Land complex environment-oriented path planning system based on three-dimensional voxel lattice

Abstract

The invention discloses a land complex environment-oriented path planning system based on three-dimensional voxel grids, which is characterized by comprising a land complex environment maneuverability evaluation and prediction module, a three-dimensional voxel grid establishment and optimization module and a three-dimensional voxel grid-based path planning module. The technical scheme disclosed by the invention mainly starts from three parts: firstly, the problem of interaction between a carrier and terramechanics is solved through regional maneuverability evaluation and prediction, and traffic capacity evaluation of the carrier in regional maneuverability is obtained after factors such as terrain, landform, soil, soil texture and vehicle power are considered; then three-dimensional voxel data of land path planning is formed through three-dimensional voxel lattice representation of the multi-dimensional environment elements; and finally, realizing a path planning task in a complex environment based on the three-dimensional voxel lattice.

Description

Technical Field

[0001] The invention relates to a novel path planning technology. Background Art

[0002] Commercial navigation software is already quite mature, such as Baidu Maps and Amap. However, these commercial navigation applications are more for civilian use, relying more on road network data and basic traffic information to support users' navigation needs, and cannot solve the problem of vehicle navigation in complex off-road environments. At present, domestic research on vehicle off-road maneuverability navigation is mostly focused on the evaluation of vehicle maneuverability performance and obstacle avoidance algorithms, and less consideration is given to vehicle path planning and navigation capabilities in complex and changeable off-road environments without road networks.

[0003] The demand for navigation in special environments and special scenarios is gradually increasing. Whether it is the NRMM (NATO Reference Mobility Model) developed and verified by NATO in the 1960s and 1970s, or the NG-NRMM (Next-Generation NRMM) developed by NATO in 2016 when the AVT-248 Committee was established, these are all multifunctional path planning technology research and development for special environments and complex scenarios. my country is also gradually realizing the importance of this aspect and focusing on the development of this new navigation technology.

[0004] In terms of off-road mobility performance evaluation, considering that one-third of my country's land area is plateau and mountainous, and some special plateau and mountainous areas have complex terrain, the Beidou Navigation Innovation Institute of Shanghai Jiao Tong University has been working in the field of Beidou navigation for a long time and has established a speed calculation model for off-road vehicles under different road conditions. The calculation model contains vehicle structural parameters, dynamic parameters and road condition parameters; Southeast University cooperated with the General Vehicle Department to conduct research on the bump passability of off-road vehicles through multi-body dynamics modeling and simulation; Ning Junshuai and others from the Armored Forces Engineering Academy proposed a comprehensive evaluation process based on the "task model" and "driving model" methods.

[0005] In terms of path planning under off-road conditions, the research of domestic scholars is relatively weak, and is relatively concentrated on the research of obstacle avoidance algorithms, and improving the shortest path planning algorithm suitable for road conditions to meet the path planning requirements under off-road conditions. For example, Tian Xiping et al. optimized the distance and direction of the classic Dijkstra algorithm, and proposed an obstacle avoidance algorithm based on direction for the analysis range of off-road maneuvering routes; Wang Qisheng et al. proposed to use an optimized maze algorithm to analyze the shortest path for off-road path planning, and proposed an optimization algorithm based on the path obtained by the maze algorithm by using the idea of ​​replacing curves with straight lines and spline models; Wu Tianyi et al. analyzed the path planning problem under the combined influence of terrain slope and surface attributes for the off-road path planning problem of different types of vehicles, constructed the evaluation function of the improved A* algorithm and the evaluation function of the ant colony algorithm, and designed a path optimization algorithm considering slope and roughness constraints; Fan Linlin et al. comprehensively considered the influence of terrain elements and landforms on off-road vehicle maneuvers, and based on this, proposed an optimization scheme for the heuristic A* algorithm, firstly expressing the complex terrain environment by terrain quantization, and then optimizing the evaluation function of the A* algorithm to explore the problem of time shortest path planning under the premise of determining the starting point and the end point;

[0006] Yan Xingyu et al. studied and designed a hierarchical off-road path planning method based on passability analysis to meet the needs of large-scale off-road path planning in complex off-road environments. This method focuses on the efficiency and feasibility of large-scale path planning, and proposes a hierarchical planning method. It realizes hierarchical path planning by constructing two map grids with different resolutions, and conducts regional passability analysis in combination with surface coverage and surface roughness, and constructs the evaluation function of the A* algorithm in a targeted manner. Tian Hongqing et al. proposed a probability map (AFP-PRM) algorithm based on the potential field model, which uses an artificial potential field algorithm to model the off-road environment, evaluates the vehicle traffic risk, and uses a probability map algorithm to optimize the multi-dimensional traffic cost between nodes for path planning. Considering the vehicle dynamics, the dynamic curvature smoothing method is used to optimize the driving trajectory. Li Kunwei et al. proposed a two-way path planning solution by analyzing the elements of path planning in a dynamic environment. In view of the particularity of off-road path planning in a dynamic environment, the search method and search step size of the two-way path planning are set, and a cost estimation function is constructed. Finally, a two-way A* search algorithm suitable for dynamic path planning is proposed. Liang Xiaohui and others from the Beijing University of Aeronautics and Astronautics used genetic algorithms for trajectory planning, but the algorithm took a long time to converge, and the generated path did not fully consider terrain information such as rivers, and the generated path was not suitable for the actual complex off-road environment. Chen Qijie divided the existing path calculation methods into traditional methods and intelligent methods.

[0007] Ji Qing sorted out the corresponding path calculation methods based on different modeling methods; the most widely used classification method is to divide it into global path calculation method and local path calculation method according to the degree of mastery of environmental information. The global path calculation method is to select a suitable path calculation method after preprocessing and modeling when the environmental map information is completely known. Its efficiency depends entirely on the choice of algorithm, taking into account both optimality and accessibility. This method may not take into account details from a global perspective. The environmental information faced by the local path calculation method is unknown, so modeling and search are carried out simultaneously. It is mostly used in mobile robots and unmanned vehicles with autonomous navigation information. This method requires real-time processing of environmental information and timely response to various emergencies through fast computing power. At the same time, the path searched by this method based on local information may not be very accurate, or even no path can be found. The two models are essentially the same, but compared with the global model, the local path calculation is more complicated. After improvement, the two model methods can be converted to each other. Wei Guanwei et al. established a penalty function neural network to quantify the collision between the path and the obstacle and provide path guidance for the robot. The neural network algorithm involves many parameters, but at this stage there is no clear theoretical basis for the design of the network model, and it is still in a black box state. Duguleana and Mogan et al. proposed combining neural networks and reinforcement learning to improve the ability of mobile robots and other moving bodies to deal with static and dynamic obstacles. In the widely used deep reinforcement learning method, neural networks can handle path finding tasks well on their own. Studies have shown that moving bodies using deep Q-learning algorithms can robustly learn to achieve goals on maps and show good performance on maps that the moving body has never appeared before.

[0008] There is a large increase in navigation needs in special environments and complex scenes in my country's non-civilian fields, and work in this area has also begun. The present invention has developed and designed a path planning technology for complex land environments based on three-dimensional voxels in response to the above needs. The path planning technology will be able to generate an integrated three-dimensional voxel grid with multiple attributes and support path planning capabilities in complex land environments.

[0009] The technical problems in off-road mobile navigation in China can be summarized as follows:

[0010] (1) Single factor evaluation. Mobility evaluation mainly relies on the commander's experience, with insufficient quantitative decision-making support and single factors. For example, only terrain factors are considered, without considering the combined impact of weather, soil quality, natural disasters, emergencies and other factors. The increase in the types and functions of ground vehicles, as well as the diversity of geological types, geometric complexity and other complex environments have brought high uncertainty to the model parameters and performance evaluation of vehicles, resulting in an urgent need for adaptive and extensible improvements to the system mechanics model.

[0011] (2) Insufficient consideration of real-time situation information. Current research mainly focuses on static information such as roads and terrain, and does not comprehensively consider the impact of real-time situation information such as weather, meteorology, and emergencies on off-road mobile navigation.

[0012] (3) Modeling and prediction capabilities are insufficient. Algorithms and models are optimized mainly based on the basic geographic information system (GIS). There is a lack of cross-domain joint modeling and analysis capabilities, such as vehicle dynamics, vehicle-soil interaction mechanics modeling, and the ability to estimate maneuvering speed, fuel consumption, and travel time in off-road environments.

[0013] (4) The algorithms in current path planning and navigation software are not very practical. They are mainly based on simple shortest path analysis based on existing road network data, which is far from practical needs. They do not take into account off-road maneuvers and complex off-road environments, such as rain, snow, ice, strong winds, heavy fog and the resulting soil moisture and slipperiness. Summary of the invention

[0014] The technical problem to be solved by the present invention is that my country started late in the research on the mobility performance evaluation and navigation of vehicles in complex off-road environments. Although there is a certain theoretical research foundation, the pertinence and systematicness are weak. Most of the algorithm verification work is based on foreign simulation software and platforms. There is a lack of comprehensive mobility evaluation products, insufficient practical testing, and the level of autonomous control needs to be improved. The acquisition of ground mechanical parameters is still at the level of following, which restricts the progress of my country's off-road mobility capability evaluation technology.

[0015] In order to solve the above technical problems, the technical solution of the present invention is to disclose a path planning system for complex land environments based on three-dimensional voxel lattices, which is characterized by comprising a complex land environment mobility evaluation and prediction module, a three-dimensional voxel lattice establishment and optimization module, and a path planning module based on three-dimensional voxel lattices, wherein:

[0016] The land complex environment mobility assessment and prediction module further includes:

[0017] The intelligent obstacle recognition unit for complex land environments builds an intelligent obstacle recognition model based on scene scale and DCNN transfer learning to quickly identify obstacles in complex land environments;

[0018] The ground mechanics parameter interpretation unit for complex land environments provides a data basis for the coupling analysis of vehicle and ground mechanics, classifies and identifies the soil, obtains soil moisture, and then realizes the construction of ground mechanics parameters;

[0019] The precise inversion unit of the ground mechanical parameters of the coupled vehicle is used to establish a knowledge base of the corresponding relationship between the mechanical properties of known soil types, so as to obtain a coupled quantitative calculation method that satisfies the ground mechanical characteristic parameters of different passing vehicles and standard classification, thereby realizing the inversion of the ground mechanical characteristic parameters of unknown areas and the evaluation of the vehicle's traffic performance;

[0020] The whole vehicle dynamics model is based on vehicle-ground interaction. The whole vehicle dynamics model is determined by the type of vehicle and ground. The vehicle type is determined by the running mechanism. Based on the interaction mechanism between different running mechanisms and ground, a multi-degree-of-freedom mechanical model of the vehicle in a three-dimensional unstructured environment is established to provide a model basis for the evaluation of vehicle passability and maneuverability.

[0021] The vehicle passability assessment unit based on multivariate analysis is used to assess the vehicle's passability in a specified area;

[0022] The maneuvering speed prediction unit, which combines vehicle performance parameters and dynamic models, is used to evaluate the vehicle's running speed. It also combines vehicle test data, ground comprehensive characterization model and vehicle maneuvering speed to obtain a driving stability classification based on deep learning for vehicle running safety. At the same time, it combines vehicle stability and speed to conduct a study on rapid maneuvering rules, form a rapid maneuvering area evaluation model, and obtain a judgment on the rapid maneuvering area.

[0023] The ground mechanics parameter interpretation semantic map construction unit uses the form of semantic maps to characterize the ground mechanics parameter characteristic inversion results of multi-source data. By referring to the semantic map construction method, the process of estimating the ground pressure and shear characteristics based on vision is integrated into the ground modeling process to construct a ground mechanics parameter characteristic inversion semantic map that integrates ground geometry and mechanical characteristics, thereby providing terrain geometry and force for vehicle traffic simulation, path planning and mobile control;

[0024] The 3D volume grid building and optimization module further includes:

[0025] Based on the three-dimensional traversable area map construction unit of multi-element voxel grid, an abstract representation model is established for the core elements of complex off-road environment related to the mobility characteristics of ground vehicles and the interaction between vehicles and the environment. The spatial topological map structure based on the voxel model is used for the digital representation of unstructured complex terrain and complex off-road environment;

[0026] Complex off-road environmental element abstract representation unit, used to realize the abstract representation of complex off-road environmental elements;

[0027] The map data structure unit based on the spatial topological relationship of the voxel model is used to establish the spatial topological relationship based on the voxel model, and cluster the spatial information based on the semantic point cloud by using the spatial continuity between different instance objects. Each cluster obtained will be considered as an independent object in the space. According to the number of categories of semantic segmentation, semantic map objects under different categories are established respectively; in the map data structure, the map of the semantic segmentation part is combined with the global voxel grid map; the voxel grid object will be managed by the global voxel grid set and organized according to the object instance, which is used to predict the vehicle inaccessible area in the subsequent path planning. At the same time, in order to realize the classification of each category The perception of object instances in the semantic map is analyzed. In each category of the spatial semantic map of complex off-road environments, an incremental density clustering algorithm is designed based on voxel grid elements. All qualified voxel grids are retrieved within a certain neighborhood according to the voxel grid coordinates. The topological relationship is generated by using the intrinsic association between the voxel grid instance information after semantic segmentation. The voxel model reflects the topological relationship between spatial objects through the adjacency relationship between cubes. The adjacency relationship describes the topological relationship between two adjacent voxels. The topological relationship between multiple continuous voxels is described by using the path and connectivity relationship, thereby generating a complex terrain map data structure of a complex off-road environment.

[0028] Path planning module based on 3D voxel grid:

[0029] Utilizing a large amount of empirical data from ground vehicle passability tests, a data-driven global path planning algorithm is obtained based on a three-dimensional traversable domain map of multi-element voxel lattices. According to differentiated application needs and ground vehicle mobility requirements, an optimization objective function with multi-state constraints is established. A hierarchical search is performed on the voxelized complex off-road environment map, and a heuristic search algorithm is used to iteratively search for traversable node calculations. Combined with a deep neural network, empirical traversable path results with different rewards are obtained to achieve the optimal set of path points that meet the objective function.

[0030] Preferably, in the ground mechanical parameter interpretation unit for complex onshore environments, a rock and soil classification system for ground mechanical properties is constructed, the ground rock and soil characteristic types in complex off-road environments are rapidly surveyed and identified, a soil moisture inversion method based on remote sensing data is performed, and ground soil mechanical property parameters in complex onshore environments are constructed.

[0031] Preferably, in the whole vehicle dynamics model, the vehicle is divided into two types of walking mechanisms: tracked and wheeled.

[0032] Preferably, in the vehicle passability assessment unit: the ground is parametrically characterized, road surface information, terrain information, geomorphic information and climate information are integrated, and feature extraction and selection are performed on various types of information based on a combination of subjective and objective methods to obtain feature types with a high correlation with vehicle passability, thereby constructing a ground comprehensive characterization parameter model covering geometric and geological characteristics;

[0033] In terms of the vehicle's own characteristics, the theoretical design or experimental experience of structural dimensions, dynamic characteristics, and performance indicators is combined to deconstruct and analyze the vehicle's own characteristics. The factors and indicators that affect the passability are studied in combination with the ground comprehensive characterization parametric model. The mutual influence mechanism between vehicle parameters, ground characterization and passability is revealed, a quantitative analysis relationship is established, and a multi-dimensional judgment evaluation model is formed.

[0034] Preferably, the maneuvering speed prediction unit studies a maneuvering speed prediction method of obtaining a power flow model based on a passability evaluation and on different types of vehicles and ground, based on vehicle power transmission characteristics and a vehicle-ground coupling dynamics model.

[0035] Preferably, in the abstract representation unit of complex off-road environment elements, for terrain soil voxels, it is necessary to expand the solid voxels, add the element attributes of the soil elements, and use the octree data structure to compress the terrain with the same type of attributes, and then realize multi-resolution terrain isosurface extraction through indirect visualization methods.

[0036] Preferably, the map data structure unit utilizes the characteristics of Hash mapping and uses the std::map object provided by the C++STL library to construct a key-value pair of voxel grid ID and voxel grid object. The voxel grid does not need to be created in advance until the sensor observation of the system platform is converted to the reference coordinate system. If a certain observation falls into the voxel grid object created in the complex off-road environment space, the voxel grid at that position will be dynamically created.

[0037] Preferably, the three-dimensional voxel-based path planning module further provides a LSTM-based neural network path planning method, which reorganizes the imported data into a three-dimensional voxel form with multiple element attributes based on an octree structure, and finally forms a three-dimensional traversable area map;

[0038] The generated three-dimensional traversable area map is used as the basic map for the path planning part. When the user needs path planning, he / she inputs the corresponding instructions into the system according to his / her mobile starting point, end point and the application requirements with the highest priority. After reading the code corresponding to the instruction, the system immediately initiates the path planning:

[0039] After receiving the user input command, the subsystem first establishes a specific path optimization function f(x). The path optimization function f(x) is related to the application requirements, navigation unit model and traversable area map. According to the generated three-dimensional traversable area map and f(x), the vehicle field maneuvering area map is divided into several voxel grids V of different scales. i In the three-dimensional feasible domain map, each voxel has a separate number and rich attribute parameters. The specific path optimization process is as follows:

[0040] a) The subsystem first searches for the trajectory on a three-dimensional feasible domain map with a larger resolution to determine a rough path trajectory;

[0041] b) Based on the rough path trajectory, the path is accurately optimized again on a map with a higher resolution according to the optimization function;

[0042] c) After going through the above two steps multiple times, a precise path trajectory P is finally determined ab and its coordinate points.

[0043] In order to solve the problems pointed out in the background technology, the technical solution disclosed in the present invention mainly starts from three parts: first, the interaction problem between the carrier and the ground mechanics is solved through regional mobility evaluation and prediction, and the carrier's traffic capacity evaluation in regional mobility is obtained after considering factors such as terrain, landform, soil, soil quality, vehicle power, etc.; then, three-dimensional voxel data for land path planning is formed through the three-dimensional voxel grid representation of multi-dimensional environmental elements; finally, the path planning task in a complex environment is realized based on the three-dimensional voxel grid.

[0044] In view of the problems of complex unstructured off-road environment, insufficient consideration of real-time situation information, and weak autonomous decision-making ability during mobile navigation in complex environments, this paper proposes a task-driven three-dimensional voxel real-time path planning technology. This new path planning technology has the following two advantages:

[0045] (1) Diverse evaluation factors and strong real-time performance. Based on the accurate evaluation results of vehicle mobility in complex environments, this technology uses dynamic events such as maneuvering tasks, weather and meteorological changes, natural disasters, road traffic, electromagnetic confrontation and threat situations acquired in real time as important planning parameters for path planning. The parameter information is stored and represented in the form of three-dimensional voxels. During the path planning process, the task information and dynamic events are comprehensively evaluated to achieve real-time response to maneuvering tasks and dynamic emergencies.

[0046] (2) Sufficient modeling and prediction capabilities, excellent practicality. By constructing a multi-element voxel grid, the unstructured ground environment can be expressed, and the three-dimensional traversable domain map can be managed using a dynamic octree. Each voxel grid point can store the traversable / impassable information generated by the mobility assessment module, extended information such as terrain, obstacle semantics and fuel consumption, task attributes, navigation performance evaluation results and other dynamic event elements. The three-dimensional voxel grid integrates the mobility preparation assessment results and navigation performance assessment results, and real-time access to dynamic information will make the path planning results more reasonable and correct. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 The overall technical idea of ​​the solution of the present invention is illustrated;

[0048] Figure 2 The overall idea of ​​mobility assessment and prediction in complex land environments is illustrated;

[0049] Figure 3 The study of ground comprehensive characterization parameter model is illustrated;

[0050] Figure 4 The maneuvering speed prediction is shown;

[0051] Figure 5 The abstract process of complex off-road environment elements is illustrated;

[0052] Figure 6 Schematic diagram of terrain modeling for basic volume elements;

[0053] Figure 7 Demonstrates real-time environment understanding based on voxel grid maps;

[0054] Figure 8 It is a voxel grid topology diagram. DETAILED DESCRIPTION

[0055] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0056] like Figure 1As described above, after acquiring a large amount of geographic information data, the present invention performs mobility evaluation and navigation performance evaluation on the current complex environment based on the characteristics of the navigation carrier, obtains the carrier's trafficability data and navigation positioning performance data in the current environment respectively, and then reorganizes the data into a three-dimensional voxel grid with multiple element attributes based on the octree structure, and finally forms a three-dimensional traversable area map. The generated three-dimensional traversable area map is used as the basic map of the path planning part. After the maneuvering starting point and end point are given, path planning will be performed and the corresponding planning results will be returned.

[0057] Based on the above ideas, the present invention discloses a path planning system based on three-dimensional voxels for complex land environments, including:

[0058] 1. Mobility assessment and prediction module in complex land environments

[0059] The overall idea of ​​the module for evaluating and predicting mobility in complex land environments is as follows: Figure 2 As shown in the figure, the research content is mainly divided into the following five parts:

[0060] 1) Intelligent recognition unit for complex obstacles in land environments

[0061] In response to the demand for rapid identification of obstacles in complex land environments, research on intelligent identification methods for obstacles in complex land environments based on scene scale and DCNN transfer learning is carried out. The difficulties are: the high semantic level and clear physical boundaries make it difficult to interpret intelligently, there are few samples, and a suitable scale is required to characterize its overall characteristics. Therefore, in response to the problems of low intelligence and accuracy in obstacle identification, the present invention constructs an intelligent obstacle identification model based on scene scale and DCNN transfer learning. The present invention utilizes public data sets and DCNN transfer learning strategies; using scenes as primitives, characterizes the overall characteristics of obstacles and complex semantic pattern construction methods.

[0062] 2) Interpretation unit of ground mechanical parameters in complex land environment

[0063] The interpretation of geomechanical parameters in complex land environments provides a data basis for the coupling analysis of vehicle and ground mechanics. Because the ground mechanical properties in complex land environments are greatly affected by soil type and soil moisture. Therefore, in order to interpret the ground mechanical parameters in complex land environments, it is necessary to classify and identify the soil, obtain soil moisture, and then realize the construction of ground mechanical parameters. The content includes: constructing a rock and soil classification system for ground mechanical properties, rapid survey and identification of ground rock and soil characteristic types in complex off-road environments, soil moisture inversion method based on remote sensing data, and construction of ground soil mechanical property parameters in complex land environments.

[0064] 3) Precise inversion unit of ground mechanical parameters of coupled vehicles

[0065] In order to quantitatively describe the relationship between ground rock and soil types and their corresponding mechanical property parameters, as well as the mechanical characteristics corresponding to different vehicle types, the mechanical characteristics of soil types that affect vehicle traffic are studied, and a knowledge base of the corresponding mechanical properties of known soil types is established to obtain a coupled quantitative calculation method that satisfies the mechanical property parameters of different passing vehicles and standard classification of the ground, thereby realizing the inversion of ground mechanical property parameters in unknown areas and the evaluation of vehicle traffic performance.

[0066] 4) Full vehicle dynamics model based on vehicle-ground interaction

[0067] The vehicle-ground interaction model is mainly determined by the type of vehicle and ground. The characterization parameters of different types of ground are not the same, and the type of vehicle is mainly determined by the type of its running mechanism. Existing vehicles are mainly divided into two types of running mechanisms: tracked and wheeled, while the types of ground exist in many different ways due to their components. Therefore, studying the interaction mechanism between tracks / wheels and different types of ground can effectively reflect their interaction relationship, obtain the corresponding types of mechanical parameters of vehicles and ground, and provide guidance for the inversion of ground mechanical parameters. Based on the interaction mechanism of different running mechanisms and ground, a multi-degree-of-freedom mechanical model of vehicles in a three-dimensional unstructured environment is established to provide a model basis for the evaluation of vehicle passability and maneuverability.

[0068] 5) Vehicle passability evaluation unit based on multivariate analysis

[0069] Vehicle passability mainly refers to the ability of a vehicle to pass through a specified area, which is a GO / NOGO judgment issue. The passability assessment study must first parametrically characterize the ground, integrate road surface information, terrain information, geomorphology information and climate information, extract and select features of various types of information based on a combination of subjective and objective methods, and obtain feature types that are highly correlated with vehicle passability. In this way, a comprehensive ground characterization parameter model covering geometric and geological characteristics is constructed, such as Figure 3 shown.

[0070] In terms of the vehicle's own characteristics, the vehicle's own characteristics are deconstructed and analyzed in combination with theoretical design or experimental experience such as structural dimensions, dynamic characteristics, and performance indicators. The factors and indicators that affect passability are studied in combination with the ground comprehensive characterization parametric model. The mutual influence mechanism between vehicle parameters, ground characterization and passability is revealed, a quantitative analysis relationship is established, and a multi-dimensional judgment evaluation model is formed.

[0071] 6) Maneuvering speed prediction unit combining vehicle performance parameters and dynamics model

[0072] Maneuverability assessment is mainly aimed at the assessment of vehicle running speed. On the basis of passability assessment, for different types of vehicles and ground, based on vehicle power transmission characteristics and vehicle-ground coupling dynamics model, the maneuverability speed prediction method based on power flow model is studied, such as Figure 4 As shown. In terms of vehicle operation safety, a deep learning-based driving stability classification study is conducted by combining vehicle test data, ground comprehensive characterization model and vehicle maneuvering speed. For the judgment of rapid maneuvering areas, it is necessary to combine vehicle stability and speed, conduct a study on the rules of rapid maneuvering, form a rapid maneuvering area evaluation model, and thus realize the judgment of rapid maneuvering areas.

[0073] 7) Ground mechanics parameter interpretation semantic map construction unit

[0074] In order to better interpret the inversion results of geomechanical parameter characteristics of multi-source data, the inversion results of geomechanical parameter characteristics of multi-source data are characterized by borrowing the form of semantic maps. By borrowing the semantic map construction method, the process of estimating the ground pressure and shear characteristics based on vision is integrated into the ground modeling process to construct a semantic map of geomechanical parameter characteristics inversion that integrates ground geometry and mechanical characteristics, thereby providing terrain geometry and force for vehicle traffic simulation, path planning and mobile control.

[0075] 2. 3D volume grid establishment and optimization module

[0076] Path planning and maneuvering guidance require maps that can reflect the terrain, objects, strata, and geological features. To this end, the present invention first develops an abstract representation model for the core elements of the complex off-road environment associated with the maneuverability of ground vehicles and the interaction between vehicles and the environment; then studies the spatial topological map structure based on the voxel model for the digital representation of unstructured complex terrain and complex off-road environments.

[0077] Based on the prior knowledge of the impact of geographical environment and vehicle performance on vehicle trafficability, the present invention studies the use of voxel models to achieve comprehensive modeling and expression of various elements of ground vehicle traffic in complex off-road environments.

[0078] Abstract representation of outdoor environmental elements in complex off-road environments. Research on the abstract representation method of landforms, landforms, and underground environmental elements in a specific area. Based on the octree voxel structure, the complex off-road environment terrain modeling process uses regular voxels to realize grid segmentation, realize the discretization of elements such as terrain surface and underground geology, and record the attributes of various elements related to the current complex off-road environment in each voxel, such as terrain elevation attributes, temperature, humidity, geology and other attributes.

[0079] Map data structure based on the spatial topological relationship of voxel models. Research data structures with fast three-dimensional spatial search capabilities and data compression capabilities, establish the spatial topological relationship of voxel models, and construct a multi-element voxel grid three-dimensional traversable domain map to support subsequent research on ground vehicle path planning algorithms.

[0080] 1) 3D traversable area map construction unit based on multi-element voxel grid

[0081] Path planning and maneuvering guidance require maps that can reflect the terrain, objects, strata, and geological features. To this end, the present invention first develops an abstract representation model for the core elements of the complex off-road environment associated with the maneuverability of ground vehicles and the interaction between vehicles and the environment; then studies the spatial topological map structure based on the voxel model for the digital representation of unstructured complex terrain and complex off-road environments.

[0082] 2) Abstract representation unit of complex off-road environment elements

[0083] Research on vehicle path planning for complex terrain, the specific abstract representation process of complex field environment elements such as Figure 5 shown.

[0084] The modeling process of complex off-road environment based on voxel model is explained by taking the modeling of terrain body using octree voxel structure as an example. The terrain body can include terrain surface, underground geology and aboveground air. First, the "boundary bounding volume" of the specified area is determined, and the resolution of terrain voxels is specified. Then, the terrain surface, underground geology and aboveground air are voxel segmented, and the volume is discretized using solid, liquid (groundwater) and gas voxels respectively.

[0085] like Figure 6 As shown in the figure, for terrain soil voxels, the solid voxels need to be expanded and the element attributes of soil elements need to be added, such as soil type, water content, water infiltration capacity, compression capacity, shear resistance, etc. The octree data structure is used to compress terrains with the same type of attributes, and then the multi-resolution terrain isosurface extraction is realized through indirect visualization methods.

[0086] The volume modeling methods of various elements of complex off-road environments are essentially the same, and all use basic voxels to realize the construction of various elements of complex off-road environments. When modeling the complex off-road environment elements of a specific area, the voxels of various elements are all extensions based on the basic voxels. By using cube segmentation in space, the attribute values ​​of various elements can be unified in one voxel for modeling. Each voxel can record different land, ocean, meteorological and electromagnetic related attribute information, thereby realizing the unified modeling of various elements of complex off-road environments, as shown in the figure below. If any element is analyzed or expressed, the attribute values ​​of the same element type in the entire voxel model can be extracted to achieve this.

[0087] Theoretically, it is possible to achieve unified modeling of complex off-road environments and build a unified voxel model. Of course, in the specific implementation process, data compression, storage and indexing are the core of modeling. Voxel data itself is already very large, so how to improve the efficiency of data storage and indexing is the top priority.

[0088] 3) Map data structure unit based on the spatial topological relationship of voxel model

[0089] In order to store and retrieve the traversable domain map more efficiently, it is necessary to establish a spatial topological relationship based on the voxel model. Different from the TSDF Volume space used by traditional algorithms, the size of the voxel grid in the actual complex off-road environment is often difficult to determine, and it is also difficult to allocate enough memory in the application of traditional algorithms. The present invention utilizes the characteristics of Hash mapping and uses the std::map object provided by the C++STL library to construct the key-value pair of the voxel grid ID and the voxel grid object. The voxel grid does not need to be created in advance until the sensor observation of the system platform is converted to the reference coordinate system. If a certain observation falls into the voxel grid object created in the complex off-road environment space, the voxel grid at that position will be dynamically created. With the fast retrieval characteristics of Hash mapping O(lgn), this method saves memory and computing time, can build maps for larger scenes, and meet the real-time requirements of the Z platform.

[0090] Commonly used path planning maps only model nodes in roads or drivable roads, and cannot express targets in ZC environments by labels, that is, the map expression results cannot distinguish between Z vehicles, environmental objects and other units in complex off-road environments. Although this defect can be compensated to a certain extent by the instance segmentation algorithm, it is still powerless to segment instances of objects such as trees and shrubs in the wild environment and ruins walls that are unique to complex off-road environments. The present invention utilizes the spatial continuity between different instance objects to cluster the spatial information based on semantic point clouds. Each cluster obtained will be considered as an independent object in space. According to the number of categories of semantic segmentation, semantic map objects under different categories are established respectively.

[0091] In this data structure, the map of the semantic segmentation part is combined with the global voxel grid map; as the number of observations increases, the voxel grid labels may jump from one category to another, in order to avoid wasting time in destructing voxel grid objects and reconstructing voxel grids when each category independently stores voxel grid information, the voxel grid objects in the present invention will be managed by the global voxel grid set and organized according to object instances, and used for predicting vehicle-inaccessible areas in subsequent path planning.

[0092] In order to realize the perception of object instances in each category of semantic maps, the system refers to the DBSCAN algorithm and designs an incremental density clustering algorithm based on voxel elements in each category of complex off-road environment spatial semantic maps. The present invention will retrieve all qualified voxel grids in a certain neighborhood based on voxel grid coordinates, which can greatly reduce the computational time. First, the semantic information obtained by overall perception is clustered to obtain several clusters, and then for each newly added cluster, its adjacent relationship with the existing clusters is judged, and the clustering results of the existing clusters are merged or added.

[0093] After the termination of incremental clustering, the overall semantic topological voxel map of the complex off-road environment space can be obtained, which can accurately represent objects such as grasslands, large buildings, ruins, forests, etc. that are difficult to represent using instance information. It can also be converted into only local object instances, realizing situational awareness covering a wide range of object sizes and types in the complex off-road environment space.

[0094] In order to effectively organize the spatial relationship of each label element of the voxel model in the off-road environment, the intrinsic correlation between the voxel grid instance information after semantic segmentation is used to generate its topological relationship. The voxel model mainly reflects the topological relationship between spatial objects through the adjacency relationship between cubes. There are three common relationships between voxels-shared faces, shared edges and shared points. These three situations are called face adjacency, edge adjacency and point adjacency respectively. Each voxel has 26 adjacent voxels, namely 6 face adjacencies, 12 edges and 8 faces. The adjacency relationship describes the topological relationship between two adjacent voxels. The topological relationship between multiple continuous voxels is described by path and connectivity relationships, such as Figure 8 As shown, a complex terrain map data structure of a complex off-road environment is generated.

[0095] According to the above method, this service can generate no less than 4 types of path planning maps, including raster maps, Voronoi polygon maps, visualization maps, topological maps and 3D voxel maps, meeting the index requirements.

[0096] 3. Path planning module based on 3D voxel grid

[0097] Making full use of the large amount of empirical data from ground vehicle passability tests, we conduct data-driven global path planning algorithm research based on a three-dimensional traversable domain map of multi-element voxels. According to differentiated application needs and ground vehicle mobility requirements, we establish an optimization objective function with multi-state constraints, conduct hierarchical search on the voxelized complex off-road environment map, use a heuristic search algorithm to iteratively find traversable node calculations, and combine deep neural networks and other methods to explore empirical traversable path results with different rewards, and achieve the optimal path point set that meets the objective function.

[0098] Based on various application requirements in complex off-road environments, a fast and real-time global path planning search is performed according to the provided navigation unit mobility assessment results and navigation performance situation assessment results, providing a basis for path selection and providing a basis for land vehicle maneuvering paths.

[0099] The present invention needs to realize global real-time dynamic path planning according to different application requirements input by users. In complex off-road environments, vehicle carriers are subject to multiple constraints such as prior knowledge of terrain, climate and environmental situation, and face multiple application requirements at the same time, which increases the difficulty of trajectory prediction.

[0100] The data inputs related to different vehicle models are mainly:

[0101] 1) Vehicle mobility assessment and prediction database file;

[0102] 2) Vehicle size, weight, and dynamics-related attributes.

[0103] The methods related to data input involve the constraint functions and optimization process of path optimization, which will regenerate the three-dimensional traversable area map, and generate the mobile unit path trajectory according to application requirements through global path planning and local path planning.

[0104] According to the generated path trajectory and combined with the dynamic characteristics of different navigation unit types, different control parameters are generated to navigate and guide the maneuverable unit.

[0105] In order to meet seven major application requirements, including speed priority, distance priority, road priority, safety priority, fuel consumption priority, commander's autonomous decision-making and intelligent mode, in addition to conducting traditional path planning algorithm research, this invention will also explore a LSTM-based neural network path planning method.

[0106] The present invention uses the relevant database in the platform, as well as the data provided by the land complex environment mobility assessment and prediction subsystem and the complex environment ubiquitous navigation performance prediction and assessment subsystem as data sources, imports geographic entity data, navigation electronic map data, vehicle parameter data and satellite navigation signal monitoring perception network data, and uses mobility assessment and prediction products to determine the passability and vehicle passability in the target area. Since the topography and geology of the area through which the mobile navigation unit passes are relatively complex, the present invention reorganizes the imported data into a three-dimensional voxel form with multiple element attributes based on an octree structure, and finally forms a three-dimensional passable area map.

[0107] The generated three-dimensional passable area map is used as the basic map for the path planning part (hereinafter referred to as the "passable map"). When the user needs path planning, he inputs the corresponding instructions into the system according to his mobile starting point, end point and the highest priority application requirements. After reading the code corresponding to the instruction, the system immediately initiates path planning.

[0108] After receiving the user input, the subsystem first establishes a specific path optimization function f(x), which is related to the application requirements, navigation unit model and traversable area map. According to the generated three-dimensional traversable area map and f(x), the vehicle field maneuvering area map is divided into several voxel grids V of different scales. i The three-dimensional feasible domain map is constructed, and each voxel has a separate number and rich attribute parameters. The specific path optimization process is described as follows.

[0109] a) The subsystem first searches for the trajectory on a three-dimensional feasible domain map with a larger resolution to determine a rough path trajectory;

[0110] b) Based on the rough path trajectory, the path is accurately optimized again on a map with a higher resolution according to the optimization function;

[0111] c) After going through the above two steps multiple times, a precise path trajectory P is finally determined ab and its coordinate points.

[0112] When the subsystem searches for a path through the above steps, due to different optimization conditions generated by application requirements, more than one optimization path is obtained, and multiple paths may meet the application requirements or similarly meet the requirements. After the subsystem generates each path, it will store the time, fuel consumption, distance, safety, and maximum speed of the path as the basic attribute information of the path together with the path to form a corresponding data file.

[0113] Determine the path P ab After that, according to the GIS coordinates output by the path, the area through which the motor vehicle passes under the guidance of the path can be determined, and the mobility attribute parameters of the area through which the vehicle passes can be read, searched and determined. The main parameters of the search process are shown in Table 1.

[0114] Table 1 Input and output of intelligent path planning

[0115]

[0116] Based on the information such as the maximum speed in the area passed in the above data, the overall time T of the planned path can be calculated and estimated, and then the recommended departure time t0 can be calculated based on the set arrival time t for reference by the commander to implement the maneuver.

Claims

1. A path planning system for complex land environments based on three-dimensional voxel grids, characterized in that: It includes a module for evaluating and predicting mobility in complex land environments, a module for establishing and optimizing three-dimensional voxel grids, and a module for path planning based on three-dimensional voxel grids, among which: The land complex environment mobility assessment and prediction module further includes: The intelligent obstacle recognition unit for complex land environments builds an intelligent obstacle recognition model based on scene scale and DCNN transfer learning to quickly identify obstacles in complex land environments; The ground mechanics parameter interpretation unit for complex land environments provides a data basis for the coupling analysis of vehicle and ground mechanics, classifies and identifies the soil, obtains soil moisture, and then realizes the construction of ground mechanics parameters; The precise inversion unit of the ground mechanical parameters of the coupled vehicle is used to establish a knowledge base of the corresponding relationship between the mechanical properties of known soil types, so as to obtain a coupled quantitative calculation method that satisfies the ground mechanical characteristic parameters of different passing vehicles and standard classification, thereby realizing the inversion of the ground mechanical characteristic parameters of unknown areas and the evaluation of the vehicle's traffic performance; The whole vehicle dynamics model is based on vehicle-ground interaction. The whole vehicle dynamics model is determined by the type of vehicle and ground. The vehicle type is determined by the running mechanism. Based on the interaction mechanism between different running mechanisms and ground, a multi-degree-of-freedom mechanical model of the vehicle in a three-dimensional unstructured environment is established to provide a model basis for the evaluation of vehicle passability and maneuverability. The vehicle passability assessment unit based on multivariate analysis is used to assess the vehicle's passability in a specified area; The maneuvering speed prediction unit, which combines vehicle performance parameters and dynamic models, is used to evaluate the vehicle's running speed. It also combines vehicle test data, ground comprehensive characterization model and vehicle maneuvering speed to obtain a driving stability classification based on deep learning for vehicle running safety. At the same time, it combines vehicle stability and speed to conduct a study on rapid maneuvering rules, form a rapid maneuvering area evaluation model, and obtain a judgment on the rapid maneuvering area. The ground mechanics parameter interpretation semantic map construction unit uses the form of semantic maps to characterize the ground mechanics parameter characteristic inversion results of multi-source data. By referring to the semantic map construction method, the process of estimating the ground pressure and shear characteristics based on vision is integrated into the ground modeling process to construct a ground mechanics parameter characteristic inversion semantic map that integrates ground geometry and mechanical characteristics, thereby providing terrain geometry and force for vehicle traffic simulation, path planning and mobile control; The 3D volume grid building and optimization module further includes: Based on the three-dimensional traversable area map construction unit of multi-element voxel grid, an abstract representation model is established for the core elements of complex off-road environment related to the mobility characteristics of ground vehicles and the interaction between vehicles and the environment. The spatial topological map structure based on the voxel model is used for the digital representation of unstructured complex terrain and complex off-road environment; Complex off-road environmental element abstract representation unit, used to realize the abstract representation of complex off-road environmental elements; The map data structure unit based on the spatial topological relationship of the voxel model is used to establish the spatial topological relationship based on the voxel model, and cluster the spatial information based on the semantic point cloud by using the spatial continuity between different instance objects. Each cluster obtained will be considered as an independent object in the space. According to the number of categories of semantic segmentation, semantic map objects under different categories are established respectively; in the map data structure, the map of the semantic segmentation part is combined with the global voxel grid map; the voxel grid object will be managed by the global voxel grid set and organized according to the object instance, which is used to predict the vehicle inaccessible area in the subsequent path planning. At the same time, in order to realize the classification of each category The perception of object instances in the semantic map is analyzed. In each category of the spatial semantic map of complex off-road environments, an incremental density clustering algorithm is designed based on voxel grid elements. All qualified voxel grids are retrieved within a certain neighborhood according to the voxel grid coordinates. The topological relationship is generated by using the intrinsic association between the voxel grid instance information after semantic segmentation. The voxel model reflects the topological relationship between spatial objects through the adjacency relationship between cubes. The adjacency relationship describes the topological relationship between two adjacent voxels. The topological relationship between multiple continuous voxels is described by using the path and connectivity relationship, thereby generating a complex terrain map data structure of a complex off-road environment. Path planning module based on 3D voxel grid: Utilizing a large amount of empirical data from ground vehicle passability tests, a data-driven global path planning algorithm is obtained based on a three-dimensional traversable domain map of multi-element voxel lattices. According to differentiated application needs and ground vehicle mobility requirements, an optimization objective function with multi-state constraints is established. A hierarchical search is performed on the voxelized complex off-road environment map, and a heuristic search algorithm is used to iteratively search for traversable node calculations. Combined with a deep neural network, empirical traversable path results with different rewards are obtained to achieve the optimal set of path points that meet the objective function.

2. A path planning system for complex land environments based on three-dimensional voxel grids as claimed in claim 1, characterized in that: In the ground mechanical parameter interpretation unit for complex onshore environments, a rock and soil classification system for ground mechanical properties is constructed, ground rock and soil characteristic types in complex off-road environments are rapidly surveyed and identified, a soil moisture inversion method based on remote sensing data is performed, and ground soil mechanical property parameters in complex onshore environments are constructed.

3. A path planning system for complex land environments based on three-dimensional voxel grids as claimed in claim 1, characterized in that: In the vehicle dynamics model, the vehicle is divided into two types of walking mechanisms: tracked and wheeled.

4. A path planning system for complex land environments based on three-dimensional voxel grids as claimed in claim 1, characterized in that: In the vehicle passability evaluation unit: the ground is parametrically characterized, road surface information, terrain information, geomorphic information and climate information are integrated, and feature extraction and selection of various types of information are performed based on a combination of subjective and objective methods to obtain feature types with a high correlation with vehicle passability, thereby constructing a ground comprehensive characterization parameter model covering geometric and geological characteristics; In terms of the vehicle's own characteristics, the theoretical design or experimental experience of structural dimensions, dynamic characteristics, and performance indicators is combined to deconstruct and analyze the vehicle's own characteristics. The factors and indicators that affect the passability are studied in combination with the ground comprehensive characterization parametric model. The mutual influence mechanism between vehicle parameters, ground characterization and passability is revealed, a quantitative analysis relationship is established, and a multi-dimensional judgment evaluation model is formed.

5. The path planning system for complex land environments based on three-dimensional voxel grids according to claim 1, characterized in that: The maneuvering speed prediction unit studies the maneuvering speed prediction method of obtaining a power flow model based on the vehicle power transmission characteristics and the vehicle-ground coupling dynamics model for different types of vehicles and ground on the basis of passability evaluation.

6. A path planning system for complex land environments based on three-dimensional voxel grids as claimed in claim 1, characterized in that: In the abstract representation unit of complex off-road environment elements, for terrain soil voxels, it is necessary to expand the solid voxels, add the element attributes of the soil elements, and use the octree data structure to compress the terrain with the same type of attributes, and then realize multi-resolution terrain isosurface extraction through indirect visualization methods.

7. A path planning system for complex land environments based on three-dimensional voxel grids as claimed in claim 1, characterized in that: The map data structure unit utilizes the characteristics of hash mapping and uses the std::map object provided by the C++STL library to construct a key-value pair of voxel grid ID and voxel grid object. The voxel grid does not need to be created in advance until the sensor observation of the system platform is converted to the reference coordinate system. If an observation falls into a voxel grid object that has been created in the complex off-road environment space, the voxel grid at that position will be dynamically created.

8. The path planning system for complex land environments based on three-dimensional voxel grids according to claim 1, characterized in that: The three-dimensional voxel-based path planning module also provides a LSTM-based neural network path planning method, which reorganizes the imported data into a three-dimensional voxel form with multiple element attributes based on an octree structure, and finally forms a three-dimensional traversable area map; The generated three-dimensional traversable area map is used as the basic map for the path planning part. When the user needs path planning, he / she inputs the corresponding instructions into the system according to his / her mobile starting point, end point and the application requirements with the highest priority. After reading the code corresponding to the instruction, the system immediately initiates the path planning: After receiving the user input command, the subsystem first establishes a specific path optimization function f(x). The path optimization function f(x) is related to the application requirements, navigation unit model and traversable area map. According to the generated three-dimensional traversable area map and f(x), the vehicle field maneuvering area map is divided into several voxel grids V of different scales. i In the three-dimensional feasible domain map, each voxel has a separate number and rich attribute parameters. The specific path optimization process is as follows: a) The subsystem first searches for the trajectory on a three-dimensional feasible domain map with a larger resolution to determine a rough path trajectory; b) Based on the rough path trajectory, the path is accurately optimized again on a map with a higher resolution according to the optimization function; c) After going through the above two steps multiple times, a precise path trajectory P is finally determined ab and its coordinate points.

Images