, channel attention and compression operations, while the latter uses element-by-element difference and multi-scale convolution combined with spatial attention mechanism to construct complementary semantic features. In the fusion decision channel, a fusion weight vector is generated based on channel variance and mean difference to realize weighted fusion operation, and finally outputs an image semantic representation vector with the ability to distinguish fuzzy semantics, which significantly improves the model's perception of terrain detail differences and representation accuracy.

[0034] Optionally, the S4 specifically includes:

[0035] S41. Constructing a contact time sequence for force perception input, specifically including: continuously collecting normal force, tangential force, and friction force data for each frame from the three-axis force sensor deployed at the contact part of the mobile robot, and sorting them by timestamp to form a force data sequence T f ;

[0036] S42, set the sliding window with length w and step size s, take the time series as input, generate K continuous behavior segments, and form a behavior segment set

[0037] S43. Define and calculate the normal force fluctuation index for each behavior segment Proportional index to tangential friction The specific formula is as follows:

[0038]

[0039] Among them, ε is a minimum constant, represents the normal force at frame t, represents the tangential force, represents friction, T is the total number of frames;

[0040] S44. Setting normal force threshold θ n and the tangential friction threshold θ t , the indicator of each behavior segment and Determine the contact event label L according to the following mapping relationship k :

[0041]

[0042] S45. Constructing a contact event mapping algorithm, including the following steps:

[0043] Each behavior fragment and the corresponding event label L k As input, a category-driven contact event feature sequence is constructed. Specifically, if the segment category is a sinking event, the normal force sudden drop trend is extracted; if the category is a slip event, the tangential force increasing trend is extracted; if the category is a drift event, the triaxial force periodic unstable fluctuation trend is extracted; if the category is a stable event, the triaxial force stable trend is extracted;

[0044] The trend features in each behavior segment are represented as a feature vector sequence, which is input into the bidirectional gated recurrent unit and the hidden state h is updated at each time step. t , the state update formula is:

[0045] h t =(1-z t )⊙h t-1 +z t ⊙tanh(W h x t +U h (r t ⊙h t-1 ));

[0046] Among them, h t is the hidden state vector at time t, z t is the update gate, which controls the degree of retention of the current state, r tis the reset gate, which controls the degree of reset of historical state information, x t is the input feature vector at time t, W h 、U h is the weight matrix of the network, ⊙ represents element-by-element multiplication;

[0047] The hidden state vector of the forward recursive unit at the last moment and the hidden state vector of the backward recursive unit at the first moment are arranged in sequence to form the event feature representation vector e of the segment k ;

[0048] S46, for all event feature representation vector sets {e1, e2, ..., e K Perform a global average pooling operation to obtain the final contact event encoding vector: The contact event encoding vector V c As a characteristic representation of the dynamic unstable behavior of the current robot during contact.

[0049] The present invention constructs a contact time series and uses a sliding window to divide continuous behavior segments, calculates the normal force fluctuation and tangential friction ratio indicators for each segment, completes event type labeling in combination with set thresholds, and then extracts trend physical features according to different contact event types. The feature sequence is input into a bidirectional gated recursive unit for state modeling, and an event feature representation vector is generated by splicing forward and reverse hidden states. Finally, a contact event encoding vector is formed through global average pooling, thereby realizing structured modeling of dynamic unstable behaviors such as soft collapse and slip drift, effectively improving the recognition ability and dynamic response accuracy of visually invisible terrain.

[0050] Optionally, the S5 specifically includes:

[0051] S51, performing maximum and minimum normalization operations on the image semantic representation vector and the contact event encoding vector respectively, so that the value range of each channel dimension is mapped to the interval [0, 1];

[0052] S52, inputting the normalized image semantic representation vector and the contact event encoding vector into two parameter-sharing linear mapping networks to generate equal-dimensional vector representations, respectively. The linear mapping network is composed of a fully connected layer, and the output dimension is consistent with the preset fusion dimension;

[0053] S53, calculating the cosine similarity of each dimension between two equal-dimensional vectors, and constructing a matching weight vector, where each dimension represents the directional similarity between the image semantic channel and the contact event channel in the same dimension;

[0054] S54. Calculating an image confidence distribution vector based on the mean and standard deviation of each channel of the image semantic representation vector. The image confidence distribution vector is used to characterize the degree of concentration of the image semantic feature distribution within each channel. Calculating a contact confidence distribution vector based on the temporal variation range of each channel of the contact event encoding vector within the sliding window. The contact confidence distribution vector is used to characterize the magnitude of the contact behavior variation in each channel.

[0055] S55, performing weighted averaging on the matching weight vector, the image confidence distribution vector, and the contact confidence distribution vector to generate a fused confidence weighted vector, wherein each dimension is used to determine a weighted ratio between the image semantic representation vector and the contact event encoding vector in the dimension;

[0056] S56. Multiply the image semantic representation vector and the contact event encoding vector by the element values ​​of the corresponding dimensions in the fusion confidence weighted vector respectively, perform dimension-by-dimension weighting and then add them together, and output a joint road surface state representation vector.

[0057] The present invention performs maximum and minimum value normalization and unified linear mapping on the image semantic representation vector and the contact event encoding vector to generate equal-dimensional vectors, then calculates the cosine similarity to construct the matching weight, and further combines the image channel distribution characteristics and the amplitude of contact behavior change to generate a confidence distribution vector. The weighted fusion mechanism of the matching weight and the confidence distribution is used to dynamically adjust the fusion ratio of the two types of perception signals, and finally outputs a joint road state representation vector, realizing the adaptive collaborative fusion of image and force information in multiple dimensions, and significantly enhancing the recognition stability and robustness of the system under conditions of inconsistent information sources or uneven confidence.

[0058] Optionally, the S6 specifically includes:

[0059] S61, receiving a joint road surface state representation vector as input and inputting it into a ground trafficability scoring model, wherein the model includes a type identification branch and a trafficability evaluation branch, and the two branches operate in parallel;

[0060] S62, the type recognition branch includes a fully connected mapping layer and a normalization function, wherein the fully connected mapping layer maps the input vector to a category probability vector, and the normalization function is a Softmax function for normalizing all category probabilities and outputting the category corresponding to the maximum probability as the ground type recognition result;

[0061] S63, the pass rating branch includes a fully connected mapping layer and an activation function, wherein the fully connected mapping layer outputs a real-valued channel, and the activation function is a Sigmoid function, which is used to compress the output value to the interval [0, 1] to obtain a pass rating result;

[0062] S64: Combining the type recognition result and the traffic scoring result into a structured output result, and transmitting it to the navigation module for path control call.

[0063] The present invention constructs a ground passability scoring model that includes a type recognition branch and a passability evaluation branch, and performs classification and scoring operations on the joint road state representation vector in parallel. The type recognition branch outputs the ground type category through a fully connected layer and a Softmax function, and the passability evaluation branch generates a continuous scoring value reflecting the possibility of passability through a Sigmoid function. The results of the two constitute a structured recognition output, which is transmitted to the navigation module for path planning and decision execution, effectively realizing the closed-loop logic of recognition-scoring-decision-making, and improving the mobile robot's intelligent passage and autonomous response capabilities in complex terrain.

[0064] Optionally, the S7 specifically includes:

[0065] S71. Encapsulate the ground type recognition result and the traffic score result according to a unified coding format to form a road surface status data packet;

[0066] S72, sending the road status data packet to the mobile robot navigation module through the communication interface, the navigation module including a data receiving unit, an environment analysis unit, a path planning unit and a motion control unit;

[0067] S73: The data receiving unit is configured to receive a road surface status data packet and extract a ground type code and a traffic score value. The environment parsing unit receives the output of the data receiving unit, searches for a road condition label corresponding to the ground type code, and retrieves ground traffic restriction parameters from a preset environment rule set.

[0068] S74. The path planning unit receives the access score and the ground access restriction parameter, constructs a path cost function, wherein the path cost function includes the access score as a cost term and the access restriction parameter as a path risk term, and uses a graph search algorithm to select the minimum cost path for the candidate paths to generate a path node sequence.

[0069] S75. Each node in the path node sequence includes two-dimensional coordinates, expected speed, posture angle and path number; the motion control unit calculates the continuous target trajectory point set based on the parameters in the node and the actual posture data fed back by the current robot position sensor using a polynomial curve fitting method, and generates corresponding speed instructions and direction angle instructions by calculating the position error and angle deviation between the continuous target trajectory points and the current posture.

[0070] The present invention encapsulates the ground type identification results and the traffic score results into a road status data packet in a unified format and sends it to the navigation module. The data reception, environment analysis, path planning and motion control unit in the module work together to construct a path cost function based on the score value and preset traffic restriction parameters, and uses a graph search algorithm to select the path with the minimum cost. Finally, speed and steering control instructions are generated through curve fitting and error analysis, realizing a closed-loop linkage from perception and recognition to path decision-making and execution, significantly improving the path adaptability, traffic efficiency and dynamic decision-making ability of the mobile robot in complex or unstructured terrain.

[0071] The beneficial effects of the present invention are:

[0072] Unlike traditional road surface recognition methods that rely solely on visual imagery or force sensing signals, this method constructs a dual-channel image feature extraction network to extract both topographic structural features and semantically ambiguous regions. Furthermore, it introduces a disentangled complementary learning model to achieve efficient recognition of fine-grained topographic structure during the feature decoupling and semantic enhancement stages. This structured channel division of labor significantly improves image representation in environments with shadow interference, low contrast, and inter-class ambiguity.

[0073] This invention innovatively introduces a contact event mapping algorithm to model the time series of contact between the wheel and the ground. It then constructs a contact event map by extracting dynamic features of behavioral patterns such as sinking, slipping, and drifting. Combined with a bidirectional gated recursive structure to encode contact sequences, the system can effectively identify potentially hazardous terrain that is invisible to the naked eye, addressing blind spots in image perception and improving the robot's ability to assess stability on complex surfaces.

[0074] During the multimodal fusion phase, the present invention proposes an alignment mechanism for visual semantic vectors and contact event encodings, and introduces a dynamic confidence weighting strategy. By jointly calculating channel similarity and perceptual confidence, this achieves heterogeneous weighted fusion of image and force information. This mechanism not only maintains output stability when perceptual information is inconsistent, but also enhances the system's adaptability to changing environments.

[0075] During the decision-making output phase, this invention constructs a ground accessibility scoring model that incorporates type recognition and accessibility scoring. This enables the system to structure the output of multi-source perception results, forming a closed-loop path control system that is highly coupled with the navigation module. Unlike the traditional "recognition-alert" model, this invention integrates the entire process of recognition, assessment, planning, and control, providing mobile robots with intelligent terrain response capabilities for field deployment. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0077] Figure 1 This is a flow chart of an intelligent road surface recognition method based on force sensing and image analysis proposed by the present invention;

[0078] Figure 2 This is a structural flow chart of the disentangled complementary learning model proposed in the present invention;

[0079] Figure 3 This is a modeling flowchart of the contact event mapping algorithm proposed in the present invention. DETAILED DESCRIPTION

[0080] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0081] refer to Figure 1-3 , an intelligent road surface recognition method based on force sensing and image analysis, comprising the following steps:

[0082] S1. Obtain environmental image data and contact force perception data collected by the mobile robot during its movement as visual perception input and force perception input respectively;

[0083] S2. The visual perception input is processed through the constructed dual-channel image feature extraction network. The main channel extracts the landform structure features, and the complementary channel extracts the difference features of the semantically fuzzy area to form an image feature representation;

[0084] S3. Input the image feature representation into the disentanglement complementary learning model, perform feature disentanglement and semantic enhancement operations, and generate an image semantic representation vector;

[0085] S4. Construct a contact time series based on the force perception input, extract the dynamic unstable behavior characteristics during the contact process through the contact event mapping algorithm, and generate a contact event encoding vector;

[0086] S5. Align the image semantic representation vector with the contact event encoding vector, fuse them through a dynamic confidence weighting mechanism, and output a joint road surface state representation vector.

[0087] S6. Input the joint road surface state representation vector into the ground accessibility scoring model to generate a current ground type recognition result and a accessibility scoring result;

[0088] S7. Transmit the type recognition result and the pass score result to the mobile robot navigation module for path planning and motion decision execution.

[0089] In this embodiment, S1 specifically includes: visual perception input through a binocular camera system installed on the front of the mobile robot to collect RGB images and depth images in front of the travel path; force perception input through a three-axis force sensor arranged at the wheel or foot contact part to collect the normal force, tangential force and friction force at each frame timestamp; the collected image data and force data are format standardized and unit normalized respectively, and linear interpolation alignment is performed according to a unified time base, and finally an image input tensor and a force input sequence are constructed for subsequent network input processing.

[0090] In this embodiment, S2 specifically includes:

[0091] S21, constructing a dual-channel image feature extraction network, wherein the network includes a main feature channel and a complementary feature channel, and the image data in the visual perception input is input into the main feature channel and the complementary feature channel respectively for parallel processing;

[0092] S22. In the main feature channel, a landform structure encoding unit is set. The landform structure encoding unit includes multiple serially connected two-dimensional convolutional layers and an integrated feature compression module. The two-dimensional convolutional layer extracts terrain boundaries, texture directions, and geometric structure features in the image. The integrated feature compression module includes batch normalization and maximum pooling structures to regularize the convolution output and reduce the spatial dimension. All extracted features are integrated along the channel dimension and then flattened to form a main channel feature representation.

[0093] S23. In the complementary feature channel, a semantic fuzzy detection unit is set. The semantic fuzzy detection unit includes an attention enhancement module and a multi-scale feature extraction module. The attention enhancement module includes a spatial weighting path and a channel importance path. The spatial weighting path generates a saliency map based on the regional response. The channel importance path estimates the semantic dimension weight based on the global pooling and activation function. The multi-scale feature extraction module includes multiple two-dimensional convolution paths with different receptive fields to identify edge and texture changes in the fuzzy area. All path outputs are spliced ​​and weightedly fused in the channel dimension to generate a complementary channel feature representation.

[0094] S24. Concatenate the main channel feature representation and the complementary channel feature representation in the channel dimension to generate a fused feature tensor, and perform channel compression and spatial consistency processing to form an image feature representation.

[0095] In this embodiment, S3 specifically includes:

[0096] S31, inputting the image feature representation into the structural semantic extraction channel and the pseudo semantic difference extraction channel at the same time, and performing parallel processing in the disentanglement complementary learning model;

[0097] S32, the structural semantic extraction channel receives the image feature representation, extracts the dominant landform structural features based on the texture consistency, geometric boundary continuity and local area statistics of the image in the spatial domain, and forms the main semantic representation tensor through two-dimensional convolution, pooling and channel compression operations in sequence;

[0098] S33, the pseudo semantic difference extraction channel receives the image feature representation and the main semantic representation tensor, and constructs a residual signal by calculating the pixel-level residual map between them. The residual signal is sent to the convolution processing under the multi-scale receptive field, and combined with the spatial attention mechanism to extract the texture difference information of the semantically ambiguous area to form a complementary semantic representation tensor;

[0099] S34. The fusion decision channel receives the main semantic representation tensor and the complementary semantic representation tensor, calculates the mean difference and coefficient of variation of the two tensors in the channel dimension, constructs the channel fusion weight, performs channel weighting and summation on the two tensors, and generates a fused image semantic representation vector.

[0100] In this embodiment, the disentanglement complementary learning model includes:

[0101] The structural semantic extraction channel consists of a series of two-dimensional convolutional layers, pooling layers, channel attention structures, and channel compression structures. The two-dimensional convolutional layer contains two standard convolution operations with a kernel size of 3×3. The pooling layer uses the maximum pooling method. The channel attention structure consists of global average pooling, a fully connected layer, and a sigmoid activation function. The channel compression structure is a convolution operation with a kernel size of 1×1, which is used to output the main semantic representation tensor.

[0102] The pseudo semantic difference extraction channel includes a residual signal construction module, a multi-scale convolution module, and a spatial attention module. The residual signal construction module contains an element-by-element difference unit and an upsampling unit to construct a residual signal between the image and the main semantic representation. The multi-scale convolution module consists of three convolution kernels of sizes 1×1, 3×3, and 5×5. The spatial attention module consists of an average pooling branch, a maximum pooling branch, a convolution fusion layer, and a sigmoid activation structure, and outputs a complementary semantic representation tensor.

[0103] The fusion decision channel includes a fusion weight generation unit and a weighted fusion unit. The fusion weight generation unit constructs a channel fusion weight vector based on the channel variance and channel mean difference of the main semantic representation tensor and the complementary semantic representation tensor. The weighted fusion unit multiplies the channels of the two tensors by the fusion weights respectively and performs a sum operation to output the image semantic representation vector.

[0104] In this embodiment, the S4 specifically includes:

[0105] S41. Constructing a contact time sequence for force perception input, specifically including: continuously collecting normal force, tangential force, and friction force data for each frame from the three-axis force sensor deployed at the contact part of the mobile robot, and sorting them by timestamp to form a force data sequence T f ;

[0106] S42, set the sliding window with length w and step size s, take the time series as input, generate K continuous behavior segments, and form a behavior segment set

[0107] S43. Define and calculate the normal force fluctuation index for each behavior segment Proportional index to tangential friction The specific formula is as follows:

[0108]

[0109] Among them, ε is a minimum constant, represents the normal force at frame t, represents the tangential force, represents friction, T is the total number of frames;

[0110] S44. Setting normal force threshold θ n and the tangential friction threshold θ t , the indicator of each behavior segment and Determine the contact event label L according to the following mapping relationship k :

[0111]

[0112] S45. Constructing a contact event mapping algorithm, including the following steps:

[0113] Each behavior fragment and the corresponding event label L k As input, a category-driven contact event feature sequence is constructed. Specifically, if the segment category is a sinking event, the normal force sudden drop trend is extracted; if the category is a slip event, the tangential force increasing trend is extracted; if the category is a drift event, the triaxial force periodic unstable fluctuation trend is extracted; if the category is a stable event, the triaxial force stable trend is extracted;

[0114] The trend features in each behavior segment are represented as a feature vector sequence, which is input into the bidirectional gated recurrent unit and the hidden state h is updated at each time step. t , the state update formula is:

[0115] h t =(1-z t )⊙h t-1 +zt ⊙tanh(W h x t +U h (r t ⊙h t-1 ));

[0116] Among them, h t is the hidden state vector at time t, z t is the update gate, which controls the degree of retention of the current state, r t is the reset gate, which controls the degree of reset of historical state information, x t is the input feature vector at time t, W h 、U h is the weight matrix of the network, ⊙ represents element-by-element multiplication;

[0117] The hidden state vector of the forward recursive unit at the last moment and the hidden state vector of the backward recursive unit at the first moment are arranged in sequence to form the event feature representation vector e of the segment k ;

[0118] S46, for all event feature representation vector sets {e1, e2, ..., e K Perform a global average pooling operation to obtain the final contact event encoding vector: The contact event encoding vector V c As a characteristic representation of the dynamic unstable behavior of the current robot during contact.

[0119] In this embodiment, the S5 specifically includes:

[0120] S51, performing maximum and minimum normalization operations on the image semantic representation vector and the contact event encoding vector respectively, so that the value range of each channel dimension is mapped to the interval [0, 1];

[0121] S52, inputting the normalized image semantic representation vector and the contact event encoding vector into two parameter-sharing linear mapping networks to generate equal-dimensional vector representations, respectively. The linear mapping network is composed of a fully connected layer, and the output dimension is consistent with the preset fusion dimension;

[0122] S53, calculating the cosine similarity of each dimension between two equal-dimensional vectors, and constructing a matching weight vector, where each dimension represents the directional similarity between the image semantic channel and the contact event channel in the same dimension;

[0123] S54. Calculating an image confidence distribution vector based on the mean and standard deviation of each channel of the image semantic representation vector. The image confidence distribution vector is used to characterize the degree of concentration of the image semantic feature distribution within each channel. Calculating a contact confidence distribution vector based on the temporal variation range of each channel of the contact event encoding vector within the sliding window. The contact confidence distribution vector is used to characterize the magnitude of the contact behavior variation in each channel.

[0124] S55, performing weighted averaging on the matching weight vector, the image confidence distribution vector, and the contact confidence distribution vector to generate a fused confidence weighted vector, wherein each dimension is used to determine a weighted ratio between the image semantic representation vector and the contact event encoding vector in the dimension;

[0125] S56. Multiply the image semantic representation vector and the contact event encoding vector by the element values ​​of the corresponding dimensions in the fusion confidence weighted vector respectively, perform dimension-by-dimension weighting and then add them together, and output a joint road surface state representation vector.

[0126] In this embodiment, S6 specifically includes:

[0127] S61. Receive a joint road surface state representation vector as input and input it into a ground trafficability scoring model, wherein the model includes a type identification branch and a trafficability evaluation branch, and the two branches operate in parallel;

[0128] S62, the type recognition branch includes a fully connected mapping layer and a normalization function, wherein the fully connected mapping layer maps the input vector to a category probability vector, and the normalization function is a Softmax function for normalizing all category probabilities and outputting the category corresponding to the maximum probability as the ground type recognition result;

[0129] S63, the pass rating branch includes a fully connected mapping layer and an activation function, wherein the fully connected mapping layer outputs a real-valued channel, and the activation function is a Sigmoid function, which is used to compress the output value to the interval [0, 1] to obtain a pass rating result;

[0130] S64: Combining the type recognition result and the traffic scoring result into a structured output result, and transmitting it to the navigation module for path control call.

[0131] In this embodiment, the S7 specifically includes:

[0132] S71. Encapsulate the ground type recognition result and the traffic score result according to a unified coding format to form a road surface status data packet;

[0133] S72, sending the road status data packet to the mobile robot navigation module through the communication interface, the navigation module including a data receiving unit, an environment analysis unit, a path planning unit and a motion control unit;

[0134] S73: The data receiving unit is configured to receive a road surface status data packet and extract a ground type code and a traffic score value. The environment parsing unit receives the output of the data receiving unit, searches for a road condition label corresponding to the ground type code, and retrieves ground traffic restriction parameters from a preset environment rule set.

[0135] S74. The path planning unit receives the access score and the ground access restriction parameter, constructs a path cost function, wherein the path cost function includes the access score as a cost term and the access restriction parameter as a path risk term, and uses a graph search algorithm to select the minimum cost path for the candidate paths to generate a path node sequence.

[0136] S75. Each node in the path node sequence includes two-dimensional coordinates, expected speed, posture angle and path number; the motion control unit calculates the continuous target trajectory point set based on the parameters in the node and the actual posture data fed back by the current robot position sensor using a polynomial curve fitting method, and generates corresponding speed instructions and direction angle instructions by calculating the position error and angle deviation between the continuous target trajectory points and the current posture.

[0137] Example 1:

[0138] To verify the feasibility and engineering adaptability of this invention, we deployed it in a mountain substation inspection scenario affiliated with a power company. The region receives an average annual rainfall of over 1200 mm, and the inspection route encompasses typical unstructured and complex terrain, including concrete pavement, soft mountain roads, slippery grass, and gravel piles. Traditional quadruped robots rely on image classification networks to identify the road ahead and use empirically defined force thresholds to determine whether the road is passable. However, these robots frequently fail to accurately identify muddy, rain-slippery terrain, and gravel landslides, severely impacting inspection efficiency and the robot's stable operation.

[0139] To address these challenges, we integrated our proposed "intelligent road surface recognition method based on force sensing and image analysis" into a domestically produced medium-sized quadruped mobile robot platform. The system features an RGB-D binocular camera on the front end and high-precision three-axis force sensors at the end legs. The edge computing platform is equipped with a lightweight neural network module. During the initial deployment phase, researchers completed preprocessing steps, including standardization of image and force data acquisition formats and data alignment and interpolation, to ensure stable system operation.

[0140] During inspections, the robot collects forward visual images and contact force signals in real time. The image channels are processed by the dual-channel image feature extraction network proposed in this paper. The primary channel extracts terrain boundaries and texture features, while the complementary channel identifies areas of image blur or fine-grained differences. Together, they form a fused image tensor. This tensor is then fed into a disentangled complementary learning model for primary semantic extraction and residual signal difference analysis, ultimately forming an image semantic vector.

[0141] Simultaneously, the force sensing channel continuously collects data on normal force, tangential force, and friction, forming a contact time series. Using the proposed contact event mapping algorithm, the normal force fluctuation coefficient and tangential friction ratio are calculated using a sliding window. Behavior segments are extracted and mapped to contact labels such as stability, sinking, slipping, or drifting. These are then fed into a bidirectionally gated recursive unit and encoded into event feature vectors.

[0142] During the fusion phase, the two modalities are aligned using a dynamic confidence weighting mechanism. The resulting joint state vector is fed into the trafficability scorer, which generates a road surface recognition result and a trafficability score. The trafficability score directly contributes to the construction of the path cost function, guiding the navigation module to select the path with the least risk.

[0143] In actual testing, we selected the same inspection route and performed a complete route mission from 9:00 AM to 11:30 AM daily for 10 consecutive days. The experimental group used a robot integrated with the proposed system, while the control group used a robot equipped with a traditional image CNN + threshold judgment model. Both systems recorded objective indicators such as the success rate of the route, route deviation, and the number of mission interruptions.

[0144] The following table shows the statistical results based on a total of 153 complex road surface tests:

[0145] Table 1 Comparison of road surface type recognition and traffic performance

[0146]

[0147] Data showed that the experimental group significantly outperformed the control group on all unstructured road surfaces, with a 17.1% to 26.3% increase in successful passage. This was particularly true on muddy, muddy depressions and gravel slopes with blurry images and complex road conditions. The experimental group benefited from the dynamic feature modeling of the contact event mapping model, enabling them to predict slippage and sinking trends in advance, improving recognition accuracy.

[0148] In terms of navigation and operational performance, we further recorded the daily frequency of path adjustments, navigation deviation, and calculation delay:

[0149] Table 2 Comparison of path planning and system performance

[0150]

[0151]

[0152] The data sheet explains that thanks to the discriminative power and rapid decision-making capabilities of the present invention's combined image-force representation, the system can reduce redundant path corrections and minimize navigational drift. Furthermore, the fusion phase utilizes a single-layer linear mapping and confidence vector weighting, eliminating the need for high-complexity models. This ensures overall operational latency is kept below 200ms, meeting the robot's real-time perception and path update requirements.

[0153] From the specific operation cases, in the 6th day mission, the 21st section of the inspection path was covered with wet grass and loose gravel after the rain the night before. The control group robot mistakenly identified it as "stable grass" and continued to move forward, causing the front foot to slip and automatically trigger an emergency shutdown; while the experimental group robot extracted the non-structural texture signal in the blurred area of ​​the image, and combined with the rapid fluctuation information of friction in the contact data, accurately identified it as a "suspicious slip" area with a score of 0.32 (lower than the pass threshold of 0.45). The system determined that this section was not safe to pass, automatically detoured along the backup path, and finally completed the mission as planned.

[0154] In summary, this example fully demonstrates the technical features and system effectiveness of this invention. The complementary learning structure in the image channel and the behavioral modeling mechanism in the force channel work synergistically to ensure the system's stable and highly accurate road surface recognition and traffic assessment capabilities even in complex terrain conditions. Compared to traditional methods, this invention significantly improves traffic success rate, path stability, and response efficiency, demonstrating promising field deployment value and engineering promotion prospects.

[0155] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An intelligent road surface recognition method based on force perception and image analysis, characterized in that: The steps include: S1. Obtain environmental image data and contact force perception data collected by the mobile robot during its movement as visual perception input and force perception input respectively; S2. The visual perception input is processed through the constructed dual-channel image feature extraction network. The main channel extracts the landform structure features, and the complementary channel extracts the difference features of the semantically fuzzy area to form an image feature representation; S3. Input the image feature representation into the disentanglement complementary learning model, perform feature disentanglement and semantic enhancement operations, and generate an image semantic representation vector; S4. Construct a contact time series based on the force perception input, extract the dynamic unstable behavior characteristics during the contact process through the contact event mapping algorithm, and generate a contact event encoding vector; S5. Align the image semantic representation vector with the contact event encoding vector, fuse them through a dynamic confidence weighting mechanism, and output a joint road surface state representation vector. S6. Input the joint road surface state representation vector into the ground accessibility scoring model to generate a current ground type recognition result and a accessibility scoring result; S7. Transmit the type recognition result and the pass score result to the mobile robot navigation module for path planning and motion decision execution.

2. The intelligent road surface recognition method based on force perception and image analysis according to claim 1 is characterized in that: The S1 specifically includes: visual perception input is obtained by a binocular camera system installed on the front of the mobile robot to collect RGB images and depth images in front of the travel path; force perception input is obtained by a three-axis force sensor arranged at the wheel or foot contact part to collect the normal force, tangential force and friction force at each frame timestamp; the collected image data and force data are format standardized and unit normalized respectively, and linear interpolation alignment is performed according to a unified time base, and finally an image input tensor and a force input sequence are constructed for subsequent network input processing.

3. The intelligent road surface recognition method based on force perception and image analysis according to claim 1, characterized in that: The S2 specifically includes: S21, constructing a dual-channel image feature extraction network, wherein the network includes a main feature channel and a complementary feature channel, and the image data in the visual perception input is input into the main feature channel and the complementary feature channel respectively for parallel processing; S22. In the main feature channel, a landform structure encoding unit is set. The landform structure encoding unit includes multiple serially connected two-dimensional convolutional layers and an integrated feature compression module. The two-dimensional convolutional layer extracts terrain boundaries, texture directions, and geometric structure features in the image. The integrated feature compression module includes batch normalization and maximum pooling structures to regularize the convolution output and reduce the spatial dimension. All extracted features are integrated along the channel dimension and then flattened to form a main channel feature representation. S23. In the complementary feature channel, a semantic fuzzy detection unit is set. The semantic fuzzy detection unit includes an attention enhancement module and a multi-scale feature extraction module. The attention enhancement module includes a spatial weighting path and a channel importance path. The spatial weighting path generates a saliency map based on the regional response. The channel importance path estimates the semantic dimension weight based on the global pooling and activation function. The multi-scale feature extraction module includes multiple two-dimensional convolution paths with different receptive fields to identify edge and texture changes in the fuzzy area. All path outputs are spliced ​​and weightedly fused in the channel dimension to generate a complementary channel feature representation. S24. Concatenate the main channel feature representation and the complementary channel feature representation in the channel dimension to generate a fused feature tensor, and perform channel compression and spatial consistency processing to form an image feature representation.

4. The intelligent road surface recognition method based on force perception and image analysis according to claim 1, characterized in that: The S3 specifically includes: S31, inputting the image feature representation into the structural semantic extraction channel and the pseudo semantic difference extraction channel at the same time, and performing parallel processing in the disentanglement complementary learning model; S32, the structural semantic extraction channel receives the image feature representation, extracts the dominant landform structural features based on the texture consistency, geometric boundary continuity and local area statistics of the image in the spatial domain, and forms the main semantic representation tensor through two-dimensional convolution, pooling and channel compression operations in sequence; S33, the pseudo semantic difference extraction channel receives the image feature representation and the main semantic representation tensor, and constructs a residual signal by calculating the pixel-level residual map between them. The residual signal is sent to the convolution processing under the multi-scale receptive field, and combined with the spatial attention mechanism to extract the texture difference information of the semantically ambiguous area to form a complementary semantic representation tensor; S34. The fusion decision channel receives the main semantic representation tensor and the complementary semantic representation tensor, calculates the mean difference and coefficient of variation of the two tensors in the channel dimension, constructs the channel fusion weight, performs channel weighting and summation on the two tensors, and generates a fused image semantic representation vector.

5. The intelligent road surface recognition method based on force sensing and image analysis according to claim 1, characterized in that: The disentangled complementary learning model includes: The structural semantic extraction channel consists of a series of two-dimensional convolutional layers, pooling layers, channel attention structures, and channel compression structures. The two-dimensional convolutional layer contains two standard convolution operations with a kernel size of 3×3. The pooling layer uses the maximum pooling method. The channel attention structure consists of global average pooling, a fully connected layer, and a sigmoid activation function. The channel compression structure is a convolution operation with a kernel size of 1×1, which is used to output the main semantic representation tensor. The pseudo semantic difference extraction channel includes a residual signal construction module, a multi-scale convolution module, and a spatial attention module. The residual signal construction module contains an element-by-element difference unit and an upsampling unit to construct a residual signal between the image and the main semantic representation. The multi-scale convolution module consists of three convolution kernels of sizes 1×1, 3×3, and 5×5. The spatial attention module consists of an average pooling branch, a maximum pooling branch, a convolution fusion layer, and a sigmoid activation structure, and outputs a complementary semantic representation tensor. The fusion decision channel includes a fusion weight generation unit and a weighted fusion unit. The fusion weight generation unit constructs a channel fusion weight vector based on the channel variance and channel mean difference of the main semantic representation tensor and the complementary semantic representation tensor. The weighted fusion unit multiplies the channels of the two tensors by the fusion weights respectively and performs a sum operation to output the image semantic representation vector.

6. The intelligent road surface recognition method based on force perception and image analysis according to claim 1, characterized in that: The S4 specifically includes: S41. Constructing a contact time sequence for force perception input, specifically including: continuously collecting normal force, tangential force, and friction force data for each frame from the three-axis force sensor deployed at the contact part of the mobile robot, and sorting them by timestamp to form a force data sequence T f ; S42, set the sliding window with length w and step size s, take the time series as input, generate K continuous behavior segments, and form a behavior segment set S43. Define and calculate the normal force fluctuation index for each behavior segment Proportional index to tangential friction The specific formula is as follows: Among them, ε is a minimum constant, represents the normal force at frame t, represents the tangential force, represents friction, T is the total number of frames; S44. Setting normal force threshold θ n and the tangential friction threshold θ t , the indicator of each behavior segment and Determine the contact event label L according to the following mapping relationship k : S45. Constructing a contact event mapping algorithm, including the following steps: Each behavior fragment and the corresponding event label L k As input, a category-driven contact event feature sequence is constructed. Specifically, if the segment category is a sinking event, the normal force sudden drop trend is extracted; if the category is a slip event, the tangential force increasing trend is extracted; if the category is a drift event, the triaxial force periodic unstable fluctuation trend is extracted; if the category is a stable event, the triaxial force stable trend is extracted; The trend features in each behavior segment are represented as a feature vector sequence, which is input into the bidirectional gated recurrent unit and the hidden state h is updated at each time step. t , the state update formula is: h t =(1-z t )⊙h t-1 +z t ⊙tanh(W h x t +U h (r t ⊙h t-1 )); Among them, h t is the hidden state vector at time t, z t is the update gate, which controls the degree of retention of the current state, r t is the reset gate, which controls the degree of reset of historical state information, x t is the input feature vector at time t, W h 、U h is the weight matrix of the network, ⊙ represents element-by-element multiplication; The hidden state vector of the forward recursive unit at the last moment and the hidden state vector of the backward recursive unit at the first moment are arranged in sequence to form the event feature representation vector e of the segment k ; S46, for all event feature representation vector sets {e1, e2, ..., e K Perform a global average pooling operation to obtain the final contact event encoding vector: The contact event encoding vector V c As a characteristic representation of the dynamic unstable behavior of the current robot during contact.

7. The intelligent road surface recognition method based on force sensing and image analysis according to claim 1 is characterized in that: The S5 specifically includes: S51, performing maximum and minimum normalization operations on the image semantic representation vector and the contact event encoding vector respectively, so that the value range of each channel dimension is mapped to the interval [0, 1]; S52, inputting the normalized image semantic representation vector and the contact event encoding vector into two parameter-sharing linear mapping networks to generate equal-dimensional vector representations, respectively. The linear mapping network is composed of a fully connected layer, and the output dimension is consistent with the preset fusion dimension; S53, calculating the cosine similarity of each dimension between two equal-dimensional vectors, and constructing a matching weight vector, where each dimension represents the directional similarity between the image semantic channel and the contact event channel in the same dimension; S54. Calculating an image confidence distribution vector based on the mean and standard deviation of each channel of the image semantic representation vector. The image confidence distribution vector is used to characterize the degree of concentration of the image semantic feature distribution within each channel. Calculating a contact confidence distribution vector based on the temporal variation range of each channel of the contact event encoding vector within the sliding window. The contact confidence distribution vector is used to characterize the magnitude of the contact behavior variation in each channel. S55, performing weighted averaging on the matching weight vector, the image confidence distribution vector, and the contact confidence distribution vector to generate a fused confidence weighted vector, wherein each dimension is used to determine a weighted ratio between the image semantic representation vector and the contact event encoding vector in the dimension; S56. Multiply the image semantic representation vector and the contact event encoding vector by the element values ​​of the corresponding dimensions in the fusion confidence weighted vector respectively, perform dimension-by-dimension weighting and then add them together, and output a joint road surface state representation vector.

8. The intelligent road surface recognition method based on force sensing and image analysis according to claim 1 is characterized in that: The S6 specifically includes: S61. Receive a joint road surface state representation vector as input and input it into a ground trafficability scoring model, wherein the model includes a type identification branch and a trafficability evaluation branch, and the two branches operate in parallel; S62, the type recognition branch includes a fully connected mapping layer and a normalization function, wherein the fully connected mapping layer maps the input vector to a category probability vector, and the normalization function is a Softmax function for normalizing all category probabilities and outputting the category corresponding to the maximum probability as the ground type recognition result; S63, the pass rating branch includes a fully connected mapping layer and an activation function, wherein the fully connected mapping layer outputs a real-valued channel, and the activation function is a Sigmoid function, which is used to compress the output value to the interval [0, 1] to obtain a pass rating result; S64. Combining the type recognition result and the traffic scoring result into a structured output result.

9. The intelligent road surface recognition method based on force sensing and image analysis according to claim 1, characterized in that: The S7 specifically includes: S71. Encapsulate the ground type recognition result and the traffic score result according to a unified coding format to form a road surface status data packet; S72, sending the road status data packet to the mobile robot navigation module through the communication interface, the navigation module including a data receiving unit, an environment analysis unit, a path planning unit and a motion control unit; S73: The data receiving unit is configured to receive a road surface status data packet and extract a ground type code and a traffic score value. The environment parsing unit receives the output of the data receiving unit, searches for a road condition label corresponding to the ground type code, and retrieves ground traffic restriction parameters from a preset environment rule set. S74. The path planning unit receives the access score and the ground access restriction parameter, constructs a path cost function, wherein the path cost function includes the access score as a cost term and the access restriction parameter as a path risk term, and uses a graph search algorithm to select the minimum cost path for the candidate paths to generate a path node sequence. S75. Each node in the path node sequence includes two-dimensional coordinates, expected speed, posture angle and path number; the motion control unit calculates the continuous target trajectory point set based on the parameters in the node and the actual posture data fed back by the current robot position sensor using a polynomial curve fitting method, and generates corresponding speed instructions and direction angle instructions by calculating the position error and angle deviation between the continuous target trajectory points and the current posture.

Images