Cross-country scene human-like trajectory planning method without high-precision map
By extracting height feature points and constructing graph data in off-road scenes without high precision maps, using deep learning technology to generate safe and reliable trajectories, the problem of insufficient path rationality and safety in the existing technology is solved, and efficient and flexible autonomous driving operations in complex environments are achieved.
Patent Information
- Application Number
- CN202411735460.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-29
- Publication Date
- 2025-04-29
AI Technical Summary
Without high-precision maps, it is difficult for the prior art to plan a sufficiently reasonable and safe path in off-road scenarios, especially in complex and dynamic environments.
By extracting height feature points from each frame local elevation map of unmanned vehicles, building a frame sequence data set, and using spatiotemporal graph convolution network and timing prediction network, combining target point prediction network and clustering technology, a safe and reliable trajectory is generated.
Based on the global sparse road network, the generation of trajectories that are closer to the real humanoid trajectory and more secure is achieved, reducing dependence on preset rules, and showing higher robustness and flexibility in complex environments.
Smart Images
Figure CN119904834A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of trajectory planning technology, and in particular to a humanoid trajectory planning method for off-road scenarios without high-precision maps. Background Art
[0002] High-precision maps have important semantic value and geometric information. However, in off-road scenarios, it is not feasible to build centimeter-level high-precision maps. Therefore, under the premise of only rough road network prompts, getting rid of the dependence on high-precision maps for trajectory planning is an important autonomous driving solution.
[0003] Currently, rules and cost functions are usually constructed and the planning scheme is obtained by solving them.
[0004] However, solutions based on rule and cost function optimization are often inflexible and heavily dependent on preset rules and models. They are difficult to adapt to complex and dynamic environmental changes and may lead to limitations in path planning. At the same time, in cost function optimization, designing a reasonable cost function requires a full understanding of the environment and vehicle performance. In off-road scenarios, the diverse environmental changes make it difficult to design an accurate and applicable cost function. Incorrect design is likely to lead to an undesirable path or inefficiency. Summary of the invention
[0005] In view of the above analysis, an embodiment of the present invention aims to provide a humanoid trajectory planning method for off-road scenarios without high-precision maps, so as to solve the problems of insufficient rationality and safety of existing paths planned without high-precision maps.
[0006] The embodiment of the present invention provides a method for human-like trajectory planning in an off-road scenario without a high-precision map, comprising the following steps:
[0007] Extract height feature points from each frame of the local elevation map of the unmanned vehicle, and construct a frame sequence dataset based on the height feature points of each frame, vehicle trajectory points, and road points in the global sparse road network;
[0008] Take multiple frames of data from the frame sequence data set according to the time window to construct the graph data, pass it into the spatiotemporal graph convolutional network, extract the graph feature vector and pass it into the time series prediction network to obtain multi-frame trajectory points; pass the multi-frame trajectory points into the target point prediction network, predict multiple target points corresponding to the last frame of the multi-frame trajectory points, and construct the target thermal matrix based on the multiple target points and the height feature points of the last frame; extract the feature vector of the target thermal matrix and fuse it with the graph feature vector and pass it into the time series prediction network again to obtain the trajectory points predicted in the current time window;
[0009] The time window is slided to sequentially cover and splice the trajectory points predicted in each time window to obtain the planned trajectory.
[0010] Based on the further improvement of the above method, the height feature points are extracted from each frame of the local elevation map of the unmanned vehicle, including:
[0011] The coordinates of each pixel point in each frame of the local elevation map are converted to the global coordinate system; the pixels in each frame of the local elevation map are clustered, and each cluster is used as a height feature point. The coordinates of the center point of each cluster and the size of the minimum circumscribed rectangle are taken as the coordinates and size of the height feature point, and the average pixel value of each cluster is taken as the pixel value of the height feature point.
[0012] Based on the further improvement of the above method, the pixel points in each frame of the local elevation map are clustered according to the pixel value of the pixel point and the distance between the pixel points, and the pixel points that simultaneously meet the pixel value tolerance threshold and the distance tolerance threshold are classified into the same cluster.
[0013] Based on the further improvement of the above method, constructing a frame sequence data set includes: taking the height feature points and vehicle trajectory points of each frame, and the road points in the global sparse road network as three types of feature points, and extracting the same multi-dimensional feature attributes for each feature point in each type of feature point as each frame data, wherein the multi-dimensional feature attributes include: frame sequence number, ID identifier, type identifier, x-coordinate, y-coordinate, z-coordinate, x-direction size, y-direction size, pixel value and heading angle.
[0014] Based on the further improvement of the above method, the type identifier in the multidimensional feature attribute is used to distinguish height feature points, vehicle trajectory points and waypoints; in two adjacent frames of local elevation maps, the height feature points whose pixel value difference and coordinate difference are less than or equal to the corresponding threshold are assigned the same ID identifier, otherwise, different ID identifiers are assigned; the heading angles of the height feature points are all 0.
[0015] Based on the further improvement of the above method, multiple frames of data are taken from the frame sequence data set according to the time window to construct the graph data, including:
[0016] Multiple frames of data are taken out according to the number of frames corresponding to the time window, and the height feature points, vehicle trajectory points and waypoints in each frame of data are used as nodes in each frame of graph data; the spatial edge set is constructed according to the distance threshold between nodes in each frame, and then the spatial connection adjacency matrix is obtained; the inter-frame edge set is constructed according to the frame number threshold between the same nodes between frames, and then the time-space self-connection adjacency matrix is obtained.
[0017] Based on the further improvement of the above method, the target thermal matrix is constructed according to multiple target points and the height feature points of the last frame, including:
[0018] Select several target points with higher probability from multiple target points;
[0019] According to the radiation threshold, a grid-based method is used to obtain the radiation area of several target points, and the initial thermal matrix is obtained by superimposing the thermal values of the same position in the radiation area;
[0020] According to the coordinates of the height feature points of the last frame, that is, the height feature points in the same frame as the target point, the elements in the initial thermal matrix are corresponding, and the thermal values of the elements corresponding to the height feature points whose pixel values are greater than the maximum threshold are set to 0 to obtain the target thermal matrix.
[0021] Based on the further improvement of the above method, the eigenvector of the target thermal matrix is extracted, which is the eigenvector obtained by passing the target thermal matrix into a trained convolutional neural network for encoding.
[0022] Based on the further improvement of the above method, the timing prediction network adopts a gated recurrent unit network.
[0023] Based on the further improvement of the above method, the target point prediction network is composed of a multi-layer perceptron, which uses the vehicle trajectory points in the historical data as training samples and the driver's actual trajectory points as true values for training, and predicts the positions of multiple target points and the probability of each target point.
[0024] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:
[0025] 1. Based on the rough guidance of the global sparse road network, multiple target points are predicted based on human-like driving knowledge and highly dangerous areas are removed to achieve safe and reliable encryption of the global sparse road network and generate local thermal guidance. The trajectory generated under the dual guidance not only meets the global navigation requirements, but also meets the human-like response to the local environment, and plans a trajectory that is closer to the real human-like trajectory and has stronger safety.
[0026] 2. By clustering pixel values to extract highly characteristic points, the number of nodes in the graph network is effectively reduced, a lightweight and streamlined graph structure is constructed, and the processing efficiency of the model is greatly improved. At the same time, the highly characteristic areas obtained by clustering make the features that the time series prediction network focuses on clearer, which promotes the network learning of the understanding and interaction between the vehicle itself and the environment.
[0027] 3. The target point prediction network is trained by simulating the driving trajectory of expert drivers. It not only focuses on the vehicle's driving path, but also includes the driver's decision-making process when facing dynamic obstacles, terrain changes and environmental characteristics, and learns human scene understanding in complex unstructured environments. By imitating the behavior of expert drivers, it can not only reduce dependence on preset rules, but also show higher robustness and flexibility when facing unknown complex environments, significantly improving the operational capabilities of autonomous driving in unstructured environments.
[0028] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.
[0030] Figure 1 This is a flow chart of a method for planning human-like trajectory in an off-road scenario without a high-precision map in an embodiment of the present invention;
[0031] Figure 2 A schematic diagram of obtaining trajectory points predicted in each time window in an embodiment of the present invention. DETAILED DESCRIPTION
[0032] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.
[0033] A specific embodiment of the present invention is directed to an autonomous driving vehicle in an off-road scenario, which is equipped with sensors such as cameras, lidar, satellite navigation and inertial navigation, and only has a global sparse road network corresponding to the off-road scenario, that is, no high-precision map, which only contains a small amount of target waypoint information and the vehicle's driving decision at the target waypoint. This embodiment discloses a human-like trajectory planning method for off-road scenarios without high-precision maps, such as Figure 1 As shown, the following steps are included:
[0034] S1. Extract height feature points from the local elevation map of each frame of the unmanned vehicle, and construct a frame sequence dataset based on the height feature points of each frame, vehicle trajectory points, and road points in the global sparse road network.
[0035] It should be noted that when the vehicle is driving, the SLAM (Simultaneous Localization and Mapping) system deployed on the vehicle receives data collected by various sensors, fuses the point cloud data collected by the real-time and high-precision lidar and the inertial measurement unit IMU data collected by inertial navigation, and outputs a local elevation map for each frame, as well as optimized vehicle trajectory points (including vehicle coordinates and posture).
[0036] Preferably, the SLAM system adopts the Lio-Sam framework.
[0037] Furthermore, coordinate transformation is performed based on the relative position of the pixel point and the vehicle in each frame of the local elevation map, as well as the vehicle's own posture, and each pixel point in each frame of the local elevation map is unified into the global coordinate system to obtain the global coordinates and pixel values of each pixel point. The larger the pixel value, the higher the altitude, which is used to reflect the height characteristics of the coordinates of each pixel point. The global coordinate system selects the IMU coordinate system.
[0038] Furthermore, the pixel points in each frame of the local elevation map are clustered, and each cluster is used as a height feature point, which reflects the environmental characteristics in the map.
[0039] Specifically, this embodiment sets a double threshold value according to the pixel value of the pixel point and the distance between the pixel points, and classifies the pixel points that meet the double threshold value into the same cluster, wherein the double threshold value includes: a pixel value tolerance threshold value and a distance tolerance threshold value. Exemplarily, the pixel value tolerance threshold value is 50, and the distance tolerance threshold value is 0.5m. The pixel points that meet the double threshold value are the pixel values and the distance differences between the pixel points that are less than or equal to the corresponding tolerance threshold values.
[0040] According to the coordinates of the center point of each cluster and the minimum circumscribed rectangle, the coordinates and size of each cluster are obtained as the coordinates and size of the height feature point, and the average pixel value of each cluster is taken as the pixel value of the height feature point.
[0041] Each height feature point is assigned a unique ID. In order to reflect the correlation between height feature points between maps, the height feature points are tracked. In two adjacent local elevation maps, the height feature points whose pixel value difference and coordinate difference are less than or equal to the corresponding threshold are assigned the same ID. Otherwise, different IDs are assigned.
[0042] Exemplarily, the threshold of the pixel value difference is 5, and the threshold of the coordinate difference is 0.2 m.
[0043] Finally, a frame sequence dataset is constructed according to the height feature points, vehicle trajectory points and road points in the global sparse road network of each frame, including: taking the height feature points and vehicle trajectory points of each frame and the road points in the global sparse road network as three types of feature points, and extracting the same multidimensional feature attributes for each feature point in each type of feature points as each frame data, wherein the multidimensional feature attributes include: frame sequence number, ID identifier, type identifier, x-coordinate, y-coordinate, z-coordinate, x-direction size, y-direction size, pixel value and heading angle.
[0044] That is to say, each frame of data in the frame sequence data set is composed of the multidimensional feature attributes of three types of feature points, and the data format of the three types of feature points is consistent, that is, the same multidimensional feature attributes are extracted. Among them, the type identifier is used to distinguish between height feature points, vehicle trajectory points and waypoints. For height feature points, the heading angle is 0; for vehicle trajectory points, the ID identifier is the vehicle's own identifier, the pixel value is 0, and the x-direction size and y-direction size are calculated based on the vehicle body size; for waypoints in the global sparse road network, the ID identifier is the unique identifier of each waypoint, the pixel value is 0, and the x-direction size and y-direction size are calculated based on the size of the area extended from the waypoint as the center. In other words, the ID identifier of the height feature point of each frame in the frame sequence data set is updated as the frame sequence changes, while the ID identifier and size of the vehicle trajectory point and waypoint remain unchanged.
[0045] This step extracts height feature points with obvious height characteristics from the local elevation map through clustering, which not only captures the characteristics of terrain changes and obstacles in the off-road environment, but also reduces unnecessary data, improves subsequent processing efficiency, and promotes the understanding and interaction between the learning vehicle itself and the environment.
[0046] S2. Take out multiple frames of data from the frame sequence data set according to the time window to construct graph data, pass it into the spatiotemporal graph convolutional network, extract the graph feature vector and pass it into the time series prediction network to obtain multi-frame trajectory points; pass the multi-frame trajectory points into the target point prediction network, predict multiple target points corresponding to the last frame of the multi-frame trajectory points, and construct the target thermal matrix based on the multiple target points and the height feature points of the last frame; extract the feature vector of the target thermal matrix and fuse it with the graph feature vector and pass it into the time series prediction network again to obtain the trajectory points predicted in the current time window.
[0047] It should be noted that this step includes two processes: Figure 2 As shown, in step S21, the initial trajectory points are first predicted by extracting the graph feature vector, and then in step S22, multiple target points in the local elevation map are predicted based on the predicted initial trajectory points, so as to achieve the purpose of encrypting the road network and refining the path selection, and after shielding the dangerous height, the target guidance area feature vector is extracted, and then the final trajectory points are predicted based on the target guidance area feature vector and the graph feature vector in step S21. The implementation methods of these two processes are described in detail below.
[0048] S21. Take out multiple frames of data from the frame sequence data set according to the time window to construct graph data, pass it into the spatiotemporal graph convolutional network, extract the graph feature vector and pass it into the time series prediction network to obtain multiple frame trajectory points.
[0049] This step extracts multiple frames of data from the frame sequence data set according to the time window length, and predicts the trajectory points of multiple frames in the future. For example, the time window length is 6 frames, and the number of predicted frames is 24 frames, that is, the historical 6 frames of data are taken to predict the data of the next 24 frames, for example, the 1st to 6th frames of data are taken to predict the trajectory points of the vehicle in the local elevation map of the 7th to 30th frames.
[0050] Graph data, i.e. graph structure data, is constructed from the extracted multi-frame data to capture the relationship between nodes, including:
[0051] The height feature points, vehicle trajectory points and waypoints in each frame of data are used as nodes in each frame of graph data; the spatial edge set is constructed according to the distance threshold between nodes in each frame, and then the spatial connection adjacency matrix is obtained; the inter-frame edge set is constructed according to the frame number threshold between the same nodes between frames, and then the time-space self-connection adjacency matrix is obtained.
[0052] Exemplarily, the distance threshold between nodes is set to 10 meters. If the difference between the position coordinates of the nodes is less than or equal to 10 meters, there is a spatial edge between the nodes, which are put into the spatial edge set, and the corresponding element in the spatial connection adjacency matrix is set to 1, otherwise, it is set to 0; the frame number threshold is set to 6 frames. If the number of frames in the multi-frame data is less than or equal to 6 frames, there is an inter-frame edge between the same nodes, which are put into the inter-frame edge set, and the corresponding element in the time-space self-connection adjacency matrix is set to 1, otherwise, it is set to 0.
[0053] The constructed graph data is taken as input and passed into the Spatial Temporal Graph Convolutional Networks (ST-GCN). The spatial convolution layer is used to learn the spatial geometric relationships among height feature points, waypoints, vehicle trajectory points, etc. At the same time, the temporal convolution layer is used to process the association between different time frames, understand the movement trend of the vehicle, and combine the past movement information with the current trajectory information to generate a spatiotemporal coupled graph feature vector, providing powerful spatiotemporal analysis capabilities for trajectory planning in off-road scenarios.
[0054] Furthermore, the extracted graph feature vector is passed into the time series prediction network to obtain subsequent multi-frame trajectory points relative to the time window as the initial trajectory points to be adjusted. Among them, the time series prediction network adopts the gated recurrent unit network (Gated Recurrent Unit, referred to as GRU).
[0055] S22, passing the multi-frame trajectory points into the target point prediction network, predicting multiple target points corresponding to the last frame of the multi-frame trajectory points, and constructing a target thermal matrix based on the multiple target points and the height feature points of the same frame; extracting the feature vector of the target thermal matrix and fusing it with the graph feature vector and passing it into the time series prediction network again to obtain the trajectory points predicted in the current time window.
[0056] It should be noted that the target point prediction network is composed of a multi-layer perceptron, which uses the vehicle trajectory points in the historical data as training samples and the driver's actual trajectory points as the true value for training, and predicts the positions of multiple target points and the probability of each target point. The process of training the target point prediction network is similar to simulating the driving trajectory of an expert driver and learning human scene understanding in a complex unstructured environment. Moreover, the output of the target point prediction network is set to multiple, with the purpose of learning the effective driving strategy of the expert based on the trained target point prediction network, and generating trajectories that adapt to different environmental conditions. This can not only reduce the dependence on preset rules, but also show higher robustness and flexibility when facing unknown complex environments, significantly improving the operational capabilities of autonomous driving in unstructured environments.
[0057] It should be noted that the generated target point is the local elevation map of the last frame corresponding to the multi-frame trajectory point. For example, the predicted trajectory points of the 7th to 30th frames are input into the target point prediction network, and multiple target points and their probabilities of the 30th frame are output, that is, multiple waypoints are added to the local elevation map of the 30th frame.
[0058] Furthermore, a target thermal matrix is constructed based on multiple target points and the height feature points of the last frame, including:
[0059] ① Select several target points with higher probabilities from multiple target points; that is, select a preset number of target points from multiple target points according to the probability from high to low, for example, output 15 target points, and select 10 target points with higher probability.
[0060] ②According to the radiation threshold, a grid-based method is used to obtain the radiation areas of several target points, and the initial thermal matrix is obtained by superimposing the thermal values of the same position in the radiation area;
[0061] For example, the radiation threshold is set to 1 meter, which means that the radiation area around each target point is 1 meter × 1 meter, and the grid resolution is 0.5 meter. Then, the radiation area of each target point is a 2 × 2 grid unit.
[0062] The size of the thermal matrix (or two-dimensional array) is determined according to the radiation area and grid resolution of each target point. The value of each element in the thermal matrix is initialized to 0 as the thermal value, and the thermal value corresponding to the radiation area of each target point is set to 1. The radiation area of each target point is superimposed in the thermal matrix in turn, and the thermal values at the same position are added to obtain the initial thermal matrix.
[0063] ③ According to the coordinates of the height feature points of the last frame, that is, the height feature points in the same frame as the target point, the elements in the initial thermal matrix are corresponding, and the thermal values of the elements corresponding to the height feature points whose pixel values are greater than the maximum threshold are set to 0 to obtain the target thermal matrix.
[0064] It should be noted that if the pixel value of the height feature point is greater than the maximum threshold, such as 200, it means that the height of the position is relatively high and may be an obstacle. If the position is in the radiation area of the target point, the unsafety of vehicle operation is relatively high, and the position needs to be shielded. Therefore, in this step, the coordinates of the height feature points are first mapped to the elements in the thermal matrix, and then the thermal values of the elements corresponding to the height feature points with pixel values greater than the maximum threshold are set to 0 to obtain a safe target thermal matrix, thereby performing local thermal guidance on the planning trend of the trajectory.
[0065] Furthermore, the target thermal matrix is passed into the trained convolutional neural network for encoding to obtain its feature vector, which is fused with the graph feature vector extracted in step S22 and then passed into the time series prediction network again to predict the final multi-frame trajectory points. That is, the multi-frame trajectory points obtained in step S21 are updated to obtain the trajectory points predicted in the current time window.
[0066] It should be noted that when the feature vector of the target thermal matrix is fused with the feature vector of the graph, the fused vector can be mapped to data of the same dimension as the feature vector of the graph through a linear layer, and the time series prediction network in step S21 can be used for prediction. Alternatively, a fused vector of a new dimension can be obtained by splicing, and another time series prediction network with the same dimension as the fused vector can be trained for prediction.
[0067] S3, sliding the time window, sequentially covering and splicing the trajectory points predicted in each time window to obtain the planned trajectory.
[0068] It should be noted that this step uses a sliding time window to use the historical multi-frame data in the time window to predict future multi-frame data. Since the sliding step of the time window is 1, the trajectory points predicted in each time window will have repeated frames. Therefore, as the time window slides, the latest data is continuously predicted. This step obtains the planned trajectory by overlaying and splicing the trajectory points predicted in each time window.
[0069] Exemplarily, the first time window uses the 1st to 6th frames to predict the trajectory points of the 7th to 30th frames, and the second time window uses the 2nd to 7th frames to predict the trajectory points of the 8th to 31st frames. Then the trajectory points of the 8th to 30th frames predicted by the second time window cover the trajectory points of the 8th to 30th frames predicted by the first time window, and the trajectory points of the 7th to 31st frames are spliced together.
[0070] Compared with the prior art, the humanoid trajectory planning method for off-road scenarios without high-precision maps provided in this embodiment is based on the rough guidance of the global sparse road network, predicts multiple target points based on humanoid driving knowledge and removes highly dangerous areas, realizes safe and reliable encryption of the global sparse road network, generates local thermal guidance, and the trajectory generated under the dual guidance not only meets the global navigation requirements, but also meets the humanoid response to the local environment, and plans a trajectory that is closer to the real humanoid trajectory and safer. By clustering pixel values to extract highly characteristic points, the number of nodes in the graph network is effectively reduced, a lightweight and streamlined graph structure is constructed, and the processing efficiency of the model is greatly improved; at the same time, the highly characteristic areas obtained by clustering make the features that the time series prediction network focuses on clearer, which helps the network learn the understanding and interaction between the vehicle itself and the environment. The target point prediction network is trained by simulating the driving trajectory of expert drivers. It not only focuses on the vehicle's driving path, but also includes the driver's decision-making process when facing dynamic obstacles, terrain changes and environmental characteristics, and learns human scene understanding in complex unstructured environments. By imitating the behavior of expert drivers, it can not only reduce dependence on preset rules, but also show higher robustness and flexibility when facing unknown complex environments, significantly improving the operational capabilities of autonomous driving in unstructured environments.
[0071] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.
[0072] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.
Claims
1. A humanoid trajectory planning method for off-road scenarios without high-precision maps, characterized in that: The following steps are involved: Extract height feature points from each frame of the local elevation map of the unmanned vehicle, and construct a frame sequence dataset based on the height feature points of each frame, vehicle trajectory points, and road points in the global sparse road network; Taking multiple frames of data from the frame sequence data set according to the time window to construct graph data, passing it into the spatiotemporal graph convolutional network, extracting the graph feature vector and passing it into the time series prediction network to obtain multi-frame trajectory points; passing the multi-frame trajectory points into the target point prediction network, predicting multiple target points of the last frame corresponding to the multi-frame trajectory points, and constructing a target thermal matrix based on the multiple target points and the height feature points of the last frame; extracting the feature vector of the target thermal matrix and fusing it with the graph feature vector and passing it into the time series prediction network again to obtain the trajectory points predicted in the current time window; The time window is slided to sequentially cover and splice the trajectory points predicted in each time window to obtain the planned trajectory.
2. The human-like trajectory planning method for off-road scenarios without high-precision maps according to claim 1 is characterized in that: The step of extracting height feature points from each frame of the local elevation map of the unmanned vehicle includes: The coordinates of each pixel point in each frame of the local elevation map are converted to the global coordinate system; the pixels in each frame of the local elevation map are clustered, and each cluster is used as a height feature point. The coordinates of the center point of each cluster and the size of the minimum circumscribed rectangle are taken as the coordinates and size of the height feature point, and the average pixel value of each cluster is taken as the pixel value of the height feature point.
3. The human-like trajectory planning method for off-road scenarios without high-precision maps according to claim 2 is characterized in that: The pixel points in each frame of the local elevation map are clustered, and the pixel points that simultaneously meet the conditions of being less than or equal to the pixel value tolerance threshold and the distance tolerance threshold are classified into the same cluster according to the pixel values of the pixel points and the distances between the pixel points.
4. The human-like trajectory planning method for off-road scenarios without high-precision maps according to claim 1 is characterized in that: The frame sequence data set construction includes: taking the height feature points and vehicle trajectory points of each frame and the road points in the global sparse road network as three types of feature points, and extracting the same multi-dimensional feature attributes from each feature point in each type of feature point as each frame data, wherein the multi-dimensional feature attributes include: frame sequence number, ID mark, type mark, x coordinate, y coordinate, z coordinate, x-direction size, y-direction size, pixel value and heading angle.
5. The human-like trajectory planning method for off-road scenarios without high-precision maps according to claim 4 is characterized in that: The type identifier in the multidimensional feature attribute is used to distinguish between height feature points, vehicle trajectory points and waypoints; in two adjacent local elevation map frames, the height feature points whose pixel value difference and coordinate difference are less than or equal to the corresponding threshold are assigned the same ID identifier, otherwise, different ID identifiers are assigned; the heading angles of the height feature points are all 0.
6. The human-like trajectory planning method for off-road scenarios without high-precision maps according to claim 4 is characterized in that: The step of extracting multiple frames of data from the frame sequence data set according to the time window to construct the graph data includes: Multiple frames of data are taken out according to the number of frames corresponding to the time window, and the height feature points, vehicle trajectory points and waypoints in each frame of data are used as nodes in each frame of graph data; the spatial edge set is constructed according to the distance threshold between nodes in each frame, and then the spatial connection adjacency matrix is obtained; the inter-frame edge set is constructed according to the frame number threshold between the same nodes between frames, and then the time-space self-connection adjacency matrix is obtained.
7. The human-like trajectory planning method for off-road scenarios without high-precision maps according to claim 2 is characterized in that: The step of constructing a target thermal matrix according to the plurality of target points and the height feature points of the last frame includes: Selecting a number of target points with higher probabilities from the multiple target points; According to the radiation threshold, a grid-based method is used to obtain the radiation areas of the plurality of target points, and an initial thermal matrix is obtained by superimposing the thermal values of the same positions in the radiation areas; According to the coordinates of the height feature points of the last frame, that is, the height feature points in the same frame as the target point, the elements in the initial thermal matrix are corresponding, and the thermal values of the elements corresponding to the height feature points whose pixel values are greater than the maximum threshold are set to 0 to obtain the target thermal matrix.
8. The human-like trajectory planning method for off-road scenarios without high-precision maps according to claim 1 is characterized in that: The feature vector of the target thermal matrix is extracted by passing the target thermal matrix into a trained convolutional neural network for encoding and obtaining the feature vector.
9. The human-like trajectory planning method for off-road scenarios without high-precision maps according to claim 1 is characterized in that: The timing prediction network adopts a gated recurrent unit network.
10. The human-like trajectory planning method for off-road scenarios without high-precision maps according to claim 1, characterized in that: The target point prediction network is composed of a multi-layer perceptron, which uses the vehicle trajectory points in historical data as training samples and the driver's actual trajectory points as true values for training, and predicts the positions of multiple target points and the probability of each target point.