Complex scene path planning generation method and system based on multiple models

Abstract

The invention discloses a multi-model-based complex scene path planning generation method and system, and relates to the technical field of path planning, a static environment information set and a dynamic environment information set in a road are collected in advance, and multi-layer feature representation of the road is constructed based on the static environment information set and the dynamic environment information set; and constructing a path planning model based on the multi-layer feature representation of the road, and generating a planned path by using the optimization model based on the path planning model. And the path planning capability of the self-driving automobile in a complex scene can be obviously improved.

Description

Technical Field

[0001] The present invention relates to the field of path planning technology, and in particular to a method and system for generating complex scenario path planning based on multiple models. Background Art

[0002] Path planning in complex scenarios has always been a major challenge in the field of autonomous driving. Traditional path planning methods are mostly based on simple environmental models and are difficult to cope with complex actual road scenarios. With the continuous development of autonomous driving technology, the requirements for path planning algorithms are also increasing, and multiple complex factors such as static obstacles, dynamic obstacles, and traffic rules need to be considered.

[0003] In the existing technology, some methods only consider static obstacles and ignore the impact of dynamic obstacles, which may lead to potential collision risks in the planned path. Other methods, although considering dynamic obstacles, fail to make full use of global topological information and local topological features, and the planned path may be clumsy and lack time and energy optimization.

[0004] A Chinese patent application with publication number CN116698065A discloses a path planning method and medium for an autonomous driving vehicle that integrates motion constraints and safety constraints, the method comprising: 1) initializing path planning parameters; 2) establishing an evaluation function that introduces safety distance and motion constraints; 3) using the evaluation function to calculate the shortest path for the autonomous driving vehicle; 4) screening the path nodes in the shortest path and removing redundant path nodes; 5) using a third-order Bezier curve to smooth the path optimization nodes to generate a planned path for the autonomous driving vehicle; 6) controlling the autonomous driving vehicle to travel along the planned path, and determining whether the obstacle has changed each time the autonomous driving vehicle passes a node. If so, updating the actual moving distance km value and the current position sstart of the vehicle, setting the current position to slast, updating the cost value of the affected node, and then returning to step 4); however, this method mainly considers obstacles and fails to comprehensively consider the overall situation of the road.

[0005] To this end, the present invention proposes a complex scene path planning generation method and system based on multiple models. Summary of the invention

[0006] The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes a complex scene path planning generation method and system based on multiple models, which can significantly improve the path planning ability of an autonomous driving vehicle in complex scenes.

[0007] To achieve the above purpose, a complex scene path planning generation method and system based on multiple models are proposed. The complex scene path planning generation method based on multiple models includes the following steps: Step 1: Pre-collect static environment information set and dynamic environment information set on the road; Step 2: Based on the static environment information set and the dynamic environment information set, construct a multi-layer feature representation of the road; Step 3: Construct a path planning model based on the multi-layer feature representation of the road; Step 4: Based on the path planning model, use the optimization model to generate a planned path; The static environment information set in the road is collected in the following manner: Load cloud maps from the Internet and collect point cloud data of obstacles scanned by the LiDAR device; Then, the road connectivity matrix is ​​constructed through cloud maps and obstacle point cloud data; Cloud maps, obstacle point cloud data, and road connectivity matrix together constitute a static environment information set; The method of constructing the road connectivity matrix through the cloud map and obstacle point cloud data is as follows: In the cloud map, the centerline data of all lanes are extracted, and each centerline consists of a series of discrete points; Each discrete point is regarded as the basic node of the road where the center line is located; these basic nodes will be used to describe the discretization structure of the road network; Filter out intersection locations in the cloud map and add convergence nodes at the intersections to indicate the connection relationship between multiple roads; The two-dimensional coordinates of the basic nodes and the convergence nodes constitute a road node set; Mark any two adjacent nodes in the road node set as i and j respectively; The connectivity label values ​​of the i-th node and the j-th node in the road connectivity matrix are marked as Aij; If the i-th node and the j-th node are continuously located on the same lane, set Aij=1; if the i-th node and the j-th node are not continuously located on the same lane, set Aij=0; For each node i or node j in the road node set, calculate the distance to the nearest obstacle. If the distance is less than the distance threshold of the obstacle, the node is judged to be near the obstacle. Otherwise, it is judged that there is no obstacle near the node. For any node pair consisting of node i and node j with Aij=1, if any node between node i or node j is near an obstacle, set Aij=0; The dynamic environment information set is collected in the following manner: The laser radar and camera sensing devices are used to collect real-time point cloud data of dynamic obstacles on the road, and the detected dynamic obstacles are tracked in real time through the multi-target tracking algorithm to calculate their movement speed and direction; Obtain real-time traffic information of roads by introducing external real-time traffic data services; The real-time point cloud data of dynamic obstacles, movement speed and direction, and real-time traffic information together constitute a dynamic environment information set; The method of constructing a multi-layer feature representation of a road based on a static environment information set and a dynamic environment information set is: For any node i in the road node set: Calculate the degree centrality of each node using the node's degree centrality formula; Degree centrality is the number of nodes directly connected to each node. The degree centrality of the node is used as a supplement to the road connectivity matrix to form the global topological feature; Extract various geometric attributes of the road between every two directly connected nodes in the road node set from the cloud map to form local topological features; For each node in the road node collection: Obtain the position and motion trajectory of each dynamic obstacle from the lidar and camera perception devices; For each dynamic obstacle, use the trajectory fitting method to predict its future trajectory and generate a probability distribution; Then, based on the probability distribution, the Gaussian kernel function is applied to construct the dynamic risk field of each dynamic obstacle; The trajectory probability distribution and dynamic risk field of each dynamic obstacle corresponding to the node constitute the dynamic environment characteristics; Multi-layer feature representation includes global topological features, local topological features, and dynamic environment features; The method of using the trajectory fitting method to predict its future trajectory and generate probability distribution is as follows: The two-dimensional coordinates of the dynamic obstacle at the latest n moments in the past are extracted to form the historical running trajectory; n is the preset time parameter; The running speed and running direction of the dynamic obstacle are obtained through the two-dimensional coordinates of the dynamic obstacle at every two adjacent moments; Then, based on the running speed of the dynamic obstacle at different times, the running acceleration of the dynamic obstacle is obtained; Based on the constant acceleration model, the two-dimensional coordinates of the next moment are predicted according to the running speed, running direction and running acceleration of the dynamic obstacle to form the future trajectory; Gaussian distribution is introduced to the future trajectory to obtain the probability distribution of the future trajectory to adapt to the uncertainty of the dynamic obstacle position; The method of introducing Gaussian distribution is: The Gaussian distribution is expressed as: ; where x and y represent the horizontal and vertical coordinates of the two-dimensional coordinates, respectively, and the mean is the position of the dynamic obstacle predicted based on the constant acceleration model, the covariance matrix Represents the variance of the forecast. The diagonal items of the covariance matrix are calculated based on the volatility of the historical trajectory and the model error, thereby generating the corresponding probability distribution for each future moment. ; The method of constructing the dynamic risk field of each dynamic obstacle based on probability distribution and applying Gaussian kernel function is as follows: The risk field is marked as R(x,y), which represents the risk value of each point in the road scene space; Define a uniform set of grid points in scene space, with a resolution of the preset resolution ; Initialize the risk value of all grid points to 0, that is, set all R(x,y) to 0; For each future trajectory distribution ; Use Gaussian kernel function to calculate the risk value of grid points; The method of constructing the path planning model based on the multi-layer feature representation of the road is as follows: The node sequence included in the path to be planned is marked as P; the node sequence P includes all nodes passed by the planned path, and the number of any node is marked as k, and the number of the node after the node is marked as k+1; Based on global topological features, local topological features and dynamic environmental features, a planning objective function is constructed by weighted summation of distance, smoothness, risk and time cost, and corresponding road connectivity constraints, dynamic risk constraints and physical feasibility constraints are constructed to form a set of constraint conditions; Taking minimizing the planning objective function as the goal and the set of constraints as the constraint set, a convex optimization problem is used as the path planning model; The method of generating a planned path based on the path planning model using the optimization model is as follows: By using an optimization algorithm to solve the path planning model, the solved node sequence P is obtained as the planned path.

[0008] A complex scene path planning generation system based on multiple models is proposed, which includes an environment data collection module, a feature extraction module, a model construction module and a path planning module; wherein each module is electrically connected; An environmental data collection module collects static environmental information sets and dynamic environmental information sets on the road in advance, and sends the static environmental information sets and dynamic environmental information sets to a feature extraction module; A feature extraction module constructs a multi-layer feature representation of the road based on a static environment information set and a dynamic environment information set, and sends the multi-layer feature representation to a model construction module; A model building module builds a path planning model based on the multi-layer feature representation of the road, and sends the path planning model to the path planning module; The path planning module generates a planned path based on the path planning model using the optimization model.

[0009] Compared with the prior art, the present invention has the following beneficial effects: The present invention first collects static environment information sets and dynamic environment information sets, then constructs a multi-layer feature representation of the road based on this information, including global topological features, local topological features and dynamic environment features, and then constructs a path planning model based on the multi-layer feature representation. The objective function of the model comprehensively considers distance, smoothness, risk and time cost, and the constraints include road connectivity constraints, dynamic risk constraints and physical feasibility constraints. Finally, the model is solved using an optimization algorithm to generate the final planning path; by constructing a multi-layer feature representation, static and dynamic environment information are integrated, and global topological structure, local geometric characteristics and dynamic risks are considered at the same time, so that a safer, smoother and more efficient path planning solution can be generated, which can significantly improve the path planning ability of autonomous driving vehicles in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is a flow chart of a complex scene path planning generation method based on multiple models in Example 1 of the present invention;

[0011] Figure 2 This is a module connection relationship diagram of the complex scene path planning generation system based on multiple models in Example 2 of the present invention. DETAILED DESCRIPTION

[0012] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0013] Example 1 like Figure 1 As shown, the complex scene path planning generation method based on multiple models includes the following steps: Step 1: Pre-collect static environment information set and dynamic environment information set on the road; Step 2: Based on the static environment information set and the dynamic environment information set, construct a multi-layer feature representation of the road; Step 3: Construct a path planning model based on the multi-layer feature representation of the road; Step 4: Based on the path planning model, use the optimization model to generate a planned path; In the implementation of the present application, the static environment information set refers to the long-term fixed characteristics related to the road and environment in the path planning scenario; The dynamic environment information set refers to the information related to the dynamic changes in the path planning scene.

[0014] The static environment information set on the road is collected in the following manner: In the implementation of the present application, the main sources of the static environment information set are cloud maps loaded from the Internet and obstacle point cloud data scanned by a laser radar device; Then, the road connectivity matrix is ​​constructed through cloud maps and obstacle point cloud data; Cloud maps, obstacle point cloud data, and road connectivity matrix together constitute a static environment information set; Specifically, the cloud map includes, but is not limited to, road topology information including connectivity between roads, intersection locations, lane distribution, etc., road geometric characteristics including road curvature, slope, width, lane line location, etc., fixed obstacle information including stationary obstacle locations (such as buildings, street lights, guardrails, isolation belts, etc.), and traffic rules information including road speed limits, one-way streets, prohibited areas, signal light locations, etc.; The obstacle point cloud data generated by LiDAR generally includes the 2D or 3D model of the obstacle, the location of the obstacle, etc. It should be understood that it is a mature technology to connect to the Internet through a wireless network in current transportation vehicles (motor vehicles, aircraft, etc.) to download the latest cloud map, and it is also a conventional method in the art to carry some laser radars in the transportation vehicles to detect obstacles through radar signals and draw point cloud data of obstacles. Therefore, the collection of cloud maps and obstacle point cloud data is a mature existing technology, and the present invention will not be repeated here; By extracting the above static environment information set, long-term and reliable basic environment data can be provided for path planning, including road topology information, geometric characteristics, fixed obstacle information and traffic rules information. Specifically, the use of these features can significantly improve the stability and global optimization ability of path planning. Road topology information can describe the global connectivity of the road network, helping the path planning algorithm to quickly generate the global optimal path, while avoiding the planning of invalid areas such as dead ends. Road geometric characteristics (such as curvature, slope, width) provide physical constraints for path planning, ensuring that the generated path meets the dynamic limitations of the vehicle and avoiding the dangers of sharp turns and excessive slopes. In addition, fixed obstacle information can mark inaccessible areas, reduce the dependence on dynamic perception modules in path planning, and improve planning efficiency and accuracy. Traffic rule information (such as speed limits, one-way streets, etc.) can constrain the behavior of path planning to ensure that the path meets the requirements of traffic regulations. The extraction and comprehensive use of these features can improve the globality, stability and compliance of path planning, and lay a solid foundation for path generation in complex scenarios; In the implementation of the present application, the road connectivity matrix is ​​a mathematical representation that describes the connectivity relationship between adjacent nodes in a road; Therefore, the method of constructing the road connectivity matrix through the cloud map and obstacle point cloud data is as follows: In the cloud map, the centerline data of all lanes are extracted. Each centerline consists of a series of discrete points. For example, a centerline can be represented as {(x1, y1), (x2, y2)...}, where (x1, y1) and (x2, y2) represent two adjacent two-dimensional coordinate points on a centerline, respectively, and the distance between each two adjacent two-dimensional coordinate points is determined according to actual needs. Each discrete point is regarded as the basic node of the road where the center line is located; these basic nodes will be used to describe the discretization structure of the road network; Filter out intersection locations in the cloud map and add convergence nodes at the intersections to indicate the connection relationship between multiple roads; The two-dimensional coordinates of the basic nodes and the convergence nodes constitute a road node set; In an optional embodiment, for a multi-lane road, when the distance between adjacent lanes is small (determined by setting a lane distance threshold), a node connection may be added to indicate a switch between lanes. If the geometric center lines of two lanes are C1 and C2, a lateral connection is inserted between the nodes of the adjacent lanes. Mark any two adjacent nodes in the road node set as i and j respectively; The connectivity label values ​​of the i-th node and the j-th node in the road connectivity matrix are marked as Aij; If the i-th node and the j-th node are continuously located on the same lane, set Aij=1; if the i-th node and the j-th node are not continuously located on the same lane, set Aij=0; It should be understood that the process of setting the connectivity label value only describes the connectivity on the road centerline and has not yet considered the impact of obstacles or intersecting lanes.

[0015] To this end, a distance threshold is set for each obstacle according to the type of obstacle; it should be noted that the distance threshold means that when the distance between the vehicle and the obstacle is less than the distance threshold, it is considered that the vehicle is difficult to pass, and thus the node is judged to be impassable; For each node i or node j in the road node set, calculate the distance to the nearest obstacle. If the distance is less than the distance threshold of the obstacle, the node is judged to be near the obstacle. Otherwise, it is judged that there is no obstacle near the node. For any node pair consisting of node i and node j with Aij=1, if any node between node i or node j is near an obstacle, set Aij=0; It should be noted that the generation of the above road connectivity matrix only considers one embodiment of obstacles. In the actual road traffic process, the influence of traffic rules also needs to be considered; In a further feasible embodiment, the connectivity matrix may be further adjusted according to known traffic rules in the cloud map (such as one-way streets and no-traffic areas). For example, if the connectivity between node i and node j violates traffic rules (such as one-way traffic in opposite directions), Aij=0 is set; It can be understood that the road connectivity matrix can accurately reflect the connectivity between nodes, and it is dynamically adjusted in combination with obstacle information. When there is an obstacle in the middle of the road, the path planning algorithm will automatically bypass this area to ensure the safety of the planned path; Furthermore, the dynamic environment information set is collected in the following manner: Real-time point cloud data of dynamic obstacles on the road are collected in real time through laser radar and camera perception equipment, and the detected dynamic obstacles are tracked in real time through a multi-target tracking algorithm to calculate their movement speed and direction; specifically, the multi-target tracking algorithm is SORT or DeepSORT; Obtaining real-time traffic information of roads by introducing external real-time traffic data services; specifically, the external real-time traffic data services include traffic management platforms, navigation platform APIs, and the real-time traffic information includes but is not limited to real-time congestion conditions, construction areas, traffic accidents, etc.; The real-time point cloud data of dynamic obstacles, movement speed and direction, and real-time traffic information together constitute a dynamic environment information set; Considering that during road driving, road feasibility is also affected by weather conditions, therefore, in a preferred embodiment, the dynamic environment information set may also include real-time weather data; By extracting dynamic environmental information sets in real time, the path planning system can be provided with core information of dynamic scenes, including dynamic obstacle information, trajectory prediction information, real-time road conditions and environmental impact information. The use of this information can significantly improve the real-time and safety of path planning in dynamic environments. Dynamic obstacle information (such as the location information, speed and direction of pedestrians and vehicles) can support the obstacle avoidance algorithm to adjust the path in real time to avoid collisions with obstacles. Trajectory prediction information helps the path planning system perceive potential risks in advance and achieve safer path design by providing the future location distribution of dynamic obstacles. Real-time road conditions (such as congestion and accident areas) can provide a reference for the current optimal passage area for path planning, thereby improving overall planning efficiency and travel time. Environmental impact information (such as weather and slippery roads) further enhances the safety assessment capability of path planning and supports the system to choose a safer path under severe weather conditions. The extraction and real-time use of these dynamic environmental features enable the path planning system to have excellent flexibility and adaptability in ever-changing scenarios, providing important support for dynamic path optimization in complex scenarios.

[0016] In the implementation of the present application, the multi-layer feature representation is constructed through the following three levels: Based on the static environment information set, the global topological features describing the overall connectivity and structural characteristics of the road; Based on the static environment information set, describe the geometric properties of a single road section, such as local topological features such as curvature, slope, width, etc. Combined with the dynamic environment information set, describe the distribution of dynamic obstacles and the dynamic environment characteristics of the risk area; The features at these three levels can provide information support from the global to the local, and from the dynamic to the static. By fusing these features, a multi-level feature representation is formed.

[0017] Specifically, the method of constructing a multi-layer feature representation of a road based on a static environment information set and a dynamic environment information set is: For any node i in the road node set: Calculate the degree centrality of each node using the node's degree centrality formula; It should be noted that the degree centrality is the number of nodes directly connected to each node (i.e., Aij=1). The degree centrality of the node is used as a supplement to the road connectivity matrix to form a global topological feature, which is used to evaluate the global influence of road nodes. Extract various geometric attributes of the road between every two directly connected nodes in the road node set from the cloud map to form local topological features; Specifically, the geometric attributes include but are not limited to curvature, slope and road width. Figure 1 Generally, they are mapped in advance, and these geometric attributes can be directly extracted from the map; For each node in the road node collection: Obtain the position and motion trajectory of each dynamic obstacle from the lidar and camera perception devices; For each dynamic obstacle, use the trajectory fitting method to predict its future trajectory and generate a probability distribution; Then, based on the probability distribution, the Gaussian kernel function is applied to construct the dynamic risk field of each dynamic obstacle; The trajectory probability distribution and dynamic risk field of each dynamic obstacle corresponding to the node constitute the dynamic environment characteristics; Multi-layer feature representation includes global topological features, local topological features, and dynamic environment features; Specifically, the method of using the trajectory fitting method to predict its future trajectory and generate probability distribution is as follows: The two-dimensional coordinates of the dynamic obstacle at the latest n moments in the past are extracted to form the historical running trajectory; n is the preset time parameter; The running speed and running direction of the dynamic obstacle are obtained through the two-dimensional coordinates of the dynamic obstacle at every two adjacent moments; Then, based on the running speed of the dynamic obstacle at different times, the running acceleration of the dynamic obstacle is obtained; Based on the constant acceleration model, the two-dimensional coordinates of the next moment are predicted according to the running speed, running direction and running acceleration of the dynamic obstacle to form the future trajectory; Gaussian distribution is introduced to the future trajectory to obtain the probability distribution of the future trajectory to adapt to the uncertainty of the dynamic obstacle position; Specifically, the method of introducing Gaussian distribution is: The Gaussian distribution is expressed as: ; where x and y represent the horizontal and vertical coordinates of the two-dimensional coordinates, respectively, and the mean is the position of the dynamic obstacle predicted based on the constant acceleration model, the covariance matrix Represents the variance of the forecast, which is used to quantify the uncertainty of the forecast. The diagonal items of the covariance matrix can be calculated based on the volatility of the historical trajectory and the model error, thereby generating the corresponding probability distribution for each future moment ; In the implementation of the present application, a spatial risk field is generated based on the trajectory distribution of obstacles to represent the dynamic risk in path planning. The higher the value of the risk field, the greater the risk in the area; Furthermore, the method of constructing the dynamic risk field of each dynamic obstacle based on the probability distribution and applying the Gaussian kernel function is as follows: The risk field is marked as R(x,y), which represents the risk value of each point in the road scene space; Define a uniform set of grid points in scene space, with a resolution of the preset resolution ; Initialize the risk value of all grid points to 0, that is, set all R(x,y) to 0; For each future trajectory distribution ; Use the Gaussian kernel function to calculate the risk value of the grid points: ; Wherein, w is the preset risk weight of the obstacle, which can be set in advance for each obstacle according to the type of obstacle, and the risk weight indicates the degree of risk that may be caused by touching the obstacle of this type; ux and uy are the center point positions of the obstacles, and the obstacles include static obstacles and movable obstacles. For movable obstacles, the center point position of the obstacle is the position predicted according to its running trajectory; thus, the risk value of the grid point quantifies the risk contribution of the dynamic obstacle to the grid point, and describes the possibility and influence range of the predicted position of the obstacle in combination with the Gaussian distribution; Furthermore, the method of constructing the path planning model based on the multi-layer feature representation of the road is as follows: The node sequence included in the path to be planned is marked as P; the node sequence P includes all nodes passed by the planned path, and the number of any node is marked as k, and the number of the node after the node is marked as k+1; It can be understood that by selecting different node sequences, the path planning algorithm can explore different path solutions. The selection of each node directly affects the total distance, smoothness, risk and time cost of the path, the order and connection of the nodes, and the order of the selected node sequence determines the shape and direction of the path; therefore, it is necessary to optimize the distance, smoothness, risk, etc. of the selected path; Based on global topological features, local topological features and dynamic environmental features, a planning objective function is constructed by weighted summation of distance, smoothness, risk and time cost, and corresponding road connectivity constraints, dynamic risk constraints and physical feasibility constraints are constructed to form a set of constraint conditions; Taking minimizing the planning objective function as the goal and the set of constraints as the constraint set, a convex optimization problem is used as the path planning model; In an embodiment of the present application, the planning objective function may be constructed in the following manner: Label the planning objective function as FP(P); Then the expression of the planning objective function FP(P) is: ; Among them, a1, a2 and a3 are all preset proportional coefficients; FD(P) is the distance cost. ;in, is the distance between the kth node and the k+1th node; the distance cost represents the total distance required to plan the path; is the smoothness cost, the smoothness cost ;in, and are the curvature of the k+1th node and the curvature of the kth node respectively; is the slope of the kth node; Zk is the degree centrality of the kth node, b1, b2 and b3 are preset proportional coefficients; this cost means that the point with the smallest curvature change, the smallest slope and as close to the center as possible is selected for movement; is the risk cost, the risk cost ;in, is the risk value R(xk,yk) of the two-dimensional coordinates corresponding to the k-th node on the road, xk, yk are the horizontal and vertical coordinates of the k-th node respectively; In a further preferred embodiment, the corresponding road connectivity constraints, dynamic risk constraints and physical feasibility constraints are constructed to form a set of constraint conditions in the following manner: The road connectivity constraints are: ; is the connectivity label value between the kth node and the k+1th node; The dynamic risk constraint is ; Wherein, Rmax is the preset risk threshold; The physical feasibility constraints are: ; Qmax is the maximum curvature of the preset road that can be passed; Furthermore, the method of generating a planned path using an optimization model based on the path planning model is as follows: The path planning model is solved by using an optimization algorithm to obtain a solved node sequence P as a planned path; Specifically, the optimization algorithm includes but is not limited to ant colony algorithm, genetic algorithm, etc.

[0018] Example 2 like Figure 2 As shown, a complex scene path planning generation system based on multiple models includes an environment data collection module, a feature extraction module, a model building module and a path planning module; wherein each module is electrically connected; An environmental data collection module collects static environmental information sets and dynamic environmental information sets on the road in advance, and sends the static environmental information sets and dynamic environmental information sets to a feature extraction module; A feature extraction module constructs a multi-layer feature representation of the road based on a static environment information set and a dynamic environment information set, and sends the multi-layer feature representation to a model construction module; A model building module builds a path planning model based on the multi-layer feature representation of the road, and sends the path planning model to the path planning module; The path planning module generates a planned path based on the path planning model using the optimization model.

[0019] The specific implementation modes as described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation mode of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0020] The above preset parameters or preset thresholds are all set by technicians in this field according to actual conditions or obtained through large-scale data simulation.

[0021] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A complex scene path planning generation method based on multiple models, characterized in that: The following steps are involved: Step 1: Pre-collect static environment information set and dynamic environment information set on the road; Step 2: Based on the static environment information set and the dynamic environment information set, construct a multi-layer feature representation of the road; Step 3: Construct a path planning model based on the multi-layer feature representation of the road; Step 4: Based on the path planning model, use the optimization model to generate a planned path.

2. The method for generating complex scene path planning based on multiple models according to claim 1, characterized in that: The static environment information set in the road is collected in the following manner: Load cloud maps from the Internet and collect point cloud data of obstacles scanned by the LiDAR device; Then, the road connectivity matrix is ​​constructed through cloud maps and obstacle point cloud data; Cloud maps, obstacle point cloud data, and road connectivity matrix together constitute a static environment information set.

3. The method for generating complex scene path planning based on multiple models according to claim 2, characterized in that: The method of constructing the road connectivity matrix through the cloud map and obstacle point cloud data is as follows: In the cloud map, the centerline data of all lanes are extracted, and each centerline consists of a series of discrete points; Each discrete point is regarded as the basic node of the road where the center line is located; these basic nodes will be used to describe the discretization structure of the road network; Filter out intersection locations in the cloud map and add convergence nodes at the intersections to indicate the connection relationship between multiple roads; The two-dimensional coordinates of the basic nodes and the convergence nodes constitute a road node set; Mark any two adjacent nodes in the road node set as i and j respectively; The connectivity label values ​​of the i-th node and the j-th node in the road connectivity matrix are marked as Aij; If the i-th node and the j-th node are continuously located on the same lane, set Aij=1; if the i-th node and the j-th node are not continuously located on the same lane, set Aij=0; For each node i or node j in the road node set, calculate the distance to the nearest obstacle. If the distance is less than the distance threshold of the obstacle, the node is judged to be near the obstacle. Otherwise, it is judged that there is no obstacle near the node. For any node pair consisting of node i and node j with Aij=1, if any node between node i or node j is located near an obstacle, set Aij=0.

4. The method for generating complex scene path planning based on multiple models according to claim 3, characterized in that: The dynamic environment information set is collected in the following manner: The laser radar and camera sensing devices are used to collect real-time point cloud data of dynamic obstacles on the road, and the detected dynamic obstacles are tracked in real time through the multi-target tracking algorithm to calculate their movement speed and direction; Obtain real-time traffic information of roads by introducing external real-time traffic data services; The real-time point cloud data of dynamic obstacles, movement speed and direction, and real-time traffic information together constitute a dynamic environment information set.

5. The method for generating complex scene path planning based on multiple models according to claim 4, characterized in that: The method of constructing a multi-layer feature representation of a road based on a static environment information set and a dynamic environment information set is: For any node i in the road node set: Calculate the degree centrality of each node using the node's degree centrality formula; Degree centrality is the number of nodes directly connected to each node. The degree centrality of the node is used as a supplement to the road connectivity matrix to form the global topological feature; Extract various geometric attributes of the road between every two directly connected nodes in the road node set from the cloud map to form local topological features; For each node in the road node collection: Obtain the position and motion trajectory of each dynamic obstacle from the lidar and camera perception devices; For each dynamic obstacle, use the trajectory fitting method to predict its future trajectory and generate a probability distribution; Then, based on the probability distribution, the Gaussian kernel function is applied to construct the dynamic risk field of each dynamic obstacle; The trajectory probability distribution and dynamic risk field of each dynamic obstacle corresponding to the node constitute the dynamic environment characteristics; Multi-layer feature representation includes global topological features, local topological features and dynamic environment features.

6. The method for generating complex scene path planning based on multiple models according to claim 5, characterized in that: The method of using the trajectory fitting method to predict its future trajectory and generate probability distribution is as follows: The two-dimensional coordinates of the dynamic obstacle at the latest n moments in the past are extracted to form the historical running trajectory; n is the preset time parameter; The running speed and running direction of the dynamic obstacle are obtained through the two-dimensional coordinates of the dynamic obstacle at every two adjacent moments; Then, based on the running speed of the dynamic obstacle at different times, the running acceleration of the dynamic obstacle is obtained; Based on the constant acceleration model, the two-dimensional coordinates of the next moment are predicted according to the running speed, running direction and running acceleration of the dynamic obstacle to form the future trajectory; Gaussian distribution is introduced to the future trajectory to obtain the probability distribution of the future trajectory to adapt to the uncertainty of the dynamic obstacle position.

7. The method for generating complex scene path planning based on multiple models according to claim 6, characterized in that: The method of introducing Gaussian distribution is: The Gaussian distribution is expressed as: ; where x and y represent the horizontal and vertical coordinates of the two-dimensional coordinates, respectively, and the mean is the position of the dynamic obstacle predicted based on the constant acceleration model, the covariance matrix Represents the variance of the forecast. The diagonal items of the covariance matrix are calculated based on the volatility of the historical trajectory and the model error, thereby generating the corresponding probability distribution for each future moment. .

8. The method for generating complex scene path planning based on multiple models according to claim 7, characterized in that: The method of constructing the dynamic risk field of each dynamic obstacle based on probability distribution and applying Gaussian kernel function is as follows: The risk field is marked as R(x,y), which represents the risk value of each point in the road scene space; Define a uniform set of grid points in scene space, with a resolution of the preset resolution , ; Initialize the risk value of all grid points to 0, that is, set all R(x,y) to 0; For each future trajectory distribution ; Use Gaussian kernel function to calculate the risk value of grid points.

9. The method for generating complex scene path planning based on multiple models according to claim 8, characterized in that: The method of constructing the path planning model based on the multi-layer feature representation of the road is as follows: The node sequence included in the path to be planned is marked as P; the node sequence P includes all nodes passed by the planned path, and the number of any node is marked as k, and the number of the node after the node is marked as k+1; Based on global topological features, local topological features and dynamic environmental features, a planning objective function is constructed by weighted summation of distance, smoothness, risk and time cost, and corresponding road connectivity constraints, dynamic risk constraints and physical feasibility constraints are constructed to form a set of constraint conditions; A convex optimization problem consisting of minimizing the planning objective function is used as the path planning model.

10. The method for generating complex scene path planning based on multiple models according to claim 9, characterized in that: The method of generating a planned path based on the path planning model using the optimization model is as follows: By using an optimization algorithm to solve the path planning model, the solved node sequence P is obtained as the planned path.

11. A complex scene path planning generation system based on multiple models, which is used to implement the complex scene path planning generation method based on multiple models as described in any one of claims 1 to 10, characterized in that: It includes an environmental data collection module, a feature extraction module, a model building module and a path planning module; wherein each module is electrically connected; An environmental data collection module collects static environmental information sets and dynamic environmental information sets on the road in advance, and sends the static environmental information sets and dynamic environmental information sets to a feature extraction module; A feature extraction module constructs a multi-layer feature representation of the road based on a static environment information set and a dynamic environment information set, and sends the multi-layer feature representation to a model construction module; A model building module builds a path planning model based on the multi-layer feature representation of the road, and sends the path planning model to the path planning module; The path planning module generates a planned path based on the path planning model using the optimization model.

Images