Planning algorithm design fusing intention and target trajectory prediction

Abstract

According to the invention, a planning algorithm design fusing intention and target trajectory prediction is realized, and the state information of the vehicle under the Frenet coordinate system is obtained in real time through mutual conversion between the Cartesian coordinate system and the Frenet coordinate system; determining a driving reference line of the vehicle by using the global path, and predicting a driving track of the dynamic obstacle by fusing a driving intention and a target track prediction network model; planning a track and a speed by adopting a dynamic planning and quadratic planning algorithm; and integrating the tracks of the current frame and the planning frame through a polynomial splicing technology, and transmitting the integrated track to a track tracker.

Description

Technical Field

[0001] The present invention relates to the field of local path planning and trajectory prediction for autonomous vehicles, and particularly to a planning algorithm design that fuses intention and target trajectory prediction. Background Art

[0002] In the field of autonomous driving, trajectory planning is one of the key technologies for achieving safe and efficient driving. Traditional trajectory planning methods usually rely on static maps and preset paths. However, with the development of autonomous driving technology, trajectory planning in dynamic environments has become particularly important. In complex traffic scenarios, vehicles need to predict the trajectories of other surrounding vehicles in real time to make reasonable decisions and plans. However, existing trajectory prediction methods still have deficiencies in dealing with multi-vehicle interactions and complex scenarios, such as inaccurate modeling of traffic participant characteristics, overemphasis on single-vehicle prediction while ignoring multi-vehicle interaction games.

[0003] To improve the accuracy and adaptability of trajectory planning, researchers have begun to explore methods that fuse driving intention and target trajectory prediction. Driving intention recognition can infer future behavioral actions by analyzing the vehicle's historical behavior and current environmental information, thus providing important references for trajectory prediction. At the same time, trajectory prediction models based on deep learning, such as long short-term memory networks (LSTMs) and Transformer networks, can better capture the interaction relationships and dynamic changes between vehicles. These methods can generate more accurate and reliable trajectory prediction results by comprehensively considering the driving intention of the vehicle and the dynamic changes in the surrounding environment.

[0004] In addition, to achieve smooth and stable trajectory planning, researchers have also adopted various optimization algorithms, such as dynamic programming and quadratic programming, to optimize the shape and speed distribution of the trajectory. These algorithms can generate optimal trajectory paths on the premise of meeting safety and comfort constraints. Through the combination of these technologies, autonomous vehicles can perform trajectory planning more flexibly in dynamic environments and effectively cope with various complex traffic scenarios. Summary of the Invention

[0005] The purpose of the present invention is to provide a planning algorithm design that fuses intention and target trajectory prediction to improve the performance of vehicle local planning in order to overcome the above problems.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A planning algorithm design that fuses intention and target trajectory prediction, comprising the following steps:

[0008] Step 1: Real-time obtain the state information of the vehicle in the Frenet coordinate system through the mutual conversion between the Cartesian coordinate system and the Frenet coordinate system;

[0009] Step 2: Determine the driving reference line of the vehicle using the global path;

[0010] Step 3: Predict the driving trajectory of the dynamic obstacle through a network model that fuses driving intention and target trajectory prediction;

[0011] Step 4: Use dynamic programming and quadratic programming algorithms for trajectory and speed planning

[0012] Step 5: Integrate the trajectories of the current frame and the planned frame through polynomial splicing technology and pass them to the trajectory tracker

[0013] Optionally, step 1 specifically includes:

[0014] In the Cartesian coordinate system, the lateral and longitudinal trajectories of the vehicle are interrelated, which makes the trajectory planning problem complex. However, in the Frenet coordinate system, the position of the vehicle can be represented by the distance from the reference line, usually denoted by the symbols s and l for the longitudinal and lateral positions respectively. By differentiating these positions with respect to time, the longitudinal and lateral speeds of the vehicle can be obtained respectively. In this way, the originally coupled trajectory problem in two-dimensional space is simplified to two independent one-dimensional problems in the Frenet coordinate system: one for dealing with the longitudinal trajectory and the other for dealing with the lateral trajectory. This decoupling greatly reduces the computational complexity and makes the trajectory planning more efficient and easier to implement. Specifically, the longitudinal trajectory mainly focuses on the forward speed and acceleration of the vehicle along the reference line, while the lateral trajectory focuses on the change in the distance between the vehicle and the reference line, so as to achieve precise control and planning of the vehicle in a complex road environment.

[0015]

[0016] Where The actual state of the Cartesian coordinate system, Is the state parameter of the Frenet coordinate system, s is the longitudinal displacement, Is the longitudinal speed, Is the longitudinal acceleration, l is the lateral displacement, Is the lateral speed, Is the lateral acceleration, l' is the first derivative with respect to the arc length, and l” is the second derivative with respect to the arc length.

[0017] Optionally, step 2 specifically includes:

[0018] To improve the accuracy of the trajectory tracker and the stability of vehicle driving, the present invention optimizes the global path. Since the original global path may be too long and tortuous, which is not conducive to the accurate tracking of the trajectory tracker, a strategy of reducing the number of reference trajectory points and shortening the path length is adopted. The global path is optimized through quadratic programming to generate a smooth global reference line with an appropriate number of trajectory points, thereby effectively guiding the vehicle to drive along the predetermined reference line.

[0019] The process of obtaining the vehicle guiding driving reference line based on the global path is specifically as follows:

[0020] Definition of the reference line optimization problem:

[0021] J = C smooth + C ref + C length

[0022] Definition of the smoothing cost:

[0023]

[0024] Wherein, w smooth is the weight coefficient, P i is the current point coordinate (x i , y i ), P i-1 is the coordinate of the previous discrete point (x i-1 , y i-1 ), P i+1 is the coordinate of the next discrete point (x i+1 , y i+1 ).

[0025] Definition of the length cost:

[0026]

[0027] Definition of the offset cost:

[0028]

[0029] Wherein, (x ir , y ir ) is the discrete point of the original reference line path.

[0030] Optionally, step 3 specifically includes:

[0031] The trajectory prediction model that fuses driving intention and target first uses the multi-head attention mechanism to extract the temporal and spatial features of the vehicle. On this basis, a lightweight shifted attention module is introduced to capture the key behavioral features of the vehicle in social interactions. Then, driving intention recognition is performed through the BILSTM-ATTENTION network and the SOFTMAX function, and it is introduced into the variational autoencoder (CVAE) module to generate more guiding targets. Finally, using the diversity of the generated targets, multi-modal prediction of the vehicle is achieved.

[0032] Temporal extraction module: First, use a linear layer to convert the state information (historical trajectory) of the agent to be predicted at a certain frame into an embedding representation.

[0033]

[0034] Then activate the embedding representation through the ELU unit.

[0035]

[0036] The activated embedding representation is then input into a bidirectional long short-term memory network (BI-LSTM) for encoding, and the BI-LSTM produces a pair of forward h hidden states and backward h hidden states at each time step.

[0037]

[0038] Dynamic social interaction module: Use a linear layer to convert the historical trajectory hidden state of the agent to be predicted output by the temporal extraction module and the historical trajectory hidden states of surrounding agents into three different vectors.

[0039]

[0040] k t =LINER k (H t ,W k )

[0041] v t =LINER v (H t ,W v )

[0042] where q t 、k t and v t represent the query vector, key vector, and value vector respectively, is the historical trajectory hidden state of the agent to be predicted, is the hidden state of the historical trajectories of surrounding agents, W q 、W k and Wv are three weight matrices. Then, q t , k t and v t are split into k heads:

[0043] Q 0 = cat(split(q t )) = cat{q1, q2,..., q k}

[0044] K i = cat(split(k t )) = cat{k1, k2,..., k k}

[0045] V i = cat(split(v t )) = cat{v1, v2,…, v k}

[0046] Then, feature extraction is performed through a multi-head attention mechanism:

[0047]

[0048] Finally, GLU (Gated Linear Unit) is adopted to extract the interaction relationships between agents and normalize them.

[0049]

[0050] Spatio-temporal interaction module: The spatio-temporal interaction module follows the standard Transformer architecture, and the module takes the sequence output by the dynamic social interaction module as input.

[0051]

[0052] Its aggregated information output is expressed as:

[0053]

[0054] Then, GLU is used to project it, extract effective information, suppress irrelevant features, and normalize it.

[0055]

[0056] Intention recognition module: The model inputs the current step of the context information into a linear layer and an activation function, inputs it into a normalization layer, and finally calculates the probabilities of horizontal and vertical intentions through a Softmax activation function.

[0057]

[0058] Thus, the horizontal and vertical intention matrices at time t can be obtained:

[0059]

[0060] As time goes by, a dynamic intention weight matrix will be constructed in the future steps:

[0061]

[0062] Thus, this dynamic intention weight can be fused with the context information output by the encoder module so that the model can take into account more intention information.

[0063]

[0064] Target encoding and decoding module: The CVAE module consists of an encoder and a decoder, which inputs the feature encoding fused with intention information to enhance the expression ability of the target point latent vector distribution and improve the accuracy of the model's prediction of the future state of the target point.

[0065]

[0066] In the formula, σ is the standard deviation, μ is the mean, ε is a random noise vector sampled from the standard normal distribution, and log(σ 2 ) is the logarithmic parameter of the latent vector variance,

[0067] μ, log(σ 2 ) = MLP2(cat(dec t , MLP1(g true , W mlp1 ), W mlp2 )

[0068] g pred = MLP g (cat(C, Z), W g-mlp )

[0069] Target-intention decoder: During the decoding process, the decoder needs to fully consider information such as intention, target, and space, and integrate the g pred output by the CVAE decoder, the horizontal and vertical intention predictions, and the context information combined with the maneuver intention into a comprehensive feature representation. Then, this comprehensive feature is input into the decoder, and the decoder will output the prediction distribution parameters of the future target point at one time.

[0070]

[0071] Goal-based Multimodal Prediction: To achieve multimodality in trajectory prediction, the model uses the decoder of the CVAE module to generate K different sets of target points, each set representing possible future positions, thereby capturing the diversity and uncertainty in trajectory prediction.

[0072]

[0073] Optionally, step 4 specifically includes:

[0074] In local path planning, accurately predict the trajectories of obstacles, reasonably constrain the kinematic parameters of the vehicle's lateral and longitudinal movements, and achieve dynamic safety obstacle avoidance for the vehicle. Specifically, use the dynamic attention trajectory prediction model that fuses driving intention and goal above to predict the trajectories of dynamic obstacles, and the trajectory planning module performs dynamic planning and quadratic planning of the SL graph and ST graph (trajectory planning and speed planning) according to the issued prediction results.

[0075] Perform trajectory planning and speed planning based on the dynamic programming algorithm and the quadratic programming algorithm;

[0076] In dynamic programming, the trajectory planning cost function is defined as:

[0077] Cost tra = Cost l + Cost dl + Cost ddl + Cost obs

[0078] where l is the lateral displacement, l' is the lateral displacement speed, and l'' is the lateral displacement acceleration, w l is the lateral displacement cost weight coefficient, w dl is the lateral speed cost weight coefficient, w ddl is the lateral acceleration cost weight coefficient,

[0079] (1) The obstacle cost is:

[0080]

[0081] where l min 、l max are the maximum and minimum lateral distances of the obstacle boundary points, s min 、s max are the maximum and minimum longitudinal distances of the obstacle boundary points, and sig mod is the activation function that can map real numbers to the interval (0,1).

[0082] In dynamic programming, the speed planning cost function is defined as:

[0083] cost = cost_speed + cost_acc + cost_jerk + cost_obs

[0084] (1) Obstacle collision cost:

[0085]

[0086] Among them, s min is the minimum longitudinal distance of the vehicle at the corresponding time t, s max is the maximum longitudinal distance, s is the longitudinal distance of the sampling point, buffer is the margin, w obs is the obstacle cost coefficient.

[0087] (2) Definition of speed cost

[0088]

[0089] Among them, w speed is the speed cost coefficient, v is the actual vehicle speed, v ref is the reference vehicle speed.

[0090] (3) Definition of acceleration cost

[0091]

[0092] Among them, s(t + 1, s + k) is the state quantity at the next moment, s(t, s + k) is the state quantity at the current moment, w acc is the speed cost coefficient, a is the actual acceleration.

[0093] (4) Jerk cost

[0094] The jerk of the jerk represents an impact quantity, mainly to reduce the uncomfortable feeling brought by the sudden acceleration of the vehicle and ensure driving comfort.

[0095]

[0096] In quadratic programming, the trajectory planning cost function is defined as:

[0097]

[0098] Among them, l ref is the lateral displacement obtained by dynamic programming.

[0099] In quadratic programming, the speed planning cost function is defined as:

[0100]

[0101] Among them, s ref(i) is the longitudinal displacement of the path obtained by dynamic programming, Obtain the longitudinal speed of the path for dynamic programming, is the longitudinal acceleration, is the longitudinal jerk, w s is the reference longitudinal distance cost coefficient, w vel is the reference longitudinal speed cost coefficient, w acc is the reference longitudinal acceleration cost coefficient, w jerk is the reference longitudinal jerk cost coefficient.

[0102] Optionally, step 5 specifically includes:

[0103] In the planning algorithm, if each frame of the trajectory is replanned, it will cause additional time overhead, and the trajectory sequence is discontinuous and the vehicle is prone to jitter. The purpose of trajectory stitching is to make the trajectory tracking control algorithm better track the trajectory and ensure the continuity of the tracking trajectory.

[0104] Based on polynomial, stitch the current frame and the planned frame trajectory and send it to the trajectory tracker. Specifically:

[0105] l(s) = a0 + a1·s + a2·s 2 + a3·s 3 + a4·s 4 + a5·s 5 dl(s) = a1 + 2a2·s + 3a3·s 2 + 4a4·s 3 + 5a5·s 4

[0106] ddl(s) = 2a2 + 6a3·s + 12a4·s 2 + 20a5·s 3 dddl(s) = 6a3·s + 24a4·s + 60a5·s 2

[0107] The present invention realizes a planning algorithm design that fuses intention and target trajectory prediction. By converting between the Cartesian coordinate system and the Frenet coordinate system, the state information of the vehicle in the Frenet coordinate system is obtained in real time; the driving reference line of the vehicle is determined using the global path, and the driving trajectories of dynamic obstacles are predicted through a network model that fuses driving intention and target trajectory prediction; dynamic programming and quadratic programming algorithms are used for trajectory and speed planning; through polynomial stitching technology, the trajectories of the current frame and the planned frame are integrated and passed to the trajectory tracker; through more accurate trajectory prediction, more information about the surrounding vehicles is provided for the downstream path planning and decision-making modules, which can effectively improve the safety and reliability of autonomous driving in a dynamic environment. Description of the Drawings

[0108] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0109] Figure 1 It is a design flow chart of a planning algorithm that fuses intention and target trajectory prediction of the present invention.

[0110] Figure 2 It is a schematic diagram of obtaining a vehicle guiding driving reference line based on a global path of the present invention;

[0111] Figure 3 It is a framework diagram of the trajectory prediction of the present invention;

[0112] Figure 4 It is a schematic diagram of the dynamic programming algorithm in the trajectory planning of the present invention;

[0113] Figure 5 It is a schematic diagram of quadratic programming in the trajectory planning of the present invention;

[0114] Figure 6 It is a schematic diagram of the dynamic programming algorithm in the speed planning of the present invention.

[0115] Figure 7 It is a schematic diagram of polynomial trajectory splicing of the present invention.

[0116] Figure 8 It is an architecture diagram of the trajectory prediction module of the present invention. Specific embodiments

[0117] The following will describe in detail the embodiments of the present invention. The examples of the embodiments are shown in the drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described below with reference to the drawings are exemplary and are intended to explain the present invention and should not be construed as a limitation of the present invention.

[0118] The following are the descriptions of some nouns and terms of the present invention:

[0119] Dynamic programming algorithm (DP, Dynamic Programming)

[0120] Quadratic programming (QP, Quadratic Programming)

[0121] Please refer to Figure 1 , a planning algorithm design that fuses intention and target trajectory prediction, including the following steps:

[0122] S1: Obtain the state information of the vehicle in the Frenet coordinate system in real time through the mutual conversion between the Cartesian coordinate system and the Frenet coordinate system;

[0123] Specifically, please refer to Figure 2 the provided schematic diagram of the mutual conversion between the Cartesian coordinate system and the Frenet coordinate system. In the Cartesian coordinate system, the lateral trajectory and the longitudinal trajectory of the vehicle are coupled with each other. While in the Frenet coordinate system, the distances between the vehicle position and the reference line can be represented by s and l respectively. By taking the derivatives with respect to time respectively, the longitudinal and lateral trajectory speeds can be obtained. Finally, the two-dimensional trajectory problem of the vehicle is decoupled into two one-dimensional trajectory problems in the Frenet coordinate system, reducing the complexity of the calculation. Specifically:

[0124]

[0125] Among them the actual state of the Cartesian coordinate system is the state parameter of the Frenet coordinate system, s is the longitudinal displacement, is the longitudinal speed, is the longitudinal acceleration, l is the lateral displacement, is the lateral speed, is the lateral acceleration, l' is the first derivative with respect to the arc length, and l” is the second derivative with respect to the arc length.

[0126] S2: Determine the driving reference line of the vehicle by using the global path;

[0127] To improve the accuracy of the trajectory tracker and the stability of the vehicle driving, the present invention optimizes the global path. Since the original global path may be too long and tortuous, which is not conducive to the accurate tracking of the trajectory tracker, a strategy of reducing the number of reference trajectory points and shortening the path length is adopted. The global path is optimized through quadratic programming to generate a smooth global reference line with an appropriate number of trajectory points, thereby effectively guiding the vehicle to drive along the predetermined reference line.

[0128] The process of obtaining the vehicle guiding driving reference line based on the global path is specifically as follows:

[0129] Definition of the reference line optimization problem:

[0130] J = C smooth + C ref + C length

[0131] Define the smoothing cost:

[0132]

[0133] Among them, wsmooth is the weight coefficient, P i is the current point coordinate (x i , y i ), P i-1 is the coordinate of the previous discrete point (x i-1 , y i-1 ), P i+1 is the coordinate of the next discrete point (x i+1 , y i+1 ).

[0134] Define the length cost:

[0135]

[0136] Define the offset cost:

[0137]

[0138] where (x ir , y ir ) are the discrete points of the original reference route.

[0139] S3: Predict the driving trajectory of dynamic obstacles through a network model that fuses driving intention and target trajectory prediction;

[0140] The trajectory prediction model that fuses driving intention and target first uses the multi-head attention mechanism to extract the temporal and spatial features of the vehicle. On this basis, a lightweight shifted attention module is introduced to capture the key behavioral features of the vehicle in social interactions. Then, driving intention recognition is performed through the BILSTM-ATTENTION network and the SOFTMAX function, and it is introduced into the variational autoencoder (CVAE) module to generate more guiding targets. Finally, using the diversity of the generated targets, multi-modal prediction of the vehicle is achieved.

[0141] S4: Use dynamic programming and quadratic programming algorithms for trajectory and speed planning;

[0142] In local path planning, accurately predict the trajectory of obstacles, reasonably constrain the kinematic parameters of the vehicle's lateral and longitudinal movements, and achieve dynamic safety obstacle avoidance for the vehicle.

[0143] Perform trajectory planning and speed planning based on dynamic programming algorithms and quadratic programming algorithms;

[0144] In dynamic programming, the trajectory planning cost function is defined as:

[0145] Cost tra = Cost l + Cost dl + Cost ddl + Costobs

[0146] Among them, l is the lateral displacement, l' is the lateral displacement speed, and l'' is the lateral displacement acceleration. w l is the cost weight coefficient of the lateral displacement, w dl is the cost weight coefficient of the lateral speed, w ddl is the cost weight coefficient of the lateral acceleration.

[0147] (1) The obstacle cost is:

[0148]

[0149] Among them, l min and l max are the maximum and minimum lateral distances of the obstacle boundary points, s min and s max are the maximum and minimum longitudinal distances of the obstacle boundary points, and sig mod is the activation function that can map real numbers to the interval (0,1).

[0150] In dynamic programming, the speed planning cost function is defined as:

[0151] cos t = cos t_speed + cos t_acc + cos t_jerk + cos t_obs

[0152] (1) The obstacle collision cost:

[0153]

[0154] Among them, s min is the minimum longitudinal distance of the vehicle at the corresponding time t, s max is the maximum longitudinal distance, s is the longitudinal distance of the sampling point, buffer is the margin, and w obs is the obstacle cost coefficient.

[0155] (2) The definition of the speed cost

[0156]

[0157] Among them, w speed is the speed cost coefficient, v is the actual vehicle speed, and v ref is the reference vehicle speed.

[0158] (3) The definition of the acceleration cost

[0159]

[0160] Among them, s(t + 1, s + k) is the state quantity at the next moment, s(t, s + k) is the state quantity at the current moment, and wacc is the speed cost coefficient, and a is the actual acceleration.

[0161] (4) Jerk cost

[0162] Jerk is characterized as an impact quantity, mainly to reduce the uncomfortable body feeling caused by the sudden acceleration of the vehicle and ensure driving comfort.

[0163]

[0164] In quadratic programming, the trajectory planning cost function is defined as:

[0165]

[0166] where l ref is the lateral displacement obtained by dynamic programming.

[0167] In quadratic programming, the speed planning cost function is defined as:

[0168]

[0169] where s ref(i) is the longitudinal displacement of the path obtained by dynamic programming, is the longitudinal speed of the path obtained by dynamic programming, is the longitudinal acceleration, is the longitudinal jerk, w s is the reference longitudinal distance cost coefficient, w vel is the reference longitudinal speed cost coefficient, w acc is the reference longitudinal acceleration cost coefficient, w jerk is the reference longitudinal jerk cost coefficient.

[0170] S5: Through polynomial splicing technology, integrate the trajectories of the current frame and the planned frame and transfer them to the trajectory tracker;

[0171] In the planning algorithm, if the trajectory is replanned for each frame, it will cause additional time overhead, and the trajectory sequence is discontinuous and the vehicle is prone to jitter. The purpose of trajectory splicing is to make the trajectory tracking control algorithm better track the trajectory and ensure the continuity of the tracking trajectory.

[0172] l(s) = a0 + a1·s + a2·s 2 + a3·s 3 + a4·s 4 + a5·s 5 dl(s) = a1 + 2a2·s + 3a3·s 2 + 4a4·s 3 + 5a5·s 4

[0173] ddl(s) = 2a2 + 6a3·s + 12a4·s 2 + 20a5·s 3 dddl(s) = 6a3·s + 24a4·s + 60a5·s 2

[0174] The above-disclosed is only a preferred embodiment of the present invention. Of course, the scope of the rights of the present invention cannot be limited thereby. Those of ordinary skill in the art can understand all or part of the processes of implementing the above embodiments, and the equivalent changes made according to the claims of the present invention still fall within the scope covered by the invention.

Claims

1. Design of a planning algorithm that fuses intention and target trajectory prediction, characterized in that, Including the following steps: Step 1: Obtain the vehicle's state information in the Frenet coordinate system in real time through the mutual conversion between the Cartesian coordinate system and the Frenet coordinate system; Step 2: Determine the vehicle's driving reference line using the global path; Step 3: Predict the driving trajectory of dynamic obstacles through a network model that fuses driving intention and target trajectory prediction; Step 4: Use dynamic programming and quadratic programming algorithms for trajectory and speed planning; Step 5: Integrate the trajectories of the current frame and the planned frame through polynomial splicing technology and transfer them to the trajectory tracker.

2. The planning algorithm design for fusing intention and target trajectory prediction according to claim 1, wherein for step 1, In the Cartesian coordinate system, the lateral and longitudinal trajectories of the vehicle are interrelated, which makes the trajectory planning problem complex. However, in the Frenet coordinate system, the vehicle's position can be represented by the distance from the reference line, usually denoted by the symbols s and l for the longitudinal and lateral positions respectively. By taking the derivatives of these positions with respect to time, the vehicle's longitudinal and lateral speeds can be obtained respectively. In this way, the originally coupled trajectory problem in two-dimensional space is simplified into two independent one-dimensional problems in the Frenet coordinate system: one for dealing with the longitudinal trajectory and the other for dealing with the lateral trajectory. This decoupling greatly reduces the computational complexity and makes trajectory planning more efficient and easier to implement. Specifically, the longitudinal trajectory mainly focuses on the vehicle's forward speed and acceleration along the reference line, while the lateral trajectory focuses on the change in the distance between the vehicle and the reference line, thereby achieving precise control and planning of the vehicle in a complex road environment. Among them The actual state of the Cartesian coordinate system is the state parameter of the Frenet coordinate system, s is the longitudinal displacement is the longitudinal velocity is the longitudinal acceleration, l is the lateral displacement is the lateral velocity is the lateral acceleration, l' is the first derivative with respect to the arc length, and l” is the second derivative with respect to the arc length 3. The planning algorithm design for fusing intention and target trajectory prediction according to claim 1, wherein for step 2, To improve the accuracy of the trajectory tracker and the stability of vehicle driving, the present invention optimizes the global path. Since the original global path may be too long and tortuous, which is not conducive to the precise tracking of the trajectory tracker, a strategy of reducing the number of reference trajectory points and shortening the path length is adopted. The global path is optimized through quadratic programming to generate a smooth global reference line with an appropriate number of trajectory points, thereby effectively guiding the vehicle to drive along the predetermined reference line. The process of obtaining the vehicle's guiding driving reference line based on the global path is specifically as follows: Definition of the reference line optimization problem: J = C smooth + C ref + C length Definition of the smoothness cost: Among them, w smooth is the weight coefficient, P i is the current point coordinate (x i , y i ), P i-1 is the previous discrete point coordinate (x i-1 , y i-1 ), P i+1 is the next discrete point coordinate (x i+1 , y i+1 ). Definition of the length cost: Definition of the offset cost: Among them, (x ir , y ir ) is the discrete point of the original reference line path.

4. The planning algorithm design for fusing intention and target trajectory prediction according to claim 1, wherein for step 3, The trajectory prediction model that fuses driving intention and target first uses the multi-head attention mechanism to extract the vehicle's temporal and spatial features. On this basis, a lightweight shifted attention module is introduced to capture the key behavioral features of the vehicle in social interactions. Then, driving intention recognition is performed through the BILSTM-ATTENTION network and the SOFTMAX function, and it is introduced into the variational autoencoder (CVAE) module to generate more guiding targets. Finally, using the diversity of the generated targets, multi-modal prediction of the vehicle is achieved.

5. For the planning algorithm design that fuses intention and target trajectory prediction as described in claim 1, in step 4, it is characterized in that In local path planning, accurately predict the trajectory of obstacles, reasonably constrain the kinematic parameters of the vehicle's lateral and longitudinal movements, and achieve dynamic safety obstacle avoidance for the vehicle. Based on the dynamic programming algorithm and the quadratic programming algorithm, perform trajectory planning and speed planning: In dynamic programming, the trajectory planning cost function is defined as: Cost tra = Cost l + Cost dl + Cost ddl + Cost obs where, l is the lateral displacement, l' is the lateral displacement speed, and l'' is the lateral displacement acceleration, w l is the cost weight coefficient for lateral displacement, w dl is the cost weight coefficient for lateral speed, w ddl is the cost weight coefficient for lateral acceleration, (1) The cost of obstacles is: where l min and l max are the maximum and minimum lateral distances of the obstacle boundary points, s min and s max are the maximum and minimum longitudinal distances of the obstacle boundary points, and sigmod is an activation function that can map real numbers to the interval (0, 1). In dynamic programming, the speed planning cost function is defined as: cost = cost_speed + cost_acc + cost_jerk + cost_obs (1) The cost of obstacle collision: Among them, s min is the minimum longitudinal distance of the vehicle at the corresponding moment t, s max is the maximum longitudinal distance, s is the longitudinal distance of the sampling point, buffer is the margin, w obs is the obstacle cost coefficient. (2) The definition of speed cost Among them, w speed is the speed cost coefficient, v is the actual vehicle speed, and v ref is the reference vehicle speed. (3) The definition of acceleration cost Among them, s(t + 1, s + k) is the state quantity at the next moment, s(t, s + k) is the state quantity at the current moment, w acc is the speed cost coefficient, and a is the actual acceleration. (4) Jerk cost Jerk represents an impact quantity, mainly to reduce the uncomfortable feeling brought by the sudden acceleration of the vehicle and ensure driving comfort. In quadratic programming, the trajectory planning cost function is defined as: Among them, l ref is the horizontal displacement obtained by dynamic programming. In quadratic programming, the speed planning cost function is defined as: where s ref(i) is the longitudinal displacement for obtaining the path by dynamic programming, is the longitudinal speed for obtaining the path by dynamic programming, is the longitudinal acceleration, is the longitudinal jerk, w s is the reference longitudinal distance cost coefficient, w vel is the reference longitudinal speed cost coefficient, w acc is the reference longitudinal acceleration cost coefficient, w jerk is the reference longitudinal jerk cost coefficient.

6. For the planning algorithm design that fuses intention and target trajectory prediction as described in claim 1, in step 5, it is characterized in that In the planning algorithm, if each frame of the trajectory is re-planned, it will cause additional time overhead, and the trajectory sequence is discontinuous and the vehicle is prone to jitter. The purpose of trajectory splicing is to enable the trajectory tracking control algorithm to better track the trajectory and ensure the continuity of the tracked trajectory. Based on polynomial, splice the current frame and the planned frame trajectory and send it to the trajectory tracker. Specifically: l(s) = a0 + a1·s + a2·s 2 + a3·s 3 + a4·s 4 + a5·s 5 dl(s) = a1 + 2a2·s + 3a3·s 2 + 4a4·s 3 + 5a5·s 4 ddl(s) = 2a2 + 6a3·s + 12a4·s 2 + 20a5·s 3 dddl(s) = 6a3·s + 24a4·s + 60a5·s 2 。

Images