Method and apparatus for road understanding under satellite denial and no prior map conditions

Abstract

The present application relates to a road understanding method and device under satellite denial and without prior maps. The method includes: generating global guidance information using spline curves based on satellite map annotation points; then fusing lidar point cloud data to construct a three-dimensional terrain model, extracting roughness and normal vectors to obtain corresponding maps, and constructing an environmental passability map; then defining confidence values for each map, fusing the evidence of the roughness map and the normal vector map, and further fusing with the basic probability assignment of the passability map to obtain the final data; then determining the passable area based on the global guidance information and the final data, estimating its width, and calculating the current position and angle expectations and standard deviations in combination with local guidance information; finally, calculating the posterior estimate using these as measurement values, analyzing the boundaries, and further obtaining the corrected guidance path and passable area. Using this method can improve the accuracy of path generation.

Description

Technical Field

[0001] The present application relates to the field of autonomous driving technology, and in particular to a method and device for understanding a road under satellite denial and no prior map conditions. Background Art

[0002] In the field of robotics and autonomous driving, the environment can be roughly divided into two types: structured and unstructured. Compared with structured environments, autonomous navigation in unstructured environments is like a difficult adventure and faces many difficult challenges. First of all, the unstructured environment itself is highly uncertain, which makes the scene in dynamic change at all times. Obstacles that may appear suddenly and the ever-changing road conditions all add great difficulty to navigation. Secondly, the complex state space also greatly increases the difficulty of accurate modeling and decision-making of the navigation system. In such an environment, various interference factors are intertwined, making it difficult for the system to accurately define its own state. At the same time, limited prior knowledge and the lack of high-precision positioning information further aggravate the complexity of navigation. When performing long-distance autonomous navigation tasks, in unfamiliar unstructured environments, the global planning module plays a vital role. This module needs to generate a path that can guide the vehicle based on the prior information of the target environment and the specific task requirements of the vehicle, and provide macro-guidance for subsequent local path planning and vehicle control. However, when faced with a complex and unknown unstructured environment, the lack of prior information is like putting shackles on the global path planning module, which seriously limits its function.

[0003] In unknown unstructured environments, due to the lack of high-precision prior maps, satellite images become the main source of prior information on target scenes. Therefore, the global path planning module can only rely on satellite maps to generate a global guidance path. However, this path generated by satellite maps has great limitations and poor accuracy. Satellite maps are restricted by factors such as resolution and update frequency, and it is difficult to reflect the real-time details of the target environment, resulting in path accuracy far below the requirements of subsequent local path planning modules. With only satellite maps as the only prior information, achieving long-distance autonomous driving is undoubtedly a very challenging task, which places extremely high demands on the reliability, adaptability and intelligence level of autonomous driving technology. Summary of the invention

[0004] Based on this, it is necessary to provide a road understanding method and device under satellite denial and no prior map conditions that can improve the accuracy of path generation in response to the above technical problems.

[0005] A method for understanding a road under satellite denial and no prior map conditions, the method comprising:

[0006] Based on the acquired satellite map, the necessary points on the marked path are used to generate dense and smooth global guidance information using spline curves;

[0007] The laser radar point cloud data acquired by the ground unmanned platform at different times and locations is effectively fused and accurately inferred to construct a three-dimensional terrain model; the terrain roughness and normal vector are extracted based on the three-dimensional terrain model to obtain a roughness map and a normal vector map; the BEV calculation constructed based on the multi-frame fusion of the laser radar and the odometer obtains an environmental passability map;

[0008] According to the corresponding trust values defined by the roughness map, the normal vector map and the environment passability map, the evidence of the roughness map and the normal vector map of the same source is fused to obtain the initial fused evidence; the basic probability distribution corresponding to the environment passability map is fused with the fused evidence to obtain the final fused data;

[0009] Determine the current traversable area based on the global guidance information and the final fused data; estimate the width of the current traversable area, and calculate the expected value and standard deviation of the current position and angle based on the estimated width and the local guidance information received by the unmanned platform;

[0010] The expected value and standard deviation of the current position and angle are taken as the measurement values, and the historical frames are used as the prior values to calculate the posterior distribution of the current frame. The boundary of the passable area is analyzed according to the posterior distribution of the current frame, and the corrected guidance path and the passable area are calculated using the boundary of the passable area.

[0011] A road understanding device under satellite denial and no prior map conditions, the device comprising:

[0012] A global guidance information generation module is used to generate dense and smooth global guidance information using a spline curve based on a set of necessary points on the path marked on the acquired satellite map;

[0013] The environment accessibility map calculation module is used to effectively fuse and accurately infer the laser radar point cloud data obtained by the ground unmanned platform at different times and locations to build a three-dimensional terrain model; extract the terrain roughness and normal vector based on the three-dimensional terrain model to obtain the roughness map and normal vector map; calculate the environment accessibility map based on the BEV constructed by multi-frame fusion of laser radar and odometer;

[0014] The data fusion module is used to define corresponding trust values according to the roughness map, the normal vector map and the environmental passability map, fuse the evidence of the same source roughness map and normal vector map to obtain the initial fused evidence; fuse the basic probability distribution corresponding to the environmental passability map with the fused evidence to obtain the final fused data;

[0015] A passable area correction module is used to determine the current passable area based on the global guidance information and the finally fused data; estimate the width of the current passable area, and calculate the expected values and standard deviations of the current position and angle according to the estimated width and the local guidance information received by the unmanned platform; use the expected values and standard deviations of the current position and angle as measurement values, use the historical frame as the prior value, calculate the posterior distribution of the current frame, analyze the boundary of the passable area according to the posterior distribution of the current frame, and calculate the corrected guidance path and the passable area by using the boundary of the passable area.

[0016] The above-mentioned road understanding method under the conditions of satellite denial and no prior map. In this application, a three-dimensional terrain model is constructed by fusing lidar point cloud data, and the roughness map and normal vector map are extracted. Combining the environmental passability map constructed by multi-frame fusion of lidar and odometer, multi-source data is used to describe the environment from different angles, making up for the deficiency of satellite map information and reducing the impact of positioning uncertainty and inaccurate global guidance information. Define confidence values for each map and perform evidence fusion. Fuse the evidence of the homologous roughness map and normal vector map, and then fuse it with the basic probability assignment of the environmental passability map to obtain more comprehensive and accurate finally fused data, making the understanding of the environment more reliable and providing a solid basis for determining the passable area. Finally, generate global guidance information using spline curves based on the marked points on the satellite map, determine the passable area in combination with the finally fused data, and at the same time consider the local guidance information received by the unmanned platform, calculate the expected values and standard deviations of the position and angle by integrating global and local information, and then analyze the boundary of the passable area to obtain the corrected guidance path and passable area, effectively improving the accuracy of path generation and making it more adaptable to complex unstructured environments. Brief Description of the Drawings

[0017] Figure 1 It is a schematic flowchart of a road understanding method under the conditions of satellite denial and no prior map in one embodiment;

[0018] Figure 2 It is a schematic flowchart of the BEV generation process in one embodiment;

[0019] Figure 3 It is a schematic diagram of the passability calculation process in one embodiment;

[0020] Figure 4 It is a schematic diagram of manually drawn guidance information in another embodiment;

[0021] Figure 5 It is a schematic diagram of the trajectory traveled on the passable map in one embodiment;

[0022] Figure 6 It is a structural block diagram of a road understanding device under the conditions of satellite denial and no prior map in one embodiment. Detailed implementation manners

[0023] In order to make the objectives, technical solutions and advantages of the present application clearer and more understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0024] In one embodiment, as Figure 1 shown, a road understanding method under satellite denial and without prior maps is provided, including the following steps:

[0025] Step 102, generating dense and smooth global guidance information using a spline curve according to the obtained set of path necessary points marked on the satellite map.

[0026] When driving in an unfamiliar environment, humans usually rely on the global route planned in public software, the surrounding environment, and the general road structure to maintain an overall sense of direction, and then can quickly find the key areas in the local environment by combining local environmental information. The same can be analogized for ground unmanned platforms. In any scenario, assuming that the satellite map of the unfamiliar area where the ground unmanned platform is about to perform a task can be obtained, the global guidance information is obtained through the satellite map. A human-machine interaction module is constructed, and the operator marks the set of path necessary points based on the satellite map , and the program will automatically generate dense and smooth global guidance information using a spline curve , as Figure 4 shown. The guidance information is loaded into the ground unmanned platform as the initial reference information for its autonomous navigation. The ground unmanned platform retrieves the nearby guidance information according to the real-time positioning information as the current local subtask and executes it.

[0027] Step 104, effectively fusing and accurately inferring the lidar point cloud data obtained by the ground unmanned platform at different times and different positions to construct a three-dimensional terrain model; extracting terrain roughness and normal vectors based on the three-dimensional terrain model to obtain a roughness map and a normal vector map; calculating the environmental passability map based on the BEV constructed by multi-frame fusion of lidar and odometer.

[0028] In a complex off-road environment, the terrain has a high degree of complexity and uncertainty. In order to effectively cope with the strong vibrations and large fluctuations in pose (position and attitude) caused by terrain changes, it is necessary to analyze the environmental passability using geometric information. By effectively fusing and accurately inferring the lidar point cloud data obtained by the ground unmanned platform at different times and different positions, a three-dimensional terrain model can be generated that can objectively, completely and robustly represent the spatial characteristics of the local scene where the ground unmanned platform is located .

[0029] Based on extracting terrain roughness and normal vectors features, both of which are key parameters for evaluating passability.

[0030] For a certain grid cell , let represent the elevation value of this grid cell, and calculate the maximum elevation difference between the point cloud within this grid cell and the elevation value , and use this as the measure of the terrain roughness :

[0031] (1)

[0032] Among them, represents the elevation value of the point cloud . To reduce the influence of noise, the elevation difference is normalized, and then Gaussian blur is applied to the elevation difference grid to reduce the influence of noise on the passability analysis, and thus the environmental roughness map can be obtained.

[0033] Similarly, for a certain grid cell , this application can use to represent its three-dimensional coordinates. By calculating the cross product of two orthogonal vectors formed by four adjacent points in the grid neighborhood, the normal vector is obtained:

[0034] (2)

[0035] Among them, , , and respectively represent the three-dimensional coordinates of the upper, lower, left, and right four spatial neighborhood grids of the grid cell , represents the cross product operation. To facilitate the evaluation of passability, this application calculates the cosine similarity between the normal vector and the unit vector perpendicular to the upward direction, and uses this as the evaluation result of the normal vector, and thus the environmental normal vector map is obtained. Let represent the cosine similarity between the normal vector corresponding to the grid and .

[0036] Finally, for the roughness corresponding to any grid cell and the normal vector , this application uses the artificial potential field (APF) to characterize the impact of these two geometric factors on the environmental passability:

[0037] (3)

[0038] Among them, is the natural exponential function, are the roughness coefficient and the normal vector coefficient respectively, and are the maximum and minimum thresholds for the roughness to affect passability respectively. Similarly, and are the maximum and minimum thresholds for the normal vector to affect passability respectively.

[0039] For the environmental apparent information, the bird's-eye view (BEV) of the environment is used for the environmental passability analysis. The BEV projects the surrounding environment onto a top-down representation, which can filter out information unrelated to vehicle driving, such as the sky, and is more suitable for downstream decision-making and planning tasks. As Figure 2 shown in the process, a high-density environmental bird's-eye view is constructed through point cloud projection and multi-frame fusion: First, a spatio-temporally aligned data stream is obtained through multi-modal fusion, including calibrated LiDAR point clouds, synchronized camera images (visible light or infrared), and a tightly coupled LiDAR-inertial odometry pose sequence. Then, using the pose sequence provided by the odometry data, the accumulated multi-frame radar point cloud data with color information is used to generate the environmental bird's-eye view. From the environmental bird's-eye view collected during driving, the self-supervised labels are generated by an automatic process. During the data collection process, the areas passed by the vehicle and the LiDAR perception results are recorded. The areas passed by the vehicle can be considered safe, and the areas determined by the perception module to be obstacles are considered impassable, serving as self-supervised signals for learning environmental passability. A method for learning environmental passability features using only self-supervised labels is constructed. For the bird's-eye view input at time , a passability feature extraction network is defined to predict the passability feature map , where D represents the dimension of the passability feature. The model consists of an encoder and a decoder . Through this model, a feature map of the same size as the original image is obtained.

[0040] The unstructured environmental scene is complex, with multiple feasible road surfaces, and it is difficult to determine the passability vector benchmark through experience. Implement online clustering of passable prototype vectors, as Figure 3As shown, maintain a dynamically adaptive prototype vector queue to objectively describe the passability of the environment. By calculating the average similarity between the feature F and the passable prototype vector queue as a measure of the apparent passability of the environment, and obtain the apparent-based environmental passability map through normalization Through probabilistic environmental representation, it is easy to integrate with various planning paradigms in the form of a cost map.

[0041] Step 106: Define the corresponding confidence values according to the roughness map, normal vector map, and environmental passability map, fuse the evidence of the homologous roughness map and normal vector map to obtain the initially fused evidence; fuse the basic probability assignment corresponding to the environmental passability map with the fused evidence to obtain the finally fused data.

[0042] Using a single sensor alone is not sufficient for a complete analysis of the passability of the environment. The color and texture information of the environment is analyzed using a camera, and the geometric information of the environment is analyzed using lidar. The robustness of the system can be effectively improved through multi-sensor information fusion. Based on the bird's-eye view of the environment constructed by multi-frame fusion of lidar and odometer, the calculated from it is completely aligned with the roughness map and the normal vector map in space and time. Therefore, construct an environmental passability map of the same size For any grid cell in

[0043] First, define the frame of discernment The evidence sources include and their corresponding basic probability assignments (BPA) are respectively. For any grid cell , the confidence value corresponding to each evidence source is:

[0044] (4)

[0045] Next, for the homologous results , evidence fusion:

[0046] (5)

[0047] where is the conflict degree between evidence and :

[0048] (6)

[0049] After synthesis represents a comprehensive evaluation of roughness and normal vectors. Next, this result is further fused with the basic probability assignment of the apparent information as follows:

[0050] (7)

[0051] where is the conflict degree between comprehensive evidences, and the calculation method is as shown in Equation (6).

[0052] The application of lidar point cloud data makes it possible to construct a high-precision three-dimensional terrain model. The lidar point cloud data obtained at different times and positions can, through effective fusion and accurate inference, depict the contour and features of the terrain in detail. By extracting terrain roughness and normal vectors, the obtained roughness map and normal vector map quantitatively describe the environment from the perspective of terrain characteristics. The roughness map can reflect the undulation and flatness of the ground, and the normal vector map can reflect the inclination direction and angle of the terrain. These information are crucial for judging the passability of the environment. The bird's-eye view (BEV) environment passability map constructed by fusing multiple frames of lidar and odometer analyzes the environment from another dimension. Odometer data can provide the motion information of the unmanned platform. After being fused with lidar data, it can more accurately determine the range of the passable area. The comprehensive use of multi-source data describes the environment from multiple levels, greatly enriches the understanding of the environment, reduces the uncertainty brought by a single data source, and provides more comprehensive and accurate basic data for subsequent path planning. The traditional global path planning that only relies on satellite maps cannot accurately reflect the real-time state of the environment due to the limitations of satellite map information. This application effectively makes up for this defect through the fusion of multi-source data.

[0053] Step 108: Determine the current passable area based on the global guidance information and the finally fused data; estimate the width of the current passable area, and calculate the expected value and standard deviation of the current position and angle according to the estimated width and the local guidance information received by the unmanned platform.

[0054] The path directly obtained from the satellite image usually has a large uncertainty, and the accuracy level cannot meet the requirements of local planning. Secondly, local planning is often highly sensitive to positioning errors. If the pose estimation module fails or produces inaccurate results, the planned local path will deviate significantly, increasing the risk of incorrect routes and improper driving behavior, posing a serious threat to the safety of autonomous driving. Therefore, this application determines the current passable area based on global guidance information and the final fused data, giving full play to the synergy of global guidance information and local guidance information. Based on the satellite map, the spline curve is used to generate dense and smooth global guidance information, which provides macro-directional guidance for the driving of the unmanned platform. Combined with the final fused data, the current passable area is determined, and the width of the passable area is estimated. On this basis, the local guidance information received by the unmanned platform is fully considered to calculate the expected value and standard deviation of the current position and angle. Local guidance information can reflect the real-time environmental information around the unmanned platform, such as the location and dynamic changes of nearby obstacles. By combining global and local information, the position and driving direction of the unmanned platform in the current environment can be more accurately determined, avoiding path planning errors caused by the inaccuracy of global information or the one-sidedness of local information.

[0055] Assume that there is only one traversable area in the local environment , whose geometric structure and local guidance information The geometric structure of is consistent with that of the area with the highest probability of passage in the current environment. It is an area with a certain width that conforms to the current environmental observation. First, its width is estimated. Assuming that there is no significant mutation in the driving environment in the short term, the width information of the passable area observed in the short term is used as the basis for predicting the width of the future driving area. In order to achieve this prediction, the short-term driving trajectory of the vehicle is projected onto the current environment passable map On the top, create a sample set ,in, is a set of grids covered by short-term travel, such as Figure 5 Assume that the accessibility index follows a Gaussian distribution , and obtain the parameters through maximum likelihood estimation:

[0056] (8)

[0057] To estimate the width of the current traversable area, follow the trajectory traveled in the short term The normal direction of the two-way area expansion is carried out using Confidence testing based on principles: For extension points , if satisfied , it is determined as a passable area; otherwise, it is marked as a boundary point. Through the iterative expansion and detection process, the average Euclidean distance of the bilateral boundary point set is finally calculated:

[0058] (9)

[0059] Among them, is the estimation of the width of the current passable area, are the left and right boundary points respectively.

[0060] Due to the influence of factors such as positioning errors and satellite map distortion, taking as the width, and taking the local guidance information received by the unmanned platform as the area of the structure and the initial position, there are position and angle deviations from the true passable area . Assuming the deviations , , and follow a Gaussian distribution and are independent of each other. Using the method of random sampling, in , , respectively represent the sampling range. Taking as the width assumption and as the geometric structure, a series of hypotheses of passable areas are randomly generated, where is the coordinate transformation operator, represents the expansion along the path normal , and the th candidate area is generated.

[0061] For any sampling , calculate the average passable cost within this area, and calculate the weight of this sampling deviation based on this:

[0062] (10)

[0063] Then, by calculating the weighted average and weighted variance, the expected values and standard deviations of the position and angle are obtained:

[0064] (11)

[0065] Among them, are the means obtained by observing the current environment for the deviations respectively, and are the standard deviations obtained by observing the current environment for the deviations respectively.

[0066] Step 110: Calculate the expected value and standard deviation of the current position and angle as measurement values, use the historical frame as the prior value, calculate the posterior distribution of the current frame, analyze the boundary of the traversable area based on the posterior distribution of the current frame, and calculate the corrected guidance path and the traversable area using the boundary of the traversable area.

[0067] Assume that the current is the th frame. Through the above calculations, the measurement obtained from the current environment can be obtained, where represents the measurement value, and represents the state. Taking the posterior distribution of the th frame as the prior distribution of the current frame, since the positioning information comes from the result of the previous frame and the current motion estimation and environment matching, the positioning result of the current frame is only related to the positioning result of the previous frame. Therefore, assuming that the deviation caused by positioning follows the Markov assumption, on this premise, estimate the posterior distribution of the th frame, and there is:

[0068] (12)

[0069] According to the chain rule:

[0070] (13)

[0071] where

[0072] (14)

[0073] Therefore, it can be obtained that:

[0074] (15)

[0075] where represents the normalization factor. Thus, the posterior estimate of the current frame is obtained. Based on this, analyze the boundary of the traversable area, and its calculation process is the same as (9), and the width of the traversable area can be obtained. Thus, the corrected guidance path and the traversable area can be obtained:

[0076] (16)

[0077] Taking the expected values and standard deviations of the current position and angle as measurement values, and using historical frame information as prior values, the posterior estimation of the current frame is calculated. This process fully considers the uncertainty of the measurement data. By analyzing the probability of the posterior estimation of the current frame, the boundaries of the traversable area can be determined more accurately. As the unmanned platform moves and the environment changes, the method based on posterior estimation can update the understanding of the traversable area in real time, and timely detect and correct errors in path planning. According to the boundaries of the traversable area, the corrected guidance path and the traversable area are calculated, enabling the path planning to dynamically adapt to the changes in the environment, and further improving the accuracy and adaptability of path generation.

[0078] The above method for road understanding under satellite denial and without prior map conditions. In this application, a three-dimensional terrain model is constructed by fusing lidar point cloud data, and the roughness map and normal vector map are extracted. Combining the environmental traversability map constructed by multi-frame fusion of lidar and odometer, multi-source data is used to describe the environment from different angles, making up for the deficiency of satellite map information and reducing the impact of positioning uncertainty and inaccurate global guidance information. Confidence values are defined for each map and evidence fusion is performed. The evidence of the homologous roughness map and normal vector map is fused, and then fused with the basic probability assignment of the environmental traversability map to obtain more comprehensive and accurate final fusion data, making the understanding of the environment more reliable and providing a solid basis for determining the traversable area. Finally, global guidance information is generated using spline curves based on the marked points on the satellite map. Combining the final fusion data to determine the traversable area, and considering the local guidance information received by the unmanned platform at the same time, the expected values and standard deviations of the position and angle are calculated by integrating global and local information, and then the boundaries of the traversable area are analyzed to obtain the corrected guidance path and traversable area, effectively improving the accuracy of path generation and making it more adaptable to complex unstructured environments.

[0079] In one embodiment, extracting terrain roughness and normal vectors based on the three-dimensional terrain model includes:

[0080] For a three-dimensional terrain model of a certain grid cell , let represent the elevation value of the grid cell, and calculate the maximum elevation difference between the point cloud in the grid cell and the elevation value as the measure of the terrain roughness :

[0081] ;

[0082] where represents the elevation value of the point cloud ;

[0083] For a three-dimensional terrain model a certain grid cell , use to represent its three-dimensional coordinates. By calculating the cross product of two orthogonal vectors formed by four adjacent points in the grid neighborhood, the normal vector is obtained:

[0084] ;

[0085] Among them, , , and respectively represent the three-dimensional coordinates of the upper, lower, left, and right four spatial neighborhood grids of the grid cell , represents the cross product operation.

[0086] In one embodiment, the roughness map and the normal vector map are respectively:

[0087] ;

[0088] Among them, is the natural exponential function, are the roughness coefficient and the normal vector coefficient respectively, and are the maximum and minimum thresholds for the roughness to affect passability respectively, and are the maximum and minimum thresholds for the normal vector to affect passability respectively, is the roughness corresponding to any grid cell , represents the cosine similarity between the normal vector corresponding to the grid cell and .

[0089] Fuse the evidence of the homologous roughness map and normal vector map to obtain the initially fused evidence as:

[0090] ;

[0091] Among them, is the conflict degree of the evidence of the homologous roughness map and normal vector map, is the evidence of the roughness map, is the evidence of the normal vector map, represents the custom fusion operator.

[0092] In one embodiment, fuse the basic probability assignment corresponding to the environmental passability map with the fused evidence to obtain the finally fused data, including:

[0093] Fuse the basic probability assignment corresponding to the environmental passability map with the fused evidence to obtain the final fused data as follows:

[0094] ;

[0095] Among them, is the conflict degree between comprehensive evidences, is the evidence after initial fusion, is the basic probability assignment corresponding to the environmental passability map.

[0096] In one embodiment, estimating the width of the passable area includes:

[0097] Project the short-term vehicle driving trajectory onto the current environmental passability map to establish a sample set , among which, is the set of grids covered by short-term driving;

[0098] Assume that the passability index follows a Gaussian distribution , and obtain the parameters through maximum likelihood estimation as:

[0099] ;

[0100] Among them, n represents the total number of candidate areas, represents the th candidate area;

[0101] Perform two-way area expansion along the normal direction of the trajectory traveled in the short term , and adopt the principle for confidence detection. For the expansion point , if it satisfies , then it is determined as a passable area, otherwise it is marked as a boundary point. Through the iterative expansion and detection process, finally calculate the mean Euclidean distance of the bilateral boundary point set as the width of the current passable area.

[0102] In one embodiment, calculating the expected value and standard deviation of the current position and angle according to the estimated width and the local guidance information received by the unmanned platform includes:

[0103] Taking the estimated width as the width hypothesis and the local guidance information received by the unmanned platform as the geometric structure, randomly generate a series of hypotheses of passable areas , among which is the coordinate transformation operator, represents expansion along the path normal direction , generating the A candidate area, indicating a position deviation, indicating an angle deviation;

[0104] For any sampling calculate the average traversable cost within this area , and based on this, calculate the weight of this sampling deviation, and then obtain the expected values and standard deviations of the position and angle by calculating the weighted average and weighted variance.

[0105] In one embodiment, obtaining the expected values and standard deviations of the position and angle by calculating the weighted average and weighted variance includes:

[0106] The expected values and standard deviations of the position and angle obtained by calculating the weighted average and weighted variance are:

[0107] ;

[0108] Among them, are the expected values obtained by observing the current environment for the deviation respectively, are the standard deviations obtained by observing the current environment for the deviation respectively, represents the weight of the sampling deviation, n represents the total number of candidate areas.

[0109] In one embodiment, taking the expected values and standard deviations of the current position and angle as measurement values, using the historical frame as the prior value, and calculating the posterior distribution of the current frame includes:

[0110] Taking the expected values and standard deviations of the current position and angle as measurement values , where represents the measurement value, represents the state, and taking the posterior distribution of the frame as the prior distribution of the current frame, assuming that the deviation caused by positioning follows the Markov assumption. On this premise, estimating the posterior distribution of the frame, there is:

[0111] ;

[0112] According to the chain rule:

[0113] ;

[0114] Among them,

[0115] ;

[0116] Therefore, it can be obtained that:

[0117] ;

[0118] wherein, represents the normalization factor, and thus, the posterior estimate of the current frame .

[0119] In one embodiment, the corrected guidance path and the traversable area are calculated using the traversable area boundary, including:

[0120] The corrected guidance path and the traversable area calculated using the traversable area boundary are:

[0121] ;

[0122] wherein, represents the posterior estimate of the current frame, represents the corrected guidance path, represents the corrected traversable area, represents the traversable area boundary.

[0123] In another embodiment, under the condition of GNSS denial and lack of prior map, the satellite map is used as the positioning reference. Due to the update delay of satellite images, there is a time-varying deviation between the acquired satellite images and the real environment, resulting in significant differences in the observation results of the same area. In order to reduce the influence of factors such as seasonal changes and lighting changes, the satellite map is analyzed, and the appearance-based environmental traversability analysis method is used to slice the satellite map along the global guidance information, and the prototype vector is extracted through the global guidance information for environmental traversability analysis, mapping the satellite map to the traversability space, and obtaining the environmental traversability map based on the satellite map , realizing the mapping of environmental features to the traversability space. By projecting the real-time perceived BEV features and satellite images into a unified feature space, the matching problem that the regional features are easily affected by meteorological conditions (rain, fog, smoke, and dust) and spatio-temporal variations (seasonal lighting) is effectively overcome. After obtaining the representation in the same feature space, the relevant area of the satellite map traversability map is intercepted according to the initial pose, and the similarity between and is calculated based on the normalized cross-correlation (NCC) as the particle weight. The particle pose is updated by odometer motion prediction + Gaussian noise, and the pose probability distribution is iteratively optimized to achieve global positioning.

[0124] It should be understood that although Figure 1The steps in the flowchart are shown in sequence according to the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 at least a part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or sub-steps or stages of other steps.

[0125] In one embodiment, as Figure 6 shown, a road understanding device under satellite denial and no prior map conditions is provided, including: a global guidance information generation module 602, an environmental passability map calculation module 604, a data fusion module 606, and a passable area correction module 608, where:

[0126] The global guidance information generation module 602 is configured to generate dense and smooth global guidance information using spline curves according to the obtained set of path necessary points marked on the satellite map;

[0127] The environmental passability map calculation module 604 is configured to effectively fuse and accurately infer the lidar point cloud data obtained by the ground unmanned platform at different times and different positions to construct a three-dimensional terrain model; extract terrain roughness and normal vectors based on the three-dimensional terrain model to obtain a roughness map and a normal vector map; calculate the environmental passability map based on the BEV constructed by multi-frame fusion of lidar and odometer;

[0128] The data fusion module 606 is configured to define corresponding confidence values according to the roughness map, the normal vector map, and the environmental passability map, fuse the evidence of the homologous roughness map and normal vector map to obtain the initially fused evidence; fuse the basic probability assignment corresponding to the environmental passability map with the fused evidence to obtain the finally fused data;

[0129] The passable area correction module 608 is configured to determine the current passable area based on the global guidance information and the finally fused data; estimate the width of the current passable area, calculate the expected value and standard deviation of the current position and angle according to the estimated width and the local guidance information received by the unmanned platform; use the expected value and standard deviation of the current position and angle as measurement values, use the historical frame as the prior value, calculate the posterior distribution of the current frame, analyze the boundary of the passable area according to the posterior distribution of the current frame, and calculate the corrected guidance path and the passable area using the boundary of the passable area.

[0130] For the specific limitations of the road understanding device under satellite denial and without prior maps, reference can be made to the limitations of the road understanding method under satellite denial and without prior maps in the foregoing text, which will not be elaborated herein. Each module in the above-mentioned road understanding device under satellite denial and without prior maps can be implemented in whole or in part by software, hardware, and their combination. The above-mentioned modules can be embedded in the processor of the computer device in hardware form or be independent of it, or can be stored in the memory of the computer device in software form, so as to facilitate the processor to call and execute the operations corresponding to the above-mentioned modules.

[0131] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0132] The above-described embodiments only represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several deformations and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A road understanding method under satellite denial and without prior maps, characterized in that The method includes: Generating dense and smooth global guidance information using a spline curve according to the obtained satellite map marked with the set of path necessary points; Effectively fusing and accurately inferring the lidar point cloud data obtained by the ground unmanned platform at different times and positions to construct a three-dimensional terrain model; extracting terrain roughness and normal vectors based on the three-dimensional terrain model to obtain a roughness map and a normal vector map; calculating an environmental traversability map based on the multi-frame fusion of lidar and odometer; Defining corresponding confidence values according to the roughness map, the normal vector map, and the environmental traversability map, fusing the evidence of the homologous roughness map and normal vector map to obtain initially fused evidence; fusing the basic probability assignment corresponding to the environmental traversability map with the fused evidence to obtain finally fused data; Determining the current traversable area based on the global guidance information and the finally fused data; estimating the width of the current traversable area, and calculating the expected values and standard deviations of the current position and angle according to the estimated width and the local guidance information received by the unmanned platform; Using the expected values and standard deviations of the current position and angle as measurement values, using the historical frame as a prior value, calculating the posterior distribution of the current frame, analyzing the boundary of the traversable area according to the posterior distribution of the current frame, and calculating the corrected guidance path and the traversable area using the boundary of the traversable area; fusing the basic probability assignment corresponding to the environmental traversability map with the fused evidence to obtain finally fused data, including: Fusing the basic probability assignment corresponding to the environmental traversability map with the fused evidence, and the finally fused data is: ； Among them, is the conflict degree between comprehensive evidences, is the evidence after initial fusion, is the basic probability assignment corresponding to the environmental passability map.

2. The method according to claim 1, wherein Extracting terrain roughness and normal vectors based on the three-dimensional terrain model, including: For a three-dimensional terrain model of a certain grid cell , let represent the elevation value of the grid cell, and calculate the maximum elevation difference between the point cloud within the grid cell and the elevation value as a measure of the terrain roughness : ； Among them, represents the elevation value of the point cloud ; For a three-dimensional terrain model of a certain grid cell , use to represent its three-dimensional coordinates. By calculating the cross product of two orthogonal vectors formed by four adjacent points within the grid neighborhood, the normal vector is obtained as follows: ； Among them, , , and respectively represent the three-dimensional coordinates of the four spatial neighboring grids of the upper, lower, left, and right of the grid cell , and represents the cross product operation.

3. The method according to claim 1, wherein The roughness map and the normal vector map are respectively: ； Among them, is the natural exponential function, are the roughness coefficient and the normal vector coefficient respectively, and are the maximum and minimum thresholds for the roughness to affect passability respectively, and are the maximum and minimum thresholds for the normal vector to affect passability respectively, is an arbitrary grid cell corresponding to the roughness, represents the cosine similarity between the normal vector corresponding to the grid cell and ; Fusing the evidence of the homologous roughness map and normal vector map to obtain initially fused evidence as: ； Among them, is the conflict degree of evidence of the homologous roughness map and normal vector map, is the evidence of the roughness map, is the evidence of the normal vector map, represents a custom fusion operator.

4. The method according to claim 1, wherein Estimating the width of the traversable area, including: Project the short-term driving trajectory of the vehicle onto the passable map of the current environment to establish a sample set , where is the set of grids covered by the short-term driving; Assume that the passability index follows a Gaussian distribution , and the parameters are obtained by maximum likelihood estimation as follows: ； Among them, n represents the total number of candidate regions, represents the th candidate region; Perform bidirectional region expansion along the normal direction of the trajectory traveled in the short term and use the principle for confidence detection. For the expanded points , if it satisfies , it is determined as a passable area; otherwise, it is marked as a boundary point. Through the iterative expansion and detection process, finally calculate the mean Euclidean distance of the bilateral boundary point set as the width of the current passable area.

5. The method according to claim 1, wherein Calculating the expected values and standard deviations of the current position and angle according to the estimated width and the local guidance information received by the unmanned platform, including: Using the estimated width as the width hypothesis, and using the local guidance information received by the unmanned platform as the geometric structure, a series of hypotheses of passable regions are randomly generated , where is the coordinate transformation operator, indicating expansion along the path normal to generate the th candidate region, representing the position deviation, representing the angular deviation; For any sampling Calculate the average traversable cost within this area , and based on this, calculate the weight of this sampling deviation, and then obtain the expected values and standard deviations of the position and angle by calculating the weighted average and weighted variance.

6. The method according to claim 5, wherein Obtaining the expected values and standard deviations of the position and angle by calculating the weighted average and weighted variance, including: Obtaining the expected values and standard deviations of the position and angle by calculating the weighted average and weighted variance as: ； Among them, are the expected values obtained by observing the deviation through the current environment, are the standard deviations obtained by observing the deviation through the current environment, represents the weight of the sampling deviation, n represents the total number of candidate regions.

7. The method according to claim 1, characterized in that, Using the expected values and standard deviations of the current position and angle as measurement values, using the historical frame as a prior value, and calculating the posterior distribution of the current frame, including: Use the expected value and standard deviation of the current position and angle as measurement values , where represents the measurement value, represents the state, with the posterior distribution of the th frame used as the prior distribution of the current frame. Assuming that the deviation caused by positioning follows the Markov assumption, under this premise, estimate the posterior distribution of the th frame , we have: ； According to the chain rule: ； Wherein, ； Therefore, it can be obtained that: ； Among them, represents the normalization factor, and thus, the posterior estimate of the current frame is obtained .

8. The method according to claim 1, wherein Calculating the corrected guidance path and the traversable area using the boundary of the traversable area, including: Calculating the corrected guidance path and the traversable area using the boundary of the traversable area as: ； Among them, represents the posterior estimate of the current frame, represents the corrected leading path, represents the corrected passable area, represents the boundary of the passable area.

9. A road understanding device under satellite denial and no prior map conditions, characterized in that, The device includes: A global guidance information generation module, configured to generate dense and smooth global guidance information using a spline curve according to the obtained satellite map marked with the set of path necessary points; An environmental passability map calculation module is used to effectively fuse and accurately infer the lidar point cloud data obtained by the ground unmanned platform at different times and positions to construct a three-dimensional terrain model; extract terrain roughness and normal vectors based on the three-dimensional terrain model to obtain a roughness map and a normal vector map; calculate an environmental passability map based on the BEV constructed by multi-frame fusion of lidar and odometer; A data fusion module is used to define corresponding confidence values according to the roughness map, the normal vector map, and the environmental passability map, fuse the evidence of the homologous roughness map and normal vector map to obtain initially fused evidence; fuse the basic probability assignment corresponding to the environmental passability map with the fused evidence to obtain finally fused data, including: Fusing the basic probability assignment corresponding to the environmental passability map with the fused evidence, the finally fused data obtained is: ； Among them, is the conflict degree between comprehensive evidences, is the evidence after initial fusion, is the basic probability assignment corresponding to the environmental passability map; A passable area correction module is used to determine the current passable area based on the global guidance information and the finally fused data; estimate the width of the current passable area, calculate the expected value and standard deviation of the current position and angle according to the estimated width and the local guidance information received by the unmanned platform; use the expected value and standard deviation of the current position and angle as measurement values, use the historical frame as a prior value, calculate the posterior distribution of the current frame, analyze the boundary of the passable area according to the posterior distribution of the current frame, and calculate the corrected guidance path and the passable area by using the boundary of the passable area.

Images