Scene understanding and intelligent decision-making method for planet probe vehicle in complex unknown environment

Abstract

The invention discloses a scene understanding and intelligent decision-making method for a planet probe vehicle in a complex unknown environment, and relates to the technical field of deep space exploration and intelligent vehicles, and the method comprises the steps: obtaining multi-dimensional environment information; constructing a slump function based on the geometric morphology, motion characteristics and material attributes of the obstacle, and dynamically quantifying the risk of the obstacle; smoothing the potential field function by adopting Gaussian filtering, and dynamically adjusting the weight of a global potential field and the weight of a local potential field; in combination with an improved clustering algorithm, fuzzy regions of obstacles capable of being crossed and obstacles not capable of being crossed are accurately divided; screening passable areas by constructing a credibility mean square value function; carrying out multi-objective optimization by adopting a genetic algorithm, and generating an optimal path matched with the environment in cooperation with an offline data set and online reasoning; and inputting the optimal path into a bottom layer control system of the probe vehicle, controlling execution of a motor driving and steering mechanism, monitoring the state of the probe vehicle in real time, and triggering dynamic re-planning. And the autonomous obstacle avoidance capability and the path planning efficiency of the probe vehicle in a complex environment are improved.

Description

Technical Field

[0001] The present invention relates to the field of deep space exploration and intelligent vehicle technology, and in particular to a scene understanding and intelligent decision-making method for planetary exploration rovers in complex and unknown environments, which is suitable for autonomous navigation and decision-making control of exploration rovers on the surface of planets such as the Moon and Mars. Background Art

[0002] As deep space exploration missions advance, planetary rovers face core challenges such as polymorphic obstacle recognition, time-varying traffic area perception, and boundary fuzzy decision-making in complex and unknown environments.

[0003] Currently, the US Mars rovers Spirit and Opportunity use an alternative arc-based path planning algorithm, which enables autonomous obstacle avoidance but lacks the ability to semantically classify polymorphic obstacles. my country's Yutu and Zhurong rovers use terrain suitability assessments for local obstacle avoidance, but suffer from decision-making lags due to time-varying regional evolution.

[0004] It can be seen that traditional methods have defects in obstacle risk modeling, mathematical representation of traversable areas and multi-sensor fusion efficiency, and are unable to meet the needs of safe and efficient driving of exploration vehicles in complex and unknown environments.

[0005] Therefore, there is an urgent need for a full-chain technology system that integrates environmental perception, risk quantification and intelligent decision-making to enhance the autonomous decision-making capabilities of the exploration vehicle in complex and unknown environments. Summary of the Invention

[0006] The technical problem to be solved by the present invention is how to solve the problem of incomplete scene understanding and unclear decision-making mechanism of planetary exploration rovers in complex and unknown environments.

[0007] In order to solve the above technical problems, the present invention provides a method for scene understanding and intelligent decision-making of planetary exploration rovers in complex and unknown environments, the method comprising: Collect environmental data using the multiple sensors carried by the rover, and perform spatiotemporal alignment on the collected data to obtain multi-dimensional environmental information; Based on the geometric shape, motion characteristics, and material properties of the obstacles in the environmental information, a collapse function is constructed to dynamically quantify the obstacle risk. A Gaussian filter is used to smooth the potential field function, and the weights of the global and local potential fields are dynamically adjusted. Combined with an improved clustering algorithm, the fuzzy intersection area between surmountable and insurmountable obstacles is accurately delineated. Based on obstacle density thresholds and risk heat maps to reflect regional risk levels, the system constructs a trust mean square value function to screen traversable areas. It uses a genetic algorithm for multi-objective optimization, collaborating with offline datasets and online reasoning to generate the optimal path that matches the environment. The planning result of the optimal path is input into the underlying control system of the probe vehicle to control the execution of the motor drive and steering mechanism, monitor the status of the probe vehicle in real time and trigger dynamic replanning.

[0008] Furthermore, the multi-sensor includes at least a binocular vision camera, a laser radar and an ultrasonic sensor.

[0009] Furthermore, the environmental information includes the location, size, material, and terrain slope and roughness of the obstacle.

[0010] Furthermore, the method for performing spatiotemporal alignment processing on the collected data is to perform time synchronization and spatial calibration on multiple sensors, wherein: The data processing of the binocular vision camera includes: The Zhang Zhengyou calibration method is used to perform distortion correction and internal and external parameter calibration on the binocular vision camera, a semi-global matching algorithm is used to generate a disparity map, and dense depth information is calculated by combining the baseline distance and focal length; Through transfer learning, we fine-tune the YOLOv5s model, design a two-stream feature fusion network, use the cross-view attention mechanism to interact with the left and right eye image features, and output obstacle detection boxes with 3D coordinates. Convert the depth map into a 3D point cloud, combine statistical filtering with voxel grid downsampling to generate a dense point cloud map with category labels; The data processing of the laser radar includes: An improved multimodal DBSCAN clustering algorithm is used to accelerate neighborhood search based on KD-Tree, and the search radius is adaptively set according to the local point cloud density. A composite distance metric function is constructed, and the terrain slope angle is calculated through RANSAC plane fitting to generate an obstacle description set including position, size, slope and point cloud density.

[0011] Furthermore, the design of the dual-stream feature fusion network includes: The left eye image is used to extract features through the C3 module, and the right eye image is used to calculate the cross-correlation matrix of the left and right eye feature maps through the cross-view attention mechanism. The left and right eye features are fused to output a three-dimensional obstacle detection frame.

[0012] Furthermore, the collapse function is:

[0013] Where, is the initial collapse degree, is the compressive strength, is the obstacle volume, is the material reflectivity, P is the external pressure, is the movement speed, is the material color contrast, is the geometric influence coefficient, is the influence coefficient of the material optical properties, and Calibrated through offline testing.

[0014] Furthermore, according to the real-time calculation of the collapse function The target obstacle risk level is divided according to the results of when When it is less than 0.3, the target obstacle is marked as a stable obstacle; When 0.3≤ When ≤0.7, the target obstacle is marked as a potential risk; when When it is greater than 0.7, the target obstacle is marked as a high-risk collapse body.

[0015] Furthermore, the method of using Gaussian filtering to smooth the potential field function includes: The potential field distribution is smoothed by Gaussian filtering, where the Gaussian kernel parameters are dynamically adjusted according to the obstacle density. The potential field function after smoothing is:

[0016] in, is a two-dimensional Gaussian kernel; is the Gaussian kernel parameter, where the sparse area , dense areas .

[0017] Furthermore, regarding the overall situation: Apply attraction to the target location ,in, is the attraction coefficient, Current location With target location The Euclidean distance between Obstacles exert repulsive forces ,in, is the repulsive force coefficient, Current location With obstacles The shortest distance between For safe distance; For the local potential field: Add a virtual repulsive force when a local minimum is detected ,in, is the virtual repulsive force coefficient, are the coordinates of the local minimum point.

[0018] Furthermore, the weights of the global potential field and the local potential field are dynamically allocated according to the speed of the probe vehicle:

[0019]

[0020] Where, is the weight of the global situation, is the weight of the local potential field, is the speed of the probe car.

[0021] Furthermore, the method for processing fuzzy areas by combining the improved clustering algorithm includes: The dynamic weight of obstacle characteristics is introduced into the fuzzy C-means objective function to obtain the improved objective function:

[0022] Where, = , for the sample To cluster center The Euclidean distance of Cluster category The preset weights of 、 and Obstacles speed, height and hardness characteristics; Based on the improved objective function, adjust the membership :

[0023] Where, Cluster category The preset weights of is the fuzziness index; For 0.4< The fuzzy region with a membership deviation of less than 0.6 is combined with a boundary optimization algorithm to minimize the classification error and adjust the cluster center until the membership deviation is less than 0.05.

[0024] Furthermore, the trust mean square value function is , where: The confidence level of the sampling point The mean of is the confidence coefficient, which is set according to the risk tolerance of the detection mission; is the mean square error and satisfies the expression ,in, , where is the slope, is the slope weight, is the roughness, is the obstacle roughness weight, is the obstacle distance, is the obstacle distance weight; By filtering A threshold of >0.6 is used to generate a traversable area mask.

[0025] Furthermore, the objective function of the multi-objective optimization using the genetic algorithm is:

[0026] Where, is the path length; is the risk value; is smoothness; The constraints of the objective function include: minimum safety distance ≥0.3m, maximum slope ≤25°, steering angle ≤10° / s and speed ≤0.2m / s.

[0027] Furthermore, the method for coordinating the offline dataset with the online reasoning includes: Establish an offline dataset of an effective space with dense obstacles, including the spatial distribution, movement patterns, and corresponding safe passage paths of different types of obstacles; During online decision-making, the real-time data is matched with the offline data set to quickly generate an initial feasible solution, and local optimization is performed in combination with real-time environmental feedback.

[0028] Furthermore, the dynamic replanning mechanism includes: Real-time monitoring of the posture and wheel slip rate of the probe vehicle; When the wheel slip rate is greater than 15%, local replanning is triggered, the speed of the probe car is reduced to 0.1m / s, and the safe distance between the probe car and the obstacle is increased to 0.5m; and the closed-loop control cycle is 100ms to ensure that the response time to dynamic obstacles is less than 300ms.

[0029] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: 1. The present invention uses the collapse function and potential field function reconstruction method, combined with an improved clustering algorithm to process fuzzy areas, to achieve semantic classification and risk grading of obstacles, and solve the risk assessment problem under the coexistence of polymorphic obstacles.

[0030] 2. Construct a trust mean square value function and a regional constraint aggregate state representation. Through probabilistic evaluation and offline dataset enhancement, real-time traffic decision-making without prior knowledge is achieved, breaking through the bottleneck of mathematical modeling of traversable areas.

[0031] 3. Integrate control theory, machine learning, and operations research to establish a closed-loop system of "environmental perception-risk quantification-decision generation" to improve the efficiency of multi-sensor fusion and the real-time nature of decision-making in complex environments, providing key technical support for deep space exploration. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0033] Figure 1 This is a flowchart of the overall process disclosed in the present invention. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0035] See also Figure 1 This invention aims to provide a scenario understanding and intelligent decision-making method for planetary rovers in complex and unknown environments. This method enables accurate identification and risk assessment of polymorphic obstacles, efficient understanding of the global scenario, and autonomous decision-making regarding traversable areas, ensuring the safety, stability, and efficiency of the rover in extremely complex environments. The following describes this method in detail.

[0036] S1. Multi-sensor fusion environmental perception: Collect environmental data based on the multiple sensors carried by the rover, and perform spatiotemporal alignment on the collected data to obtain multi-dimensional environmental information.

[0037] The multi-sensor system includes at least a binocular vision camera, a laser radar, and an ultrasonic sensor. The multi-dimensional environmental information includes the location, size, material, and terrain slope and roughness of obstacles.

[0038] Multi-sensor spatiotemporal alignment uses hardware trigger signals to achieve microsecond-level time synchronization of binocular vision, lidar, and IMU, uses external parameter matrices to convert radar point clouds to camera coordinate systems, and combines IMU data to compensate for motion distortion to complete multi-sensor spatiotemporal alignment. Specifically: Multi-sensor time synchronization achieves microsecond-level synchronization of sensors through hardware trigger signals (such as GPIO pulses), records the data acquisition timestamp of each sensor, and uses linear interpolation (time difference < 50μs) to time calibrate asynchronous data.

[0039] Multi-sensor spatial alignment includes extrinsic calibration and motion distortion compensation. Extrinsic calibration uses a high-precision calibration plate (including the lidar reflective surface and visual feature points) to solve the transformation matrix from the radar coordinate system to the camera coordinate system through the hand-eye calibration method. Motion distortion compensation uses IMU data (angular velocity, acceleration) to calculate the position change of the probe vehicle during data collection through integration, and performs rigid body transformation compensation on the lidar point cloud.

[0040] The following describes the processing of binocular vision camera data and lidar data respectively.

[0041] S1.1. The following describes the data processing of the binocular vision camera.

[0042] S1.1.1. Use Zhang Zhengyou calibration method to perform distortion correction and internal and external parameter calibration on the binocular vision camera, use semi-global matching (SGM) algorithm to generate disparity map, and combine baseline distance and focal length to calculate dense depth information.

[0043] First, explanation of Zhang Zhengyou's calibration method: First, checkerboard calibration is performed. That is, a checkerboard template with 10×8 corner points is used to collect 20 sets of images at different angles, and sub-pixel coordinates are obtained through corner detection.

[0044] Then calculate the internal and external parameters, where the internal parameter matrix , optimize the focal length by the least squares method ( , )、( , ). The external parameter matrix is ​​used to calculate the rotation matrix and translation vectors , used to describe the pose of the camera coordinate system relative to the world coordinate system.

[0045] Finally, distortion correction is performed using the Brown-Conradii model to correct radial distortion ( , , ) and tangential distortion ( , ), and output an undistorted image.

[0046] Second, explanation of disparity map generation and depth calculation: Semi-Global Matching (SGM) algorithm: First, the matching cost is calculated. Based on the Census Transform and Normalized Cross Correlation (NCC), a matching cost cube with a 15×15 window is constructed. Path aggregation is then performed, using dynamic programming along eight directions (horizontal, vertical, and diagonal). The costs are aggregated and a smoothness constraint is imposed (penalizing disparity jumps greater than 1). Finally, disparity optimization is performed, with a left-right consistency check (threshold set to 1) and median filtering to generate a high-precision disparity map.

[0047] Depth calculation: Using the principle of triangulation, the depth value , where baseline distance B = 0.1m, focal length f = 500px, parallax .

[0048] S1.1.2. Fine-tune the YOLOv5s model through transfer learning, design a dual-stream feature fusion network, use the cross-view attention mechanism to interact the left and right eye image features, and output the obstacle detection box with 3D coordinates.

[0049] Among them, fine-tuning the YOLOv5s model through transfer learning specifically includes pre-training on the COCO dataset, freezing the first 5 layers of the backbone network, and replacing the last 3 layers of the detection head.

[0050] The design of the dual-stream feature fusion network includes extracting features from the left eye image through the C3 module, calculating the cross-correlation matrix of the left and right eye feature maps of the right eye image through the cross-view attention mechanism, fusing the left and right eye features, and outputting a three-dimensional obstacle detection frame ( ).

[0051] S1.1.3. Convert the depth map into a 3D point cloud and combine statistical filtering with voxel grid downsampling to generate a dense point cloud map with category labels.

[0052] It is further explained that the depth map is converted into a three-dimensional point cloud, ROI is extracted based on the detection box, and a dense point cloud map with category labels is generated through statistical filtering (eliminating outliers, mean ± 3σ) and voxel grid downsampling (voxel size 0.05m).

[0053] S1.2. The following describes the data processing of the LiDAR.

[0054] S1.2.1. For the non-uniform point cloud on the planetary surface, an improved multimodal DBSCAN clustering algorithm is used, which accelerates the neighborhood search based on KD-Tree and adaptively sets the search radius according to the local point cloud density. .

[0055] The description of the improved multimodal DBSCAN clustering algorithm is as follows.

[0056] Construct a KD-Tree, project the point cloud to the XY plane, and construct a kd tree to accelerate neighborhood search and average query time complexity. .

[0057] Adaptive search radius, query k=20 neighbors for each point, calculate density mean and standard deviation , search radius , dynamically adapting to sparse or dense point clouds.

[0058] Resultant distance metric , where: Euclidean distance ; Normal vector angle distance , the normal vector is calculated by local plane fitting; Reflection Intensity Distance ; Weight coefficient , which is optimized via grid search.

[0059] S1.2.2. Fusion of normal vector similarity and reflection intensity to construct a composite distance metric function. Calculate the terrain slope angle through RANSAC plane fitting to generate an obstacle description set containing position, size, slope, and point cloud density, thus achieving point cloud clustering and obstacle feature extraction.

[0060] First, the method of calculating the terrain slope angle by RANSAC plane fitting is: randomly select 3 points to fit the plane, calculate the distance from all points to the plane, set the threshold to 0.03m, iterate 500 times, and retain the largest inner point set as the terrain plane. , where the plane equation is .

[0061] Second, the obstacle description set includes position (x, y, z), size (w, h, l), slope , point cloud density (i.e., the number of points per unit volume), and the mean reflection intensity.

[0062] S2. Construction of a polymorphic obstacle risk assessment model: A collapse function is constructed based on the geometric shape, motion characteristics, and material properties of obstacles in the environmental information to dynamically quantify obstacle risks. Gaussian filtering is used to smooth the potential field function, and the weights of the global and local potential fields are dynamically adjusted. Combined with an improved clustering algorithm, the fuzzy intersection area between traversable and intraversable obstacles is accurately delineated.

[0063] S2.1. Collapse function:

[0064] Where, is the initial collapse degree, which is preset according to the obstacle type (such as rock , sand pile 5) and then corrected by historical collision data; is the compressive strength; is the obstacle volume; is the material reflectivity; P is the external pressure , where the rover mass m = 20 kg, the gravitational acceleration g is set according to the planet (1.62 m / s² on the moon and 3.71 m / s² on Mars), and the contact force Estimation through ultrasonic sensor pressure feedback; is the movement speed; is the material color contrast; is the geometric influence coefficient, is the influence coefficient of the material optical properties, and Calibration through offline tests, such as obstacle crossing tests with different obstacle volumes ( ), visual inspection accuracy test of materials with different reflectivity ( ).

[0065] Based on real-time calculation of collapse function The target obstacle risk level is divided according to the results of when When <0.3, the target obstacle is marked as a stable obstacle.

[0066] When 0.3≤ When ≤0.7, the target obstacle is marked as a potential risk.

[0067] when When it is greater than 0.7, the target obstacle is marked as a high-risk collapse body.

[0068] S2.1. Reconstruction and optimization of potential field function: Introduce a virtual potential field to break the local minimum point, and add a repulsive force near the local minimum point to guide the rover to break away. Use the Gaussian filter formula: , smoothing the potential field distribution and reducing gradient mutations. The potential field is divided into global potential field and local potential field. The global potential field is used for long-distance planning, while the local potential field is used for real-time obstacle avoidance. Through dynamic weight adjustment, the global and local potential fields are integrated to improve the efficiency and safety of path planning.

[0069] S2.1.1. Gaussian filtering and smoothing: Gaussian filtering and smoothing are performed on the potential field distribution. The Gaussian kernel parameters are dynamically adjusted according to the obstacle density. The potential field function after smoothing is:

[0070] in, is a two-dimensional Gaussian kernel; is the Gaussian kernel parameter, where the sparse area , dense areas .

[0071] S2.1.2 Fusion of global and local potential fields Regarding the overall situation: Apply attraction to the target location ,in, is the attraction coefficient, Current location With target location The Euclidean distance between Obstacles exert repulsive forces ,in, is the repulsive force coefficient, Current location With obstacles The shortest distance between the obstacles, the repulsive force is only effective when the distance to the obstacle is less than the safe distance effective at When the repulsive force is 0, it emphasizes close-range obstacle avoidance, which is suitable for the strict safety distance requirements of planetary exploration vehicles.

[0072] For the local potential field: When a local minimum point is detected (gradient is 0 and non-target point), a virtual repulsive force is added ,in, is the virtual repulsive force coefficient, is the position coordinate of the local minimum point; the virtual repulsive force points away from the local minimum point The direction and intensity of Control, used to guide the rover out of the trap.

[0073] The weights of the global potential field and the local potential field are dynamically allocated according to the speed of the rover:

[0074]

[0075] Where, is the weight of the overall situation; is the weight of the local potential field; is the speed of the probe car, where the focus is on local obstacle avoidance at low speeds. For example, the probe car =0.05m / s when approaching the crater, the local potential field prioritizes calculating the repulsive force at the edge of the crater to adjust the direction in real time to avoid falling; at high speeds, the focus is on global planning, for example, the rover =0.2m / s when driving on flat sand, the shortest path is planned based on the overall situation, avoiding distant obstacles and reducing losses caused by frequent steering.

[0076] S2.2. Improved clustering algorithm and risk classification: Improved fuzzy C-means (FCM) and DBSCAN clustering algorithms, in the FCM objective function , introduce dynamic weights and adjust the membership according to the speed, height, hardness and other characteristics of the obstacle Build a risk probability model ,Combining multi-sensor data with boundary optimization algorithms (such as genetic algorithms) to process fuzzy areas,,and achieve obstacle crossable / non-crossable classification and risk level,division.

[0077] Specifically, the dynamic weight of obstacle characteristics is introduced into the fuzzy C-means objective function to obtain the improved objective function:

[0078] Where, = , for the sample To cluster center The Euclidean distance of Cluster category Preset weights, such as "crossable obstacles" =0.2, "uncrossable obstacle" =0.8; 、 and Obstacles speed, height and hardness characteristics.

[0079] Based on the improved objective function, according to the obstacles Speed ,high and hardness characteristics Dynamically adjust the degree of membership of data points to cluster centers :

[0080] Where, Cluster category The preset weights of is the fuzziness index.

[0081] Using logistic regression model Calculate the collision risk probability. =0.5s / m, =0.3 , =0.2, =-1, which is a parameter obtained through training of historical collision data; hardness It is estimated by the attenuation degree of the ultrasonic echo signal.

[0082] For 0.4< For the fuzzy region with a membership deviation of <0.6, a boundary optimization algorithm (such as a genetic algorithm, a population size of 50, and 30 iterations) is used to minimize the classification error and adjust the cluster center until the membership deviation is <0.05.

[0083] S3. Global scene understanding and autonomous decision-making on traversable areas: Based on the obstacle density threshold division and risk heat map to reflect the regional risk level, the traversable area is screened by constructing a trust mean square value function; a genetic algorithm is used for multi-objective optimization, and offline data sets and online reasoning are coordinated to generate the optimal path that matches the environment.

[0084] S3.1. Obstacle area division and heat map construction: Based on the density and distribution characteristics of obstacles, the environment is divided into dense obstacle areas and sparse obstacle areas. Through multi-sensor data fusion, an obstacle distribution heat map is constructed to intuitively reflect the regional risk level.

[0085] Among them, density threshold division: calculate the number of obstacles within a unit area (1m²), and a density greater than 5 obstacles / m² is considered a "dense area", and a density ≤5 obstacles / m² is considered a "sparse area".

[0086] Heat map generation: Gaussian kernel density estimation is used to estimate the risk value contributed by each obstacle. , accumulate to generate the risk heat map, and normalize it to the grayscale value of 0-255.

[0087] S3.2. Construction of Trust Mean Square Value Function The trust mean square value function is , where The confidence level of the sampling point The mean ; is the confidence coefficient, which is set according to the risk tolerance of the detection mission; is the mean square error and satisfies the expression ,in, , where is the slope, is the slope weight, is the roughness, is the obstacle roughness weight, is the obstacle distance, is the obstacle distance weight; in this embodiment, , , . By filtering A threshold of >0.6 is used to generate a passable area mask; a 5×5 Gaussian convolution kernel is applied to smooth the mask, and adjacent passable areas are connected through morphological dilation (kernel size 3×3) to eliminate isolated noise points.

[0088] S3.3. Boundary Constraint Optimization and Path Solution Set Solving: Establish regional boundary constraints and use a genetic algorithm (GA / NSGA-II) to solve a multi-objective optimization problem. Objective functions include path length, risk value, and smoothness. Constraints include obstacle avoidance, terrain passability, and rover kinematic limitations. Dynamically adjust weight coefficients to balance path safety and efficiency and generate an optimal solution set.

[0089] S3.3.1. The objective function for multi-objective optimization using genetic algorithm is:

[0090] In the formula, the path length ; Value at Risk , normalized after negation; smoothness ; Determine weights through Pareto optimal frontier analysis .

[0091] The constraints of the objective function include the minimum safe distance ≥0.3m, maximum slope ≤25°, steering angle ≤10° / s and speed ≤0.2m / s.

[0092] S3.3.2. Genetic Algorithm Implementation The path point coordinates are coded with real numbers, and the chromosome length is dynamically adjusted to 5-10. Its fitness function is , , avoiding division by zero. Through selection (roulette wheel, probability 0.8), crossover (single-point crossover, probability 0.7), mutation (Gaussian mutation, standard deviation 0.1m), and iteration, the optimal solution set is gradually found. The genetic algorithm terminates when the optimal solution remains unchanged for 10 consecutive generations or when 50 generations have been iterated and the Pareto optimal solution set is output.

[0093] S3.4. Collaboration between Offline Datasets and Online Reasoning: Build an offline dataset for an effective space densely populated with obstacles. This dataset includes the spatial distribution, movement patterns, and corresponding safe paths of different obstacle types. During online decision-making, real-time data is matched with the offline dataset to quickly generate an initial feasible solution. This solution is then optimized locally based on real-time environmental feedback, improving real-time decision-making.

[0094] S3.4.1. Offline dataset construction: In a simulated lunar / Mars environment, collect 100,000 sets of dense obstacle scene data, including point cloud maps, obstacle attributes (speed, volume, hardness), and safe paths (manual annotation + simulation verification); mark the boundaries of the traversable area, risk level, and key points of the optimal path (at 0.5m intervals).

[0095] S3.4.2. Online reasoning process: Extract features such as obstacle density, average volume, and slope distribution from real-time scenarios, and use the k-NN algorithm (k=5) to match the five most similar scenarios in the offline dataset. The optimal path of the weighted average matching scenario is used as the initial population of the genetic algorithm (accounting for 30%), and the remaining 70% of the population is randomly generated. The environmental data is updated every 200ms, and the 30% nodes at the end of the path are locally replanned to ensure real-time performance (single planning time <100ms).

[0096] S4. Decision-making, execution, and closed-loop control: The optimal path planning results are input into the rover's underlying control system to control the execution of the motor drive and steering mechanisms. At the same time, the rover's status (speed, posture, sensor data, etc.) is monitored in real time and fed back to the scene understanding and decision-making module to dynamically adjust the path, forming a "perception-decision-making-execution-feedback" closed loop to ensure the rover's safe and efficient driving in complex and unknown environments.

[0097] Specifically, the path planning results are converted into motor speed and steering angle, and sent to the drive module via the CAN bus. The control accuracy includes: speed ±0.01m / s, steering angle ±1°.

[0098] The dynamic replanning mechanism includes real-time monitoring of the rover's posture and wheel slip rate. When the wheel slip rate is greater than 15%, local replanning is triggered, reducing the rover's speed to 0.1m / s and increasing the safe distance between the rover and obstacles to 0.5m. The closed-loop control cycle is 100ms to ensure that the response time to dynamic obstacles is less than 300ms.

[0099] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for scene understanding and intelligent decision-making of planetary exploration rovers in complex and unknown environments, characterized by: The method comprises: Collect environmental data using the multiple sensors carried by the rover, and perform spatiotemporal alignment on the collected data to obtain multi-dimensional environmental information; Based on the geometric shape, motion characteristics, and material properties of the obstacles in the environmental information, a collapse function is constructed to dynamically quantify the obstacle risk. A Gaussian filter is used to smooth the potential field function, and the weights of the global and local potential fields are dynamically adjusted. Combined with an improved clustering algorithm, the fuzzy intersection area between surmountable and insurmountable obstacles is accurately delineated. Based on obstacle density thresholds and risk heat maps to reflect regional risk levels, the system constructs a trust mean square value function to screen traversable areas. It uses a genetic algorithm for multi-objective optimization, collaborating with offline datasets and online reasoning to generate the optimal path that matches the environment. The planning result of the optimal path is input into the underlying control system of the probe vehicle to control the execution of the motor drive and steering mechanism, monitor the status of the probe vehicle in real time and trigger dynamic replanning.

2. The method for scene understanding and intelligent decision-making of planetary exploration rovers in complex and unknown environments according to claim 1 is characterized in that: The multiple sensors include at least a binocular vision camera, a laser radar and an ultrasonic sensor.

3. The method for scene understanding and intelligent decision-making of planetary exploration rovers in complex and unknown environments according to claim 1 is characterized in that: The environmental information includes the location, size, material, and terrain slope and roughness of obstacles.

4. The method for scene understanding and intelligent decision-making of planetary exploration rovers in complex and unknown environments according to claim 2 is characterized in that: The method for performing spatiotemporal alignment processing on the collected data is to perform time synchronization and spatial calibration on multiple sensors, wherein: The data processing of the binocular vision camera includes: The Zhang Zhengyou calibration method is used to perform distortion correction and internal and external parameter calibration on the binocular vision camera, a semi-global matching algorithm is used to generate a disparity map, and dense depth information is calculated by combining the baseline distance and focal length; Through transfer learning, we fine-tune the YOLOv5s model, design a two-stream feature fusion network, use the cross-view attention mechanism to interact with the left and right eye image features, and output obstacle detection boxes with 3D coordinates. Convert the depth map into a 3D point cloud, combine statistical filtering with voxel grid downsampling to generate a dense point cloud map with category labels; The data processing of the laser radar includes: An improved multimodal DBSCAN clustering algorithm is used to accelerate neighborhood search based on KD-Tree, and the search radius is adaptively set according to the local point cloud density. A composite distance metric function is constructed, and the terrain slope angle is calculated through RANSAC plane fitting to generate an obstacle description set including position, size, slope and point cloud density.

5. The method for scene understanding and intelligent decision-making of planetary exploration rovers in complex and unknown environments according to claim 4 is characterized in that: The design of the dual-stream feature fusion network includes: The left eye image is used to extract features through the C3 module, and the right eye image is used to calculate the cross-correlation matrix of the left and right eye feature maps through the cross-view attention mechanism. The left and right eye features are fused to output a three-dimensional obstacle detection frame.

6. The method for scene understanding and intelligent decision-making of planetary exploration rovers in complex and unknown environments according to claim 1 is characterized in that: The collapse function is: Where, is the initial collapse degree, is the compressive strength, is the obstacle volume, is the material reflectivity, P is the external pressure, is the movement speed, is the material color contrast, is the geometric influence coefficient, is the influence coefficient of the material optical properties, and Calibrated through offline testing.

7. The method for scene understanding and intelligent decision-making of planetary exploration rovers in complex and unknown environments according to claim 6 is characterized in that: Based on real-time calculation of collapse function The target obstacle risk level is divided according to the results of when When it is less than 0.3, the target obstacle is marked as a stable obstacle; When 0.3≤ When ≤0.7, the target obstacle is marked as a potential risk; when When it is greater than 0.7, the target obstacle is marked as a high-risk collapse body.

8. The method for scene understanding and intelligent decision-making of planetary exploration rovers in complex and unknown environments according to claim 1 is characterized in that: The method of using Gaussian filtering to smooth the potential field function includes: The potential field distribution is smoothed by Gaussian filtering, where the Gaussian kernel parameters are dynamically adjusted according to the obstacle density. The potential field function after smoothing is: in, is a two-dimensional Gaussian kernel; is the Gaussian kernel parameter, where the sparse area , dense areas .

9. The method for scene understanding and intelligent decision-making of planetary exploration rovers in complex and unknown environments according to claim 1 or 8, characterized in that: For the global situation: Apply attraction to the target location ,in, is the attraction coefficient, Current location With target location The Euclidean distance between Obstacles exert repulsive forces ,in, is the repulsive force coefficient, Current location With obstacles The shortest distance between For safe distance; For the local potential field: Add a virtual repulsive force when a local minimum is detected ,in, is the virtual repulsive force coefficient, are the coordinates of the local minimum point.

10. The method for scene understanding and intelligent decision-making of planetary exploration rovers in complex and unknown environments according to claim 9, characterized in that: The weights of the global potential field and the local potential field are dynamically allocated according to the speed of the probe vehicle: Where, is the weight of the global situation, is the weight of the local potential field, is the speed of the probe car.

11. The method for scene understanding and intelligent decision-making of planetary exploration rovers in complex and unknown environments according to claim 1, characterized in that: The method for processing fuzzy areas by combining the improved clustering algorithm includes: The dynamic weight of obstacle characteristics is introduced into the fuzzy C-means objective function to obtain the improved objective function: Where, = , for the sample To the cluster center The Euclidean distance of Cluster category The preset weights of 、 and Obstacles speed, height and hardness characteristics; Based on the improved objective function, adjust the membership : Where, Cluster category The preset weights of is the fuzziness index; For 0.4< The fuzzy region with a membership deviation of less than 0.6 is combined with a boundary optimization algorithm to minimize the classification error and adjust the cluster center until the membership deviation is less than 0.

05.

12. The method for scene understanding and intelligent decision-making of planetary exploration rovers in complex and unknown environments according to claim 1, characterized in that: The trust mean square value function is: , where: The confidence level of the sampling point The mean of is the confidence coefficient, which is set according to the risk tolerance of the detection mission; is the mean square error and satisfies the expression ,in, , where is the slope, is the slope weight, is the roughness, is the obstacle roughness weight, is the obstacle distance, is the obstacle distance weight; By filtering A threshold of >0.6 is used to generate a traversable area mask.

13. The method for scene understanding and intelligent decision-making of planetary exploration rovers in complex and unknown environments according to claim 12, characterized in that: The objective function of the multi-objective optimization using genetic algorithm is: Where, is the path length; is the risk value; is smoothness; The constraints of the objective function include: minimum safety distance ≥0.3m, maximum slope ≤25°, steering angle ≤10° / s and speed ≤0.2m / s.

14. The method for scene understanding and intelligent decision-making of planetary exploration rovers in complex and unknown environments according to claim 1, characterized in that: The method for coordinating the offline dataset and online reasoning includes: Establish an offline dataset of an effective space with dense obstacles, including the spatial distribution, movement patterns, and corresponding safe passage paths of different types of obstacles; During online decision-making, the real-time data is matched with the offline data set to quickly generate an initial feasible solution, and local optimization is performed in combination with real-time environmental feedback.

15. The method for scene understanding and intelligent decision-making of planetary exploration rovers in complex and unknown environments according to claim 1, characterized in that: The dynamic replanning mechanism includes: Real-time monitoring of the posture and wheel slip rate of the probe vehicle; When the wheel slip rate is greater than 15%, local replanning is triggered, the speed of the probe car is reduced to 0.1m / s, and the safe distance between the probe car and the obstacle is increased to 0.5m; and the closed-loop control cycle is 100ms to ensure that the response time to dynamic obstacles is less than 300ms.

Images