Intelligent Driving Multi-Step Long-Term Behavior Decision-Making Method and Device Combining Spatiotemporal Information

Abstract

The present invention discloses an intelligent driving multi-step long-term behavior decision-making method and device combining spatio-temporal information. First, according to road structure information and traffic participant information, a driving state set is generated. Taking each driving state as a node, based on the driving states of adjacent lanes, directed edges between nodes are generated to construct a driving state transition graph. Then, based on the driving state transition graph and the spatio-temporal information of traffic participants, through a first-in-first-out queue and a list of explored driving states, a spatio-temporal trajectory tree of the ego vehicle is constructed. Finally, candidate spatio-temporal trajectories are extracted from the spatio-temporal trajectory tree, and all candidate spatio-temporal trajectories are scored. Ultimately, the spatio-temporal trajectory with the highest score is selected as the decision result. The present invention enables the intelligent driving platform to perform multi-step and long-term spatio-temporal joint decision-making in a dynamic traffic environment, guiding the intelligent driving platform to obtain more reasonable driving behaviors.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent driving behavior decision-making, and in particular to an intelligent driving multi-step long-term behavior decision-making method and device that combines spatio-temporal information. Background Art

[0002] In recent years, intelligent driving and its related research have made great progress and shown great potential in improving traffic efficiency and driving safety. As an important module in the intelligent driving system, the performance of the behavior decision-making module directly determines the intelligence level of intelligent driving vehicles and is also one of the most core indicators for evaluating intelligent driving capabilities. The intelligent driving behavior decision-making module gives a decision result based on the upper-layer perception and prediction results and the driving task, and guides the human driver or the lower-layer trajectory planning and control to generate driving behaviors.

[0003] Traditional driving behavior decision-making methods are based on rules to select each decision primitive (for example: lane keeping, following, lane changing, overtaking, accelerating, decelerating). For example, the Chinese patent application for invention "Lane Decision Method Based on Multi-Objective Decision Matrix for Autonomous Driving Vehicles" with the publication number CN108583578A discloses a method for a three-lane model, which constructs a decision matrix based on the current position and speed information of traffic participants and lane speed limit information to achieve real-time autonomous lane decision-making for autonomous driving vehicles; the Chinese patent application for invention "Lateral Decision System and Lateral Decision Determination Method for Autonomous Driving Vehicles" with the publication number CN110667578A discloses a method for evaluating the target lane and lane abnormal conditions required for an autonomous driving vehicle to make a lateral decision through road feature information and a pre-selected target line and environmental object target, and making a lateral decision that conforms to the road features accordingly.

[0004] Existing methods are based on road structure information and the current position and speed information of traffic participants, and select the optimal decision primitive for the next step according to the designed rules or evaluation functions, which is a one-step short-term behavior decision-making. The one-step short-term behavior decision-making evaluates the current selection of a certain decision primitive based on short-term information, which may lead to short-sighted decision results such as local minimum or instability (frequent switching of decision results in adjacent decision cycles). In addition, the decision results output by existing methods are biased towards semantic-level information (such as left lane change, lane keeping, right lane change, acceleration, deceleration, etc.) and contain less quantitative information (such as the starting and ending points of lane change, speed expectation value, etc.). On the premise of meeting real-time decision-making, decision results containing more quantitative information can provide more explicit and effective guidance for human drivers or lower-layer trajectory planning and control. Summary of the Invention

[0005] To address the deficiencies of the prior art, achieve a decision result that can span multiple decision-making elements, consider more long-term environmental information in the spatio-temporal dimension, and obtain a multi-step long-term decision result containing a richer sequence of trajectory points, thereby helping human drivers or intelligent driving platforms obtain safer, more stable, and more efficient driving behaviors, the present invention adopts the following technical solutions:

[0006] An intelligent driving multi-step long-term behavior decision-making method that combines spatio-temporal information. This method constructs a driving state transition graph based on information such as the position distribution of traffic participants and the road structure. That is, the intelligent vehicle may reach between two adjacent traffic participants (including two virtual traffic participants, the extreme front and the extreme rear, established for each adjacent lane) on a specified lane according to this driving state transition graph. Then, forward simulation is performed based on the spatio-temporal information of traffic participants (e.g., predicted trajectories) to construct the spatio-temporal trajectory tree of the intelligent vehicle, thereby finding the complete sequence of trajectory points that constitutes the decision result. Finally, all spatio-temporal trajectory trees are comprehensively scored according to the designed metrics, and the spatio-temporal trajectory with the highest comprehensive score is used as the current decision result. Specifically, it includes the following steps:

[0007] An intelligent driving multi-step long-term behavior decision-making method that combines spatio-temporal information, including the following steps:

[0008] Step S1: Generate a set of driving states based on road structure information and traffic participant information. Use each driving state as a node, and generate directed edges between nodes based on the driving states of adjacent lanes to construct a driving state transition graph;

[0009] Step S2: Construct a spatio-temporal trajectory tree based on the driving state transition graph and the spatio-temporal information of traffic participants. Take the current state of the intelligent vehicle as the root node of the spatio-temporal trajectory tree and add it to the end of the first-in-first-out queue. Add the driving states associated with the root node to the list of explored driving states. If the time step of the head node in the first-in-first-out queue is less than the decision-making step length of the list of explored driving states, then take out the head node. Add the direct child nodes of the head node to the end of the first-in-first-out queue. Except for leaf nodes, other nodes have exactly 1 direct child node, and the driving state of the direct child node is the same as that of the parent node. If there is a directed edge in the driving state transition graph from the driving state of the head node to the driving state of its candidate collateral child node, and the driving state of the candidate collateral child node does not belong to the list of explored driving states, then add the candidate collateral child node to the end of the first-in-first-out queue and add the driving state of the candidate collateral child node to the list of explored driving states. Except for leaf nodes, other nodes have at most 2 collateral child nodes, and the driving states of the collateral child nodes are different from that of the parent node and do not belong to the same lane. If the time step of the head node in the first-in-first-out queue is greater than or equal to the decision-making step length of the list of explored driving states, terminate the iteration and output the spatio-temporal trajectory tree;

[0010] Step S3: Extract spatio-temporal trajectories from the spatio-temporal trajectory tree, score all candidate spatio-temporal trajectories, and finally select the spatio-temporal trajectory with the highest score as the decision result.

[0011] Further, the step S1 includes the following steps:

[0012] Step S1.1: Determine the reference lane M according to the lane where the intelligent vehicle V is located; if the driving direction of the intelligent vehicle V is opposite to the driving direction of the lane where it is located, then use the first lane in the same direction as the intelligent vehicle V adjacent to the lane where the intelligent vehicle V is located as the reference lane M; if the driving direction of the intelligent vehicle V is the same as the driving direction of the lane where it is located, then use the lane where the intelligent vehicle V is located as the reference lane M;

[0013] Step S1.2: Construct a lane set S according to the reference lane M and the lanes adjacent to the reference lane M lane ;

[0014] Step S1.3: Sort the other traffic participants except the intelligent vehicle V on the lane X ∈ S lane in the reverse direction of the driving direction of the reference lane M to obtain a list of other traffic participants, and merge the driving state sets of each lane in S lane into a total driving state set S state ;

[0015] Step S1.4: Use all the driving states in the total driving state set S state as the nodes of the driving state transition graph STG. For any two driving states x i and x j , if and only if the lanes to which x i and x j belong are adjacent, then add a directed edge from x i to x j and a directed edge from x j to x i in the driving state transition graph STG.

[0016] Further, in the step S1.2, and M ∈ S lane , and L ∈ S lane is mutually exclusive with C ∈ S lane . L represents the same-direction lane on the left side of M, R represents the same-direction lane on the right side of M, and C represents the reverse-direction lane on the left side of M;

[0017] In the step S1.3, the list of other traffic participants S X traffic = {t X 0, t X 1, …, tX n ,t X n+1}, where t X i , i = 1, …, n are other traffic participants on lane X, t X 0 is a virtual traffic participant set at INF behind the driving direction of the reference lane M on lane X, t X n+1 is a virtual traffic participant set at INF in front of the driving direction of the reference lane M on lane X, INF is a set first threshold (a large value), then the driving state set on lane X is S X state = {x0 = f(V, A(t X 0) ∩ B(t X 1)), …, x n = f(V, A(t X n ) ∩ B(t X n+1 ))}, where, A(t X ) represents the front area of other traffic participant t X on lane X, B(t X ) represents the rear area of other traffic participant t X on lane X, A(t X i ) ∩ B(t X i+1 ) represents the area between other traffic participants t X i and t X i+1 on lane X, f represents a function that maps the area A(t X i ) ∩ B(t X i+1 ) where the intelligent vehicle V travels on lane X to the driving state, X ∈ S lane ;

[0018] The driving state sets and driving state representations corresponding to lanes with different markings are as follows:

[0019] If the lane marking is M, then S M state = {m0, m1, …, m n(M)}, where n(M) is the number of other traffic participants m on M;

[0020] If the lane marking is L, then S L state = {l0, l1, …, l n(L)}, where n(L) is the number of other traffic participants l on L;

[0021] If the lane marking is R, then S R state ={r0,r1,…,r n(R)}}, where n(R) is the number of other traffic participants r on R;

[0022] If the lane marking is C, then S C state ={c0,c1,…,c n(C)}}, where n(C) is the number of other traffic participants c on C.

[0023] Furthermore, the step S2 includes the following steps:

[0024] Step S2.1: Initialize an empty first-in-first-out queue An empty list of explored driving states And the decision step size D;

[0025] Step S2.2: According to the current state of the intelligent vehicle V, construct the root node root of the spatio-temporal trajectory tree T. Among them, each node node of the spatio-temporal trajectory tree is represented by a six-tuple <d,X,s,v,x,F>, where d, X, s, v, and x respectively represent the time step, lane, position, speed, driving state, and parent node of the intelligent vehicle V at this node, etc.; assume that the driving state associated with the root node root is x i ∈S state , add the root node root to the first-in-first-out queue Q FIFO At the end of the queue, add the driving state x i To the list of explored driving states ClosedList;

[0026] Step S2.3: Iteratively construct the spatio-temporal trajectory tree according to the following steps:

[0027] Step S2.3.1: The head node in the first-in-first-out queue Q FIFO Is the node i =<d i ,X,s i ,v i ,x i ,F i >>. If the time step d i Where the node is located <D, then take out the head node of the queue for the subsequent steps, otherwise terminate the iteration and output the spatio-temporal trajectory tree T; i <D, then take out the head node of the queue for the subsequent steps, otherwise terminate the iteration and output the spatio-temporal trajectory tree T;

[0028] Step S2.3.2: Use the longitudinal one-step speed planning model P 纵向, obtain node i 's direct child node node iic , add node iic to the first-in-first-out queue Q FIFO at the end of the queue; except for leaf nodes, other nodes have exactly 1 direct child node, and the driving state of the direct child node is the same as that of the parent node;

[0029] Step S2.3.3: Use the lateral one-step lane change decision-making model P 横向 , obtain all possible candidate collateral child nodes node i of node isc : Assume that the driving state of the candidate collateral child node is x nisc , if there is a directed edge from x ni to x nisc in the driving state transition graph STG and when, then add node isc to the first-in-first-out queue Q FIFO at the end of the queue, and add x nisc to the explored driving state list ClosedList; except for leaf nodes, other nodes have at most 2 collateral child nodes, and the driving states of the collateral child nodes are different from those of the parent node and do not belong to the same lane;

[0030] Step S2.3.4: Jump to Step S2.3.1.

[0031] Furthermore, the longitudinal one-step speed planning model P 纵向 in Step S2.3.2 is specifically: Assume that the node node i of the spatio-temporal trajectory tree is used as the starting state of the longitudinal one-step speed planning model P 纵向 , and the predicted state of the traffic participant at the d i th step is used as the environmental state information of P 纵向 . P 纵向 plans and outputs the speed and position considering the dynamics and kinematics constraints of the intelligent vehicle V body based on the starting state and environmental state information; P 纵向 is based on rule construction or is a learnable network model.

[0032] Furthermore, the longitudinal one-step speed planning model P 纵向 is: v d = max{v min , min{v max , dist(node i , t X aheadi(di) ) / HT}} and s d = s i + vd × dt, where v d is the expected speed value of the direct child node of node i , v min and v max are, respectively, the minimum speed and the maximum speed that can be achieved by the direct child nodes of node i after comprehensively considering the speed of node i , the maximum acceleration / deceleration of V, and the lane speed limit, etc. dist(node i , t X aheadi(di) ) is the distance between the position of node i and the first traffic participant in front of it in its lane. HT is the shortest headway time (generally, for safety, HT = 2 s), s i is the position of node i , s d is the position of node i , and s

[0033] Furthermore, the lateral one-step lane-changing decision model P 横向 in step S2.3.2 is specifically as follows: Assume that the node node i of the spatio-temporal trajectory tree is used as the starting state of the lateral one-step lane-changing decision model P 横向 . The predicted state of the d i -th step of the traffic participant is used as the environmental state information of P 横向 . The adjacent lane X ∈ S i of the lane where node lane is located is used as the target lane. P 横向 outputs decision information such as whether to change lanes in the next step according to the starting state, the environmental state, and the target lane; P 横向 is constructed based on rules or is a learnable network model.

[0034] Furthermore, the lateral one-step lane-changing decision model P 横向 is: If dist(node i , t P behind(di) ) > d1 and dist(node i , t P ahead(di) ) > d2, that is, if the distances between the position of node n i and the first traffic participants behind and in front of the adjacent lane P are greater than d1 and d2 respectively, then v d = max{v min , min{v max , dist(node i,t P aheadi(di) ) / HT}} and s d = s i + v d × dt, where d1 and d2 are the set safety intervals in the workshop, and v min and v max are respectively the minimum speed and the maximum speed that can be achieved by the direct child nodes of node i after comprehensively considering the speed of node i , the maximum acceleration / deceleration of V, and the speed limit of the lane, etc. dist(node i ,t P aheadi(di) ) is the distance between the position of node i and the first traffic participant in front of the adjacent lane P. HT is the shortest headway time, s i is the position of node i , s d is the expected position of the direct child node of node i , and dt is the time of one-step planning.

[0035] Furthermore, step S3 includes the following steps:

[0036] Step S3.1: Follow the spatio-temporal trajectory tree T and collect all nodes from the root node to the leaf nodes to form a trajectory point sequence Trace i , i = 1, 2,..., h, where h is the number of leaf nodes in T and also the number of trajectory point sequences finally collected;

[0037] Step S3.2: Score all trajectory point sequences according to the designed evaluation indicators:

[0038] The evaluation indicators include but are not limited to guidance C guide , timeliness C effic , safety C safe and consistency C consist , and the interpretations of each evaluation indicator are as follows:

[0039] C guide : The guidance evaluation measures the proximity between the lane where V finally arrives through the spatio-temporal trajectory and the target lane given by the global route information, and is evaluated by the center line distance between the two lanes;

[0040] C effic : The timeliness evaluation measures that V can complete the driving task faster through the spatio-temporal trajectory, and can be evaluated according to the speed information of the trajectory point sequence associated with the spatio-temporal trajectory;

[0041] C safe: Safety assessment V The probability of avoiding collisions with other traffic participants when driving along the spatio-temporal trajectory can be evaluated based on the spatio-temporal distances between the sequence of trajectory points associated with the spatio-temporal trajectory and the surrounding traffic participants;

[0042] C consist : Consistency assessment V The number of lane changes when driving according to the decision motion can be evaluated based on the event count of two consecutive nodes not being in the same lane in the sequence of trajectory points associated with the spatio-temporal trajectory;

[0043] Step S3.3: Calculate the total score C through weighted summation total = w effic × C effic + w safe × C safe + w consist × C consist , where w effic , w safe , w consist are the weights of the corresponding indicators. Select the one with the highest total score C from the valid spatio-temporal trajectories with the highest guiding C guide as the decision result. total

[0044] An intelligent driving multi-step long-term behavior decision-making device combining spatio-temporal information, including a memory and one or more processors. Executable code is stored in the memory. When the one or more processors execute the executable code, it is used to implement the intelligent driving multi-step long-term behavior decision-making method combining spatio-temporal information.

[0045] The advantages and beneficial effects of the present invention are as follows:

[0046] Compared with the prior art, the present invention breaks through the limitation of one-step short-term behavior decision-making, uses information such as the position distribution of traffic participants and road structure to obtain decision results that can span multiple decision primitives, and the spatio-temporal trajectories obtained based on considering the spatio-temporal prediction information of traffic participants contain richer quantitative information, enabling intelligent vehicles to quantitatively evaluate the decision results of multi-step long-term (evaluation indicators include but are not limited to guidance, timeliness, safety, and consistency, etc.). In addition, the present invention uses an embedded low-level one-step speed planning model and one-step lane change decision model to achieve real-time behavior decision-making in the spatio-temporal dimension of multi-step long-term. The more visionary decision results obtained by the present invention can be directly used as an assisted driving technology to provide decision-making suggestions for human drivers, or can be used as a decision-making module of an unmanned driving system to guide the underlying motion planning and control, further improving traffic efficiency and driving safety. Brief Description of the Drawings

[0047] Figure 1 is the method flow chart in the embodiment of the present invention.

[0048] Figure 2a It is the driving state transition diagram constructed in Example 1 in the embodiments of the present invention.

[0049] Figure 2b It is the driving state transition diagram constructed in Example 2 in the embodiments of the present invention.

[0050] Figure 3a It is a schematic diagram of the spatio-temporal trajectory tree constructed in Example 2 in the embodiments of the present invention;

[0051] Figure 3b It is a schematic diagram of extracting spatio-temporal trajectories from the spatio-temporal trajectory tree in Example 2 in the embodiments of the present invention.

[0052] Figure 4 It is a schematic diagram of the structure of the device in the embodiments of the present invention. Detailed implementation manners

[0053] The following will describe the detailed implementation manners of the present invention with reference to the accompanying drawings. It should be understood that the detailed implementation manners described herein are only for explaining and illustrating the present invention, and are not used to limit the present invention.

[0054] As Figure 1 shown, an intelligent driving multi-step long-term behavior decision-making method combining spatio-temporal information. First, construct a driving state transition diagram according to road structure information and traffic participant information; then construct a spatio-temporal trajectory tree based on the driving state transition diagram and the spatio-temporal information of traffic participants; then, extract spatio-temporal trajectories from the spatio-temporal trajectory tree, and score all candidate spatio-temporal trajectories, and finally select the spatio-temporal trajectory with the highest score as the decision result; specifically, it includes the following steps:

[0055] Step S1: According to road structure information and traffic participant information, generate a driving state set, use each driving state as a node, and generate directed edges between nodes based on the driving states of adjacent lanes to construct a driving state transition diagram, which specifically includes the following steps:

[0056] Step S1.1: Determine the reference lane (marked as M) according to the lane where the intelligent vehicle (marked as V) is located, that is: if the driving direction of V is opposite to the driving direction of the lane where it is located, then use the first lane adjacent to the lane where V is located and in the same driving direction as V as M; if the driving direction of V is the same as the driving direction of the lane where it is located, then use the lane where V is located as M;

[0057] Step S1.2: Construct a lane set S according to the reference lane M and the lanes adjacent to the reference lane M lane ;

[0058] Specifically, according to M and the lanes adjacent to M (which may include the left same-direction lane of M (labeled as L), the right same-direction lane of M (labeled as R), and the left reverse lane of M (labeled as C))

[0059] Construct a lane set S lane , where and M ∈ S lane and L ∈ S lane and C

[0060] ∈ S lane is mutually exclusive;

[0061] Step S1.3: Sort the traffic participants other than V (referred to as other traffic participants) on the lane X ∈ S lane in the reverse direction of M's driving direction to obtain a list of other traffic participants (labeled as S X traffic = {t X 0, t X 1, …, t X n , t X n+1}, where t X i , i = 1, …, n are other traffic participants on X, t0 is a virtual traffic participant set at INF behind along the driving direction of M on X, t n+1 is a virtual traffic participant set at INF in front along the driving direction of M on X, and INF is a set larger value), then the driving state set on X is S X state = {x0 = f(V, A(t X 0) ∩ B(t X 1)), …, x n = f(V, A(t X n ) ∩ B(t X n+1 ))}, where A(t X ) represents the front area of t X on the lane X, B(t X ) represents the rear area of t X on the lane X, A(t X i ) ∩ B(t X i+1 ) represents the area between t X i and t X i+1 on the lane X, and f represents the intersection of the V row and the area A(t X i ) ∩ B(tX i+1 ) The function mapped to the driving state, X ∈

[0062] S lane ;

[0063] For the convenience of description, the driving state sets and driving state representations corresponding to lanes with different markings are as follows:

[0064] If the lane marking is M, then S M state = {m0, m1, …, m n(M)}, where n(M) is the number of other traffic participants on M;

[0065] If the lane marking is L, then S L state = {l0, l1, …, l n(L)}, where n(L) is the number of other traffic participants on L;

[0066] If the lane marking is R, then S R state = {r0, r1, …, r n(R)}, where n(R) is the number of other traffic participants on R;

[0067] If the lane marking is C, then S C state = {c0, c1, …, c n(C)}, where n(C) is the number of other traffic participants on C;

[0068] Merge the driving state sets of each lane in S lane into the total driving state set (marked as S state ).

[0069] Step S1.4: Use all the driving states in S state as the nodes of the STG. For any two driving states x i and x j , if and only if the lanes to which x i and x j belong are adjacent, then add a directed edge from x i to x j and a directed edge from x j to x i in the STG.

[0070] Figure 2a , Figure 2b gives a schematic diagram of constructing the driving state transition graph STG for two examples. Figure 2a Example one of Figure 2bExample 2 is a one-way three-lane road (as shown in the following figure). Here, the two-way single-lane example in the figure is used to illustrate the construction of the driving state transition graph according to step S1: The driving direction of the intelligent vehicle V is the same as the driving direction of the lane it is in. Mark the lane where V is located as the reference lane M. The lane on the left side of the driving direction of V is the only lane adjacent to M and has the opposite direction to M. Mark this adjacent lane as C. Therefore, in this example, the lane set S lane ={M, C}. Collect the traffic participants on M along the driving direction of V as S M traffic ={t M 0, t M 1, t M 2}. Collect the traffic participants on C as S C traffic ={t C 0, t C 1}, where t M 0, t M 2, t C 0, t C 1 are virtual traffic participants. Obtain the driving state set S M state ={m0 = f(V, <A(t M 0) ∩ B(t M 1)>), m 10 = f(V, <A(t M 1) ∩ B(t M 2)>)} on M. Obtain the driving state set S C state ={c0 = f(V, <A(t C 0) ∩ B(t C 1)>)}, on C. Merge S M state and S C state to obtain the total driving state set S state ={m0, m1, c0}; Finally, take m0, m1, c0 as the nodes of the STG, and since m0 and c0, m1 and c0 are adjacent in the lanes they belong to, add a directed edge from m0 to c0, a directed edge from c0 to m0, a directed edge from m1 to c0, and a directed edge from c0 to m1. The obtained STG is shown on the right side of Figure 2a the figure.

[0071] According to the same step S1, the STG of Example 2 can be obtained (as shown on the right side of Figure 2b the figure).

[0072] Step S2: Based on the driving state transition graph and the spatio-temporal information of traffic participants, construct a spatio-temporal trajectory tree; use the current state of the intelligent vehicle as the root node of the spatio-temporal trajectory tree and add it to the end of the first-in-first-out queue, and add the driving state associated with the root node to the explored driving state list; if the time step of the head node in the first-in-first-out queue is less than the decision step size of the explored driving state list, then take out the head node; add the direct child nodes of the head node to the end of the first-in-first-out queue. Except for leaf nodes, other nodes have exactly 1 direct child node, and the driving state of the direct child node is the same as that of the parent node; if in the driving state transition graph, there is a directed edge from the driving state of the head node to the driving state of its candidate collateral child node, and the driving state of the candidate collateral child node does not belong to the explored driving state list, then add the candidate collateral child node to the end of the first-in-first-out queue, and add the driving state of the candidate collateral child node to the explored driving state list; except for leaf nodes, other nodes have at most 2 collateral child nodes, and the driving states of the collateral child nodes are different from that of the parent node and do not belong to the same lane; if the time step of the head node in the first-in-first-out queue is greater than or equal to the decision step size of the explored driving state list, terminate the iteration and output the spatio-temporal trajectory tree, which specifically includes the following steps:

[0073] Step S2.1: Initialize an empty first-in-first-out queue (marked as ), an empty explored driving state list (marked as ) and the decision step size D;

[0074] Step S2.2: Construct the root node (marked as root) of the spatio-temporal trajectory tree (marked as T) according to the current state of V. Among them, each node (marked as node) of the spatio-temporal trajectory tree is represented by a six-tuple

[0075] <d, X, s, v, x, F>, where d, X, s, v, x respectively represent information such as the time step, lane, position, speed, driving state and parent node of V at this node; assume that the driving state associated with root is x i ∈S state , add root to the FIFO end of Q i , and add x

[0076] to ClosedList;

[0077] Step S2.3: Iteratively construct the spatio-temporal trajectory tree according to the following steps:

[0077] Step S2.3.1: The head node in Q FIFO is node i =<d i , X, s i , v i , x i , Fi >, if the time step d where the node node i is located i < D, then take out the head node of the queue and proceed with the following steps; otherwise, terminate the iteration and output the spatio-temporal trajectory tree T;

[0078] Step S2.3.2: Use the longitudinal one-step speed planning model (labeled as P 纵向 ) to obtain the direct child node of node i (labeled as node iic ), and add node iic to the end of the Q FIFO queue; except for leaf nodes, other nodes have exactly 1 direct child node, and the driving state of the direct child node is the same as that of the parent node;

[0079] The longitudinal one-step speed planning model P 纵向 is specifically as follows: Assume that the node node of the spatio-temporal trajectory tree i is used as the starting state of the longitudinal one-step speed planning model P 纵向 , and the predicted state of the traffic participant at the d i th step is used as the environmental state information of P 纵向 . P 纵向 plans and outputs the speed and position considering the V body dynamics and kinematic constraints for the next step according to the starting state and environmental state information, etc.; P 纵向 can be constructed based on rules or can be a learnable network model.

[0080] A preferred longitudinal one-step speed planning model P longitudinal is v d = max{v min , min{v max , dist(node i , t X aheadi(di) ) / HT}} and s d = s i + v d × dt, where v d is the expected speed value of the direct child node of node i , v min and v max are respectively the minimum speed and the maximum speed that can be achieved by the direct child node of node i considering the speed of node i , the maximum acceleration / deceleration of V, and the lane speed limit, etc. constraints, dist(node i , t X aheadi(di) ) is node iThe distance between the position and the first traffic participant in front of the lane it is in. HT is the shortest headway time (generally, for safety, HT = 2 s), s i is the position of node i , s d is the expected position of the direct children nodes of node i . dt is the time for one-step planning.

[0081] Step S2.3.3: Use the lateral one-step lane change decision model (labeled as P 横向 ) to obtain all possible candidate collateral children nodes of node i (labeled as node isc ): Assume the driving state of the candidate collateral children node is x nisc . If there is a directed edge from x ni to x nisc in the STG and at this time, then add node isc to the end of the Q FIFO queue, and add x nisc to the ClosedList; Except for leaf nodes, other nodes have at most 2 collateral children nodes. The driving states of the collateral children nodes are different from those of the parent node and do not belong to the same lane;

[0082] The lateral one-step lane change decision model P 横向 is specifically as follows: Assume that the node of the spatio-temporal trajectory tree node i is used as the starting state of the lateral one-step lane change decision model P 横向 . Use the predicted state of the traffic participant at the d i th step as the P 横向 environmental state information. Use the adjacent lane X ∈ S i of the lane where node lane is located as the target lane. P 横向 outputs decision information such as whether the lane can be changed in the next step according to the starting state, environmental state, and target lane; P 横向 can be constructed based on rules or can be a learnable network model.

[0083] A preferred lateral one-step lane change decision model P 横向 is that if dist(node i , t P behind(di) ) > d1 and dist(node i , t P ahead(di) ) > d2, that is, if the distances between the position of node i and the first traffic participants behind and in front of the adjacent lane P are greater than d1 and d2 respectively, then vd = max{v min , min{v max , dist(node i , t P aheadi(di) ) / HT}} and s d = s i + v d × dt, where d1 and d2 are the set safety intervals between workshops, and the meanings of other symbolic notations are the same as those in the described preferred longitudinal one-step speed planning model.

[0084] Step S2.3.4: Jump to Step S2.3.1.

[0085] Figure 3a Give Figure 2b the schematic diagram of the spatio-temporal trajectory tree. Given the driving state set S of this scenario state = {m0, m1, l0, l1, r0}, and the driving state transition graph STG is shown in Figure 2b as shown. Initialize D = 6, and initialize the root node (labeled with integer 0) of the spatio-temporal trajectory tree T with the current state (including position, heading, and speed). Node 0 is in the position interval of A(t M 0) ∩ B(t M 1) of lane M. Therefore, the driving state of the root node is m0. Finally, simulate forward according to the longitudinal one-step speed planning model and the lateral one-step lane change decision model to obtain the spatio-temporal trajectory tree as shown in Figure 3a as shown.

[0086] Step S3: Extract spatio-temporal trajectories from the described spatio-temporal trajectory tree, score all candidate spatio-temporal trajectories, and finally select the spatio-temporal trajectory with the highest score as the decision result, which specifically includes the following steps:

[0087] Step S3.1: Follow T to collect all nodes from the root node to the leaf nodes to form a trajectory point sequence (labeled as Trace i , i = 1, 2,..., h, where h is the number of leaf nodes in T and also the number of trajectory point sequences finally collected); Figure 3b Give the spatio-temporal trajectory extracted from the spatio-temporal trajectory tree of Example 2 shown in Figure 3a as shown.

[0088] Step S3.2: Score all trajectory point sequences according to the designed evaluation indicators. A preferred evaluation indicator includes but is not limited to guidance (labeled as C guide ), timeliness (labeled as C effic ), safety (labeled as C safe ), and consistency (labeled as C consist), the definitions of each evaluation index are as follows:

[0089] C guide : The guidance evaluation measures the proximity of the lane that V finally reaches through the spatio-temporal trajectory to the target lane given by the global route information, and is evaluated by the center line distance between the two lanes;

[0090] C effic : The timeliness evaluation measures that V can complete the driving task faster through the spatio-temporal trajectory, and can be evaluated according to the speed information of the trajectory point sequence associated with the spatio-temporal trajectory;

[0091] C safe : The safety evaluation measures the probability that V drives according to the spatio-temporal trajectory and avoids collisions with other traffic participants, and can be evaluated according to the spatio-temporal distance between the trajectory point sequence associated with the spatio-temporal trajectory and the surrounding traffic participants;

[0092] C consist : The consistency evaluation measures the number of lane changes when V drives according to the decision motion, and can be evaluated according to the event count of two consecutive nodes not being in the same lane in the trajectory point sequence associated with the spatio-temporal trajectory.

[0093] Step S3.3: Calculate the total score C through weighted summation total = w effic × C effic + w safe × C safe + w consist × C consist , where w effic 、w safe 、w consist are the weights of the corresponding indicators, and the one with the highest total score C guide is selected from the effective spatio-temporal trajectories with the highest guidance C total as the decision result.

[0094] Corresponding to the embodiments of the intelligent driving multi-step long-term behavior decision-making method with joint spatio-temporal information described above, the present invention also provides embodiments of an intelligent driving multi-step long-term behavior decision-making device with joint spatio-temporal information.

[0095] See Figure 4 , the intelligent driving multi-step long-term behavior decision-making device with joint spatio-temporal information provided by the embodiments of the present invention includes a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the intelligent driving multi-step long-term behavior decision-making method in the above embodiments.

[0096] Embodiments of the intelligent driving multi-step long-term behavior decision-making device that combines spatio-temporal information of the present invention can be applied to any device with data processing capabilities, and such a device with data processing capabilities can be a device or apparatus such as a computer. The device embodiments can be implemented through software, or through hardware, or a combination of software and hardware. Taking software implementation as an example, as a logically meaningful device, it is formed by the processor of any device with data processing capabilities reading the corresponding computer program instructions in the non-volatile memory into the memory for operation. At the hardware level, as Figure 4 shown, it is a hardware structure diagram of any device with data processing capabilities where the intelligent driving multi-step long-term behavior decision-making device that combines spatio-temporal information of the present invention is located. In addition to Figure 4 the processor, memory, network interface, and non-volatile memory shown, generally, according to the actual functions of any device with data processing capabilities where the device in the embodiment is located, other hardware may also be included, which will not be elaborated here.

[0097] For the implementation processes of the functions and roles of each unit in the above device, please refer to the implementation processes of the corresponding steps in the above method for details, which will not be elaborated here.

[0098] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can refer to the partial descriptions of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they can be located in one place, or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present invention. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0099] Embodiments of the present invention also provide a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, it implements the intelligent driving multi-step long-term behavior decision-making method that combines spatio-temporal information in the above embodiments.

[0100] The computer-readable storage medium may be an internal storage unit of any data processing-capable device described in any of the foregoing embodiments, such as a hard disk or a memory. The computer-readable storage medium may also be an external storage device of any data processing-capable device, such as a plug-in hard disk, a Smart Media Card (SMC), an SD card, a Flash Card, etc. equipped on the device. Further, the computer-readable storage medium may also include both an internal storage unit and an external storage device of any data processing-capable device. The computer-readable storage medium is used to store the computer program and other programs and data required by any data processing-capable device, and may also be used to temporarily store the data that has been output or is to be output.

[0101] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent driving multi-step long-term behavior decision-making method combining spatio-temporal information, characterized in that It includes the following steps: Step S1: Generate a driving state set according to road structure information and traffic participant information. Take each driving state as a node, and generate directed edges between nodes based on the driving states of adjacent lanes to construct a driving state transition graph; Step S2: Construct a spatio-temporal trajectory tree based on the driving state transition graph and the spatio-temporal information of traffic participants; Take the current state of the intelligent vehicle as the root node of the spatio-temporal trajectory tree and add it to the end of the first-in-first-out queue. Add the driving states associated with the root node to the explored driving state list. If the time step of the head node in the first-in-first-out queue is less than the decision step length of the explored driving state list, then take out the head node. Add the direct child nodes of the head node to the end of the first-in-first-out queue. Except for leaf nodes, other nodes have exactly 1 direct child node, and the driving state of the direct child node is the same as that of the parent node. If there is a directed edge from the driving state of the head node to the driving state of its candidate collateral child node in the driving state transition graph, and the driving state of the candidate collateral child node does not belong to the explored driving state list, then add the candidate collateral child node to the end of the first-in-first-out queue and add the driving state of the candidate collateral child node to the explored driving state list. Except for leaf nodes, other nodes have at most 2 collateral child nodes, and the driving states of the collateral child nodes are different from those of the parent node and do not belong to the same lane. If the time step of the head node in the first-in-first-out queue is greater than or equal to the decision step length of the explored driving state list, terminate the iteration and output the spatio-temporal trajectory tree. Specifically, it includes the following steps: Step S2.1: Initialize an empty first-in-first-out queue An empty list of explored driving states And a decision step size D; Step S2.2: Construct the root node root of the spatio-temporal trajectory tree T according to the current state of the intelligent vehicle V. Each node node of the spatio-temporal trajectory tree is represented by a six-tuple <d, X, s, v, x, F>, where d, X, s, v, and x represent the time step, lane, position, speed, driving state of the intelligent vehicle V at this node, and the parent node information respectively; assume that the driving state associated with the root node root is x i ∈S state , add the root node root to the first-in-first-out queue Q FIFO At the end of the queue, add the driving state x i to the explored driving state list ClosedList; Step S2.3: Iteratively construct the spatio-temporal trajectory tree according to the following steps: Step S2.3.1: First-In-First-Out queue Q FIFO The head node in the queue is node i =<d i ,X,s i ,v i ,x i ,F i >, if the time step d where the node node i is located < i <D, then take out the head node of the queue for subsequent steps, otherwise terminate the iteration and output the spatio-temporal trajectory tree T; Step S2.3.2: Use the longitudinal one-step speed planning model P 纵向 , to obtain the direct child node node i of node iic , and add node iic to the first-in-first-out queue Q FIFO at the end of the queue; except for leaf nodes, other nodes have exactly one direct child node, and the driving state of the direct child node is the same as that of the parent node; Step S2.3.3: Use the lateral one-step lane change decision-making model P 横向 , to obtain all possible candidate collateral child nodes node i of isc : Assume that the driving state of the candidate collateral child node is x nisc . If there is a directed edge from x ni to x nisc in the driving state transition graph STG and at this time, then add node isc to the end of the first-in-first-out queue Q FIFO , and add x nisc to the explored driving state list ClosedList; except for leaf nodes, other nodes have at most 2 collateral child nodes, and the driving states of the collateral child nodes are different from those of the parent node and do not belong to the same lane; Step S2.3.4: Jump to Step S2.3.1; Step S3: Extract spatio-temporal trajectories from the spatio-temporal trajectory tree and score all candidate spatio-temporal trajectories, and finally select the spatio-temporal trajectory with the highest score as the decision result.

2. The intelligent driving multi-step long-term behavior decision-making method combining spatio-temporal information according to claim 1, characterized in that: The said Step S1 includes the following steps: Step S1.1: Determine the reference lane M according to the lane where the intelligent vehicle V is located. If the driving direction of the intelligent vehicle V is opposite to the driving direction of the lane where it is located, then use the first lane adjacent to the lane where the intelligent vehicle V is located and in the same driving direction as the intelligent vehicle V as the reference lane M. If the driving direction of the intelligent vehicle V is the same as the driving direction of the lane where it is located, then use the lane where the intelligent vehicle V is located as the reference lane M; Step S1.2: Construct a lane set S based on the reference lane M and the lanes adjacent to the reference lane M lane ; Step S1.3: Sort other traffic participants except the intelligent vehicle V on the lane X ∈ S in the driving direction of the reference lane M from the back to the front to obtain a list of other traffic participants, and merge the driving state sets of each lane in S into a total driving state set S lane ; lane Merge the driving state sets of each lane in S into a total driving state set S state ; Step S1.4: Using all driving states in the total driving state set S state as the nodes of the driving state transition graph STG, for any two driving states x i and x j , if and only if the lanes to which x i and x j belong are adjacent, then in the driving state transition graph STG, add a directed edge from x i to x j and a directed edge from x j to x i respectively.

3. The intelligent driving multi-step long-term behavior decision-making method combining spatio-temporal information according to claim 2, characterized in that: In the step S1.2, and M ∈ S lane , and L ∈ S lane is mutually exclusive with C ∈ S lane . L represents the same-direction lane on the left side of M, R represents the same-direction lane on the right side of M, and C represents the reverse lane on the left side of M; in the step S1.3, the list S X traffic = {t X 0, t X 1, …, t X n , t X n+1}, where t X i , i = 1, …, n are other traffic participants on lane X, t X 0 is a virtual traffic participant set at INF behind the driving direction of the reference lane M on lane X, t X n+1 is a virtual traffic participant set at INF in front of the driving direction of the reference lane M on lane X, and INF is a set first threshold value. Then the driving state set on lane X is S X state = {x0 = f(V, A(t X 0) ∩ B(t X 1)), …, x n = f(V, A(t X n )) ∩ B(t X n+1 ))}, where A(t X ) represents the front area of other traffic participants t on lane X X and B(t X ) represents the rear area of other traffic participants t on lane X X A(t X i ) ∩ B(t X i+1 ) represents the area between other traffic participants t X i and t X i+1 on lane X, and f represents a function that maps the area A(t X i ) ∩ B(t X i+1 ) where the intelligent vehicle V travels on lane X to the driving state, X ∈ S lane ; The driving state sets and driving state representations corresponding to lanes with different markings are as follows: If the lane marking is M, then S M state ={m0, m1, …, m n(M)}, where n(M) is the number of other traffic participants m on M; If the lane marking is L, then S L state ={l0, l1, …, l n(L)}, where n(L) is the number of other traffic participants l on L; If the lane marking is R, then S R state ={r0,r1,…,r n(R)}, where n(R) is the number of other traffic participants r on R; If the lane marking is C, then S C state = {c0, c1, …, c n(C)}, where n(C) is the number of other traffic participants c on C.

4. The intelligent driving multi-step long-term behavior decision-making method combining spatio-temporal information according to claim 1, characterized in that: The longitudinal one-step speed planning model P in the step S2.3.2 纵向 Specifically: Assume that the node of the spatio-temporal trajectory tree i is used as the starting state of the longitudinal one-step speed planning model P 纵向 and the predicted state of the traffic participant at the d i -th step is used as the environmental state information of P 纵向 Based on the starting state and the environmental state information, P 纵向 plans and outputs the speed and position considering the dynamics and kinematics constraints of the intelligent vehicle V itself for the next step; P 纵向 It is constructed based on rules or is a learnable network model.

5. The intelligent driving multi-step long-term behavior decision-making method combining spatio-temporal information according to claim 4, characterized in that: The longitudinal one-step speed planning model P 纵向 is: v d = max{v min , min{v max , dist(node i , t X aheadi(di) ) / HT}} and s d = s i + v d × dt, where v d is the expected speed value of the direct child node of node i , v min and v max are respectively the minimum speed and the maximum speed that can be achieved by the direct child nodes of node i , dist(n i , t X aheadi(di) ) is the distance between the position of node i and the first traffic participant in front of it on its lane, HT is the shortest headway time, s i is the position of node i , s d is the expected position of the direct child node of node i , and dt is the time of one-step planning.

6. The intelligent driving multi-step long-term behavior decision-making method combining spatio-temporal information according to claim 1, characterized in that: The lateral one-step lane change decision model P in step S2.3.2 横向 Specifically: Assume that the node node of the space-time trajectory tree i As a lateral one-step lane change decision model P 横向 The starting state of the traffic participant d i The predicted state of the step is taken as P 横向 Environment status information, in node i The adjacent lane of the lane X∈S lane is the target lane, P 横向 Output the decision information of whether to change lanes in the next step according to the starting state, environment state and target lane; P 横向 Based on rule construction, or a learnable network model.

7. The intelligent driving multi-step long-term behavior decision-making method combining spatio-temporal information according to claim 6, characterized in that: The lateral one-step lane-changing decision-making model P 横向 is: If dist(node i ,t P behind(di) ) > d1 and dist(node i ,t P ahead(di) ) > d2, that is, if the distances between the position of node i and the first traffic participants behind and in front of the adjacent lane P are greater than d1 and d2 respectively, then v d = max{v min , min{v max , dist(node i ,t P aheadi(di) ) / HT}} and s d = s i + v d × dt, where d1 and d2 are the set safe inter-vehicle intervals, v min and v max are respectively the minimum speed and the maximum speed that can be achieved by the direct children nodes of n i , dist(node i ,t P aheadi(di) ) is the distance between the position of node i and the first traffic participant in front of the adjacent lane P, HT is the shortest headway time, s i is the position of node i , s d is the expected position of the direct child node of node i , and dt is the time for one-step planning.

8. The intelligent driving multi-step long-term behavior decision-making method combining spatio-temporal information according to claim 1, characterized in that: The said Step S3 includes the following steps: Step S3.1: Follow the spatio-temporal trajectory tree T to collect all nodes from the root node to the leaf nodes, forming a trajectory point sequence Trace i , i = 1, 2, ..., h, where h is the number of leaf nodes in T and also the number of finally collected trajectory point sequences; Step S3.2: Score all trajectory point sequences according to the designed evaluation index: Evaluation metrics include but are not limited to guiding C guide , timeliness C effic , security C safe and consistency C consist , and the interpretations of each evaluation metric are as follows: C guide : The guiding evaluation assesses the proximity between the lane where V finally arrives through the spatio-temporal trajectory and the target lane given by the global route information, and is evaluated by the center line distance between the two lanes; C effic : The timeliness evaluation enables V to complete the driving task faster through the spatio-temporal trajectory, and can be evaluated according to the speed information of the sequence of trajectory points associated with the spatio-temporal trajectory; C safe : The safety assessment V is the probability of avoiding collisions with other traffic participants when driving along the spatio-temporal trajectory, which can be evaluated based on the spatio-temporal distances between the sequence of trajectory points associated with the spatio-temporal trajectory and the surrounding traffic participants. C consist : The consistency evaluation V is based on the number of lane changes when driving according to the decision motion, and can be evaluated by counting the events where two consecutive nodes in the sequence of trajectory points associated with the spatio-temporal trajectory are not in the same lane; Step S3.3: Calculate the total score C by weighted summation total = w effic × C effic + w safe × C safe + w consist × C consist , where w effic , w safe , w consist are the weights corresponding to the indicators. Select the one with the highest total score C guide from the valid spatio-temporal trajectories with the highest guiding C total as the decision result.

9. An intelligent driving multi-step long-term behavior decision-making device combining spatio-temporal information, characterized in that It includes a memory and one or more processors. Executable code is stored in the memory. When the one or more processors execute the executable code, it is used to implement the intelligent driving multi-step long-term behavior decision method combining spatio-temporal information according to any one of claims 1-8.

Images