A Multi-Objective Road 3D Intelligent Alignment Method and System Based on Deep Reinforcement Learning

Abstract

This application relates to the field of neural network technology, and in particular to a multi-objective intelligent 3D road alignment method and system based on deep reinforcement learning. The proposed scheme involves acquiring 3D terrain data and geographic information data of the candidate route area and constructing a 3D environment dataset. Combined with road design specification parameters, multi-channel input parameters are generated, including the terrain forward profile, longitudinal feasible envelope, and cut-fill pressure difference. Based on these input parameters, the decision-making agent iteratively generates multiple candidate road centerlines in the 3D alignment environment. Furthermore, by comprehensively evaluating candidate routes using multi-dimensional indicators such as engineering cost, ecological avoidance, and alignment smoothness, multi-objective ranking and scheme selection are achieved. This application can proactively perceive engineering constraints and cost trends during the alignment process, improving the feasibility and scheme diversity of 3D road alignment under complex terrain conditions.

Description

Technical Field

[0001] This application relates to the field of neural network technology, and in particular to a multi-objective road 3D intelligent route selection method and system based on deep reinforcement learning. Background Technology

[0002] In the planning and design phase of road engineering, route selection typically requires a comprehensive balance of multiple objectives, including engineering costs, ecological impacts, and alignment quality, under complex terrain conditions. Current technologies for 3D road alignment largely rely on rule-based search methods, traditional optimization algorithms, or longitudinal and transverse corrections based on 2D planar alignment results. These methods often require multiple rounds of manual intervention and repeated comparisons, resulting in low overall efficiency. While some automated solutions incorporate machine learning or reinforcement learning models to generate alignment paths, these methods usually use 3D environmental data or local terrain features as input, lacking prior consideration of road design specifications and engineering cost trends.

[0003] In practical applications, the aforementioned existing route selection technologies are prone to generating routes that appear reasonable in local areas but have excessively high engineering costs, abnormal bridge-to-tunnel ratios, or insufficient alignment continuity on an overall scale. Furthermore, candidate routes generated under complex terrain conditions tend to be highly homogeneous, making it difficult to cover feasible solutions with different objectives. In addition, existing methods often conduct engineering quantity and ecological impact assessments only after route generation, resulting in the need to eliminate a large number of routes that do not meet engineering requirements in the post-processing stage, thus limiting both overall route selection efficiency and solution quality.

[0004] Therefore, there is an urgent need for a technical solution that can comprehensively consider multiple objective constraints and improve the three-dimensional road alignment effect during the alignment process. Summary of the Invention

[0005] To address the above issues, this application designs a multi-objective intelligent 3D road alignment method and system based on deep reinforcement learning. By acquiring 3D terrain and geographic information data of the candidate route area and constructing a 3D environment dataset, and combining road design specification parameters, multi-channel input parameters are generated, including the terrain forward profile, longitudinal feasible envelope, and cut-fill pressure difference. Based on these input parameters, the decision-making agent iteratively generates multiple candidate road centerlines in the 3D alignment environment. Then, by comprehensively evaluating the candidate routes based on multi-dimensional indicators such as engineering cost, ecological avoidance, and alignment smoothness, multi-objective ranking and scheme selection are achieved. This application can proactively perceive engineering constraints and cost trends during the alignment process, improving the feasibility and scheme diversity of 3D road alignment under complex terrain conditions.

[0006] To achieve the above objectives, this application provides the following technical solution:

[0007] A multi-objective intelligent 3D road alignment method based on deep reinforcement learning, the method comprising:

[0008] Acquire 3D terrain and geographic information data of the candidate line area, and generate a 3D environment dataset through voxelization processing;

[0009] Based on the three-dimensional environment dataset and combined with preset road design specification parameters, input parameters adapted to the preset decision-making agent are generated.

[0010] The input parameters are input to the decision-making agent, which is then trained to output actions and generate candidate road centerlines in a 3D route selection environment based on the current state, resulting in multiple 3D candidate routes. The main body of the decision-making agent is a deep neural network.

[0011] According to the preset evaluation strategy, the multiple three-dimensional candidate routes are calculated in parallel to obtain the multi-dimensional evaluation index vector corresponding to each candidate route. Based on the multi-dimensional evaluation index vector, the three-dimensional candidate routes are sorted and clustered to generate a set of route selection schemes covering different target orientations.

[0012] The three-dimensional terrain data includes digital elevation model data and three-dimensional point cloud data. The digital elevation model data is acquired through remote sensing mapping, and the three-dimensional point cloud data is acquired through lidar scanning. The geographic information data includes ecological zone data, land type data, and geological condition data. The geographic information data is imported through a preset geographic information database.

[0013] The process of generating a 3D environment dataset through voxelization includes:

[0014] Spatial registration and unified coordinate transformation are performed on the digital elevation model data and the 3D point cloud data;

[0015] The registered 3D space is divided into voxels according to the preset spatial resolution to generate a regular voxel mesh.

[0016] The digital elevation model data and 3D point cloud data are mapped to corresponding voxel units to obtain voxel features that characterize terrain height and space occupancy.

[0017] The geographic information data is mapped to the voxel unit to label the corresponding attribute information in the voxel unit, wherein the attribute information includes ecological attributes, land type attributes and geological condition attributes;

[0018] A 3D environment dataset is generated based on voxel units, voxel features, and attribute information.

[0019] The road design specifications include design speed, maximum longitudinal slope, minimum vertical curve radius, and roadbed width. The input parameters for generating a decision-making agent adapted to the preset parameters include:

[0020] The forward sight length and sampling interval are determined based on the design speed, and the profile sampling range is determined based on the roadbed width;

[0021] Based on the forward look length, sampling interval, and profile sampling range, multiple candidate action directions are determined;

[0022] For each candidate action direction, extract the corresponding terrain forward profile sequence from the three-dimensional environment dataset;

[0023] Based on the maximum longitudinal slope and the minimum vertical curve radius, construct a longitudinal feasible envelope sequence corresponding to the terrain forward profile sequence;

[0024] Calculate the difference sequence of the topographic foresight profile sequence relative to the longitudinal feasible envelope sequence, wherein the difference sequence includes a first difference sequence characterizing cut pressure and a second difference sequence characterizing fill pressure;

[0025] The terrain forward profile sequence, the longitudinal feasible envelope sequence, and the difference sequence are used as input parameters for the decision agent.

[0026] The determination of multiple candidate action directions based on the forward look length, sampling interval, and profile sampling range includes:

[0027] Starting from the current node of the road centerline to be generated, the radius of the forward search sector is determined according to the forward length;

[0028] The lateral coverage width of the forward-looking search sector is determined based on the profile sampling range;

[0029] Using the current node's direction of travel as the central axis, the direction angles are discretized within the horizontal coverage width of the forward-looking search sector according to a preset angle step size, resulting in multiple direction angles.

[0030] Multiple candidate directional rays are determined based on the direction angle and the forward look length, serving as the multiple candidate action directions.

[0031] Based on the maximum longitudinal slope and the minimum vertical curve radius, a longitudinal feasible envelope sequence corresponding to the topographic foresight profile sequence is constructed, including:

[0032] Obtain the initial elevation and initial longitudinal slope corresponding to the current node of the road centerline to be generated, and discretize the forward length according to the sampling interval to obtain multiple mileage sampling points;

[0033] The allowable range of longitudinal slope at each mileage sampling point is determined based on the maximum longitudinal slope, and the upper limit of longitudinal slope variation per unit mileage is determined based on the minimum vertical curve radius.

[0034] Under the conditions of satisfying the allowable range of longitudinal slope and the upper limit of longitudinal slope change, the upper limit sequence of reachable elevation and the lower limit sequence of reachable elevation at each mileage sampling point are calculated by forward recursion.

[0035] The consistency correction of the upper and lower bound sequences of reachable elevations is performed by backward contraction.

[0036] The upper bound sequence of reachable elevations after consistency correction is used as the upper envelope sequence of the longitudinal feasible envelope, and the lower bound sequence of reachable elevations after consistency correction is used as the lower envelope sequence of the longitudinal feasible envelope, so as to obtain the longitudinal feasible envelope sequence corresponding to the terrain foresight profile sequence.

[0037] The decision-making intelligent agent includes:

[0038] An input encoding layer is used to encode features of the terrain forward profile sequence, the longitudinal feasible envelope sequence, and the difference sequence to obtain a state feature vector. The input encoding layer includes a convolutional feature extraction sub-layer and a recurrent neural network sub-layer along the sequence dimension, as well as a fully connected layer for feature fusion.

[0039] The strategy layer, connected to the input encoding layer, is used to output action selection based on the state feature vector. The action selection includes a discrete selection probability distribution for characterizing multiple candidate action directions, and an extension parameter vector corresponding to the selected candidate action direction for generating the road centerline.

[0040] The value layer, connected to the input encoding layer, is used to output the action value corresponding to the state feature vector for training and updating the decision agent.

[0041] The output layer is used to determine the target candidate action direction based on the discrete selection probability distribution output by the strategy layer in the current round, and to generate a centerline increment generation result corresponding to the target candidate action direction based on the extended parameter vector. The centerline increment generation result includes the three-dimensional coordinates of the next control point and the geometric parameters of the next line segment. The output layer is also used to evaluate and filter the centerline increment generation result based on the action value output by the value layer, and to superimpose the filtered centerline increment generation result onto the current road centerline, so as to output the updated road centerline as part of the input of the decision agent in the next round, until the termination condition is met and the candidate road centerline is output.

[0042] The process involves enabling a decision-making agent to output actions based on its current state and generate candidate road centerlines in a 3D route selection environment, resulting in multiple 3D candidate routes, including:

[0043] Obtain the start and end constraints for generating the road centerline;

[0044] Perform iterative operations, calculate random sampling seeds based on the starting point constraints and ending point constraints, initialize the starting control point of the current road centerline and the initial travel direction and initial longitudinal slope corresponding to the starting control point based on the random sampling seeds, call the decision agent to generate candidate road centerlines, until the preset maximum number of iteration steps is reached, and output multiple three-dimensional candidate routes corresponding to the maximum number of iteration steps.

[0045] The evaluation strategy includes:

[0046] The multiple three-dimensional candidate routes are obtained, and each candidate route is segmented and sampled to obtain a route sampling sequence composed of multiple three-dimensional sampling points;

[0047] Based on the route sampling sequence and the three-dimensional environmental dataset, a set of engineering cost indicators is calculated. The set of engineering cost indicators includes at least earthwork volume indicators and bridge-tunnel ratio indicators.

[0048] Based on the route sampling sequence and the geographic information data, an ecological avoidance index set is calculated, which includes at least an ecological zone crossing length index and a land type occupation index.

[0049] The set of alignment smoothness indices is calculated based on the route sampling sequence, and the set of alignment smoothness indices includes at least the curvature change rate index and the longitudinal slope change rate index.

[0050] The engineering cost index set, the ecological avoidance index set, and the linear smoothness index set are normalized and combined to obtain a multi-dimensional evaluation index vector corresponding to each candidate route.

[0051] A multi-objective road 3D intelligent alignment system based on deep reinforcement learning, the system comprising:

[0052] The data processing module is used to acquire three-dimensional terrain data and geographic information data of the area to be selected, and to perform voxelization processing on the three-dimensional terrain data and the geographic information data to generate a three-dimensional environment dataset.

[0053] The input parameter generation module is used to generate input parameters adapted to the preset decision-making agent based on the three-dimensional environment dataset and in combination with preset road design specification parameters.

[0054] The route selection decision module is used to input the input parameters into the decision agent, and select and iteratively generate candidate road centerlines through the output actions of the decision agent, thereby obtaining multiple three-dimensional candidate routes, wherein the main body of the decision agent is a deep neural network;

[0055] The scheme evaluation module is used to perform parallel calculations on the multiple three-dimensional candidate routes according to a preset evaluation strategy to obtain the multi-dimensional evaluation index vector corresponding to each candidate route, and to perform multi-objective ranking and cluster analysis on the three-dimensional candidate routes based on the multi-dimensional evaluation index vector, thereby filtering and generating a set of route selection schemes covering different objective orientations.

[0056] Compared with the prior art, the beneficial effects of this application are:

[0057] This application introduces a decision-making mechanism based on deep reinforcement learning into the three-dimensional road alignment process, and explicitly integrates information such as the terrain forward profile, longitudinal feasible envelope, and cut-fill pressure at the input level. This enables the alignment decision to perceive engineering constraints and cost trends during the generation stage, thereby avoiding the problem of route infeasibility or excessive cost caused by relying solely on post-evaluation.

[0058] By using multi-directional parallel input and incremental centerline generation, a diverse set of candidate routes can be formed under complex terrain conditions. Furthermore, a comprehensive evaluation can be conducted by combining multiple dimensions such as engineering costs, ecological avoidance, and alignment smoothness, thereby improving the feasibility and overall quality of the route selection results. Attached Figure Description

[0059] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0060] Figure 1 An exemplary application scenario diagram provided for an embodiment of this application;

[0061] Figure 2 This is a schematic diagram of the structure of the line selection device provided in the embodiments of this application;

[0062] Figure 3 A flowchart illustrating the multi-objective road 3D intelligent alignment method based on deep reinforcement learning provided in this application embodiment;

[0063] Figure 4 This is a schematic diagram of the structure of the decision-making agent provided in the embodiments of this application. Detailed Implementation

[0064] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0065] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0066] This application proposes a multi-objective intelligent 3D road alignment method and system based on deep reinforcement learning, primarily targeting preliminary alignment work in road engineering under typical complex terrain conditions such as mountainous areas, hills, and highly undulating terrain. In these areas, design units typically have access to digital elevation models, point cloud mapping results, and multi-source constraint layers such as ecological red lines, basic farmland, and geological hazard zones. While these data are sufficient in form to support automated calculations, in engineering practice, data availability does not equate to route decision-making. The fundamental reason is that road alignment is not a static shortest path problem, but a continuous 3D alignment generation process driven by design specifications, terrain constructability, and multi-dimensional objectives. A route that appears feasible locally often encounters problems such as excessive longitudinal slope, insufficient vertical curves, cut-fill imbalance, or a surge in the proportion of bridges and tunnels at later mileages, leading to extensive rework of preliminary schemes, low comparison efficiency, and difficulty in forming a set of schemes covering different value orientations.

[0067] In existing engineering projects, the common practice is to rely on manual experience, rule constraints, and comparison of several candidate solutions. Designers first provide several channels on a two-dimensional plane, and then continuously revise them through longitudinal and transverse sections and quantity verification; or they use heuristic search / grid cost methods to aggregate indicators such as earthwork, crossing sensitive areas, and route length into a single cost for optimization in a weighted manner. The above methods are feasible when the data conditions are relatively simple, but under the complex terrain and strict constraints addressed in this application, two typical dilemmas are easily exposed: First, single-objective or weighted single-solution modes often compress the minimum engineering workload, ecological avoidance, and linear smoothness into a mixed compression, resulting in an optimal solution that is not interpretable in terms of engineering semantics and is difficult to output a representative multi-scheme solution set; Second, and more importantly, traditional feature representations mostly stay at the level of local elevation, slope, or grid crossing penalties, making it difficult to characterize the longitudinal feasibility and engineering workload trend of a future distance in advance at the decision-making time. As a result, misleading decisions that are locally economical but globally expensive often occur during the automatic generation process. For example, in deep valleys or continuously undulating terrain, the route maintains a low cost in the short distance, but as the mileage progresses, large-scale cutting and filling or the use of bridge and tunnel structures are inevitably introduced, leading to a surge in the a posteriori engineering workload of the scheme, which in turn causes the automated optimization to fall into a large number of ineffective explorations.

[0068] Based on the commonalities of the aforementioned engineering projects, the application scenario of this application does not rely on pre-dividing the terrain or fixed channel divisions. Instead, it addresses a more fundamental form of route selection and imaging constraints: each extension of the road centerline is limited by regulatory constraints such as longitudinal slope and vertical curve smoothness, and the engineering costs and environmental impacts exhibit significant mileage accumulation and non-locality. Especially in terrains with high-altitude canyons, dense gullies, and alternating slopes, although the external input data is complete, if there is a lack of state information that can express the feasible longitudinal section space, any automatic route selection based on a general learning model or general search strategy is easily attracted by local terrain features in the early decision-making stage, and cannot identify potential bridge and tunnel tendencies and fill and cut pressures before the action occurs. Therefore, the starting point of the processing logic of this application is: how to enable the decision-making process to have a priori perception of future longitudinal feasibility and engineering volume trends in the generation stage without relying on post-event engineering quantity verification, and on this basis, stably generate multiple candidate routes that meet regulatory constraints and have different trade-offs between economy, ecology, and comfort.

[0069] Understandably, the core concept of this application is to abstract the spatial growth of the road centerline into a sequential decision-making process of an agent in a three-dimensional environment, and at the same time, through the engineering construction of the input parameters, the deviation pressure between the longitudinal feasible space under the normative constraints and the terrain profile can be explicitly encoded as a state input before each round of decision-making.

[0070] Specifically, for the current centerline endpoint and multiple candidate extension directions, this application extracts the terrain forward profile within the forward-looking window and constructs a corresponding longitudinal feasible envelope based on standard parameters such as the maximum longitudinal slope and the minimum vertical curve radius. Then, the deviation of the terrain profile from the envelope forms a differential sequence of excavation and embankment pressures, enabling the agent to obtain a learnable representation of future construction cost trends before outputting actions. Thus, decision-making no longer relies solely on local terrain or post-hoc penalty feedback, but can suppress seemingly feasible directions that inevitably incur high costs later during the generation phase, thereby more effectively exploring and outputting a set of route options covering different objective orientations.

[0071] refer to Figure 1 , Figure 1 This is an exemplary application scenario diagram provided for an embodiment of this application.

[0072] like Figure 1As shown, the road alignment object exists in a complex three-dimensional environment containing various constraints and obstacles. These obstacles can be natural barriers formed by topographical undulations, ecologically sensitive areas, geologically risky areas, areas impassable by engineering work, or spaces where alignment is not permitted by design specifications. Spatially, they are irregularly distributed, uncertainly located, and significantly constrain the continuous generation of the route. These obstacles are not limited to a single type, nor do they require pre-defined structural zoning; rather, they exist in the alignment environment as unified constraint elements.

[0073] It is understood that the alignment selection device of this application is deployed in a three-dimensional alignment selection environment to perform unified decision-making and control over the generation process of road centerlines. The alignment selection device does not simply perform static path search, but rather, based on a continuous decision-making mechanism, comprehensively considers the current route status, the terrain conditions ahead, and feasibility information under regulatory constraints in each alignment selection round, dynamically selecting the extension direction and geometric shape of the road centerline. Under the continuous decision-making action of the alignment selection device, the system gradually generates multiple candidate road centerlines and further forms an alignment scheme set. This alignment scheme set is not limited to a single optimal result, but includes multiple representative candidate routes under different objective orientations, reflecting different trade-offs in terms of engineering cost, environmental impact, and alignment smoothness.

[0074] It should be noted that, Figure 1 The application scenarios shown are only for illustrating the basic applicable environment and working method of the technical solution of this application, and do not limit the types, quantities, spatial distribution forms of obstacles, or the specific implementation structure of the alignment device. The technical solution of this application is also applicable to other road three-dimensional alignment scenarios with complex constraints, continuous spatial decision-making needs, and multi-objective alignment requirements.

[0075] refer to Figure 2 , Figure 2 This is a schematic diagram of the line selection device provided in an embodiment of this application.

[0076] like Figure 2 As shown, the route selection device of this application includes an aerial mapping sensor, a laser scanning sensor, a processor, a memory, and an output interface. The components interact and work together through a bus or communication link.

[0077] The aerial mapping sensor is used to perform aerial remote sensing mapping of the selected route area to obtain topographic image data covering the selected route area, and then processes the data to form digital elevation model data. The aerial mapping sensor can be mounted on a satellite platform, a manned aerial platform, or an unmanned aerial platform, and its specific implementation is not limited.

[0078] The laser scanning sensor is used to perform laser scanning on the selected area to obtain three-dimensional point cloud data reflecting the surface and its structural features. The laser scanning sensor can be one or more of airborne lidar, vehicle-mounted lidar, or ground-based laser scanning equipment, and the acquisition method can be configured according to engineering needs.

[0079] The processor is communicatively connected to the aerial mapping sensor and the laser scanning sensor, and is used to receive the digital elevation model data and the three-dimensional point cloud data, and execute program instructions stored in the memory to realize functions such as three-dimensional environment dataset construction, input parameter generation based on road design specifications, decision-making agent reasoning and training, and iterative generation of candidate road centerlines.

[0080] The memory is used to store program instructions as well as intermediate and result data generated or used during the route selection process. The output interface is connected to the processor and is used to output the generated multiple three-dimensional candidate routes or route selection scheme sets to external devices or systems. The external devices may include display terminals, design platforms, or other engineering application systems to support subsequent scheme comparison, visualization, or detailed design.

[0081] It should be noted that, Figure 2 The line selection device structure shown is a schematic structure used to illustrate the composition relationship and data flow between the functional entities in this application. It is not intended to limit the specific number, connection method or implementation form of each sensor, processor, memory and interface. The line selection device of this application can also be deployed modularly or in a distributed manner according to the actual application scenario.

[0082] This application embodiment also provides a multi-objective road 3D intelligent alignment system based on deep reinforcement learning. The system can be configured within the processor, and the system includes:

[0083] The data processing module is used to acquire three-dimensional terrain data and geographic information data of the area to be selected, and to perform voxelization processing on the three-dimensional terrain data and the geographic information data to generate a three-dimensional environment dataset.

[0084] The input parameter generation module is used to generate input parameters adapted to the preset decision-making agent based on the three-dimensional environment dataset and in combination with preset road design specification parameters.

[0085] The route selection decision module is used to input the input parameters into the decision agent, and select and iteratively generate candidate road centerlines through the output actions of the decision agent, thereby obtaining multiple three-dimensional candidate routes, wherein the main body of the decision agent is a deep neural network;

[0086] The scheme evaluation module is used to perform parallel calculations on the multiple three-dimensional candidate routes according to a preset evaluation strategy to obtain a multi-dimensional evaluation index vector, and to perform multi-objective ranking and cluster analysis on the three-dimensional candidate routes based on the multi-dimensional evaluation index vector, thereby filtering and generating a set of route selection schemes covering different objective orientations and outputting them.

[0087] Next, with reference to the accompanying drawings, the multi-objective road 3D intelligent alignment method based on deep reinforcement learning provided in the embodiments of this application will be further described. Figure 3 The methods shown include:

[0088] S1: Obtain the 3D terrain data and geographic information data of the area to be selected, and generate a 3D environment dataset through voxelization processing;

[0089] In this embodiment, the three-dimensional terrain data of the area to be selected includes digital elevation model data obtained through remote sensing mapping and three-dimensional point cloud data obtained through laser scanning. The geographic information data includes constraint information such as ecological zone distribution, land use type, and geological conditions.

[0090] The aforementioned data originates from existing surveying and mapping results or engineering databases. The specific resolution, accuracy level, and acquisition method can be flexibly determined according to the engineering stage and data conditions, as long as they meet the minimum requirements of reflecting the terrain undulations and constraint distribution characteristics. By spatially registering multi-source data, unifying coordinate transformation, and mapping it to a regular voxel grid, continuous and irregular 3D terrain and constraint information is transformed into a structured 3D environment dataset, thereby providing a unified and computable environmental representation for subsequent route selection decisions.

[0091] S2: Based on the three-dimensional environment dataset and combined with the preset road design specification parameters, generate input parameters adapted to the preset decision-making agent;

[0092] In this embodiment, the road design specifications include at least the design speed, maximum longitudinal slope, minimum vertical curve radius, and roadbed width, which are used to constrain the feasibility of the road at the longitudinal and transverse levels. Based on the current generation state of the road centerline, a sequence of terrain forward profiles is extracted within the forward-looking window in multiple candidate action directions, and a corresponding longitudinal feasible envelope sequence is further constructed based on the maximum longitudinal slope and minimum vertical curve radius. By comparing the actual terrain profile with the feasible envelope, a difference sequence representing the cut pressure and fill pressure is formed, so that the input parameters not only reflect the terrain features of the current location, but also explicitly characterize the engineering cost trend that may occur over a future distance under specification constraints.

[0093] S3: Input the input parameters into the decision-making agent, and train the input parameters through the decision-making agent so that the decision-making agent outputs actions and generates candidate road centerlines in the three-dimensional route selection environment according to the current state, thereby obtaining multiple three-dimensional candidate routes;

[0094] In this embodiment, the decision-making process is modeled as a sequential decision-making process in a three-dimensional route selection environment. The decision agent takes the terrain forward profile sequence, longitudinal feasible envelope sequence, and difference sequence as state inputs, and outputs the selection result of the candidate action direction and the corresponding centerline extension parameters in each round. By iteratively executing this process, control points or line segments of the road centerline are gradually generated, and a complete candidate route is output when the termination condition is met.

[0095] S4: Perform parallel calculations on the multiple three-dimensional candidate routes according to the preset evaluation strategy to obtain the multi-dimensional evaluation index vector corresponding to each candidate route. Based on the multi-dimensional evaluation index vector, perform multi-objective ranking and cluster analysis on the three-dimensional candidate routes to filter and generate a set of route selection schemes covering different objective orientations.

[0096] In this embodiment, each generated candidate route is sampled in segments, and the corresponding engineering cost index, ecological avoidance index, and linear smoothness index are calculated in combination with the three-dimensional environmental dataset to form a multi-dimensional evaluation index vector.

[0097] The specific calculation method of the above indicators can be adjusted according to the needs of the project, such as earthwork volume, bridge-tunnel ratio, ecological zone crossing length or curvature change rate, etc. Those skilled in the art can select appropriate indicator types and weight configurations according to the actual project conditions. This embodiment does not make further limitations.

[0098] By sorting and clustering the multidimensional evaluation index vectors, the candidate routes are divided into several scheme categories with different objective orientations, thereby forming a set of route selection schemes covering different design focuses such as prioritizing engineering economics, ecological avoidance, or alignment quality, providing a basis for subsequent scheme comparison and engineering decision-making.

[0099] Before delving into the specific technical details of the steps, this application's embodiments need to reiterate that three-dimensional road alignment in engineering practice presents a constraint pattern significantly different from two-dimensional path planning: the road centerline cannot be established simply by free connectivity within a plane; rather, it must simultaneously meet the continuity and constructability requirements stipulated by design specifications at both the longitudinal and transverse levels. Especially in mountainous and highly undulating terrain areas, a seemingly passable route in some local terrains often leads to irreversible expansion of engineering costs in subsequent mileages due to accumulated longitudinal slopes, insufficient vertical curves, or superimposed terrain undulations. This phenomenon does not manifest as an anomaly at a single isolated point, but rather as a gradually accumulating trend deviation along the mileage direction: if effective suppression is not achieved early on, even if corrections are made later, it will come at the cost of significantly increased cut-and-fill, bridge-and-tunnel, or detour lengths.

[0100] Therefore, in the route selection generation stage, the key to the quality of automated decision-making is not whether the immediate feasibility of a certain step can be calculated, but whether a stable and learnable prior representation of the longitudinal feasible space of a future distance can be formed before the action occurs.

[0101] Understandably, if the common approach in existing technologies is to superimpose standard parameters as hard constraints or penalties into the decision-making process—that is, to first generate routes based on local costs or heuristic strategies, and then correct them through longitudinal section verification and engineering quantity estimation—this approach is interpretable in traditional manual design processes. However, it is prone to introducing two types of systematic biases under an automated generation framework. First, standard constraints have obvious path-dependent characteristics, and single-step judgment is insufficient to cover future connectivity, leading to the concentrated failure of a large number of candidate routes in the later stages of generation. Second, engineering costs and structural tendencies are non-local indicators. If they are only evaluated after generation, the decision-making process cannot obtain timely feedback on early deviation trends. This manifests as the strategy being overly sensitive to local low-cost areas, while the generated routes frequently exhibit fill-cut imbalances or abnormal bridge-tunnel ratios at a macro scale.

[0102] Next, we will further elaborate on the technical content of the method in this application regarding the three-dimensional environment dataset.

[0103] In one example, the three-dimensional terrain data includes digital elevation model data and three-dimensional point cloud data. The digital elevation model data is acquired through remote sensing mapping, and the three-dimensional point cloud data is acquired through lidar scanning. The geographic information data includes ecological zone data, land type data, and geological condition data, and the geographic information data is imported through a preset geographic information database.

[0104] In yet another example, the generation of a 3D environment dataset through voxelization includes:

[0105] Spatial registration and unified coordinate transformation are performed on the digital elevation model data and the 3D point cloud data;

[0106] The registered 3D space is divided into voxels according to the preset spatial resolution to generate a regular voxel mesh.

[0107] The digital elevation model data and 3D point cloud data are mapped to corresponding voxel units to obtain voxel features that characterize terrain height and space occupancy.

[0108] The geographic information data is mapped to the voxel unit to label the corresponding attribute information in the voxel unit, wherein the attribute information includes ecological attributes, land type attributes and geological condition attributes;

[0109] A 3D environment dataset is generated based on voxel units, voxel features, and attribute information.

[0110] It is understood that the aforementioned method for generating a 3D environment dataset through voxelization is a data processing approach commonly used by those skilled in the art to achieve unified 3D spatial modeling after acquiring multi-source terrain data and geographic information data. Through steps such as spatial registration, coordinate unification, voxel partitioning, and attribute mapping, the original digital elevation model data, 3D point cloud data, and geographic information data can be converted into a structured 3D environment representation, facilitating subsequent spatial analysis and computation. This processing flow has mature applications in existing geographic information systems, 3D modeling, and engineering simulation fields. Its specific implementation, parameter selection, and data accuracy can be adjusted according to actual engineering needs, and this application does not further limit these aspects.

[0111] Next, we will further elaborate on the technical content of parameter preprocessing in the method of this application.

[0112] In one example, the road design specification parameters include design speed, maximum longitudinal slope, minimum vertical curve radius, and roadbed width. These parameters are derived from road engineering design specifications, industry technical standards, or design conditions already determined in the project. These parameters are typically defined during the road alignment stage, or can be selected based on road grade, functional positioning, and regional conditions. The methods of obtaining these parameters are understandable to those skilled in the art, and this application does not limit the specific specification version or value range.

[0113] In another example, generating input parameters adapted to a preset decision agent includes:

[0114] S2.1: Determine the forward sight length and sampling interval based on the design speed, and determine the profile sampling range based on the roadbed width;

[0115] Specifically, if the decision to extend the road centerline lacks a spatial scale matching the road grade, two types of biases are likely to occur: one is that an excessively short foreground scale leads to decisions that are only sensitive to local terrain, making it difficult to perceive subsequent longitudinal constraint pressures in a timely manner; the other is that an excessively long foreground scale results in the input containing a large amount of information irrelevant to the current decision, diluting the state representation and introducing unnecessary computational burdens. Therefore, incorporating the design speed into the determination of the foreground length and sampling interval can ensure that the decision scale is consistent with the road operating conditions, allowing the input to cover sufficient terrain trends while maintaining computability. Simultaneously, if the profile sampling range is disconnected from the roadbed width, the characterization of the terrain-affected zone will deviate from the actual engineering land occupation and construction disturbance range, thus lacking an engineering correspondence in the subsequent assessment of cut-and-fill pressure.

[0116] In this embodiment, the design speed is used to determine a forward-looking window. This window extends forward from the current centerline end along the candidate extension direction. The window length follows the conventional correlation between road grade and design speed, covering several potential vertical undulation units, making longitudinal constraints observable at this scale. The sampling interval is set to ensure a stable spatial resolution of the profile sequence within the forward-looking window, allowing elevation changes, undulation transitions, and valley depth changes to be continuously expressed in the sequence, while avoiding sequence noise and redundancy caused by overly dense sampling. The profile sampling range is based on the roadbed width and superimposed with a certain engineering influence width to cover areas on both sides of the roadbed where fill and cut and slope disturbances may occur. The profile sampling range is set as a strip-shaped area around the candidate direction ray, so that the profile expression in the same candidate direction reflects both the topographic elevation at the centerline position and the impact information of the lateral topography on construction feasibility.

[0117] S2.2: Determine multiple candidate action directions based on the forward look length, sampling interval, and profile sampling range;

[0118] Specifically, centerline growth is a continuous space decision-making problem, and decision-making agents typically use a finite set of candidate directions as the entry point for the action space to balance search coverage and computational efficiency. If too few candidate directions are set, feasible paths are easily missed, causing the strategy to converge prematurely to local paths; if too many candidate directions are set, a profile needs to be extracted, an envelope constructed, and a difference formed in each direction, significantly increasing the input construction overhead and introducing repetitive state information between highly similar directions. Therefore, the determination of candidate action directions needs to form a consistent geometric relationship with the look-ahead length, sampling interval, and profile sampling range, ensuring sufficient distinguishability of each direction within the look-ahead window, while guaranteeing that the strip-shaped regions covered by the profile extraction do not excessively overlap or have gaps.

[0119] In this embodiment, the starting point is the control point at the end of the current road centerline. The central axis of the forward-looking search sector is determined using the current direction of travel, and the sector radius is determined by the forward-looking length. The lateral coverage width is calculated from the profile sampling range and is used to limit the maximum deflection range of the candidate directions, ensuring that the lateral offset of the candidate directions does not exceed the terrain area that the profile sampling can stably represent. Subsequently, the direction angle is discretized within the sector at a preset angle step size, forming multiple candidate direction rays. Each candidate direction ray corresponds to a set of forward-looking sampling point positions arranged along the sampling interval, thereby enabling the subsequent profile sequence to be generated with a consistent structure of direction-mileage sampling points. The preset angle step size can be set according to the terrain complexity, road grade, and computing resources. Those skilled in the art will understand that the angle step size only needs to ensure that the lateral spacing between adjacent direction rays within the forward-looking length scale can reflect significant terrain differences.

[0120] In one example, determining multiple candidate action directions based on the forward look length, sampling interval, and profile sampling range includes:

[0121] Starting from the current node of the road centerline to be generated, the radius of the forward search sector is determined according to the forward length;

[0122] The lateral coverage width of the forward-looking search sector is determined based on the profile sampling range;

[0123] Using the current node's direction of travel as the central axis, the direction angles are discretized within the horizontal coverage width of the forward-looking search sector according to a preset angle step size, resulting in multiple direction angles.

[0124] Multiple candidate directional rays are determined based on the direction angle and the forward look length, serving as the multiple candidate action directions.

[0125] S2.3: For each candidate action direction, extract the corresponding terrain forward profile sequence from the three-dimensional environment dataset;

[0126] Specifically, if a decision-making agent only perceives the elevation or slope of the current point, it cannot distinguish between a direction that is flat in the short term but undulating sharply in the distance and a direction that is generally smooth and suitable for construction. This makes it easy for action choices to be misled by local information. The purpose of introducing a terrain forward profile sequence is not to perform detailed longitudinal design, but to express the terrain trend over a future mileage in a sequential form, enabling decisions to make directional judgments based on the trend. At the same time, if the profile only takes single-line sampling along the centerline, it is easy to ignore the impact of lateral terrain changes on cuts, fills, and slopes, resulting in insufficient perception of the scope of engineering disturbances.

[0127] In this embodiment, for each candidate action direction, multiple mileage sampling points are arranged at sampling intervals along the corresponding candidate direction ray within the forward-looking length. A transverse strip-shaped cross-section is constructed at each mileage sampling point using a profile sampling range. This cross-section is represented in the voxel mesh as a set of voxels perpendicular or approximately perpendicular to the candidate direction, used to extract representative features of the terrain at that location. Representative features include at least terrain elevation features and spatial occupancy features: terrain elevation can be taken as the statistical value of the voxel elevation within the cross-section or the estimated value of the main surface, to reduce the influence of point cloud noise and vegetation interference; spatial occupancy features can reflect whether there are steep walls, cavities, or impassable occupant voxels within the cross-section. By arranging the representative features at each mileage sampling point in mileage order, a terrain forward-looking profile sequence is formed, ensuring that the input sequence corresponding to each candidate action direction has consistent dimensions and order.

[0128] S2.4: Construct a longitudinal feasible envelope sequence corresponding to the terrain forward profile sequence based on the maximum longitudinal slope and the minimum vertical curve radius;

[0129] Specifically, longitudinal feasibility is not a single-point constraint, but a continuous set of constraints jointly limited by the upper limit of the longitudinal slope and the smoothness of the vertical curve. Simply making a static judgment on whether the longitudinal slope exceeds the limit for each sampling point cannot express the continuity of the longitudinal slope change caused by the vertical curve, nor can it reflect the range of elevations achievable within future mileages starting from the current longitudinal slope state. Using the maximum longitudinal slope and the minimum vertical curve radius together to construct the longitudinal feasibility envelope sequence means that, within the forward window, a set of allowable upper and lower elevation limits is provided for each mileage sampling point. This describes the range of elevations achievable at that mileage location under the condition that the longitudinal slope does not exceed the limit and the longitudinal slope change is not too rapid, thus providing a reference benchmark for subsequent differential pressure construction.

[0130] Understandably, under longitudinal constraints, the elevation achievable by the road centerline over a future mileage is not arbitrary but rather constrained by the current longitudinal slope, maximum longitudinal slope limits, and vertical curve continuity requirements. Judging whether the longitudinal slope exceeds limits only at a single mileage point fails to reflect the cumulative effect of slope changes and also fails to determine whether a connectable longitudinal solution space still exists under continuous mileage conditions. Therefore, in this embodiment, a "longitudinal feasible envelope sequence" is constructed to transform the specification constraints from discrete judgments into a continuous elevation interval that varies with mileage. This sequence describes the overall elevation range achievable by the centerline longitudinal profile within the forward-looking window under given initial conditions. This envelope is essentially an accessibility boundary defined by specification constraints; its function is not to provide specific longitudinal solutions but rather to offer a stable reference framework for subsequent comparisons with topographic profiles.

[0131] In this embodiment, at the end of the current road centerline, the actual elevation of that location and the current longitudinal slope determined by the previous alignment are known. This longitudinal slope is introduced into the envelope construction process as the starting condition for the longitudinal profile state. Subsequently, the foresight length is discretized according to the sampling interval, forming a sequence of mileage sampling points that correspond one-to-one with the terrain foresight profile sequence. At each mileage sampling point, the range of longitudinal slope values ​​is first determined based on the maximum longitudinal slope, ensuring that the instantaneous longitudinal slope of any feasible longitudinal profile at that location does not exceed the upper or lower limit allowed by the specifications. Simultaneously, the rate of longitudinal slope change is constrained based on the minimum vertical curve radius, i.e., limiting the maximum amplitude of longitudinal slope adjustment between adjacent sampling points, thereby avoiding drastic changes in longitudinal slope over short distances. This constraint corresponds to the lower limit requirement of the vertical curve length in engineering, and its introduction ensures that the envelope not only satisfies the single-point longitudinal slope constraint but also embeds the continuity condition of longitudinal profile smoothness.

[0132] Furthermore, the reachable elevation intervals at each mileage sampling point are calculated progressively using a forward recursion method. During forward recursion, based on the elevation and slope of the initial sampling point, the upper and lower bounds of the reachable elevation at the previous sampling point are propagated forward to the current sampling point, under the condition of satisfying the allowable slope range and the upper limit of slope change, forming new upper and lower bounds of the reachable elevation. This propagation process considers at each step which elevations are reachable from the previous step and whether these elevations still satisfy the slope and vertical curve constraints in the next step, thus ensuring that the obtained elevation intervals are physically and conceptually feasible. Since forward recursion may produce overly lenient intervals at a local scale, to avoid inconsistencies or irreversibility of the envelope at the overall scale, a backward contraction process is performed after the forward calculation is completed. This involves checking backward from the end of the forward window whether the reachable elevation intervals of each sampling point can be traced back to the initial state under continuous constraints, and contracting and correcting intervals that do not meet the backtracking conditions. Through the combination of forward recursion and backward contraction, an upper envelope sequence and a lower envelope sequence that maintain consistency throughout the entire forward viewing window are ultimately formed.

[0133] S2.5: Calculate the difference sequence of the topographic foresight profile sequence relative to the longitudinal feasible envelope sequence, wherein the difference sequence includes a first difference sequence characterizing cut pressure and a second difference sequence characterizing fill pressure;

[0134] Specifically, the topographic forward profile sequence describes what the terrain is, while the longitudinal feasible envelope sequence describes what can be achieved under the constraints of the specifications. The deviation between the two corresponds to the engineering adjustment pressure that will be faced if progress is made in that direction. Directly using the profile elevation as input requires the decision-making agent to learn the implicit effects of the specifications during training, which is difficult and relies on a large number of samples. Introducing the differential sequence can make the deviation trend explicit, enabling the decision to identify the continuous cut and fill pressure earlier, thereby avoiding continuous progress in the direction that will inevitably have a high cost in the later action selection stage. The differential sequence is split into two sequences, cut pressure and fill pressure, because they correspond to different construction risks and cost structures in engineering treatment, and often show different distribution patterns in the same direction. The split representation is more conducive to subsequent feature encoding.

[0135] In this embodiment, for each candidate action direction, the elevation represented by the terrain profile is aligned and compared with the upper and lower boundaries of the envelope at each mileage sampling point: when the profile elevation is higher than the upper boundary of the envelope, the excess is recorded as the sampled value of the cut pressure, forming the first differential sequence; when the profile elevation is lower than the lower boundary of the envelope, the excess is recorded as the sampled value of the fill pressure, forming the second differential sequence; when the profile elevation falls within the envelope interval, the cut pressure and fill pressure at the corresponding sampling point are both recorded as zero or near-zero values, so that the differential sequence only presents a non-zero response at the location where "adjustment is unavoidable". The continuity of the differential sequence in the mileage direction can reflect the pressure accumulation section. For example, a continuous cut pressure section may correspond to mountain cutting or tunnel tendency, and a continuous fill pressure section may correspond to bridge lifting or high fill tendency; the peak value and duration of the differential sequence provide an intuitive trend signal for subsequent action selection.

[0136] S2.6: Use the terrain forward profile sequence, the longitudinal feasible envelope sequence, and the difference sequence as input parameters for the decision agent;

[0137] Specifically, if only terrain profile sequences are input, the decision-making agent needs to learn the implicit relationships between terrain morphology, normative constraints, and engineering costs simultaneously, which can easily lead to a dispersion of learning objectives. If only difference sequences are input, there is a lack of supplementary descriptions of the terrain background and the width of the envelope space that generate pressure, which may lead to misjudgments under different terrain morphologies. Combining the profile sequence, envelope sequence, and difference sequence as input parameters can provide three types of complementary information simultaneously: the profile sequence provides terrain trends, the envelope sequence provides the reachability space under normative constraints, and the difference sequence provides the engineering pressure trends formed by the deviation between the two. This allows the decision-making agent to obtain a structured representation of terrain-feasibility-deviation in the same round, thereby more stably completing the comparison and selection between candidate directions.

[0138] In this embodiment, for each candidate action direction, the terrain forward profile sequence, longitudinal feasible envelope sequence, and the first and second difference sequences are concatenated in a preset channel order to form an input tensor of fixed length and fixed number of channels. The envelope sequence may include two channels, an upper envelope and a lower envelope, so that the envelope width information is explicitly present in the input. Subsequently, the input tensors corresponding to all candidate action directions are stacked according to the candidate direction index to form an input parameter set that corresponds one-to-one with the action space, for subsequent input encoding layers to perform convolution and cyclic encoding.

[0139] For example, the following numerical example is provided to illustrate how to convert the input parameters into calculations. This example is only used to explain the relationship between the calculation link and the units. The selected parameters and values ​​are illustrative and do not represent the actual calibration results or engineering recommendations.

[0140] In one example, the design speed is set to 80 km / h, the maximum longitudinal slope is set to ±4% (i.e., uphill slope not exceeding +4% and downhill slope not exceeding -4%), the minimum vertical curve radius is set to 5000 m, the roadbed width is set to 12 m, and the current state of the end of the road centerline (current node) is: initial elevation of 1000.00 m, initial longitudinal slope of +2%, and the current travel direction (azimuth) is set to 0 degrees (for ease of understanding only).

[0141] In this example, a forward-looking time window (to cover sufficient longitudinal trend) is set at 13.5 s, resulting in a forward-looking length of 300 m. The sampling interval needs to be compatible with the engineering scale and data resolution; in this example, it is set to 10 m, resulting in 31 mileage sampling points (including the starting point). The roadbed width in this example is 12 m, but the profile sampling range typically covers not only the roadbed but also areas related to the engineering volume trend, such as slopes and cut-fill influence zones. In this example, the half-width of the roadbed plus the engineering influence margin (exemplarily 26 m) is used to determine the half-width of the profile as 32 m, meaning the profile sampling range is a transverse ±32 m strip area along the candidate direction ray (total width approximately 64 m). To ensure that the lateral offset of the candidate direction at the foresight end still falls within the profile sampling zone, the example uses a lateral offset at the foresight end not exceeding 32 m as the control. The maximum deflection angle is selected to be approximately 6 degrees (at a foresight distance of 300 m, a lateral offset of 6 degrees is on the order of 31 m, still within ±32 m). For example, the fan radius is 300 m, and the lateral coverage width is represented by the fan angle corresponding to the half-width of the profile sampling. The angle step is 2 degrees. The candidate direction angle set is: centered on the current travel direction 0°, discretized into -6 degrees, -4 degrees, -2 degrees, 0 degrees, +2 degrees, +4 degrees, and +6 degrees (a total of 7 candidate action directions). For each candidate direction ray, a sampling point is taken every 10 m along the mileage from 0 m to 300 m for subsequent profile extraction and envelope calculation.

[0142] Furthermore, this example uses the candidate direction of 0 degrees (along the current travel direction) as an example to illustrate how the profile sequence is obtained. At each mileage sampling point, voxel features are extracted from the 3D environment dataset within a horizontal ±32 m strip area centered on the candidate direction ray to form the representative terrain elevation of that mileage point. The selection of the representative terrain elevation can be achieved in engineering using one of the following methods: taking the median / quantile of the terrain elevation within the strip area to suppress outliers; or taking the estimated value of the main surface to reduce the impact of point cloud sparsity and noise; for missing voxels, neighborhood interpolation or neighborhood statistical completion can be used, as long as the profile sequence is continuously usable. An example provides an extracted terrain foresight profile elevation sequence (in meters), corresponding one-to-one with each mileage sampling point (31 points in total, the example is only partial): 0 m, 1000.0; 10 m, 1000.4; 20 m, 1000.9…300 m, 1006.0.

[0143] This example exemplifies the setting of a minimum vertical curve radius of 5000 m to reflect the gentleness of the longitudinal slope change. This translates to: at each 10 m sampling step, the allowable longitudinal slope adjustment is no more than 10 m, which is 0.002 relative to 5000 m. This can be understood as the maximum change in longitudinal slope every 10 m is 0.2%, with an initial longitudinal slope of +2%. The longitudinal slope must not exceed ±4% at each 10 m step, and the change in longitudinal slope relative to the previous sampling point must not exceed 0.2%. Therefore, in the example: at the first sampling point (10 m): the upper limit of the longitudinal slope can increase from 2.0% to 2.2%, the lower limit of the longitudinal slope can decrease from 2.0% to 1.8%, the upper limit of elevation is 1000.22, and the lower limit of elevation is 1000.18. At the second sampling point (20 m): the upper limit of the longitudinal slope continues to increase to 2.4%, the lower limit of the longitudinal slope decreases to 1.6%, the upper limit of elevation increases again based on the previous upper limit, resulting in 1000.46, and the lower limit of elevation increases again based on the previous lower limit, resulting in 1000.34. And so on, the sequence of reachable upper and lower limits of elevation within the entire forward window can be obtained.

[0144] For each mileage sampling point: if the terrain elevation is higher than the upper envelope, the excess portion is recorded as "cutting pressure"; if the terrain elevation is lower than the lower envelope, the deficiency portion is recorded as "filling pressure"; if the terrain elevation falls within the envelope interval, both are recorded as 0. Taking the candidate direction 0 degrees as an example, it can be organized into the following channels (the channel order is not limited in the example): Channel 1: Terrain forward profile sequence (31 points); Channel 2: Upper envelope of longitudinal section (31 points); Channel 3: Lower envelope of longitudinal section (31 points); Channel 4: Cutting pressure sequence (i.e., the first difference sequence) (31 points); Channel 5: Filling pressure sequence (i.e., the second difference sequence) (31 points). The remaining 6 candidate directions are processed in the same way to obtain their respective 5-channel sequences, which are then stacked according to the candidate direction index to form the final input parameter set, which is used by the input encoding layer for convolutional / recurrent encoding and output action selection probability and extension parameters.

[0145] In the specific implementation not shown in the example, the profile extraction results for the candidate direction of 0 degrees show that the terrain elevation generally rises within the forward window, but there is also a pattern of low elevation in the middle section followed by a rise in the later section; after constructing the envelope based on normative constraints and performing difference, the difference sequence exhibits two characteristic intervals:

[0146] In the 10-100 m range, the excavation pressure remained positive and the peak value was relatively high, indicating that if longitudinal continuity is to be maintained in this direction, strong cutting and excavation adaptation is required in the early stage.

[0147] In the 160-230m range, the fill pressure is continuously positive and shows a tendency to accumulate, indicating that in the middle and later sections, the longitudinal section may reach a lower limit higher than the terrain, which may easily lead to high fill or bridge uplift. After 250m, the cut pressure rises again, indicating that the rising terrain in the later section exerts continuous pressure on the smooth longitudinal section.

[0148] It is understandable that if a 3D environmental dataset (terrain height, occupancy, ecological / land / geological attributes, etc. in a voxel grid) is directly cropped into fixed-size 3D blocks or multi-channel voxel tensors in the neighborhood of the current node and used as the state input of the decision-making agent, the agent directly obtains the elevation distribution, slope geometry, and attribute voxels within a fixed spatial range. Even if the voxel block contains slope undulation information, it tends to reflect immediate features such as whether the local slope is steep or whether there are prohibited attributes in the surrounding area. However, the continuous fill pressure accumulation in the 160–230m section is a trend information that gradually manifests as it progresses along the direction. Without explicitly encoding the forward profile and the standard reachable space, the agent needs to learn on its own through a large number of samples during training: the correspondence between a certain local 3D morphological combination and the inevitable high fill in the future longitudinal section. In engineering practice, this mapping relationship not only highly depends on the current longitudinal slope state, sampling scale, and road grade, but is also affected by the vertical curve continuity constraint, resulting in a large demand for training samples, slow convergence, and a tendency to produce action sequences that appear feasible locally but have a sudden increase in subsequent costs.

[0149] In a specific embodiment of this application, the difference sequence has already converted the deviation between the canonical reachable space and the terrain profile into a directly learnable trend signal during the generation stage:

[0150] The continuous positive filling pressure zone allows the agent to perceive the continuous increase in pressure in the subsequent mileage before selecting the direction;

[0151] The continuous positive excavation pressure zone enables the agent to distinguish between short, localized undulations and situations requiring long-distance cutting to maintain longitudinal continuity.

[0152] Under the same model structure, the input organization method of this application makes it easier for the strategy output to tend to choose the direction with less pressure clustering or shorter clustering sections. This reduces the large fluctuation of the a posteriori engineering volume, reduces the proportion of abnormal sections such as bridges and tunnels or high fills and high cuts in the candidate route generation, and increases the proportion of effective routes that reach the termination conditions. It does not rely on additional post-processing corrections, but makes the decision avoid high-risk directions earlier and reduces invalid exploration rounds by making the trend explicit at the input level.

[0153] It should be noted that the upper limit of the longitudinal slope and the smoothness of the vertical curve are not implicit rules that need to be learned from the data, but rather deterministic boundaries given by the design specifications. Implementing them as externally computable envelopes and differences ensures that the input of each candidate direction carries clear engineering semantics. If this process is left to the network submodules to spontaneously form equivalent expressions during training, it usually requires a large number of training samples covering different longitudinal slope states and different terrain morphologies, and the constraint expression is prone to shift when the sample distribution changes, leading to unstable responses to the specification boundaries during the inference stage.

[0154] In engineering implementation, to enhance the comparability between different directions, the above channels can also be scaled. For example, the elevation channels can be converted into relative elevations (increments relative to the current node elevation), and the pressure channels can be clipped and normalized according to a preset amplitude range. This processing does not change the relative shape of the sequence, but is only used to improve the numerical stability of the training and inference process. The specific normalization strategy can be determined according to the data resolution and the range of terrain undulations of the project. This application does not limit it here.

[0155] Next, we will further elaborate on the technical content of the method of this application regarding the decision-making intelligent agent.

[0156] refer to Figure 4 , Figure 4 This is a schematic diagram of the structure of the decision-making agent provided in the embodiments of this application.

[0157] In one example, the decision-making agent includes:

[0158] An input encoding layer is used to encode features of the terrain forward profile sequence, the longitudinal feasible envelope sequence, and the difference sequence to obtain a state feature vector. The input encoding layer includes a convolutional feature extraction sub-layer and a recurrent neural network sub-layer along the sequence dimension, as well as a fully connected layer for feature fusion.

[0159] The strategy layer, connected to the input encoding layer, is used to output action selection based on the state feature vector. The action selection includes a discrete selection probability distribution for characterizing multiple candidate action directions, and an extension parameter vector corresponding to the selected candidate action direction for generating the road centerline.

[0160] The value layer, connected to the input encoding layer, is used to output the action value corresponding to the state feature vector for training and updating the decision agent.

[0161] The output layer is used to determine the target candidate action direction based on the discrete selection probability distribution output by the strategy layer in the current round, and to generate a centerline increment generation result corresponding to the target candidate action direction based on the extended parameter vector. The centerline increment generation result includes the three-dimensional coordinates of the next control point and the geometric parameters of the next line segment. The output layer is also used to evaluate and filter the centerline increment generation result based on the action value output by the value layer, and to superimpose the filtered centerline increment generation result onto the current road centerline, so as to output the updated road centerline as part of the input of the decision agent in the next round, until the termination condition is met and the candidate road centerline is output.

[0162] It should be noted that the input received by the decision-making agent is not a state in a single direction, but rather a set of states corresponding to multiple candidate action directions. Specifically, after the input parameters are generated, each candidate action direction corresponds to a set of multi-channel sequence features arranged along the mileage. These multi-channel sequence features include at least: a terrain forward profile sequence, an upper envelope sequence of the longitudinal feasible envelope, a lower envelope sequence of the longitudinal feasible envelope, a first difference sequence representing excavation pressure, and a second difference sequence representing embankment pressure. These features correspond one-to-one with the sampling points along the mileage dimension and are aligned and combined according to a preset channel order to form an input unit of "one multi-channel sequence per direction". The input units of multiple candidate action directions are then stacked according to the candidate direction index to form an input tensor set, enabling the input encoding layer to encode and compare all candidate directions in parallel within the same round.

[0163] In this embodiment, the input encoding layer includes a convolutional feature extraction sublayer along the sequence dimension, a recurrent neural network sublayer, and a fully connected layer for feature fusion. Its internal processing is organized according to the principle of first extracting local patterns, then building long-range dependencies, and finally compressing and fusing. The convolutional feature extraction sublayer performs one-dimensional convolution operations on the multi-channel sequence for each candidate action direction. The convolutional kernel slides along the mileage sampling point direction to extract local morphological features such as terrain undulations, pressure accumulation, pressure abrupt changes, and envelope contraction. The convolutional layer can include multi-level convolution and downsampling to suppress short-scale noise and enhance trend structures. It can also extract composite patterns of short-term and medium-term undulations simultaneously through convolutional kernels with different receptive fields. The recurrent neural network sublayer receives the sequence features output by the convolutional sublayer and updates the state according to mileage order. This establishes long-range dependencies such as pressure accumulation with mileage, gradual envelope contraction, and alternating cut-fill sections, so that the encoding result no longer reflects only a local point but can characterize the longitudinal adaptation trend within the entire forward window. The recurrent neural network sublayers can employ a unidirectional structure to maintain causality consistent with the road's direction of travel, or a gating structure to suppress irrelevant fluctuations from interfering with the memory state. Fully connected layers compress the output of the recurrent sublayers into a fixed-dimensional state feature vector and fuse feature contributions from different channels, resulting in a final state feature vector that simultaneously incorporates terrain trend information, normalized reachability information, and deviation pressure information. To facilitate comparison of candidate action directions, the state feature vector can maintain a one-vector-per-direction structure during generation; that is, the input encoding layer outputs a directional feature vector for each candidate direction, and the dimension of the directional feature vector is consistent across all candidate directions.

[0164] Furthermore, in the construction of the strategy layer and value layer, this embodiment adopts a shared encoding and branch output structure to improve feature reuse efficiency and maintain training stability. The strategy layer receives the directional feature vectors of each candidate direction and outputs a discrete selection probability distribution corresponding one-to-one with the candidate direction set. This probability distribution is used to characterize the relative tendency of each candidate action direction to be selected in the current state, and its output dimension is consistent with the number of candidate action directions. In addition to direction selection, the strategy layer also outputs an extension parameter vector corresponding to the selected candidate action direction, which is used to describe the incremental generation method of the centerline in that direction. The extension parameter vector may include at least one of the following: parameters for determining the next advancement step size, parameters for determining the degree of planar linear bending, and parameters for determining the longitudinal climbing or descending trend. The specific parameter form can be selected according to the centerline representation method. Those skilled in the art will understand that it is only necessary to ensure that the extension parameters can drive the determination of the next control point and the geometry of the next line segment. The value layer receives the same directional feature vector as the policy layer and outputs action value. This action value can be an estimate of the expected return for each candidate direction or an estimate of the value of the selected action, which is used to guide policy convergence during training and updates. In implementation, the output of the value layer can take the form of a vector with the same number of candidate directions, so as to evaluate the risk and return of different directions and provide a basis for subsequent confidence screening.

[0165] Furthermore, the output layer is responsible for transforming the discrete selection probability distribution and extended parameter vector of the strategy layer into geometric results that can be directly used to update the road centerline, thus forming the final engineering output of the decision-making agent. Specifically, the output layer determines the target candidate action direction based on the discrete selection probability distribution. The determination method can be maximum probability selection or sampling selection according to the probability distribution to form diverse candidate routes. After determining the target direction, the output layer reads the extended parameter vector corresponding to that direction and performs geometric mapping on the extended parameters based on the three-dimensional coordinates of the current road centerline end control point, the end travel direction, and the end longitudinal slope to generate the centerline incremental generation result. The centerline incremental generation result includes at least the three-dimensional coordinates of the next control point and the geometric parameters of the next line segment connecting the current end control point and the next control point. The geometric parameters of the next line segment can be used to characterize the spatial orientation, length, and linear change characteristics of the segment, enabling the road centerline to be incrementally constructed in the form of a control point sequence or a line segment sequence. To improve the reliability of the generation process, the output layer also combines the action value output by the value layer to perform confidence evaluation and screening of the centerline incremental generation results: when the action value is lower than a preset threshold or shows a significant disadvantage among candidate directions, it can trigger resampling of the target candidate action direction or regeneration of the extension parameters, thereby avoiding the direct superposition of high-risk increments onto the centerline. After the above screening, the centerline incremental generation results are superimposed onto the current road centerline to obtain the updated road centerline; the end control point, end travel direction, and end longitudinal slope of the updated road centerline are used to constitute the state conditions required for the generation of input parameters in the next round, so that the decision-making process can maintain the continuity of the line shape and the consistency of the longitudinal profile in continuous iteration. The iteration ends when the termination condition is met, which may include at least one of the following: reaching the endpoint constraint, reaching the preset maximum number of iteration steps, reaching the preset centerline length, and triggering the specification violation condition; when the iteration ends, the cumulative generated control point sequence or line segment sequence is output as the candidate road centerline, thereby forming a three-dimensional candidate route.

[0166] It should be noted that the decision-making agent of this application can be understood as a combination of a deep reinforcement learning policy network / value network for sequence decision-making, and its network body is not limited to a specific structural form. Those skilled in the art will understand that, without departing from the overall concept of this application, the decision-making agent can be implemented or replaced by mature deep neural network structures in the prior art. For example, a cascaded structure of one-dimensional convolutional networks and recurrent neural networks can be used to encode multi-channel sequences, or a Transformer-type sequence encoding network can be used to model the attention of mileage sequences, or a graph neural network can be used to aggregate features of the relationships between candidate directions; the corresponding policy learning framework can also be an existing Actor-Critic method, a proximal policy optimization method, a soft actor / critic method, or a discrete action extension of deep Q-learning.

[0167] The above substitutions only change the implementation form or training framework of the network and do not affect the core technical idea emphasized in this application: that is, through the input parameter generation mechanism, the terrain forward profile, longitudinal feasible envelope and cut-fill pressure difference of each candidate action direction are explicitly organized into a multi-channel sequence, so that the strategy can obtain a state representation tightly coupled with the normative constraints and engineering cost trend before the action occurs, thereby driving the incremental generation of the centerline without relying on the post-event engineering quantity verification.

[0168] In one example, the decision-making agent is trained using a reinforcement learning approach based on interactive sampling: During the training phase, a 3D environment dataset is used as the environment state space, a set of candidate action directions and their corresponding extended parameters are used as the action space, and the current state of the road centerline endpoint and a set of multi-directional input parameters are used as the agent's observations. In each training round, the centerline is initialized from the starting point constraints. In each round, the agent outputs the discrete selection results of the candidate action directions and their extended parameters. The environment generates a centerline increment based on this action and updates the centerline state, while simultaneously performing a standard consistency check and constraint crossing check on this increment. The round terminates when conditions such as exceeding longitudinal slope limits, violating vertical curve smoothness, crossing impassable areas, or reaching the maximum number of steps occur. Reward signals are calculated and accumulated on a round-by-round basis. Rewards include at least engineering costs, ecological avoidance, and alignment smoothness. Engineering costs can be determined by the clustering degree of the cut-fill pressure differential sequence, the length and magnitude of continuous exceeding limits, etc., forming a phased cost feedback. Ecological avoidance can be determined by the overlap degree or crossing length of the centerline increment and ecological attribute voxels, forming a penalty feedback. Alignment smoothness can be determined by the curvature and longitudinal slope changes of adjacent segments, forming a penalty feedback. For violations of design specifications or boundary crossings, additional penalties can be applied to encourage the strategy to converge within the feasible region. Training updates can employ a joint update approach using policy gradient and value estimation: the value layer regresses action values ​​to reduce policy update variance, and the policy layer optimizes the output distribution of discrete direction selection and continuous extension parameters based on round-sampling trajectories. Simultaneously, experience replay or parallel sampling can be set to improve sample efficiency.

[0169] Those skilled in the art will understand that the number of training rounds, batch size, learning rate, sampling parallelism, and termination condition threshold can be determined based on terrain complexity and computing resource configuration, as long as they enable the strategy to converge to a stable and feasible route selection behavior under given specification parameters and environmental constraints. This application does not impose any further limitations.

[0170] In one optional implementation, training is implemented using a parallel sampling Actor-Critic-type method, with the number of candidate action directions set to 7 to 15, the look-forward length set to 200 m to 400 m, and the sampling interval set to 5 m to 10 m. The number of training epochs can be set to 10,000 to 60,000 epochs, where a higher number of epochs can be used in areas with strong terrain undulations and dense constraint layers to cover more state combinations. The maximum number of iteration steps per epoch (i.e., the upper limit of the centerline incremental generation steps) can be set to 300 to 1200 steps; for example, when the advancement distance per step is approximately 5 m to 10 m, the corresponding maximum generation length is approximately 1.5 km to 12 km. The number of parallel sampling environments (sampling parallelism) can be set to 8 to 64, for example, commonly 16 or 32 parallel environments are used to obtain sufficiently diverse trajectory samples before a single update.

[0171] In a sample configuration, policy updates employ a batch optimization approach. The sampled trajectory samples are expanded over time steps to form training batches. The batch size for a single parameter update can be set to 16384 to 262144 time-step samples; for example, 65536 time steps can be used as the sample size for one round of updates. With this sample size, mini-batch iterations can be used during optimization, with the mini-batch size set to 512 to 4096, for example, 2048. The number of optimization rounds per update can be set to 3 to 10 rounds, for example, 5 rounds. The learning rate can be set to... to The discount factor can be set to 0.95 to 0.995, for example, 0.99; the time scale parameter for advantage estimation can be set to 0.90 to 0.98, for example, 0.95. To improve training stability, the gradient clipping threshold can be set to 0.5 to 2.0, and the entropy regularization coefficient can be set to 0.001 to 0.02 to maintain exploration.

[0172] Regarding termination threshold conditions, the following examples can be set: termination occurs when the cumulative generated length reaches 100% to 120% of the endpoint constraint distance but still fails to meet the endpoint constraint; termination occurs when there are 3 to 10 consecutive violations of the specification, for example, if the centerline increment generation results fail the longitudinal slope or vertical curve consistency check 5 times consecutively, the round is considered a failure; termination occurs when the single-step longitudinal slope exceeds the allowable upper limit by 0.5% to 2.0%, for example, if the maximum longitudinal slope is 4%, and the longitudinal slope calculated in a certain step exceeds 4.8%, it will terminate directly; termination occurs when the envelope differential pressure sequence has consecutive over-threshold segments within the forward window whose length exceeds 20% to 50% of the forward length, for example, if the forward length is 300 m, and the consecutive over-threshold segments exceed 120 m, it will terminate; termination occurs when the length of the centerline increment crossing the inaccessible attribute voxel exceeds 10% to 30% of the single-step length, for example, if more than 3 m of the single-step advancement falls into the prohibited area after advancing 10 m, it will terminate. The aforementioned thresholds can be used in conjunction with rewards and penalties to truncate obviously infeasible trajectories as early as possible during the training phase, thereby improving sample efficiency.

[0173] In one example, the process of enabling a decision-making agent to output an action based on its current state and generate candidate road centerlines in a 3D route selection environment, resulting in multiple 3D candidate routes, includes:

[0174] Obtain the start and end constraints for generating the road centerline;

[0175] Perform iterative operations, calculate random sampling seeds based on the starting point constraints and ending point constraints, initialize the starting control point of the current road centerline and the initial travel direction and initial longitudinal slope corresponding to the starting control point based on the random sampling seeds, call the decision agent to generate candidate road centerlines, until the preset maximum number of iteration steps is reached, and output multiple three-dimensional candidate routes corresponding to the maximum number of iteration steps.

[0176] It is understandable that the automatic generation of 3D road alignment does not yield a complete centerline in a single step, but rather is a continuous growth process influenced by the starting point, ending point, and multiple constraints. If centerline generation is based solely on a single deterministic initial condition, the generated result may become overly reliant on the initial direction or local terrain features, leading to path homogenization or local optima in complex terrain conditions. Therefore, in this embodiment, a random sampling seed is introduced to perturb the initial state while satisfying the starting point and ending point constraints, allowing for a degree of diversity in the initial stage of centerline generation. This random sampling seed does not alter the geometric positional constraints of the starting and ending points but is used to generate multiple valid initial travel directions and initial longitudinal slope combinations within the starting point's neighborhood, thus providing different search starting points for subsequent decision-making. This approach enables the decision-making agent to explore in different directions under the same 3D environmental dataset and design specification parameters, creating conditions for forming multiple differentiated candidate routes.

[0177] It is understandable that, given different starting control points on the current road centerline and corresponding initial travel directions and initial longitudinal slopes, the state feature vectors received by the decision-making agent will exhibit significant differences. These differences stem not only from the spatial variations in the terrain forward profile and longitudinal feasible envelope, but also from the differences in the constraints imposed by the longitudinal slope and travel direction on the reachable space. Based on this, the responses of the policy layer and value layer to the same 3D environmental dataset are not fixed outputs, but rather dynamically adjusted according to state changes, resulting in different action selection and value assessment results under different initial conditions.

[0178] In this embodiment, the strategy layer jointly evaluates multiple candidate action directions based on the state feature vector output by the input encoding layer, and outputs a discrete selection probability distribution corresponding one-to-one with the set of candidate action directions. When the starting control point is located in different spatial positions or the initial direction of travel changes, the distribution characteristics of the terrain profile morphology, longitudinal feasible envelope width, and cut-fill pressure differential sequence within the forward window will change, causing each candidate action direction to exhibit different comprehensive performances in terms of "feasibility margin," "pressure concentration degree," and "longitudinal adaptation continuity." Based on these differences, the strategy layer reorders the relative merits of the candidate action directions, causing the discrete selection probability distribution to change: under certain initial conditions, the strategy layer may tend to select directions with smaller deflection angles and gentler longitudinal pressure changes; while under other initial conditions, it may increase the selection probability of directions with larger deflection angles but more sufficient envelope margin. Meanwhile, the extension parameter vector corresponding to the selected candidate action direction will also be adjusted with the state change. For example, when the longitudinal feasible space is relatively spacious, a larger advancement step or a gentler longitudinal change parameter will be output, while when the longitudinal constraints become tighter or the terrain undulations intensify, a more conservative advancement and longitudinal slope adjustment parameter will be output, so that the centerline generation behavior can adapt to the engineering conditions under different initial states.

[0179] In another example, the evaluation strategy is used to establish a comparable quantitative evaluation basis among multiple generated 3D candidate routes. Its overall calculation process is based on existing road engineering evaluation and geospatial analysis technologies and does not rely on a specific route generation method. Those skilled in the art will understand that the above evaluation strategy essentially transforms the manual experience-based process of ex-post comparison in traditional road engineering into a programmable process of index calculation and vectorized expression. The calculation methods for each index have clear implementation paths in existing engineering practices and related technical literature; this application only describes its organization within a multi-objective joint evaluation framework.

[0180] Specifically, after obtaining multiple 3D candidate routes, each candidate route is first subjected to segmented sampling. This segmented sampling can be performed by setting the sampling point density at fixed intervals along the route mileage direction or adaptively according to the curve's geometric characteristics, thus obtaining a route sampling sequence composed of multiple sequentially connected 3D sampling points. This sampling method has been widely adopted in existing road surveying and design software and engineering analysis systems to discretize continuous curves into a computable set of spatial points to support subsequent engineering quantity calculations and spatial overlay analysis. In this embodiment, the specific value of the sampling interval is not limited; it is only necessary to ensure that the sampling accuracy can reflect the main changing trends of the route in the horizontal and vertical directions.

[0181] In calculating the engineering cost index, the spatial relationship between the route sampling sequence and the three-dimensional environmental dataset is analyzed. The earthwork volume index can be understood as comparing the difference between the route design elevation and the actual terrain elevation segment by segment along the longitudinal profile of the route, and combining the lateral influence range determined by the roadbed width, the cut volume and fill volume of each sampling segment are accumulated. This calculation method is consistent with the cross-section method or volume integration method commonly used in the longitudinal profile design of existing roads. The bridge-tunnel ratio index is calculated based on preset engineering triggering conditions. For example, when the terrain undulation exceeds a certain threshold, the cut and fill height continuously exceeds the limit, or the route crosses a geological or ecological area that cannot be directly cut or filled, it can be determined that the corresponding route segment needs to adopt a bridge or tunnel structure. By statistically analyzing the lengths of route segments that meet the bridge triggering conditions and tunnel triggering conditions, and calculating the ratio with the total route length, the bridge-tunnel ratio index can be obtained. The above-mentioned bridge-tunnel determination logic has been maturely applied in the existing route scheme comparison and feasibility study stages. This embodiment does not further limit the specific thresholds or determination models.

[0182] In calculating ecological avoidance indicators, the spatial overlay relationship between route sampling sequences and geographic information data is utilized for analysis. The ecological zone crossing length indicator is obtained by determining whether route sampling points or adjacent sampling segments fall within the spatial range defined by the ecological zone data and by statistically analyzing the length of consecutive crossing segments. This method is consistent with the existing line-area intersection length calculation method in GIS spatial overlay analysis. The land type occupancy indicator is obtained by generating buffer zones corresponding to the roadbed width or design influence range on both sides of the route centerline, spatially overlaying these buffer zones with land type data, and statistically analyzing the proportion of occupied area or length of different land types within the buffer zones, thereby reflecting the occupancy of different land resources by the candidate routes. The above calculation methods for ecological and land type indicators are widely used in environmental impact assessment and land use analysis, and their specific implementation methods are readily understood by those skilled in the art.

[0183] In calculating the route alignment smoothness index, the analysis is based on the geometric continuity of the route sampling sequence. The curvature change rate index measures the smoothness of the route in the horizontal direction by calculating the planar curvature of adjacent route segments and statistically analyzing the curvature difference or its changing trend. The longitudinal slope change rate index reflects the continuity in the longitudinal direction by comparing the longitudinal slope difference between adjacent sampling segments. These indices are consistent with the evaluation approach for curve continuity and comfort in existing road alignment design specifications, and their calculation process can be directly completed based on the coordinates and mileage information of the sampling points without introducing additional complex models.

[0184] After calculating the aforementioned sets of engineering cost indicators, ecological avoidance indicators, and linear smoothness indicators, the indicators are normalized to allow for the combination of indicators with different dimensions and orders of magnitude within the same evaluation space. Normalization methods can include linear normalization, interval mapping, or standardization based on historical statistics, as long as the normalized indicators are comparable. Subsequently, the normalized indicators are combined in a preset order to form a multi-dimensional evaluation indicator vector corresponding to each candidate route, thus providing a unified input representation for subsequent multi-objective ranking, cluster analysis, and scheme selection.

[0185] In one example, multi-objective ranking and clustering analysis are performed on 3D candidate routes to generate a set of route selection schemes covering different objective orientations, including:

[0186] First, the multidimensional evaluation index vectors of all candidate routes undergo a consistent preprocessing process to eliminate the impact of differences in the dimensions and scales of different indicators on the ranking. Preprocessing methods can employ interval normalization or standardization based on sample distribution. Indicators where smaller values ​​are considered better, and those where larger values ​​are considered better, are uniformly converted into unidirectional indicators. For example, indicators such as bridge-tunnel ratio, earthwork volume, ecological crossing length, curvature change rate, and longitudinal slope change rate are uniformly converted into "cost-based" indicators, allowing subsequent comparisons to be conducted in the same direction of merit. Based on this, a dominance relationship is constructed between candidate routes: for any two candidate routes, if one is not inferior to the other in all indicators and is superior in at least one indicator, then it is determined to dominate the other. A dominance hierarchy is established by traversing all candidate routes. The first-level non-dominated solution set is extracted as the preferred candidate set in a multi-objective context. Furthermore, the congestion or diversity measure of each route can be calculated to measure its distribution density in the indicator space, thus avoiding over-concentration of the preferred set on a single objective orientation.

[0187] Furthermore, after obtaining the non-dominated candidate set, cluster analysis is performed on this set to form scheme categories covering different target orientations. During clustering, a multi-dimensional evaluation index vector is used as feature input, and a distance metric is employed to calculate the similarity of candidate routes. The distance metric can be any of Euclidean distance, Mahalanobis distance, or cosine distance. The clustering algorithm can be any of K-means, hierarchical clustering, or density clustering; this embodiment does not limit the algorithm type. To enhance the engineering interpretability of the clustering results, dimensionality reduction or feature filtering can be performed on the index vector before clustering. For example, core indicators that contribute significantly to the differences between schemes, such as engineering cost, ecological avoidance, and linear smoothness, can be retained, so that the cluster centers correspond to clear engineering orientations. After clustering, routes within each cluster are sorted according to comprehensive priority. The comprehensive priority can be determined based on the dominance level, the distance from the cluster center, or a weighted score, thereby selecting one or more representative routes within each cluster as recommended schemes for that target orientation.

[0188] Furthermore, to ensure that the output scheme set covers different engineering preferences, output quotas or representativeness constraints can be set for each cluster. For example, a preset number of representative routes can be selected from different clusters such as the engineering cost priority cluster, the ecological avoidance priority cluster, and the linear smoothness priority cluster to form the final route selection scheme set.

[0189] The selection of representative routes can adopt one of the following rules: within the cluster, prioritize routes with the highest dominance level and greater congestion to balance the superiority of indicators and the diversity of solutions; or prioritize routes near the cluster center to represent the typical orientation of the cluster.

[0190] Through the above multi-objective ranking and clustering screening process, the output route selection scheme set can avoid the solidification of preferences caused by outputting only a single optimal route, while reducing the homogeneity among candidate routes and improving the coverage and comparability of the scheme set for different engineering objectives.

[0191] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A multi-objective intelligent 3D road selection method based on deep reinforcement learning, characterized in that, The method includes: Acquire 3D terrain and geographic information data of the candidate line area, and generate a 3D environment dataset through voxelization processing; Based on the three-dimensional environment dataset and combined with preset road design specification parameters, input parameters adapted to the preset decision-making agent are generated. The input parameters are input to the decision-making agent, which is then trained to output actions and generate candidate road centerlines in a 3D route selection environment based on the current state, resulting in multiple 3D candidate routes. The main body of the decision-making agent is a deep neural network. According to the preset evaluation strategy, the multiple three-dimensional candidate routes are calculated in parallel to obtain the multi-dimensional evaluation index vector corresponding to each candidate route. Based on the multi-dimensional evaluation index vector, the three-dimensional candidate routes are sorted and clustered to generate a set of route selection schemes covering different target orientations. The road design specifications include design speed, maximum longitudinal slope, minimum vertical curve radius, and roadbed width; The input parameters for generating a decision-making agent adapted to a preset model include: The forward sight length and sampling interval are determined based on the design speed, and the profile sampling range is determined based on the roadbed width; Based on the forward look length, sampling interval, and profile sampling range, multiple candidate action directions are determined; For each candidate action direction, extract the corresponding terrain forward profile sequence from the three-dimensional environment dataset; Based on the maximum longitudinal slope and the minimum vertical curve radius, construct a longitudinal feasible envelope sequence corresponding to the terrain forward profile sequence; Calculate the difference sequence of the topographic foresight profile sequence relative to the longitudinal feasible envelope sequence, wherein the difference sequence includes a first difference sequence characterizing cut pressure and a second difference sequence characterizing fill pressure; The terrain forward profile sequence, the longitudinal feasible envelope sequence, and the difference sequence are used as input parameters for the decision-making agent. Based on the maximum longitudinal slope and the minimum vertical curve radius, a longitudinal feasible envelope sequence corresponding to the topographic foresight profile sequence is constructed, including: Obtain the initial elevation and initial longitudinal slope corresponding to the current node of the road centerline to be generated, and discretize the forward length according to the sampling interval to obtain multiple mileage sampling points; The allowable range of longitudinal slope at each mileage sampling point is determined based on the maximum longitudinal slope, and the upper limit of longitudinal slope variation per unit mileage is determined based on the minimum vertical curve radius. Under the conditions of satisfying the allowable range of longitudinal slope and the upper limit of longitudinal slope change, the upper limit sequence of reachable elevation and the lower limit sequence of reachable elevation at each mileage sampling point are calculated by forward recursion. The consistency correction of the upper and lower bound sequences of reachable elevations is performed by backward contraction. The upper bound sequence of reachable elevations after consistency correction is used as the upper envelope sequence of the longitudinal feasible envelope, and the lower bound sequence of reachable elevations after consistency correction is used as the lower envelope sequence of the longitudinal feasible envelope, so as to obtain the longitudinal feasible envelope sequence corresponding to the terrain foresight profile sequence.

2. The multi-objective road 3D intelligent alignment method based on deep reinforcement learning according to claim 1, characterized in that, The three-dimensional terrain data includes digital elevation model data and three-dimensional point cloud data; wherein, the digital elevation model data is acquired through remote sensing mapping, and the three-dimensional point cloud data is acquired through lidar scanning. The geographic information data includes ecological zone data, land type data, and geological condition data, which are imported through a preset geographic information database.

3. The multi-objective road 3D intelligent alignment method based on deep reinforcement learning according to claim 2, characterized in that, The process of generating a 3D environment dataset through voxelization includes: Spatial registration and unified coordinate transformation are performed on the digital elevation model data and the 3D point cloud data; The registered 3D space is divided into voxels according to the preset spatial resolution to generate a regular voxel mesh. The digital elevation model data and 3D point cloud data are mapped to corresponding voxel units to obtain voxel features that characterize terrain height and space occupancy. The geographic information data is mapped to the voxel unit to label the corresponding attribute information in the voxel unit, wherein the attribute information includes ecological attributes, land type attributes and geological condition attributes; A 3D environment dataset is generated based on voxel units, voxel features, and attribute information.

4. The multi-objective road 3D intelligent alignment method based on deep reinforcement learning according to claim 1, characterized in that, The determination of multiple candidate action directions based on the forward look length, sampling interval, and profile sampling range includes: Starting from the current node of the road centerline to be generated, the radius of the forward search sector is determined according to the forward length; The lateral coverage width of the forward-looking search sector is determined based on the profile sampling range; Using the current node's direction of travel as the central axis, the direction angles are discretized within the horizontal coverage width of the forward-looking search sector according to a preset angle step size, resulting in multiple direction angles. Multiple candidate directional rays are determined based on the direction angle and the forward look length, serving as the multiple candidate action directions.

5. The multi-objective road 3D intelligent alignment method based on deep reinforcement learning according to claim 1, characterized in that, The decision-making agent includes: An input encoding layer is used to encode features of the terrain forward profile sequence, the longitudinal feasible envelope sequence, and the difference sequence to obtain a state feature vector. The input encoding layer includes a convolutional feature extraction sub-layer and a recurrent neural network sub-layer along the sequence dimension, as well as a fully connected layer for feature fusion. The strategy layer, connected to the input encoding layer, is used to output action selection based on the state feature vector. The action selection includes a discrete selection probability distribution for characterizing multiple candidate action directions, and an extension parameter vector corresponding to the selected candidate action direction for generating the road centerline. The value layer, connected to the input encoding layer, is used to output the action value corresponding to the state feature vector for training and updating the decision agent. The output layer is used to determine the target candidate action direction based on the discrete selection probability distribution output by the strategy layer in the current round, and to generate a centerline increment generation result corresponding to the target candidate action direction based on the extended parameter vector. The centerline increment generation result includes the three-dimensional coordinates of the next control point and the geometric parameters of the next line segment. The output layer is also used to evaluate and filter the centerline increment generation result based on the action value output by the value layer, and to superimpose the filtered centerline increment generation result onto the current road centerline, so as to output the updated road centerline as part of the input of the decision agent in the next round, until the termination condition is met and the candidate road centerline is output.

6. The multi-objective road 3D intelligent alignment method based on deep reinforcement learning according to claim 5, characterized in that, The process involves enabling a decision-making agent to output actions based on its current state and generate candidate road centerlines in a 3D route selection environment, resulting in multiple 3D candidate routes, including: Obtain the start and end constraints for generating the road centerline; Perform iterative operations, calculate random sampling seeds based on the starting point constraints and ending point constraints, initialize the starting control point of the current road centerline and the initial travel direction and initial longitudinal slope corresponding to the starting control point based on the random sampling seeds, call the decision agent to generate candidate road centerlines, until the preset maximum number of iteration steps is reached, and output multiple three-dimensional candidate routes corresponding to the maximum number of iteration steps.

7. The multi-objective road 3D intelligent alignment method based on deep reinforcement learning according to claim 1, characterized in that, The evaluation strategy includes: The multiple three-dimensional candidate routes are obtained, and each candidate route is segmented and sampled to obtain a route sampling sequence composed of multiple three-dimensional sampling points; Based on the route sampling sequence and the three-dimensional environmental dataset, a set of engineering cost indicators is calculated. The set of engineering cost indicators includes at least earthwork volume indicators and bridge-tunnel ratio indicators. Based on the route sampling sequence and the geographic information data, an ecological avoidance index set is calculated, which includes at least an ecological zone crossing length index and a land type occupation index. The set of alignment smoothness indices is calculated based on the route sampling sequence, and the set of alignment smoothness indices includes at least the curvature change rate index and the longitudinal slope change rate index. The engineering cost index set, the ecological avoidance index set, and the linear smoothness index set are normalized and combined to obtain a multi-dimensional evaluation index vector corresponding to each candidate route.

8. A multi-objective road 3D intelligent alignment system based on deep reinforcement learning, implemented based on the multi-objective road 3D intelligent alignment method based on deep reinforcement learning as described in any one of claims 1-7, characterized in that, The system includes: The data processing module is used to acquire three-dimensional terrain data and geographic information data of the area to be selected, and to perform voxelization processing on the three-dimensional terrain data and the geographic information data to generate a three-dimensional environment dataset. The input parameter generation module is used to generate input parameters adapted to the preset decision-making agent based on the three-dimensional environment dataset and in combination with preset road design specification parameters. The route selection decision module is used to input the input parameters into the decision agent, and select and iteratively generate candidate road centerlines through the output actions of the decision agent, thereby obtaining multiple three-dimensional candidate routes, wherein the main body of the decision agent is a deep neural network; The scheme evaluation module is used to perform parallel calculations on the multiple three-dimensional candidate routes according to a preset evaluation strategy, obtain the multi-dimensional evaluation index vector corresponding to each candidate route, and perform multi-objective ranking and cluster analysis on the three-dimensional candidate routes based on the multi-dimensional evaluation index vector to filter and generate a set of route selection schemes covering different objective orientations.

Images