Efficient trajectory planning method suitable for field unmanned carrying platform

Abstract

The invention relates to an efficient trajectory planning method suitable for a field unmanned carrying platform. The method comprises the following steps: S1, constructing a local elevation map and a passable domain map; s2, through a limit performance boundary model, outputting a limit boundary used for representing the limit capacity of the vehicle, constraining a coarse trajectory generated based on a deep learning model, and guiding the generation of a trajectory in combination with the autonomous operation attribute of a carrying platform during local optimization; s3, outputting a coarse trajectory point planning sequence, storing the coarse trajectory point planning sequence as long-time global planning information, and providing a high-quality initial value for a parallel local optimization trajectory planning module; s4, defining a generalized performance margin matrix, carrying out local convex optimization on a current track section by utilizing a main thread of a dual-thread track optimization strategy, taking the local convex optimization as controller execution input, asynchronously optimizing a next track section by an auxiliary thread, and carrying out rolling updating and continuous splicing on the track; and the real-time performance, the rationality, the safety and the generalization of the unmanned carrying platform trajectory planning method in the field non-structural environment are improved.

Description

Technical Field

[0001] The present invention relates to the fields of intelligent unmanned transport platforms, autonomous driving, and trajectory planning technologies, and more particularly to an efficient trajectory planning method suitable for field unmanned transport platforms, in particular an efficient trajectory planning method that considers external characteristics of vehicle dynamics and combines deep learning sampling with variable step-size parallel optimization. Background Art

[0002] With the continuous development of intelligent vehicle technology, unmanned transport platforms for off-road applications have garnered increasing attention, driven by practical needs in fields such as emergency rescue, scientific exploration, and resource exploration. Trajectory generation is crucial for enabling unmanned transport platforms to operate in all weather conditions. In unstructured off-road scenarios, roads often feature significant three-dimensional terrain features such as potholes, slopes, bumps, and insurmountable obstacles. Since search algorithms significantly increase computational complexity as the problem dimension increases, and directly applying high-dimensional dynamic models to global optimization often results in low computational efficiency and struggles to meet real-time requirements, traditional two-dimensional trajectory planning methods struggle to rapidly generate reasonable trajectories while ensuring safety and controllability, resulting in poor applicability in off-road scenarios. Furthermore, when solving trajectory planning problems in three-dimensional scenarios, they fail to fully consider vehicle dynamics, such as vehicle passability, stability, and safety. Consequently, they often produce overly conservative or highly risky planning results in complex environments. Furthermore, they neglect the execution capabilities of the underlying actuators, leading to system disconnects. Summary of the Invention

[0003] In view of the above problems, the purpose of the present invention is to provide an efficient trajectory planning method suitable for field unmanned carrier platforms, which is used to solve the problems of real-time, rationality, safety and generalization of existing unmanned carrier platform trajectory planning methods in field unstructured environments.

[0004] The present invention provides an efficient trajectory planning method applicable to an unmanned transport platform in the field, comprising the following steps:

[0005] Step S1: construct a local elevation map and a traversable area map based on multi-source data collected by the laser radar, visual sensor, and GNSS-INS inertial navigation sensor carried by the unmanned carrier platform;

[0006] Step S2: Using the vehicle's structural parameters, dynamic capabilities, and external environmental characteristics as input, the extreme performance boundary model is used to output a limit boundary representing the vehicle's extreme capabilities, i.e., the vehicle's extreme performance boundary. A coupling relationship is established between the vehicle's own driving capabilities and the parameters representing the common characteristics of the environment. This constrains the coarse trajectory generated by the deep learning model and, during local optimization, guides the trajectory generation by combining the autonomous operating properties of the carrier platform.

[0007] Step S3: The local elevation map, the traversable area map, the navigation target point, and the vehicle status information are used as inputs to the neural network. The output is a coarse trajectory point planning sequence, which is stored as long-term global planning information to provide high-quality initial values ​​for the parallel local optimization trajectory planning module.

[0008] Step S4: Based on the complexity of the terrain environment, a non-uniform time step partitioning strategy is adopted for the coarse sampling trajectory set, the vehicle extreme performance boundary is parameterized, and a generalized performance margin matrix is ​​defined. At the same time, according to the current working mode, the vehicle performance utilization and safety expectations in the generalized performance margin matrix are adjusted as constraints for local trajectory optimization to guide trajectory optimization; the main thread of the dual-thread trajectory optimization strategy is used to perform local convex optimization on the current trajectory segment as the controller execution input, and the auxiliary thread asynchronously optimizes the next trajectory segment, and performs rolling updates and continuous splicing on the trajectory.

[0009] As a preferred embodiment of the present invention, step S1 further includes the following steps:

[0010] Step S11: Use the upstream environmental perception module to process the collected multi-source data to complete the construction of the local elevation map and the traversable domain map and the extraction of high-dimensional semantics, wherein the local elevation information of the local elevation map and the traversable domain information of the traversable domain map are unified into a multi-channel image with semantic labels; specifically: Channel 1 elevation map H(x,y), which represents the absolute height of the ground unit; Channel 2 traversability score map T(x,y), which represents the vehicle traversability evaluation; Channel 3 obstacle distance map D(x,y), which represents the Euclidean distance to the nearest obstacle; Channel 4 terrain roughness map R(x,y), which reflects the degree of ground undulation; Channel 5 semantic code S(x,y), which is the surface category label, which is directly used as the input tensor of the deep learning network through a unified data format;

[0011] Step S12: The local elevation map is projected onto the BEV plane using the laser point cloud data and the Z value is averaged and presented as an accessibility score map. The height variance is statistically calculated as the roughness of the terrain. The semantic label is generated by the Deeplabv3 or PSPNet semantic segmentation network to generate a numerical code uniform identification.

[0012] As a preference of the present invention, step S2 also includes the following steps:

[0013] Step S21: Vehicle structural parameters specifically include: minimum ground clearance, approach / departure angle, longitudinal passability angle, maximum climbable height, maximum safe rollover ratio, and maximum longitudinal slope passability; dynamic capabilities include: lateral and longitudinal tire friction limits, minimum turning radius, and critical rollover acceleration; external environmental characteristics include: road slope, road friction coefficient, and curvature; vehicle structural parameters, dynamic capabilities, and external environmental characteristics constitute the vehicle's ultimate performance boundaries at the current moment in a certain driving state, a certain specific vehicle structural configuration, and a certain specific environment. The vehicle's ultimate performance boundaries are specifically characterized by: handling stability boundary, lateral and longitudinal coupled acceleration boundary, passability boundary, and lateral stability boundary.

[0014] Step S22: After the deep learning model generates a rough trajectory, it must be verified against the vehicle's extreme performance boundaries to ensure that the trajectory meets the constraints of the vehicle's dynamic capabilities.

[0015] Step S23: A generalized performance margin matrix is ​​defined using the vehicle's extreme performance boundary. The generalized performance margin matrix is ​​used to describe the vehicle's performance utilization and safety expectations. A parameterized mathematical model of the vehicle's extreme performance boundary is obtained through particle swarm optimization and multiple targeting methods. When the area or volume occupied by the set of vehicle driving states described by the state quantities in the vehicle's trajectory planning sequence within the boundary is larger, it indicates that the vehicle's performance utilization is higher and the safety expectations are lower. The generalized performance margin matrix dynamically adjusts the vehicle's trajectory planning strategy through the vehicle's performance utilization, establishes a mapping relationship between task-working mode-strategy, and guides the vehicle to generate a trajectory sequence.

[0016] As a preference of the present invention, step S3 also includes the following steps:

[0017] Step S31: Use a neural network to simultaneously pre-plan the vehicle speed and acceleration information corresponding to each trajectory point. The local optimization trajectory planning module adopts an encoder-decoder plus multi-head output mechanism network structure, and decouples the prediction of the spatial trajectory from the vehicle state.

[0018] As a preferred embodiment of the present invention, the encoder in step S31 uses a CNN encoder, the CNN encoder uses ResNet18 or UNet to extract environmental spatial features, and the decoder uses a Transformer decoder to predict the output of the rough trajectory based on the vehicle state information.

[0019] As a preferred embodiment of the present invention, the CNN encoder inputs a multi-channel image, uses a multi-layer perceptron to splice and fuse the vehicle's current and historical 5-frame states, the navigation target point, and the extracted environmental feature tensor, and feeds the fused features into the Transformer decoder, while outputting a motion parameter sequence, which includes: a coordinate sequence, a velocity sequence, and an acceleration sequence;

[0020] The output of the Transformer decoder is a sequence of trajectory points of length T, including three-dimensional coordinate information, velocity and acceleration information. The feature vector generated by the multi-layer perceptron splicing encodes the global state of the vehicle-environment-target and serves as the key / value input of the Transformer decoder.

[0021] As a preferred embodiment of the present invention, a supervised learning dataset is constructed within the neural network to form a data closed loop, so that the coarse trajectory continuously approaches the planning performance of the local trajectory optimizer, and then the guidance characteristics of the local optimization trajectory planning module are implicitly integrated into the neural network through the generalized performance margin matrix; finally, the output trajectory sequence is collected multiple times to form a candidate trajectory cluster, and the trajectories outside the preset range are eliminated through vehicle extreme performance boundary verification, wherein the supervised learning dataset includes: real vehicle trajectory playback, offline optimizer trajectory generation, and trajectory set output by the local optimization trajectory planning module.

[0022] As a preference of the present invention, step S4 also includes the following steps:

[0023] Step S41: The non-uniform time step division strategy dynamically adjusts the density of trajectory points according to the slope, obstacle density, and trajectory point curvature;

[0024] Step S42: Using the generalized performance margin matrix to normalize all extreme performance boundaries as input, the performance utilization is defined as the value of each item in the generalized performance margin matrix, ranging from [0, 1]. The closer the value is to 1, the higher the risk.

[0025] Step S43: Using performance utilization for local trajectory optimization includes the following two methods:

[0026] Method 1: For soft constraint settings: When the value is close to 1, the relaxation factor penalty is increased to form a dynamic safety margin;

[0027] Method 2: Used to form the loss term in the cost function: Convert the strategic choice of the working mode into the weight coefficient of each loss in performance utilization.

[0028] As a preferred embodiment of the present invention, in step S4, the local trajectory optimization and the coarse trajectory in step S3 are combined in parallel and asynchronously. When the vehicle is at time T0, a coarse trajectory within the next 10 seconds is generated. The coarse trajectory is specifically divided into the first to fifth segments. At the same time, the local trajectory optimizer optimizes the first two segments and performs trajectory splicing and fitting. When the vehicle has passed the first segment at time T1, a new coarse trajectory within the next 10 seconds is generated in a rolling manner. At the same time, the second segment processed by the local trajectory optimizer replaces the coarse trajectory. Repeated iterations are performed to improve the efficiency and quality of trajectory generation.

[0029] As a preferred embodiment of the present invention, in step S4, the dual-thread trajectory optimization strategy satisfies the architecture of the trajectory splicing through a dual-thread shared cache area. The main thread outputs the current local optimized trajectory segment to provide it to the vehicle for immediate trajectory tracking control input. The auxiliary thread processes the advance optimization of the next trajectory segment in parallel and stores the optimized trajectory segment into the shared splicing cache area, so that the main thread execution is stable.

[0030] The beneficial effects of the present invention are as follows:

[0031] Compared with the existing technology, the present invention introduces an extreme performance boundary model to couple the vehicle structural parameters, dynamic capabilities and external environment characteristics for modeling, which not only improves the physical consistency and feasibility of the trajectory, but also reduces the dependence on a large amount of manually labeled data, and enhances the generalization ability and interpretability of the network. The deep network architecture combining the CNN encoder, MLP fusion mechanism and Transformer multi-branch decoder can simultaneously output a variety of motion information such as spatial trajectory points, velocity, acceleration, etc., providing a high-quality reference for subsequent local trajectory optimization. At the same time, a dual-thread trajectory optimization strategy based on non-uniform step size division is proposed, which realizes the rapid optimization of the control segment by the main thread and the asynchronous preprocessing of the prediction segment by the auxiliary thread, while ensuring real-time performance while improving trajectory continuity and system response speed. In addition, by constructing a generalized performance margin matrix, it supports dynamic strategy adjustment under different task modes (such as energy saving, efficiency, and stability), making trajectory planning more consistent with the operation objectives and vehicle capability boundaries, and has good versatility, scalability and applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] By referring to the following description in conjunction with the accompanying drawings, and with a more complete understanding of the present invention, other objects and results of the present invention will become more clear and easy to understand. In the accompanying drawings:

[0033] Figure 1 This is a logic flow chart of an efficient trajectory planning method applicable to an unmanned vehicle platform in the field according to an embodiment of the present invention;

[0034] Figure 2 Schematic diagram of the principle of the method proposed in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] See Figure 1-2A specific embodiment of the present invention is directed to an unmanned autonomous working platform in a field environment. The platform is equipped with sensors such as lidar, visual sensors, and GNSS-INS inertial navigation. At the same time, there are obstacles in the working scene that the unmanned platform cannot cross, and the platform has a variety of different work tasks and corresponding different driving requirements, such as energy saving, efficiency, and stability. The present invention aims to provide an effective trajectory planning solution to cope with the above-mentioned challenging working conditions. This embodiment discloses an efficient trajectory planning method suitable for field unmanned platforms, including the following steps:

[0036] Step S1: construct a local elevation map and a traversable area map based on multi-source data collected by the laser radar, visual sensor, and GNSS-INS inertial navigation sensor carried by the unmanned carrier platform;

[0037] Step S11: Use the upstream environmental perception module to process the collected multi-source data, complete the construction of the local elevation map and the traversable domain map, and extract high-dimensional semantics, which are used as one of the inputs of the trajectory sampling network in step S2. The local elevation information of the local elevation map and the traversable domain information of the traversable domain map are unified into a multi-channel image with semantic labels; specifically: Channel 1: elevation map H(x,y), which represents the absolute height of the ground unit; Channel 2: traversability score map T(x,y), which represents the vehicle passability evaluation; Channel 3: obstacle distance map D(x,y), which represents the Euclidean distance to the nearest obstacle; Channel 4: terrain roughness map R(x,y), which reflects the degree of ground undulation; Channel 5: semantic code S(x,y), which is the surface category label. Through the unified data format, it is directly used as the input tensor of the deep learning network; it is easy to expand and does not require additional preprocessing.

[0038] Step S12: The local elevation map is projected onto the BEV (bird's eye view) plane using laser point cloud data and the Z value is averaged and presented. The accessibility score map is generated based on logical rules by comprehensively calculating the slope, obstacle height, obstacle labeling, and neighborhood undulation. The obstacle distance map uses a binary image for distance transformation. The roughness map uses a 4×4 neighborhood as a unit, and the height variance is statistically calculated as the roughness of the terrain. The semantic label is generated by the Deeplabv3 or PSPNet semantic segmentation network to generate a numerical code uniform identification.

[0039] Step S2: Using the vehicle's structural parameters, dynamic capabilities, and external environmental characteristics as input, the extreme performance boundary model is used to output a limit boundary representing the vehicle's extreme capabilities, i.e., the vehicle's extreme performance boundary. A coupling relationship is established between the vehicle's own driving capabilities and the parameters representing the common characteristics of the environment. This constrains the coarse trajectory generated by the deep learning model and, during local optimization, guides the trajectory generation by combining the autonomous operating properties of the carrier platform.

[0040] Step S21: It should be noted that the vehicle structural parameters related to passability and stability in field trajectory planning specifically include: minimum ground clearance, approach angle / departure angle, longitudinal passability angle, maximum climbable height, maximum safe rollover ratio, and maximum longitudinal slope passability; the above parameters and indicators need to be calculated. Dynamic capabilities include: lateral and longitudinal tire friction limits, minimum turning radius, and critical rollover acceleration; external environmental characteristics include: road slope, road friction coefficient, and curvature; vehicle structural parameters, dynamic capabilities, and external environmental characteristics constitute the vehicle's current limit performance boundary under a certain driving state, a certain specific vehicle structural configuration, and a certain specific environment. The vehicle's limit performance boundary is specifically characterized by: handling stability boundary, lateral and longitudinal coupling acceleration boundary, passability boundary, and lateral stability boundary.

[0041] In step S22, after the deep learning model generates a rough trajectory, it is verified against the vehicle's performance limits to ensure that the trajectory meets the constraints of the vehicle's dynamic capabilities. This improves applicability while also avoiding the online computational burden of calculating complex vehicle dynamics models.

[0042] Step S23: defining a generalized performance margin matrix using the vehicle's extreme performance boundaries. The generalized performance margin matrix is ​​used to describe the vehicle's performance utilization and safety expectations. The above four extreme performance boundaries are all closed plane figures or three-dimensional surfaces. A parameterized mathematical model of the vehicle's extreme performance boundaries is obtained through particle swarm optimization and multiple targeting methods. When the area or volume occupied by the set of vehicle driving states described by the state quantities in the vehicle's trajectory planning sequence within the boundary is larger, it indicates that the vehicle's performance utilization is higher and the safety expectations are lower. The generalized performance margin matrix dynamically adjusts the vehicle's trajectory planning strategy through the vehicle's performance utilization, establishes a mapping relationship between task-working mode-strategy, and more flexibly guides the vehicle to generate a trajectory sequence.

[0043] Step S3: The local elevation map, the traversable area map, the navigation target point (obtained by the GNSS-INS inertial navigation sensor), and the vehicle status information are used as inputs to the neural network. The output is a coarse trajectory point planning sequence, which is stored as long-term global planning information to provide high-quality initial values ​​for the parallel local optimization trajectory planning module.

[0044] Step S31: Use a neural network to simultaneously pre-plan the vehicle speed and acceleration information corresponding to each trajectory point. The local optimization trajectory planning module adopts an encoder-decoder plus multi-head output mechanism network structure, and decouples the prediction of the spatial trajectory from the vehicle state.

[0045] Furthermore, the encoder uses a CNN encoder, which selects ResNet18 or UNet to extract environmental spatial features, and the decoder uses a Transformer decoder to predict the output of the coarse trajectory based on the vehicle state information.

[0046] Furthermore, the CNN encoder inputs a multi-channel image and uses a multi-layer perceptron (MLP) to concatenate and fuse the vehicle's current and historical five-frame states, the navigation target point, and the extracted environmental feature tensor. The fused features are then fed into the Transformer decoder (multi-branch decoder), which also outputs a motion parameter sequence, including a coordinate sequence, a velocity sequence, and an acceleration sequence.

[0047] Furthermore, the output of the Transformer decoder (multi-branch decoder) is a trajectory point sequence of length T, including three-dimensional coordinate information, velocity and acceleration information. The feature vector generated by the multi-layer perceptron splicing encodes the global state of the vehicle-environment-target and serves as the key / value input of the Transformer decoder.

[0048] It should be noted that the neural network used in this embodiment needs to construct a supervised learning dataset. The data set mainly comes from: real vehicle trajectory playback, offline optimizer trajectory generation, and the trajectory set output by the local optimization trajectory planning module. This forms a data closed loop, so that the coarse trajectory continuously approaches the planning performance of the local trajectory optimizer. Then, the guidance characteristics of the local optimization trajectory planning module are implicitly integrated into the neural network through the generalized performance margin matrix; finally, the output trajectory sequence is collected multiple times to form a candidate trajectory cluster. Through the vehicle limit performance boundary check, the trajectories outside the preset range are eliminated.

[0049] Step S4: Based on the complexity of the terrain environment, a non-uniform time step partitioning strategy is adopted for the coarse sampling trajectory set, the vehicle extreme performance boundary is parameterized, and a generalized performance margin matrix is ​​defined. At the same time, according to the current working mode, the vehicle performance utilization and safety expectations in the generalized performance margin matrix are adjusted as constraints for local trajectory optimization to guide trajectory optimization; the main thread of the dual-thread trajectory optimization strategy performs local convex optimization on the current trajectory segment as the controller execution input, and the auxiliary thread asynchronously optimizes the next trajectory segment, and performs rolling updates and continuous splicing on the trajectory;

[0050] Step S41: The non-uniform time step division strategy dynamically adjusts the density of trajectory points according to the slope, obstacle density, and trajectory point curvature to reduce the computational burden in flat areas and improve spatial recognition capabilities;

[0051] Step S42: Using the generalized performance margin matrix to normalize all extreme performance boundaries as input, the performance utilization is defined as the value of each item in the generalized performance margin matrix, ranging from [0, 1]. The closer the value is to 1, the higher the risk.

[0052] Step S43: Using performance utilization for local trajectory optimization includes the following two methods:

[0053] Method 1: For soft constraint settings: When the value is close to 1, the relaxation factor penalty is increased to form a dynamic safety margin;

[0054] Method 2: Used to form the loss term in the cost function: The operating mode strategy is converted into weight coefficients for each loss in performance utilization. For example, a higher weight coefficient can be selected in the stable mode than in the efficient mode to increase the penalty for vehicle performance utilization, making the trajectory strategy more conservative. It should be noted that the above two methods can be used in combination.

[0055] It should be pointed out that in local trajectory optimization, other influencing factors need to be considered, such as comfort, safety, time efficiency, and trajectory smoothness. Since traditional trajectory optimization methods also take this into consideration, it is not considered an innovative contribution of the present invention.

[0056] Furthermore, local trajectory optimization and the coarse trajectory in step S3 are combined in parallel and asynchronously. When the vehicle is at time T0, a coarse trajectory for the next 10 seconds is generated. The coarse trajectory is specifically divided into segments 1 to 5. At the same time, the local trajectory optimizer optimizes the first two segments and performs trajectory splicing and fitting. When the vehicle has passed the first segment at time T1, a new coarse trajectory for the next 10 seconds is generated in a rolling manner. At the same time, the second segment processed by the local trajectory optimizer replaces the coarse trajectory. Repeated iterations improve the efficiency and quality of trajectory generation.

[0057] Furthermore, the dual-thread trajectory optimization strategy satisfies the trajectory splicing architecture through a dual-thread shared cache. The main thread outputs the current locally optimized trajectory segment to provide the vehicle with immediate trajectory tracking control input, and the auxiliary thread parallelly processes the advance optimization of the next trajectory segment and stores the optimized trajectory segment in the shared splicing cache, making the main thread execution stable and unaffected by large step size calculation blocking.

[0058] Those skilled in the art will appreciate that all or part of the process steps of the above-described embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, such as a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0059] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. An efficient trajectory planning method suitable for field unmanned transport platforms, characterized in that: The following steps are involved: Step S1: construct a local elevation map and a traversable area map based on multi-source data collected by the laser radar, visual sensor, and GNSS-INS inertial navigation sensor carried by the unmanned carrier platform; Step S2: Using the vehicle's structural parameters, dynamic capabilities, and external environmental characteristics as input, the extreme performance boundary model is used to output a limit boundary representing the vehicle's extreme capabilities, i.e., the vehicle's extreme performance boundary. A coupling relationship is established between the vehicle's own driving capabilities and the parameters representing the common characteristics of the environment. This constrains the coarse trajectory generated by the deep learning model and, during local optimization, guides the trajectory generation by combining the autonomous operating properties of the carrier platform. Step S3: The local elevation map, the traversable area map, the navigation target point, and the vehicle status information are used as inputs to the neural network. The output is a coarse trajectory point planning sequence, which is stored as long-term global planning information to provide high-quality initial values ​​for the parallel local optimization trajectory planning module. Step S4: Based on the complexity of the terrain environment, a non-uniform time step partitioning strategy is adopted for the coarse sampling trajectory set, the vehicle extreme performance boundary is parameterized, and a generalized performance margin matrix is ​​defined. At the same time, according to the current working mode, the vehicle performance utilization and safety expectations in the generalized performance margin matrix are adjusted as constraints for local trajectory optimization to guide trajectory optimization; the main thread of the dual-thread trajectory optimization strategy is used to perform local convex optimization on the current trajectory segment as the controller execution input, and the auxiliary thread asynchronously optimizes the next trajectory segment, and performs rolling updates and continuous splicing on the trajectory.

2. The efficient trajectory planning method for an unmanned transport platform in the field according to claim 1, characterized in that: Step S1 also includes the following steps: Step S11: Use the upstream environmental perception module to process the collected multi-source data to complete the construction of the local elevation map and the traversable domain map and the extraction of high-dimensional semantics, wherein the local elevation information of the local elevation map and the traversable domain information of the traversable domain map are unified into a multi-channel image with semantic labels; specifically: Channel 1 elevation map H(x,y), which represents the absolute height of the ground unit; Channel 2 traversability score map T(x,y), which represents the vehicle traversability evaluation; Channel 3 obstacle distance map D(x,y), which represents the Euclidean distance to the nearest obstacle; Channel 4 terrain roughness map R(x,y), which reflects the degree of ground undulation; Channel 5 semantic code S(x,y), which is the surface category label, which is directly used as the input tensor of the deep learning network through a unified data format; Step S12: The local elevation map is projected onto the BEV plane using the laser point cloud data and the Z value is averaged and presented as an accessibility score map. The height variance is statistically calculated as the roughness of the terrain. The semantic label is generated by the Deeplabv3 or PSPNet semantic segmentation network to generate a numerical code uniform identification.

3. The efficient trajectory planning method for an unmanned transport platform in the field according to claim 1, characterized in that: Step S2 also includes the following steps: Step S21: Vehicle structural parameters specifically include: minimum ground clearance, approach / departure angle, longitudinal passability angle, maximum climbable height, maximum safe rollover ratio, and maximum longitudinal slope passability; dynamic capabilities include: lateral and longitudinal tire friction limits, minimum turning radius, and critical rollover acceleration; external environmental characteristics include: road slope, road friction coefficient, and curvature; vehicle structural parameters, dynamic capabilities, and external environmental characteristics constitute the vehicle's ultimate performance boundaries at the current moment in a certain driving state, a certain specific vehicle structural configuration, and a certain specific environment. The vehicle's ultimate performance boundaries are specifically characterized by: handling stability boundary, lateral and longitudinal coupled acceleration boundary, passability boundary, and lateral stability boundary. Step S22: After the deep learning model generates a rough trajectory, it must be verified against the vehicle's extreme performance boundaries to ensure that the trajectory meets the constraints of the vehicle's dynamic capabilities. Step S23: A generalized performance margin matrix is ​​defined using the vehicle's extreme performance boundary. The generalized performance margin matrix is ​​used to describe the vehicle's performance utilization and safety expectations. A parameterized mathematical model of the vehicle's extreme performance boundary is obtained through particle swarm optimization and multiple targeting methods. When the area or volume occupied by the set of vehicle driving states described by the state quantities in the vehicle's trajectory planning sequence within the boundary is larger, it indicates that the vehicle's performance utilization is higher and the safety expectations are lower. The generalized performance margin matrix dynamically adjusts the vehicle's trajectory planning strategy through the vehicle's performance utilization, establishes a mapping relationship between task-working mode-strategy, and guides the vehicle to generate a trajectory sequence.

4. The efficient trajectory planning method for an unmanned transport platform in the field according to claim 1, characterized in that: Step S3 also includes the following steps: Step S31: Use a neural network to simultaneously pre-plan the vehicle speed and acceleration information corresponding to each trajectory point. The local optimization trajectory planning module adopts an encoder-decoder plus multi-head output mechanism network structure, and decouples the prediction of the spatial trajectory from the vehicle state.

5. The efficient trajectory planning method for an unmanned transport platform in the field according to claim 4, characterized in that: The encoder in step S31 uses a CNN encoder, which uses ResNet18 or UNet to extract environmental spatial features, and the decoder uses a Transformer decoder to predict the output of the rough trajectory based on the vehicle state information.

6. The efficient trajectory planning method for an unmanned transport platform in the field according to claim 5, characterized in that: The CNN encoder takes a multi-channel image as input and uses a multi-layer perceptron to concatenate and fuse the vehicle's current and historical five-frame states, the navigation target point, and the extracted environmental feature tensor. The fused features are fed into the Transformer decoder, which simultaneously outputs a motion parameter sequence, including a coordinate sequence, a velocity sequence, and an acceleration sequence. The output of the Transformer decoder is a sequence of trajectory points of length T, including three-dimensional coordinate information, velocity and acceleration information. The feature vector generated by the multi-layer perceptron splicing encodes the global state of the vehicle-environment-target and serves as the key / value input of the Transformer decoder.

7. The efficient trajectory planning method for an unmanned transport platform in the field according to claim 1, characterized in that: A supervised learning dataset is constructed within the neural network, forming a closed data loop. This allows the coarse trajectory to continuously approach the planning performance of the local trajectory optimizer. The guidance characteristics of the local optimization trajectory planning module are then implicitly integrated into the neural network through a generalized performance margin matrix. Finally, the output trajectory sequence is collected multiple times to form a cluster of candidate trajectories. Trajectories outside the preset range are eliminated through vehicle extreme performance boundary verification. The supervised learning dataset includes: real-vehicle trajectory playback, trajectory generation by the offline optimizer, and trajectory sets output by the local optimization trajectory planning module.

8. The efficient trajectory planning method for an unmanned transport platform in the field according to claim 1, characterized in that: Step S4 also includes the following steps: Step S41: The non-uniform time step division strategy dynamically adjusts the density of trajectory points according to the slope, obstacle density, and trajectory point curvature; Step S42: Using the generalized performance margin matrix to normalize all extreme performance boundaries as input, the performance utilization is defined as the value of each item in the generalized performance margin matrix, ranging from [0, 1]. The closer the value is to 1, the higher the risk. Step S43: Using performance utilization for local trajectory optimization includes the following two methods: Method 1: For soft constraint settings: When the value is close to 1, the relaxation factor penalty is increased to form a dynamic safety margin; Method 2: Used to form the loss term in the cost function: Convert the strategic choice of the working mode into the weight coefficient of each loss in performance utilization.

9. The efficient trajectory planning method for an unmanned transport platform in the field according to claim 1, characterized in that: In step S4, local trajectory optimization and the coarse trajectory in step S3 are combined in parallel and asynchronously. When the vehicle is at time T0, a coarse trajectory for the next 10 seconds is generated. The coarse trajectory is specifically divided into segments 1 to 5. At the same time, the local trajectory optimizer optimizes the first two segments and performs trajectory splicing and fitting. When the vehicle has passed the first segment at time T1, a new coarse trajectory for the next 10 seconds is generated in a rolling manner. At the same time, the second segment processed by the local trajectory optimizer replaces the coarse trajectory. Repeated iterations improve the efficiency and quality of trajectory generation.

10. The efficient trajectory planning method for an outdoor unmanned transport platform according to claim 9, characterized in that: In step S4, the dual-thread trajectory optimization strategy satisfies the trajectory splicing architecture through a dual-thread shared cache. The main thread outputs the current local optimized trajectory segment to provide the vehicle with immediate trajectory tracking control input. The auxiliary thread processes the advance optimization of the next trajectory segment in parallel and stores the optimized trajectory segment in the shared splicing cache, making the main thread execution stable.

Images