A negative obstacle risk map construction method and system

Abstract

The application provides a negative obstacle risk map construction method and system, and relates to the technical field of robot environment perception. The method comprises the following steps: acquiring multi-modal sensing data of a legged robot in a current working environment to obtain the type of negative obstacles, a plurality of candidate areas, and the confidence and uncertainty corresponding to the candidate areas in the current working environment; performing ink halo processing on the type and the confidence and uncertainty by using a risk diffusion algorithm to generate a continuous risk field corresponding to the candidate areas; dividing the continuous risk field to generate a continuous risk gradient distribution corresponding to each candidate area; modeling the terrain based on laser point clouds, updating the continuous risk gradient distribution by using a Bayesian fusion method to obtain a multi-level grid map, and performing touch closed-loop correction on the multi-level grid map to generate a risk cost map. The application improves the construction accuracy of the negative obstacle risk map.

Description

Technical Field

[0001] This invention relates to the field of robot environmental perception technology, and more specifically, to a method and system for constructing a negative obstacle risk map. Background Technology

[0002] Legged mobile robots are widely used in many scenarios due to their ability to cross obstacles and adapt to complex terrains. However, in complex unstructured environments, there are often negative obstacles such as pits, ditches, collapses, the lower edge of steps, and the edge of cliffs. These obstacles are geometrically represented as cavities, depressions, or height differences. When observed by sensors, they are prone to problems such as occlusion, reflection, shadows, sparse point clouds, or parallax instability, which can easily lead to the robot's feet slipping, instability, mechanical damage, or even tipping over.

[0003] In related technologies, the detection of negative obstacles typically involves increasing the density of ground point clouds by installing LiDAR at an angle, and then combining geometric features with multi-frame fusion to identify trenches / potholes, or performing feature detection under sparse point cloud conditions to identify negative obstacles. However, these methods focus more on detecting the presence of negative obstacles and lack a continuous expression of the degree of danger and safety buffer of negative obstacles. Furthermore, in long-term tasks, extrinsic parameter drift, changes in ambient lighting, dust, rain, and snow can cause systematic biases in visual or geometric observations, leading to the gradual solidification of errors in the risk map, which in turn can cause planning failures or safety risks. Summary of the Invention

[0004] The problem addressed by this invention is how to improve the accuracy of constructing negative obstacle risk maps.

[0005] To address the above problems, this invention provides a method and system for constructing a negative obstacle risk map.

[0006] In a first aspect, the present invention provides a method for constructing a negative obstacle risk map, comprising: Acquire multimodal sensing data of the legged robot in the current working environment. The multimodal sensing data includes visual images of the current working environment, laser point clouds, and contact perception data of the legged robot. Based on the visual image, semantic segmentation is performed to obtain the type of negative obstacle in the current working environment, multiple candidate regions of the negative obstacle, and the confidence and uncertainty corresponding to each candidate region; Using a risk diffusion algorithm, ink smudging is performed based on the type of negative obstacle and the confidence and uncertainty of each candidate region to generate a continuous risk field corresponding to each candidate region. The continuous risk field is divided to generate a continuous risk gradient distribution corresponding to each candidate region; Terrain modeling is performed based on the laser point cloud to obtain geometric risk cues indicating the presence of negative obstacles in the current working environment; By using Bayesian fusion, the continuous risk gradient distribution is updated based on the geometric risk cues to obtain a multi-level grid map; Based on the tactile perception data, the multi-level grid map is corrected using tactile closed-loop technology to generate a risk cost map for the current working environment.

[0007] Optionally, acquiring multimodal sensing data of the legged robot in the current working environment includes: The legged robot uses its visual sensor, lidar sensor, and foot-end tactile sensing component to collect initial visual images, initial laser point clouds, and initial contact perception data of the current working environment, as well as the contact between the legged robot's foot and the ground, at preset frequencies corresponding to the visual sensor, lidar sensor, and foot-end tactile sensing component, respectively. The initial visual image, the initial laser point cloud, and the initial contact sensing data are time-synchronized and calibrated to obtain the visual image, the laser point cloud, and the contact sensing data of the current working environment.

[0008] Optionally, the step of performing semantic segmentation based on the visual image to obtain the type of negative obstacle in the current working environment, multiple candidate regions of the negative obstacle, and the confidence and uncertainty corresponding to each candidate region includes: The visual image is classified at the pixel level using a lightweight semantic segmentation network to obtain the category mask of the visual image and the pixel confidence and pixel uncertainty corresponding to each pixel position in the visual image; Connectivity aggregation and contour extraction are performed on the category mask to obtain the initial candidate region of the negative obstacle and the type of the negative obstacle; Perform geometric consistency filtering on each of the initial candidate regions to obtain multiple candidate regions for the negative obstacle; The confidence level and uncertainty corresponding to each candidate region are obtained by statistically calculating the pixel-level confidence level and pixel uncertainty of each pixel position within each candidate region.

[0009] Optionally, the step of generating a continuous risk field corresponding to each candidate region by performing ink blurring processing based on the type of the negative obstacle and the confidence and uncertainty of each candidate region using a risk diffusion algorithm includes: Determine the category risk coefficient for the type of negative obstacle; Based on the category risk system, and combining the confidence level and uncertainty of each candidate region, the risk intensity coefficient of the candidate region is obtained; Feature extraction is performed on the candidate region to obtain the center position, main morphological direction, and scale features of the candidate region; Based on the center location, the main orientation of the shape, and the scale characteristics, combined with the terrain aspect and slope characteristics of the laser point cloud, a covariance matrix is ​​generated. Ink smudging is performed based on the risk intensity coefficient and the covariance matrix to generate the continuous risk field corresponding to each candidate region.

[0010] Optionally, the step of dividing the continuous risk field to generate a continuous risk gradient distribution corresponding to each candidate region includes: Based on a preset multi-level risk classification threshold, the continuous risk field is divided into intervals to obtain multiple risk zones corresponding to the candidate region. The risk zones include fatal zones, buffer zones, and safe zones. Based on the risk distribution characteristics of each risk zone, a continuous risk gradient distribution with gradient decay from the lethal zone to the safe zone is generated.

[0011] Optionally, the step of performing terrain modeling based on the laser point cloud to obtain geometric risk cues indicating the presence of the negative obstacle in the current operating environment includes: Voxel downsampling and ground segmentation are performed on the laser point cloud to obtain the ground point cloud in the laser point cloud; Based on the ground point cloud, according to the preset grid parameters, an elevation map of the current working environment is constructed, and the slope and height change values ​​of each grid in the elevation map are obtained; Based on the slope and height abrupt change value of each grid in the elevation map, the cavity features in the elevation map, the cliff edge features of the ground point cloud, and the distance jump features of the laser point cloud are obtained. The slope, the height abrupt change value, the cavity feature, the cliff edge feature, and the distance jump feature are used as the geometric risk clues.

[0012] Optionally, the step of updating the continuous risk gradient distribution based on the geometric risk cues using a Bayesian fusion method to obtain a multi-level grid map includes: Based on the slope, the height abrupt change value, the cavity feature, the cliff edge feature, and the distance jump feature, the geometric risk distribution value of each grid in the elevation map is generated, and the laser observation uncertainty corresponding to the geometric risk distribution value is obtained; Based on the continuous risk gradient distribution, the visual risk distribution value is obtained, and the visual observation uncertainty corresponding to the visual risk distribution value is acquired. The visual risk distribution value and the geometric risk distribution value are weighted and fused to obtain risk layer data; The occupancy observation data corresponding to the laser point cloud are accumulated to obtain occupancy layer data; Based on the visual observation uncertainty and the laser observation uncertainty, uncertainty layer data is generated, and freshness layer data is generated through a time decay model. The risk layer data, the occupation layer data, the uncertainty layer data, and the freshness layer data are used as levels for raster index fusion to obtain the multi-level raster map.

[0013] Optionally, the step of performing tactile closed-loop correction on the multi-level grid map based on the tactile sensing data to generate a risk cost map for the current working environment includes: Based on the contact sensing data, the actual landing point of the foot of the legged robot and the corresponding ground contact state of the foot are determined. Based on the ground contact state and the actual landing point, the multi-level grid map is subjected to tactile closed-loop correction to obtain the tactile closed-loop corrected multi-level grid map. Based on the multi-level grid map corrected by the tactile closed loop, hierarchical fusion is performed, and the risk cost map in the current working environment is generated by combining the layer weight coefficients corresponding to each level.

[0014] Optionally, the ground contact state includes safe ground contact and abnormal ground contact. The step of performing tactile closed-loop correction on the multi-level grid map based on the ground contact state and the actual landing point location to obtain a tactile closed-loop corrected multi-level grid map includes: If the ground contact state is the safe ground contact, then the grid corresponding to the actual landing point and the neighboring grid in the multi-level grid map are subjected to risk dynamic erasure processing to obtain the multi-level grid map after tactile closed-loop correction. If the ground contact state is an abnormal ground contact, then the grid corresponding to the actual landing point and the neighboring grids in the multi-level grid map are subjected to risk deepening processing to obtain the multi-level grid map after tactile closed-loop correction.

[0015] Secondly, the present invention provides a negative obstacle risk map construction system, comprising: The data acquisition unit is used to acquire multimodal sensing data of the legged robot in the current working environment. The multimodal sensing data includes visual images of the current working environment, laser point clouds, and contact perception data of the legged robot. A semantic segmentation unit is used to perform semantic segmentation based on the visual image to obtain the type of negative obstacle in the current working environment, multiple candidate regions of the negative obstacle, and the confidence and uncertainty corresponding to each candidate region. The ink smudging unit is used to perform ink smudging processing based on the type of the negative obstacle and the confidence and uncertainty of each candidate region through a risk diffusion algorithm, and generate a continuous risk field corresponding to each candidate region. A partitioning unit is used to partition the continuous risk field and generate a continuous risk gradient distribution corresponding to each candidate region; The modeling unit is used to perform terrain modeling based on the laser point cloud to obtain geometric risk clues indicating the existence of the negative obstacle in the current working environment. A Bayesian fusion unit is used to update the continuous risk gradient distribution based on the geometric risk cues using a Bayesian fusion method to obtain a multi-level grid map. The map generation unit is used to perform tactile closed-loop correction on the multi-level raster map based on the tactile perception data, and generate a risk cost map for the current working environment.

[0016] The negative obstacle risk map construction method and system of the present invention firstly acquires multimodal sensing data, including visual images, laser point clouds, and contact perception data, to comprehensively collect multi-dimensional information on appearance, topographic geometry, and physical contact in negative obstacle perception. This firstly solves the limitations of single sensors in negative obstacle observation, such as occlusion, reflection, and sparse sampling, by using multi-source data to lay a foundation for risk map construction and reduces construction errors caused by missing information at the data acquisition level. Next, semantic segmentation is performed on the visual image to obtain the negative obstacle type, candidate region, and corresponding confidence and uncertainty, thereby accurately identifying the category and location range of the negative obstacle. Simultaneously, the reliability of the visual recognition results is quantified through confidence and uncertainty, avoiding misjudgments caused by directly using indiscriminate visual recognition results for mapping. Subsequently, a risk diffusion algorithm is used to combine the confidence and uncertainty of the negative obstacle type and candidate region to generate a continuous risk field through ink diffusion and to delineate a continuous risk gradient distribution. This transforms the negative obstacle from a discrete binary label into a continuous risk representation with gradient decay from the core to the periphery, thus conforming to the actual distribution characteristics of the negative obstacle and solving the problem of overly conservative or overly risky traditional binary labeling, improving the accuracy of risk representation. Finally, terrain modeling is performed based on laser point clouds to obtain geometric risk cues indicating the presence of negative obstacles. Features such as elevation, slope, voids, or cliffs of negative obstacles can be extracted from the terrain geometry perspective, providing objective geometric evidence for the existence of negative obstacles and compensating for the shortcomings of visual recognition. The system overcomes the limitations of being susceptible to environmental lighting and material variations, enabling geometric verification of visual semantic risks. Subsequently, a multi-level grid map is generated by updating the continuous risk gradient distribution using a Bayesian fusion approach combined with geometric risk cues. This integrates continuous risks at the visual semantic level with risk cues at the laser geometric level, achieving complementarity and verification of multi-source information. This allows each level of the risk map to be updated with both visual and geometric information, reducing mapping errors caused by biases from a single information source. Finally, tactile closed-loop correction is performed on the multi-level grid map based on contact perception data, generating a risk cost map. The actual physical feedback from foot contact with the ground serves as the direct basis for risk map correction, dynamically verifying and correcting the risk map obtained from the fusion of visual and laser data. This eliminates observational biases caused by external parameter drift and environmental changes during long-term operations, allowing the risk map to continuously evolve with actual ground contact experience. This improves the accuracy of negative obstacle risk map construction, making the risk map more closely reflect the true risk distribution of negative obstacles in the actual working environment. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the negative obstacle risk map construction method according to an embodiment of the present invention; Figure 2 This is a risk distribution diagram of ink smudging anisotropy in an embodiment of the present invention; Figure 3 This is a schematic diagram of the negative obstacle risk map construction system according to an embodiment of the present invention. Detailed Implementation

[0018] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0019] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0020] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first," "second," etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0021] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0022] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties. The collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0023] Combination Figure 1 As shown, an embodiment of the present invention provides a method for constructing a negative obstacle risk map, comprising: Acquire multimodal sensing data of the legged robot in the current working environment. The multimodal sensing data includes visual images of the current working environment, laser point clouds, and contact perception data of the legged robot.

[0024] Specifically, the legged robot includes a body (body and joint actuators), feet (which may include flexible footpads or replaceable contact parts), sensor components (cameras, LiDAR, IMU, foot force / torque or tactile arrays, etc.), and computing and communication components (CPU / GPU / accelerator, storage, Ethernet / serial port / USB / CAN, etc.). By simultaneously acquiring visual images, laser point clouds, and tactile perception data, a multi-dimensional data foundation for negative obstacle perception is constructed. Visual images provide the semantic features of the appearance of negative obstacles, such as the shadows of potholes and the texture of ravines; laser point clouds provide the three-dimensional terrain geometry of the legged robot in the current operating environment; and tactile perception data provides the physical feedback of the interaction between the robot's feet and the ground. Through the collaborative acquisition of visual images, laser point clouds, and tactile perception data, the limitations of LiDAR's blind spots in observing the near-field ground and the interference of light on visual sensors are overcome. This also reserves core data sources for subsequent tactile closed-loop error correction, laying a data foundation for high-precision mapping.

[0025] Based on the visual image, semantic segmentation is performed to obtain the type of negative obstacle in the current working environment, multiple candidate regions of the negative obstacle, and the confidence and uncertainty corresponding to each candidate region.

[0026] Specifically, pixel-level semantic segmentation is used to accurately distinguish different types of negative obstacles such as potholes, ravines, and cliff edges from visual images, and their candidate region contours are extracted. Meanwhile, in this embodiment of the invention, the confidence and uncertainty of each candidate region can be calculated using softmax entropy or Monte Carlo dropout methods, optimizing the visual recognition results from qualitative labeling to quantitative characterization. This solves the problem of the lack of reliability measurement in traditional visual recognition results, provides a quantitative basis for calculating the risk intensity of ink smudging, and effectively reduces the interference of visual false detections on mapping accuracy.

[0027] Using a risk diffusion algorithm, ink smudging is performed based on the type of the negative obstacle and the confidence and uncertainty of each candidate region to generate a continuous risk field corresponding to each candidate region.

[0028] Specifically, a continuous risk field is generated through ink diffusion processing using a risk diffusion algorithm. This represents the core transformation from discrete candidate regions to a continuous risk distribution. The intuitive logic of ink diffusion is combined with the risk diffusion algorithm, using negative obstacle type, confidence level, and uncertainty as core inputs to determine the intensity and scope of risk diffusion. This approach overcomes the limitations of traditional discrete risk labeling, constructing a continuous field that conforms to the actual environmental risk propagation patterns. This allows risk assessment to cover the potential danger range surrounding the candidate region, providing a more comprehensive risk reference for sports planning.

[0029] The continuous risk field is divided to generate a continuous risk gradient distribution corresponding to each candidate region.

[0030] Specifically, the continuous risk field is divided to generate a continuous risk gradient distribution, transforming the abstract continuous risk field into directly usable hierarchical risk information. In particular, through gradient partitioning, clear risk level attributes are assigned to continuous risk values, achieving quantitative hierarchical risk levels. This quantitative hierarchical processing establishes a communication bridge between the perception layer and the decision-making layer, enabling the legged robot to quickly formulate differentiated movement strategies based on different risk gradients.

[0031] Terrain modeling is performed based on the laser point cloud to obtain geometric risk clues indicating the presence of negative obstacles in the current working environment.

[0032] Specifically, terrain modeling is performed on laser point clouds to obtain geometric risk cues, thereby mining the essential geometric features of the environment and supplementing the visual semantic recognition results. Specifically, spatial localization is performed using laser point clouds to construct a terrain model of the operational environment, and then geometric cues strongly correlated with negative obstacles, such as missing ground point clouds and abrupt height changes, are extracted. Since these geometric cues are generated entirely based on the physical structure of the environment, they are objective and stable, effectively verifying the rationality of the visual recognition results and providing a geometric basis for subsequent multi-source information fusion.

[0033] By using Bayesian fusion, the continuous risk gradient distribution is updated based on the geometric risk cues to obtain a multi-level grid map.

[0034] Specifically, a multi-level grid map is obtained by updating the continuous risk gradient distribution through Bayesian fusion. The core advantage of Bayesian fusion lies in its ability to combine the uncertainties of different data sources, achieving probabilistic integration of information. This method organically combines visually derived risk gradients with laser-derived geometric risk cues. During the fusion process, the weights of the two types of information are adaptively adjusted. The resulting multi-level grid map retains both semantic and geometric core information while achieving precise optimization of risk assessment, significantly improving map reliability. In a preferred embodiment of the invention, occupancy, risk, uncertainty, and freshness can be maintained in multiple layers using log-odds or probabilistic methods, and updated adaptively with weighted updates based on the uncertainty of each observation source.

[0035] Based on the tactile perception data, the multi-level grid map is corrected using tactile closed-loop technology to generate a risk cost map for the current working environment.

[0036] Specifically, tactile closed-loop correction is performed on a multi-level grid map based on tactile perception data. This tactile perception data, as directly correlated data between the legged robot and its environment, accurately reflects the actual terrain. By matching and verifying this data with the grid map, deviations in the map can be dynamically corrected. This invention effectively solves the potential misjudgment problems of visual and laser perception through a closed-loop mechanism, allowing the risk-cost map to adapt to the actual working environment in real time. The resulting risk-cost map thus possesses higher practicality and accuracy.

[0037] The negative obstacle risk map construction method and system of the present invention firstly acquires multimodal sensing data, including visual images, laser point clouds, and contact perception data, to comprehensively collect multi-dimensional information on appearance, topographic geometry, and physical contact in negative obstacle perception. This firstly solves the limitations of single sensors in negative obstacle observation, such as occlusion, reflection, and sparse sampling, by using multi-source data to lay a foundation for risk map construction and reduces construction errors caused by missing information at the data acquisition level. Next, semantic segmentation is performed on the visual image to obtain the negative obstacle type, candidate region, and corresponding confidence and uncertainty, thereby accurately identifying the category and location range of the negative obstacle. Simultaneously, the reliability of the visual recognition results is quantified through confidence and uncertainty, avoiding misjudgments caused by directly using indiscriminate visual recognition results for mapping. Subsequently, a risk diffusion algorithm is used to combine the confidence and uncertainty of the negative obstacle type and candidate region to generate a continuous risk field through ink diffusion and to delineate a continuous risk gradient distribution. This transforms the negative obstacle from a discrete binary label into a continuous risk representation with gradient decay from the core to the periphery, thus conforming to the actual distribution characteristics of the negative obstacle and solving the problem of overly conservative or overly risky traditional binary labeling, improving the accuracy of risk representation. Finally, terrain modeling is performed based on laser point clouds to obtain geometric risk cues indicating the presence of negative obstacles. Features such as elevation, slope, voids, or cliffs of negative obstacles can be extracted from the terrain geometry perspective, providing objective geometric evidence for the existence of negative obstacles and compensating for the shortcomings of visual recognition. The system overcomes the limitations of being susceptible to environmental lighting and material variations, enabling geometric verification of visual semantic risks. Subsequently, a multi-level grid map is generated by updating the continuous risk gradient distribution using a Bayesian fusion approach combined with geometric risk cues. This integrates continuous risks at the visual semantic level with risk cues at the laser geometric level, achieving complementarity and verification of multi-source information. This allows each level of the risk map to be updated with both visual and geometric information, reducing mapping errors caused by biases from a single information source. Finally, tactile closed-loop correction is performed on the multi-level grid map based on contact perception data, generating a risk cost map. The actual physical feedback from foot contact with the ground serves as the direct basis for risk map correction, dynamically verifying and correcting the risk map obtained from the fusion of visual and laser data. This eliminates observational biases caused by external parameter drift and environmental changes during long-term operations, allowing the risk map to continuously evolve with actual ground contact experience. This improves the accuracy of negative obstacle risk map construction, making the risk map more closely reflect the true risk distribution of negative obstacles in the actual working environment.

[0038] Optionally, acquiring multimodal sensing data of the legged robot in the current working environment includes: The legged robot uses its visual sensor, lidar sensor, and foot-end tactile sensing component to collect initial visual images, initial laser point clouds, and initial contact perception data of the current working environment, as well as the contact between the legged robot's foot and the ground, at preset frequencies corresponding to the visual sensor, lidar sensor, and foot-end tactile sensing component, respectively. The initial visual image, the initial laser point cloud, and the initial contact sensing data are time-synchronized and calibrated to obtain the visual image, the laser point cloud, and the contact sensing data of the current working environment.

[0039] Specifically, firstly, considering the differences in hardware characteristics between the visual sensor, LiDAR sensor, and foot-end tactile sensing component, initial visual images, initial laser point clouds, and initial contact perception data are acquired using preset frequencies matched to the working capabilities of each sensor. For example, the visual sensor acquires images at 10–30Hz, the LiDAR acquires point clouds at 5–20Hz, and the foot-end tactile sensing component acquires contact data at 200–1000Hz, achieving efficient raw acquisition of data from different modalities. Subsequently, a time synchronization calibration operation is performed. For those with hardware triggering conditions, hardware triggering or Precise Time Protocol (PTP) is used to achieve microsecond-level alignment. When there is no hardware synchronization, the motion state data of the legged robot, such as the correlation between IMU angular velocity and image / point cloud motion estimation, is used to estimate the time offset between the visual image and the laser point cloud through software algorithms. At the same time, the laser point cloud is interpolated to a unified reference time, and row-level motion compensation is performed on the initial visual image of the rolling shutter camera. Finally, the initial data with different timestamps and coordinate systems are mapped to the same spatiotemporal reference, resulting in visual images, laser point clouds, and contact perception data that can be directly used for subsequent fusion processing.

[0040] In a preferred embodiment of the invention, in addition to time synchronization, distortion correction and motion compensation can also be performed. Specifically, for rotating 3D LiDAR, the point cloud can be interpolated to a unified reference time. For rolling shutter cameras, row-level compensation can be performed using IMU angular velocity to reduce geometric distortion under high-speed motion. Extrinsic parameter calibration can also be performed; specifically, laser coordinate system L, camera coordinate system C, body coordinate system B, and world coordinate system W are defined. The extrinsic parameter between the camera and the laser is T. CL =[R CL |t CL For any laser point X L ∈R 3 Satisfying X C =R CL ×X L +t CL ; (X) L X is a three-dimensional point in the lidar coordinate system. CTo transform to a 3D point in the camera coordinate system, R CL A 3×3 rotation matrix is ​​used to describe the attitude relationship between the lidar and the camera, t CL (This is a 3×1 translation vector used to describe the positional offset between the LiDAR and the camera.) Pixel projection satisfies u ~ K·X C Where K is the camera intrinsic parameter matrix, ~ represents homogeneous equivalence, and u is the pixel coordinate on the image plane. In another preferred embodiment of the invention, when the robot operates for a long time, the extrinsic parameters may slowly drift due to vibration, collision, and temperature drift. The present invention can optimize the changes in extrinsic parameters between the camera and the laser within a short sliding window using an online calibration mechanism, aiming at semantic and geometric consistency. This maximizes the overlap between the projected point cloud ground boundary and the image ground / pit boundary; the projection residual is minimized under a robust kernel function. This optimization can be performed at low frequencies (e.g., 0.1-1Hz), and stability constraints are set to avoid false convergence.

[0041] In this embodiment of the invention, by acquiring data at differentiated preset frequencies, the image frame rate required for visual semantic recognition and the point cloud density required for laser terrain modeling are ensured, while also taking into account the high-frequency response requirements of foot contact perception. This avoids the waste of computing power or loss of key contact information caused by a single acquisition frequency. Furthermore, by using time synchronization calibration, the problem of spatiotemporal misalignment caused by different acquisition sequences and motion distortion of multimodal data is completely solved. This ensures that the negative obstacle candidate region of visual recognition, the geometric risk cues extracted by laser, and the foot contact feedback can accurately correspond to the same spatiotemporal location in the working environment. This provides high-quality and highly consistent basic data for subsequent core steps such as Bayesian fusion and tactile closed-loop correction, eliminating the problem of risk map misjudgment and accuracy reduction caused by data spatiotemporal mismatch from the source.

[0042] Optionally, the step of performing semantic segmentation based on the visual image to obtain the type of negative obstacle in the current working environment, multiple candidate regions of the negative obstacle, and the confidence and uncertainty corresponding to each candidate region includes: The visual image is classified at the pixel level using a lightweight semantic segmentation network to obtain the category mask of the visual image and the pixel confidence and pixel uncertainty corresponding to each pixel position in the visual image; Connectivity aggregation and contour extraction are performed on the category mask to obtain the initial candidate region of the negative obstacle and the type of the negative obstacle; Perform geometric consistency filtering on each of the initial candidate regions to obtain multiple candidate regions for the negative obstacle; The confidence level and uncertainty corresponding to each candidate region are obtained by statistically calculating the pixel-level confidence level and pixel uncertainty of each pixel position within each candidate region.

[0043] Specifically, this invention uses an encoder-decoder structure combined with a lightweight semantic segmentation network optimized by knowledge distillation to perform pixel-level classification reasoning on visual images to obtain category masks that distinguish different types of negative obstacles such as potholes, ravines, and the lower edge of steps from the background. At the same time, it outputs the pixel confidence and pixel uncertainty corresponding to each pixel position in the image, and normalizes the uncertainty to the [0,1] interval. Specifically, for pixel (i,j), let the network output the softmax probability of each category as p. k (i,j). Pixel confidence is represented as: c(i,j)=max k {p k (i,j)}; Where c(i,j) is the confidence level of pixel (i,j), and p k (i,j) represents the softmax probability that pixel (i,j) in the network output belongs to the k-th class. k { } represents taking the maximum value among all class probabilities.

[0044] Uncertainty can be expressed as entropy: u(i,j)= -Σ k [p k (i,j)log(p k [(i,j)+ε)]; Where u(i,j) is the uncertainty of pixel (i,j), Σ k This represents the summation over all categories k, where ε is the local minimum constant, such as ε=10. 8 , used to avoid p k The value of log(0) is abnormal when (i,j)=0.

[0045] To facilitate engineering implementation, the uncertainty can be normalized to [0,1] and used as a weight decay factor in subsequent fusion.

[0046] Next, a connected component aggregation operation is performed on the category mask to integrate connected pixels of the same type of negative obstacle into a whole region. Then, the initial candidate regions of negative obstacles are obtained through contour extraction. At the same time, the specific type of negative obstacle corresponding to each initial candidate region is determined based on the annotation information of the category mask. Subsequently, geometric consistency filtering is performed on each initial candidate region. The pixels of the initial candidate region are projected onto the elevation map constructed by the laser point cloud to verify whether the geometric features such as height abrupt changes, slope, and point cloud holes in the corresponding region match the inherent features of the type of negative obstacle. If there is a significant contradiction, the region is downweighted or eliminated to filter out effective negative obstacle candidate regions. Finally, the pixel confidence and pixel uncertainty corresponding to all pixel positions in each effective candidate region are calculated by mean statistical calculation to obtain the overall confidence and uncertainty of each candidate region, completing the quantization conversion from pixel-level features to region-level features.

[0047] In a preferred embodiment of the present invention, the encoder-decoder structure can be a MobileNet / ShuffleNet backbone and a lightweight decoder head, while a knowledge distillation strategy is introduced for model optimization. The backbone network replaces standard convolution with depthwise separable convolution and enhances the extraction capability of negative obstacle edge features by combining SE attention mechanism. The decoder head uses a pyramid pooling module to fuse features from different receptive fields to solve the problem of missed detection of small-sized pit targets. In terms of uncertainty calculation, this embodiment uses the Monte Carlo dropout method, embedding dropout layers in the convolutional layers of the encoder and decoder head of the network, setting the dropout probability to 0.15, and performing 10 consecutive forward inferences on the same visual image. The pixel uncertainty is obtained by calculating the variance of the class probability at each pixel position in the 10 inferences, and then normalizing it to the [0,1] interval using the sigmoid function. The pixel confidence is directly taken as the average probability of the class with the highest proportion in the 10 inferences. In the connected component aggregation stage, the 8-neighbor connected component analysis algorithm is used to eliminate tiny connected components with fewer than 20 pixels based on a preset pixel connectivity threshold (e.g., 0.8) to avoid noise interference. The Douglas-Puk algorithm is used to simplify the pixel contours of the initial candidate regions into polygons, while recording the bounding rectangle size and center pixel coordinates of each region. The pixel coordinates of the initial candidate region are then projected onto the lidar coordinate system using the camera intrinsic and extrinsic parameter matrices. A subset of the lidar point cloud corresponding to this region is sampled, and the average height change value and slope of the lidar point cloud subset are calculated. The height change threshold for potholes is set to 0.15m and the slope threshold to 60°, while the height change threshold for ravines is set to 0.2m and the slope threshold to 45°. If the geometric features of the projected region do not match the inherent features of the corresponding type of negative obstacle, i.e., the set threshold is not met, the confidence of the initial candidate region is reduced. If there is a significant contradiction between the geometric features and the negative obstacle assumption, the initial candidate region is directly eliminated. Finally, multiple effective negative obstacle candidate regions are selected, significantly reducing the probability of visual false detection. Finally, for each effective candidate region after selection, all pixel positions within it are traversed, and the pixel confidence and pixel uncertainty of each pixel position are calculated by mean statistical calculation to obtain the overall confidence and uncertainty of each candidate region. At the same time, the pixel set, center pixel coordinates, and scale features of each candidate region are output, realizing the accurate conversion from pixel-level features to region-level features.

[0048] In this embodiment of the invention, a lightweight semantic segmentation network is used to achieve pixel-level fine classification of negative obstacles. This ensures the accuracy of negative obstacle type recognition and location localization, and optimizes the deployment requirements of legged robots with limited onboard computing power through knowledge distillation and network pruning, effectively reducing inference latency and meeting real-time requirements. Scattered pixels are integrated into complete negative obstacle regions through connected component aggregation and contour extraction, avoiding region segmentation errors caused by misjudgment of a single pixel, accurately defining the initial range and type of negative obstacles. Geometric consistency filtering incorporates terrain geometric features from LiDAR for cross-validation, effectively avoiding visual image errors caused by lighting conditions, etc. The system significantly improves the effectiveness and authenticity of candidate regions by addressing false detection issues caused by environmental factors such as shadows and material reflections. Regional mean statistics are performed on pixel-level confidence and uncertainty, transforming pixel-level feature indicators into overall quantitative features of candidate regions. This provides a precise and unified basis for calculating risk intensity in the subsequent generation of a continuous risk field through ink smudging, ensuring that risk representation aligns with the actual recognition reliability of negative obstacle regions. Furthermore, the output candidate region center and scale features provide a foundation for configuring the core location and range of subsequent risk diffusion. This approach enhances the accuracy, rationality, and robustness of subsequent negative obstacle risk map construction from the source of visual semantic feature extraction.

[0049] Optionally, the step of generating a continuous risk field corresponding to each candidate region by performing ink blurring processing based on the type of the negative obstacle and the confidence and uncertainty of each candidate region using a risk diffusion algorithm includes: Determine the category risk coefficient for the type of negative obstacle; Based on the category risk system, and combining the confidence level and uncertainty of each candidate region, the risk intensity coefficient of the candidate region is obtained; Feature extraction is performed on the candidate region to obtain the center position, main morphological direction, and scale features of the candidate region; Based on the center location, the main orientation of the shape, and the scale characteristics, combined with the terrain aspect and slope characteristics of the laser point cloud, a covariance matrix is ​​generated. Ink smudging is performed based on the risk intensity coefficient and the covariance matrix to generate the continuous risk field corresponding to each candidate region.

[0050] Specifically, firstly, based on the inherent hazard levels of different negative obstacles such as pits, gullies, cliff edges, and the lower edges of steps, differentiated category risk coefficients are assigned to each category. High-risk negative obstacles such as cliff edges and deep gullies are assigned high category risk coefficients, while low-risk negative obstacles such as shallow pits and the lower edges of gentle slope steps are assigned low category risk coefficients. The coefficient values ​​are normalized to the [0,1] interval. Then, combined with these category risk coefficients, the risk intensity coefficient for each candidate area is calculated using the following formula: α m =k cls ×c m ×(1-u m ); Where, k cls c is the category risk coefficient. m u is the candidate region confidence score for the m-th negative obstacle candidate region. m Let α be the candidate region uncertainty for the m-th negative obstacle candidate region. m The risk intensity coefficient of the m-th negative obstacle candidate region is used to strongly correlate the risk intensity with the danger level and recognition reliability of the negative obstacle. Next, feature extraction is performed on each candidate region. Based on the candidate region contour and center pixel coordinates obtained from visual semantic segmentation, combined with the ground projection results of the laser point cloud elevation map, the center position of the candidate region in the robot's local coordinate system is accurately determined. The principal axis analysis of the contour is used to obtain the principal direction of the shape, and the length, width, and equivalent radius of the candidate region are extracted. Then, the covariance matrix of the Gaussian diffusion kernel is initialized based on the center position, principal direction of the shape, and scale features of the candidate region. The terrain aspect and slope features obtained from the laser point cloud terrain modeling are incorporated into the matrix optimization, so that the principal direction of the covariance matrix matches the terrain drop direction and the diffusion range matches the terrain slope. For long, narrow negative obstacles such as gullies, an elliptical covariance matrix with a larger principal direction major axis is set; for approximately circular negative obstacles such as pits, an isotropic covariance matrix is ​​set. Finally, using the center position of the candidate region as the kernel, an anisotropic Gaussian smudging formula is used for ink smudging. R m (x)=α m ×exp(-0.5×(x-μ m ) (x-μ m )); Where, α m Let μ be the risk intensity coefficient of the m-th negative obstacle candidate region, and let μ be the peak baseline of the risk field. m Let m be the center coordinate vector of the m-th negative obstacle candidate region. Covariance matrix The inverse matrix is ​​used to calculate the Mahalanobis distance to achieve anisotropic diffusion, where x is the spatial position vector and R is the inverse matrix.m (x) represents the risk value of the m-th negative obstacle candidate region at its spatial location vector x, exp( The natural exponential function (F) achieves a smooth decay of risk values ​​from the core region to the periphery. The above formula completes the modeling from discrete candidate regions to a continuous risk distribution, generating a continuous risk field corresponding to each candidate region. If multiple candidate regions exist, the risk fields are superimposed, and the risk values ​​are limited to the [0,1] interval using the clip function or Noisy-OR method to avoid risk value oversaturation. Combined with... Figure 2 As shown, the risk field at the center of the gully μ1 spreads in a long, elliptical shape along the main direction (white arrow), with a narrower diffusion range in the secondary direction. This reflects the morphological characteristics of gully-type negative obstacles and the constraint of terrain slope on risk propagation. In contrast, the risk field at the center of the pit μ2 spreads in an approximately circular, isotropic manner, matching the spatial morphology of the pit. The core peak value of the risk field is determined by the risk intensity coefficient α, and the diffusion range and direction are smoothly transformed from discrete candidate regions to continuous risk distribution by the anisotropic Gaussian shading formula mentioned above. The color bar on the right quantifies the gradual change of risk value from 0 to 0.8, showing the law of risk gradually decreasing from the center of the candidate region to the edge.

[0051] In a preferred embodiment of the present invention, a continuous risk field is generated for the effective candidate regions of two types of negative obstacles, gullies and potholes, obtained from visual semantic segmentation. The robot's local grid map resolution is 0.05m, the terrain is a sloping outdoor environment, the confidence level of the gully candidate region is 0.92 and the uncertainty is 0.05, and the confidence level of the pothole candidate region is 0.89 and the uncertainty is 0.07. First, according to the category risk coefficient standard preset in the document, a category risk coefficient of 0.9 is configured for the gully and a category risk coefficient of 0.75 is configured for the pothole, both of which are normalized to the interval [0,1]. Then, the risk field is generated using formula α. m =k cls ×c m ×(1-u m The risk intensity coefficients are calculated, where the risk intensity coefficient for gullies is 0.9 × 0.92 × (1 - 0.05) = 0.7866, and the risk intensity coefficient for pits is 0.75 × 0.89 × (1 - 0.07) = 0.6143, thus quantifying the risk intensity. Next, feature extraction is performed on the two candidate regions, and the center pixels obtained from visual segmentation are processed by the extrinsic parameter T. CLProjecting onto the lidar coordinate system, the center position of the candidate gully region is obtained as (2.5m, 1.2m). Through contour principal axis analysis, its main morphological direction is found to be 30° along the slope direction, with a scale feature of 8m in length and 0.8m in width. The center position of the candidate pit region is (4.1m, 0.5m), with a main morphological direction of isotropic and no obvious orientation, and a scale feature of equivalent radius of 0.6m. Combining the terrain aspect (downhill along the 30° direction, slope of 15°) and slope feature obtained from lidar point cloud terrain modeling, a covariance matrix is ​​generated. The covariance matrix of the gully is set with a major axis of 2.0m and a minor axis of 0.3m according to the matching of the main morphological direction and the terrain aspect. The covariance matrix of the pit is set according to isotropic... Both the horizontal and vertical axes are set to 0.8m, and the diffusion range in the downhill direction is appropriately expanded based on the terrain slope. Finally, using the center position of the two candidate areas as the core, the anisotropic Gaussian diffusion formula is applied for ink diffusion. The gully area undergoes elliptical risk diffusion along the main direction of 30°, while the pit area undergoes circular risk diffusion. The risk intensity coefficient of 0.7866 for the gully determines that its core risk value is higher, while the risk intensity coefficient of 0.6143 for the pit indicates a relatively lower core risk value. Furthermore, the risk values ​​of both types of areas decrease smoothly from the center to the edge. Finally, continuous risk fields corresponding to the candidate areas of gullies and pits are generated respectively. The two risk fields are superimposed in the overlapping area using the Noisy-OR method, where the superposition formula is: R=1-Π m (1-R m ); Where R is the total risk value after aggregation, Π m For all m negative obstacle candidate regions, the corresponding (1 R m ) terms are multiplied consecutively, R m This represents the risk value of the m-th negative obstacle candidate region. After superposition, the risk value remains confined to the [0,1] interval, thus completing the generation of the continuous risk field for this negative obstacle candidate region.

[0052] In this embodiment of the invention, by configuring differentiated category risk coefficients for different types of negative obstacles, the generation of the risk field is made to match the actual danger level of the negative obstacles, avoiding planning misjudgments caused by indiscriminate risk labeling. The category risk coefficient is combined with the confidence and uncertainty of the candidate area to calculate the risk intensity coefficient, so that the core strength of the risk field is strongly bound to the reliability of the identification result. The more accurate the identification and the higher the risk level of the area, the greater the risk intensity, thus improving the rationality of risk expression. The center position, main morphological direction and scale features of the candidate area are extracted and combined with the terrain aspect and slope to optimize the covariance matrix, so that the diffusion direction and range of the ink stain match both the actual shape of the negative obstacle and the terrain features of the working environment. This approach achieves anisotropic risk diffusion, which better aligns with the risk propagation patterns of negative obstacles in unstructured environments. The Gaussian-based ink smudging model transforms discrete candidate regions into a continuous risk field that smoothly decays from the core to the edge, abandoning the traditional binary risk labeling model and naturally forming a safety buffer zone. This solves the problem of traditional methods being either overly conservative or overly risky. Furthermore, this risk diffusion algorithm performs smudging calculations only within a local window of the candidate region, resulting in low computational complexity. It is compatible with the onboard computing power constraints of legged robots, meets real-time requirements, and provides a precise and continuous risk representation carrier for subsequent risk gradient distribution generation. This improves the accuracy and practical adaptability of the negative obstacle risk map construction from the perspective of risk expression.

[0053] Specifically, the step of dividing the continuous risk field to generate a continuous risk gradient distribution corresponding to each candidate region includes: Based on a preset multi-level risk classification threshold, the continuous risk field is divided into intervals to obtain multiple risk zones corresponding to the candidate region. The risk zones include fatal zones, buffer zones, and safe zones. Based on the risk distribution characteristics of each risk zone, a continuous risk gradient distribution with gradient decay from the lethal zone to the safe zone is generated.

[0054] Specifically, firstly, based on the legged robot's motion state parameters (current walking speed, load status), the type of negative obstacle, and its inherent hazard level, and referring to the terrain slope features obtained from laser point cloud modeling, a multi-level risk classification threshold is preset to adapt to the operational scenario. Among these, the fatal risk threshold θ D The value range is 0.75-0.9, and the buffer risk threshold θ B The value range is 0.3-0.5, and the threshold can be dynamically adjusted according to the working scenario. For example, when the robot is moving at high speed, performing heavy load operations, or when the terrain slope is large, θ can be appropriately increased. B To expand the buffer zone, the threshold is appropriately reduced when passing through at low speeds with precision, and then a preset θ is used. D and θ BTo define the boundaries, the risk values ​​of a continuous risk field are strictly divided into intervals, with risk values ​​≥ θ being defined as... D The area was designated as a lethal zone, a core danger zone containing negative obstacles. Legged robots entering this zone are prone to accidents such as missteps and tipping over. The θ B ≤ Risk value < θ D The area is designated as a buffer zone, a safety buffer zone for negative obstacles, with a risk value <θ. B The area is designated as a safe zone, a region where robots can safely pass, thus completing the hierarchical division of the risk field. Next, the risk distribution characteristics of each risk zone are extracted, including the core risk peak of the lethal zone, the risk decay rate of the buffer zone, and the basic risk value of the safe zone. Combining the natural risk decay law formed by ink smudging, the risk values ​​of each risk zone are smoothly connected to eliminate the problem of abrupt risk value changes caused by interval division. According to the rule that "the core risk value of the lethal zone is the highest, gradually decays linearly or non-linearly towards the buffer zone, and then continues to decay towards the safe zone to the basic risk value", a continuous risk gradient distribution with continuous and smooth gradient decay from the lethal zone to the safe zone is generated. At the same time, the gradient distribution is matched with the shape of the negative obstacle and the slope of the terrain. For example, the gradient distribution of long strip-shaped negative obstacles such as ravines extends along the main direction of the shape, and the gradient distribution in the sloping environment appropriately expands the decay range along the fall direction to ensure that the gradient distribution conforms to the actual hazard propagation law.

[0055] In this embodiment of the invention, the continuous risk field is divided into three levels: a lethal zone, a buffer zone, and a safe zone. This allows for a clear hierarchical expression of the risk level of negative obstacles, providing a clear risk level reference for robot planning and control, and facilitating subsequent linkage with control strategies. By extracting the distribution features of each risk zone and performing smooth connection processing, a continuous risk gradient distribution with gradient decay from the lethal zone to the safe zone is generated. This not only preserves the continuous characteristics of the ink-smeared risk field, but also makes the risk distribution more interpretable through hierarchical division. It eliminates the defects of traditional binary annotation, and the naturally formed buffer zone allows the robot to automatically maintain a safe distance during planning, avoiding the safety risks caused by walking close to the edge.

[0056] Optionally, the step of performing terrain modeling based on the laser point cloud to obtain geometric risk cues indicating the presence of the negative obstacle in the current operating environment includes: Voxel downsampling and ground segmentation are performed on the laser point cloud to obtain the ground point cloud in the laser point cloud; Based on the ground point cloud, according to the preset grid parameters, an elevation map of the current working environment is constructed, and the slope and height change values ​​of each grid in the elevation map are obtained; Based on the slope and height abrupt change value of each grid in the elevation map, the cavity features in the elevation map, the cliff edge features of the ground point cloud, and the distance jump features of the laser point cloud are obtained. The slope, the height abrupt change value, the cavity feature, the cliff edge feature, and the distance jump feature are used as the geometric risk clues.

[0057] Specifically, the preset grid parameters include grid resolution. First, voxel downsampling is performed on the original laser point cloud, setting the voxel resolution to 0.02~0.10m to remove redundant point cloud data and reduce computational complexity. Then, a ground segmentation algorithm based on height threshold combined with local plane fitting is adopted, while IMU attitude compensation is introduced to rotate the point cloud to a gravity-aligned coordinate system, effectively separating the ground point cloud from the non-ground point cloud and solving the ground segmentation deviation problem in sloping environments. Next, preset grid parameters are set according to the operational requirements of the legged robot (grid resolution 0.02~0.10m, establishing a local grid centered on the robot). Based on the accumulated ground point height of each grid i, the average height h of each grid i is calculated. i σ 2 {h,i}, highest / lowest altitude, number of ground points n i By using statistical measures, a dense elevation map of the current working environment is constructed. Then, the slope of each grid is obtained by gradient calculation of the elevation map. The height change value is calculated by the height difference between adjacent grids to accurately represent the undulation and abrupt change characteristics of the terrain. Specifically, the formula for calculating slope is: s i =|| h i ||; Among them, s i Let be the slope of the i-th grid cell. h i This represents the gradient of the elevation map at grid i. It also detects abrupt changes in elevation Δh. i The edge region of a cliff. The edge of a negative obstacle often corresponds to Δh. i Significantly large and the number of ground points n i The area that suddenly drops.

[0058] Subsequently, using the slope and height abrupt change values ​​of each grid as the core judgment criteria, areas where the number of ground points within a grid suddenly drops, the height abrupt change value exceeds the threshold, and the slope increases significantly are identified as void features in the elevation map. Edge areas where the height abrupt change value and slope change in a stepwise manner between continuous grids are extracted as cliff edge features of the ground point cloud. At the same time, in the distance image of the laser point cloud, the features of areas where the distance value shows a continuous break or jump are extracted as distance jump features of the laser point cloud. The extraction of these three types of features are strongly bound to the geometric characteristics of negative obstacles. Void features correspond to pit-type negative obstacles, while cliff edge and distance jump features correspond to ravine and cliff edge-type negative obstacles. Finally, the directly calculated slope and height abrupt change values ​​are integrated with the void features, cliff edge features, and distance jump features obtained through feature recognition. Among them, the slope and height abrupt change values ​​are quantitative geometric indicators, while void, cliff edge, and distance jump features are geometric features that combine qualitative and quantitative methods. Together, they serve as geometric risk clues indicating the existence of negative obstacles, providing objective geometric basis for subsequent multi-source fusion.

[0059] In a preferred embodiment of the present invention, terrain modeling is performed based on laser point clouds collected by a 32-line 3D LiDAR in an environment with interspersed gravel roads, ditches, and potholes. Geometric risk cues are extracted. The preset grid resolution is 0.05m, the voxel downsampling resolution is 0.05m, the height abrupt change threshold is set to 0.15m, and the slope threshold is set to 45°. First, voxel downsampling is performed on the original laser point cloud to remove redundant point clouds and retain core terrain features. Then, a ground segmentation algorithm based on height threshold and local RANSAC plane fitting is used, combined with IMU gravity alignment attitude compensation, to rotate the point cloud to a gravity coordinate system, separating the ground point cloud from the non-ground point cloud, and removing non-ground interference points such as gravel and weeds. Next, a local grid is established with the robot as the center. Based on the ground point cloud, the ground point height is accumulated for each 0.05m resolution grid, the average height of each grid is calculated, and an elevation map of the field operation environment is constructed. Then, the slope of each grid is calculated through the height gradient of the neighboring grids, and the height abrupt change value is calculated through the difference of the average height of adjacent grids, accurately obtaining the slope. The topographic indicators for each grid cell were quantified. Then, based on the slope and height abrupt changes in each grid cell, areas with a sudden drop in the number of ground points, a height abrupt change value >0.15m, and a slope >45° were identified as cavities in the elevation map. Continuous edge areas where the height abrupt changes abruptly on both sides of the ditch, and the slope abruptly increases from 10° to 50°, were extracted as cliff edge features. Simultaneously, in the laser point cloud distance image, continuous break areas where the distance value of the ditch location suddenly jumps from 2.0m to 5.0m were extracted as distance jump features. Finally, the slope and height abrupt change values ​​of each grid were used as two quantitative indicators to integrate with the three types of geometric features identified: cavity features, cliff edge features, and distance jump features. Among them, the cavity feature was labeled as a geometric clue of negative obstacles, and the cliff edge and distance jump features were labeled as geometric clues of negative obstacles. All features corresponded one-to-one with the elevation map grid, and finally a complete geometric risk clue indicating the existence of negative obstacles in the current field operation environment was obtained, which provided grid-level objective geometric basis for subsequent Bayesian fusion.

[0060] In this embodiment of the invention, voxel downsampling and ground segmentation with IMU attitude compensation reduce the computational complexity of point cloud processing, adapt to the onboard computing power of legged robots, and improve the accuracy of ground point cloud extraction in complex sloping environments, laying a high-quality data foundation for subsequent terrain modeling. Based on preset grid parameters, an elevation map is constructed and the slope and height abrupt change values ​​of each grid are calculated, achieving grid-level quantitative representation of terrain features. This allows geometric risk cues to be integrated with visual semantic risks and multi-level grid maps in the same grid coordinate system, ensuring the spatiotemporal consistency of multi-source information. Based on slope and height abrupt change values, three types of features—holes, cliff edges, and distance jumps—are extracted, accurately capturing the core geometric attributes of different types of negative obstacles. This achieves targeted geometric representation of negative obstacles, compensating for the shortcomings of visual recognition which is susceptible to interference from lighting, shadows, and materials. The integrated geometric risk cues possess both quantitative and feature indicators, enabling objective geometric verification of candidate negative obstacle regions identified by visual recognition and providing independent geometric detection basis for visually missed negative obstacles, thus achieving complementarity with visual semantic information.

[0061] Optionally, the step of updating the continuous risk gradient distribution based on the geometric risk cues using a Bayesian fusion method to obtain a multi-level grid map includes: Based on the slope, the height abrupt change value, the cavity feature, the cliff edge feature, and the distance jump feature, the geometric risk distribution value of each grid in the elevation map is generated, and the laser observation uncertainty corresponding to the geometric risk distribution value is obtained; Based on the continuous risk gradient distribution, the visual risk distribution value is obtained, and the visual observation uncertainty corresponding to the visual risk distribution value is acquired. The visual risk distribution value and the geometric risk distribution value are weighted and fused to obtain risk layer data; The occupancy observation data corresponding to the laser point cloud are accumulated to obtain occupancy layer data; Based on the visual observation uncertainty and the laser observation uncertainty, uncertainty layer data is generated, and freshness layer data is generated through a time decay model. The risk layer data, the occupation layer data, the uncertainty layer data, and the freshness layer data are used as levels for raster index fusion to obtain the multi-level raster map.

[0062] Specifically, firstly, quantization weights are set based on the slope and height abrupt changes of each grid cell in the elevation map. Combining the existence probability of voids, cliff edges, and distance jumps, a laser geometric risk distribution value for each grid cell is generated through numerical mapping. Simultaneously, the laser observation uncertainty corresponding to each grid cell is calculated based on the sparsity of the laser point cloud and the pixel variance constructed from the elevation map, and both are normalized to the [0,1] interval. Next, the risk values ​​of each grid cell in the continuous risk gradient distribution are extracted to form a visual risk distribution value in the same grid coordinate system as the elevation map. Simultaneously, the visual observation uncertainty obtained during the visual semantic segmentation stage is retrieved to complete the grid-level quantization representation of visual risk. Then, a weighted fusion under a Bayesian framework is performed, calculating adaptive weights based on the observation uncertainties of both visual and laser data. Lower uncertainties correspond to higher weights. The visual risk distribution value and the geometric risk distribution value are fused and updated using an exponential moving average formula. After limiting the risk value range with a saturation function, the risk layer data is obtained. The risk layer is driven by two parts: visual ink blot risk R. V With geometric void risk R L For grid i, the expression is: R i ←sat( (1-γ)·R i +γ·( · + · ) ); Among them, R i Let γ be the risk value of grid i after fusion, γ∈[0,1] be the fusion step size, and sat(·) be the saturation function restricted to [0,1]. The visual risk weight for grid i, Let i be the laser geometric risk weight. Let i be the visual risk distribution value of grid i. Let be the laser geometric risk distribution value for grid i.

[0063] Then, within the log-odds framework, the occupancy observation data of the laser point cloud is cumulatively updated, the occupancy log-odds value of each grid is calculated and converted into occupancy probability, generating occupancy layer data, where the formula is: l i =log(P occ (i) / (1-P occ (i))); Among them, l i Let P be the occupancy probability of the i-th grid cell. occ (i) represents the occupancy probability of the i-th grid cell.

[0064] When radar or depth camera provides occupancy observation Updated to: li ←clip(l i + ×Δ , l min , l max ); Among them, l i Let Δ be the odds of the i-th grid cell being occupied. 为 The log-probability increment of sensor observations for the i-th grid, given by the sensor inverse model. For the visual risk weight of grid i, l min , l max These are the lower and upper bounds of the logarithmic probability, respectively. Finally, the logarithmic probability is converted to the probability of occupancy using the following formula: P occ (i)=1 / (1+exp(-l i )).

[0065] Simultaneously, the observation uncertainties from both visual and laser observations are integrated to obtain the comprehensive uncertainty of each grid, constructing uncertainty layer data. The uncertainty can be obtained by weighting multiple sources of uncertainty and used to plan the cost (the more uncertain, the more conservative). Based on a time decay model, the freshness value of each grid is calculated, generating freshness layer data. Freshness decreases as the time since the last observation increases. The expression for the time decay model is: F i (t)=exp(-(tt last (i)) / τ F ); Among them, F i (t) represents the information freshness value of the i-th grid at time t, where t last (i) represents the last time the i-th grid was effectively observed by the LiDAR / vision camera / tactile sensor, t represents the current time, and τ represents the current time. F This represents the freshness decay constant. Finally, the data from the risk layer, occupation layer, uncertainty layer, and freshness layer are aligned based on the same raster index, and the quantitative indicators of each layer are bound to the corresponding raster cells. This achieves the integrated fusion of multi-layer data, resulting in a multi-level raster map containing multi-dimensional environmental information. The data of each layer is stored independently and can be retrieved as needed.

[0066] In this embodiment of the invention, Bayesian fusion is used as the core. The risk gradient at the visual semantic level and the risk clue at the laser geometric level are fused in a grid-level weighted manner. By adaptively allocating weights based on observation uncertainty, visual and laser information can be mutually verified and complementary, effectively eliminating the problems of false detection and missed detection caused by illumination interference and sparse point cloud due to a single information source. This significantly improves the accuracy and robustness of the risk layer data.

[0067] Optionally, the step of performing tactile closed-loop correction on the multi-level grid map based on the tactile sensing data to generate a risk cost map for the current working environment includes: Based on the contact sensing data, the actual landing point of the foot of the legged robot and the corresponding ground contact state of the foot are determined. Based on the ground contact state and the actual landing point, the multi-level grid map is subjected to tactile closed-loop correction to obtain the tactile closed-loop corrected multi-level grid map. Based on the multi-level grid map corrected by the tactile closed loop, hierarchical fusion is performed, and the risk cost map in the current working environment is generated by combining the layer weight coefficients corresponding to each level.

[0068] Specifically, based on the contact perception data collected by the foot tactile sensing component, the grid coordinates of the actual landing point of the foot are extracted. At the same time, combined with indicators such as peak contact force, slip rate, drop impact value and contact duration, the ground contact state is divided according to the above indicators, and the multi-level grid map is corrected according to the ground contact state and the actual landing point position. Specifically, for different ground contact states, differentiated correction strategies are adopted for the grid corresponding to the actual landing point and its neighboring grids. The risk value, uncertainty value, and freshness value of the corresponding grid in the map are adjusted synchronously. In some ground contact states, the map's exclusive feature layer is also updated in a targeted manner and linked with the robot's motion control strategy to complete the accurate and adaptive correction of the data at each level of the multi-level grid map. Finally, the multi-level grid map after tactile closed-loop correction is subjected to hierarchical fusion processing. First, based on the actual operation scenario factors such as the type of task, walking speed, and terrain complexity of the legged robot, adaptive layer weight coefficients are configured for each level, such as the risk layer, occupation layer, uncertainty layer, and freshness layer. Then, through a preset cost function, the grid quantification index of each level is weighted and calculated with the corresponding layer weight coefficients. A unique comprehensive cost value is obtained for each grid. Finally, a risk cost map with grid as the basic unit and full coverage of the operation environment is generated and provided to the global / local planner, foot landing point planner, or MPC controller. Closed-loop operation is achieved through safety constraints.

[0069] In a preferred embodiment of the present invention, gait planning, landing point planning, and MPC constraints can be performed based on the obtained risk cost map.

[0070] Gait planning is divided into global planning and local planning. For global planning, at a larger scale, this embodiment can use search algorithms such as A* / D* / Hybrid-A* to plan the path on the cost map. Since the risk is continuous, the search can shorten the path as much as possible while ensuring a safe distance, avoiding the over-conservatism caused by traditional inflated barriers. For local planning, this embodiment can use optimization-based (TEB, MPC) or sampling-based (DWA) methods to delineate the torso trajectory and the desired landing area, and the landing planner selects the specific landing point on the risk map.

[0071] For landing point planning, we define candidate landing points f and their landing costs J(f), expressed as follows: J(f)=w1·R(f)+w2·S(f)+w3·U(f)+w4·D(f); Where J(f) is the comprehensive cost of candidate landing point f, R(f) is the negative obstacle risk cost of candidate landing point f, S(f) is the kinematic / stability cost of candidate landing point f, U(f) is the information uncertainty cost of candidate landing point f, w1, w2, w3, and w4 are the weight coefficients of each cost dimension, and D(f) is the distance cost term of candidate landing point f. Landing point planning can evaluate multiple candidate landing points in each cycle and select the landing point with the minimum cost that satisfies the kinematic constraints.

[0072] For MPC constraints, in model predictive control, the risk constraint can be written as: R(x t ) ≤θ safe (R(x) t Let x be the robot state at time t. t The corresponding negative obstacle risk value, θ safe (as a safety risk threshold), or by adding a risk term to the objective function: J risk = R(x t (N is the time domain length of MPC prediction).

[0073] For the landing point decision variable, R(f) can be added. k )≤θ step Hard constraints (R(f) k w represents the candidate landing point f in the k-th step. k The corresponding negative obstacle risk value, θ step (This is the safety risk threshold for a single-step landing point). If the controller cannot find a feasible solution within the constraints, it triggers a replanning or shutdown at the upper level.

[0074] In this embodiment of the invention, by analyzing multiple indicators of contact perception data, the foot contact state and actual landing point are accurately determined. The physical interaction feedback between the robot and the ground is transformed into a direct basis for map correction, effectively solving the problem of map error accumulation caused by external parameter drift, changes in ambient lighting, and sparse point clouds during long-term operations. At the same time, differentiated grid-level correction strategies are implemented for different contact states. The dynamic erasure mechanism for safe contact can eliminate the risk of false detection caused by the fusion of vision and laser, the risk deepening mechanism for abnormal contact can make up for the missed detection problem of vision and laser, and the separate handling of friction anomalies avoids overly conservative planning caused by misjudging low-friction ground as negative obstacles, which greatly improves the accuracy and robustness of multi-level grid maps.

[0075] Optionally, the ground contact state includes safe ground contact and abnormal ground contact. The step of performing tactile closed-loop correction on the multi-level grid map based on the ground contact state and the actual landing point location to obtain a tactile closed-loop corrected multi-level grid map includes: If the ground contact state is the safe ground contact, then the grid corresponding to the actual landing point and the neighboring grid in the multi-level grid map are subjected to risk dynamic erasure processing to obtain the multi-level grid map after tactile closed-loop correction. If the ground contact state is an abnormal ground contact, then the grid corresponding to the actual landing point and the neighboring grids in the multi-level grid map are subjected to risk deepening processing to obtain the multi-level grid map after tactile closed-loop correction.

[0076] Specifically, firstly, based on foot contact sensing data, the accurate extraction of the actual landing point grid coordinates and the binary determination of the ground contact state are completed, clearly defining the landing point grid and its neighborhood range corresponding to safe and abnormal ground contact. The selection radius of the neighborhood grid can be adaptively adjusted according to the type of negative obstacle and the complexity of the terrain. If the landing is determined to be safe, risk dynamic erasure processing is performed on the grid corresponding to the actual landing point and its neighborhood grids, the expression of which is: R i ←(1-β)·R i Among them, R i Let β be the risk value of grid i after fusion, and β be the erasure rate. Specifically, the risk value of the corresponding grid needs to be reduced according to the preset erasure rate. At the same time, the uncertainty layer and freshness layer of the multi-level grid map are updated simultaneously to reduce the observation uncertainty of the corresponding grid and increase the freshness value. Moreover, the erasure rate can be gradually increased with the number of times the foot touches the area for repeated verification, so that the degree of risk erasure matches the credibility of the actual ground touch verification.

[0077] If an abnormal ground contact is identified, risk deepening processing is applied to the grid cell corresponding to the actual ground contact location and its neighboring grid cells. The expression is: R i ← min(1,R i +β plus); where R i β is the fused risk value of grid i. plus The risk amplification rate is defined as follows: Specifically, the risk value of the corresponding grid needs to be increased according to the preset risk amplification rate. After the risk value is increased, it should not exceed the quantization range of [0,1]. At the same time, the update radius of the neighboring grid is appropriately expanded to match the spatial correlation characteristics of the negative obstacle danger area. The uncertainty layer of the multi-level grid map is updated synchronously. The risk amplification rate can be adjusted in gradient according to the deviation of the abnormal ground contact indicator. The greater the deviation of the indicator from the threshold, the greater the amplification. Through the above two targeted processing methods, the risk layer, uncertainty layer, and freshness layer of the multi-level grid map are accurately corrected, and finally, the multi-level grid map after tactile closed-loop correction is obtained.

[0078] In a preferred embodiment of the present invention, a tactile closed-loop correction is performed on the constructed 0.05m resolution multi-level grid map to generate a risk cost map. The robot's current travel speed is 0.6m / s, and it performs an inspection task of ditches and pits in the field. The erasure rate β is set to 0.15 (usually 0.05-0.2), and the risk amplification rate β is... plus Set to 0.2 (usually 0.1-0.4), the drop impact threshold is set to 50N, the slip rate threshold is set to 15%, and the preset layer weight coefficient is: risk layer weight coefficient λ. risk =0.4, Occupying layer weight coefficient λ occ =0.2, Uncertainty layer weight coefficient λ unc =0.2, Freshness layer weight coefficient λ fresh =0.2, the cost function expression is: C i =λ risk R(i)+λ occ P occ (i)+λ unc U(i)+λ fresh (1-F(i)); Among them, C i Let R(i) be the total risk value of grid i, and P be the risk value of grid i. occ (i) represents the occupancy probability of grid i, U(i) represents the uncertainty of grid i, and F(i) represents the freshness value of grid i.

[0079] In this embodiment of the invention, the ground contact state is divided into safe ground contact and abnormal ground contact, and a differentiated grid-level correction strategy is executed. This strongly binds the correction of multi-level grid maps with the actual ground contact feedback of the foot, transforming the robot's physical interaction experience into a direct basis for map correction. This effectively solves the problems of false detection and missed detection caused by illumination interference, sparse point clouds, and extrinsic parameter drift in vision and laser fusion mapping, and realizes the self-correction and dynamic evolution of the map. Among them, the dynamic erasure mechanism for safe ground contact can gradually eliminate false risk markings in the map, reduce the uncertainty of the corresponding area and improve freshness, making the map more consistent with the real safety status of the actual working environment and avoiding the reduction in passage efficiency caused by overly conservative planning. The risk deepening mechanism for abnormal ground contact can promptly strengthen the risk areas that are missed or underscaled in the map, and expanding the neighborhood update radius can further improve the safety buffer zone, effectively avoiding the safety risks of the robot's subsequent landing points. At the same time, the gradient deepening rate makes the risk adjustment more consistent with the actual severity of abnormal ground contact.

[0080] Combination Figure 3 As shown, a negative obstacle risk map construction system of the present invention includes: The data acquisition unit is used to acquire multimodal sensing data of the legged robot in the current working environment. The multimodal sensing data includes visual images of the current working environment, laser point clouds, and contact perception data of the legged robot. A semantic segmentation unit is used to perform semantic segmentation based on the visual image to obtain the type of negative obstacle in the current working environment, multiple candidate regions of the negative obstacle, and the confidence and uncertainty corresponding to each candidate region. The ink smudging unit is used to perform ink smudging processing based on the type of the negative obstacle and the confidence and uncertainty of each candidate region through a risk diffusion algorithm, and generate a continuous risk field corresponding to each candidate region. A partitioning unit is used to partition the continuous risk field and generate a continuous risk gradient distribution corresponding to each candidate region; The modeling unit is used to perform terrain modeling based on the laser point cloud to obtain geometric risk clues indicating the existence of the negative obstacle in the current working environment. A Bayesian fusion unit is used to update the continuous risk gradient distribution based on the geometric risk cues using a Bayesian fusion method to obtain a multi-level grid map. The map generation unit is used to perform tactile closed-loop correction on the multi-level raster map based on the tactile perception data, and generate a risk cost map for the current working environment.

[0081] The negative obstacle risk map construction system of the present invention has the same advantages over the prior art as the negative obstacle risk map construction method described above, and will not be repeated here.

[0082] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for constructing a negative obstacle risk map, characterized in that, include: Acquire multimodal sensing data of the legged robot in the current working environment. The multimodal sensing data includes visual images of the current working environment, laser point clouds, and contact perception data of the legged robot. Based on the visual image, semantic segmentation is performed to obtain the type of negative obstacle in the current working environment, multiple candidate regions of the negative obstacle, and the confidence and uncertainty corresponding to each candidate region; Using a risk diffusion algorithm, ink smudging is performed based on the type of negative obstacle and the confidence and uncertainty of each candidate region to generate a continuous risk field corresponding to each candidate region. The continuous risk field is divided to generate a continuous risk gradient distribution corresponding to each candidate region; Terrain modeling is performed based on the laser point cloud to obtain geometric risk cues indicating the presence of negative obstacles in the current working environment; By using Bayesian fusion, the continuous risk gradient distribution is updated based on the geometric risk cues to obtain a multi-level grid map; Based on the tactile perception data, the multi-level grid map is corrected using tactile closed-loop technology to generate a risk cost map for the current working environment.

2. The method for constructing a negative obstacle risk map according to claim 1, characterized in that, The acquisition of multimodal sensing data of the legged robot in the current working environment includes: The legged robot uses its visual sensor, lidar sensor, and foot-end tactile sensing component to collect initial visual images, initial laser point clouds, and initial contact perception data of the current working environment, as well as the contact between the legged robot's foot and the ground, at preset frequencies corresponding to the visual sensor, lidar sensor, and foot-end tactile sensing component, respectively. The initial visual image, the initial laser point cloud, and the initial contact sensing data are time-synchronized and calibrated to obtain the visual image, the laser point cloud, and the contact sensing data of the current working environment.

3. The method for constructing a negative obstacle risk map according to claim 1, characterized in that, The semantic segmentation based on the visual image, to obtain the type of negative obstacle in the current working environment, multiple candidate regions of the negative obstacle, and the confidence and uncertainty corresponding to each candidate region, includes: The visual image is classified at the pixel level using a lightweight semantic segmentation network to obtain the category mask of the visual image and the pixel confidence and pixel uncertainty corresponding to each pixel position in the visual image; Connectivity aggregation and contour extraction are performed on the category mask to obtain the initial candidate region of the negative obstacle and the type of the negative obstacle; Perform geometric consistency filtering on each of the initial candidate regions to obtain multiple candidate regions for the negative obstacle; The confidence level and uncertainty corresponding to each candidate region are obtained by statistically calculating the pixel-level confidence level and pixel uncertainty of each pixel position within each candidate region.

4. The method for constructing a negative obstacle risk map according to claim 1, characterized in that, The process involves using a risk diffusion algorithm to perform ink blurring based on the type of the negative obstacle and the confidence and uncertainty of each candidate region, generating a continuous risk field corresponding to each candidate region, including: Determine the category risk coefficient for the type of negative obstacle; Based on the category risk coefficient, and combined with the confidence level and uncertainty of each candidate region, the risk intensity coefficient of the candidate region is obtained; Feature extraction is performed on the candidate region to obtain the center position, main morphological direction, and scale features of the candidate region; Based on the center location, the main orientation of the shape, and the scale characteristics, combined with the terrain aspect and slope characteristics of the laser point cloud, a covariance matrix is ​​generated. Ink smudging is performed based on the risk intensity coefficient and the covariance matrix to generate the continuous risk field corresponding to each candidate region.

5. The method for constructing a negative obstacle risk map according to claim 4, characterized in that, The step of dividing the continuous risk field and generating a continuous risk gradient distribution corresponding to each candidate region includes: Based on a preset multi-level risk classification threshold, the continuous risk field is divided into intervals to obtain multiple risk zones corresponding to the candidate region. The risk zones include fatal zones, buffer zones, and safe zones. Based on the risk distribution characteristics of each risk zone, a continuous risk gradient distribution with gradient decay from the lethal zone to the safe zone is generated.

6. The method for constructing a negative obstacle risk map according to claim 1, characterized in that, The terrain modeling based on the laser point cloud, to obtain geometric risk cues indicating the presence of negative obstacles in the current operating environment, includes: Voxel downsampling and ground segmentation are performed on the laser point cloud to obtain the ground point cloud in the laser point cloud; Based on the ground point cloud, according to the preset grid parameters, an elevation map of the current working environment is constructed, and the slope and height change values ​​of each grid in the elevation map are obtained; Based on the slope and height abrupt change value of each grid in the elevation map, the cavity features in the elevation map, the cliff edge features of the ground point cloud, and the distance jump features of the laser point cloud are obtained. The slope, the height abrupt change value, the cavity feature, the cliff edge feature, and the distance jump feature are used as the geometric risk clues.

7. The method for constructing a negative obstacle risk map according to claim 6, characterized in that, The method of updating the continuous risk gradient distribution based on the geometric risk cues using Bayesian fusion to obtain a multi-level grid map includes: Based on the slope, the height abrupt change value, the cavity feature, the cliff edge feature, and the distance jump feature, the geometric risk distribution value of each grid in the elevation map is generated, and the laser observation uncertainty corresponding to the geometric risk distribution value is obtained; Based on the continuous risk gradient distribution, the visual risk distribution value is obtained, and the visual observation uncertainty corresponding to the visual risk distribution value is acquired. The visual risk distribution value and the geometric risk distribution value are weighted and fused to obtain risk layer data; The occupancy observation data corresponding to the laser point cloud are accumulated to obtain occupancy layer data; Based on the visual observation uncertainty and the laser observation uncertainty, uncertainty layer data is generated, and freshness layer data is generated through a time decay model. The risk layer data, the occupation layer data, the uncertainty layer data, and the freshness layer data are used as levels for raster index fusion to obtain the multi-level raster map.

8. The method for constructing a negative obstacle risk map according to claim 1, characterized in that, The step of performing tactile closed-loop correction on the multi-level grid map based on the tactile sensing data to generate a risk cost map for the current working environment includes: Based on the contact sensing data, the actual landing point of the foot of the legged robot and the corresponding ground contact state of the foot are determined. Based on the ground contact state and the actual landing point, the multi-level grid map is subjected to tactile closed-loop correction to obtain the tactile closed-loop corrected multi-level grid map. Based on the multi-level grid map corrected by the tactile closed loop, hierarchical fusion is performed, and the risk cost map in the current working environment is generated by combining the layer weight coefficients corresponding to each level.

9. The method for constructing a negative obstacle risk map according to claim 8, characterized in that, The ground contact states include safe ground contact and abnormal ground contact. The step of performing tactile closed-loop correction on the multi-level grid map based on the ground contact states and the actual landing point location to obtain a tactile closed-loop corrected multi-level grid map includes: If the ground contact state is the safe ground contact, then the grid corresponding to the actual landing point and the neighboring grid in the multi-level grid map are subjected to risk dynamic erasure processing to obtain the multi-level grid map after tactile closed-loop correction. If the ground contact state is an abnormal ground contact, then the grid corresponding to the actual landing point and the neighboring grids in the multi-level grid map are subjected to risk deepening processing to obtain the multi-level grid map after tactile closed-loop correction.

10. A negative obstacle risk map construction system, characterized in that, include: The data acquisition unit is used to acquire multimodal sensing data of the legged robot in the current working environment. The multimodal sensing data includes visual images of the current working environment, laser point clouds, and contact perception data of the legged robot. A semantic segmentation unit is used to perform semantic segmentation based on the visual image to obtain the type of negative obstacle in the current working environment, multiple candidate regions of the negative obstacle, and the confidence and uncertainty corresponding to each candidate region. The ink smudging unit is used to perform ink smudging processing based on the type of the negative obstacle and the confidence and uncertainty of each candidate region through a risk diffusion algorithm, and generate a continuous risk field corresponding to each candidate region. A partitioning unit is used to partition the continuous risk field and generate a continuous risk gradient distribution corresponding to each candidate region; The modeling unit is used to perform terrain modeling based on the laser point cloud to obtain geometric risk clues indicating the existence of the negative obstacle in the current working environment. A Bayesian fusion unit is used to update the continuous risk gradient distribution based on the geometric risk cues using a Bayesian fusion method to obtain a multi-level grid map. The map generation unit is used to perform tactile closed-loop correction on the multi-level raster map based on the tactile perception data, and generate a risk cost map for the current working environment.

Images