A motion planning method based on safe trajectory tree network

Abstract

This invention discloses a motion planning method based on a safe trajectory tree network, comprising: generating a trajectory tree that represents a discretized feasible space for all traffic participants using a TTNET network; obtaining deep feature encodings through multilayer perceptron (MLP) encoding, including the bounding box features of the autonomous vehicle and its surrounding traffic participants, and the deep feature encodings of local reference lines; these deep feature encodings are concatenated and fed into a spatial Transformer encoder, where they undergo sufficient feature fusion to output an overall interactive encoding; the overall interactive encoding is input into a trajectory prediction aid and a safe trajectory tree network model to output the final motion planning trajectory. Compared to existing methods, the TTNET network improves the driving score by 8.3% and route completion by 39.2% on the CARLA Town05 closed-loop benchmark; on the CARLA Longest6 benchmark, the TTNET network improves the traffic rule compliance score by 10.6%. Simultaneously, the TTNET network also has an inference speed of at least 28 Hz, which is 1.5 times faster than the best methods.

Description

Technical Field

[0001] The invention belongs to the field of neural networks and path planning, and in particular relates to a motion planning method based on a safety trajectory tree network. Background Art

[0002] Autonomous vehicle technology has the advantages of reducing traffic accidents, improving traffic efficiency, reducing environmental pollution and providing mobility for the disabled. After nearly a decade of rapid development, this technology has begun to be integrated into people's daily lives. When current autonomous vehicles perform point-to-point navigation tasks in different traffic scenarios, the overall framework still follows the basic design of "perception-planning-control" of mobile robot systems. After determining the global route, the motion planning part is responsible for generating short-term collision-free trajectories required for real-time operation of the vehicle. Motion planning often plays a key role in the entire framework after being coupled with behavioral decision-making. Solving the motion planning problem requires considering multiple constraints at the same time, mainly including feasibility constraints in vehicle dynamics, real-time interaction constraints in the surrounding dynamic environment, traffic rules constraints such as traffic lights, and time efficiency constraints required to complete the task. The superposition of these different constraints makes the motion planning task very complicated and makes it difficult for autonomous vehicles to make reasonable decisions. Therefore, there are still many open challenges in the motion planning task that need to be solved.

[0003] In addition, motion planning methods based on imitation learning still cannot solve the "inertia problem", which is the phenomenon that self-driving cars remain stationary for long periods of time in complex scenarios. Summary of the invention

[0004] The present invention focuses on the safety of the planning module and the completeness of the overall autonomous driving task, and provides a motion planning method based on a safe trajectory tree network, which uses a safe trajectory tree network (TTNET) to solve the above problems.

[0005] The present invention is achieved by adopting the following technical solutions:

[0006] A motion planning method based on a safe trajectory tree network is proposed. The method uses a combination of trajectory trees and neural networks to complete the planning task of an autonomous vehicle and the multimodal trajectory prediction task of surrounding traffic participants. The method specifically includes the following steps:

[0007] 1) The trajectory tree set of the autonomous vehicle and surrounding traffic participants is generated through the TTNET network during the offline data processing stage; the trajectory tree set can represent the discretized feasible space of all traffic participants, and the trajectory tree of the autonomous vehicle is encoded by the multi-layer perceptron MLP to obtain the deep feature encoding of the trajectory tree;

[0008] 2) These deep feature codes are cascaded and sent to the Transformer encoder of the TTNET network. After sufficient feature fusion between the three types of information of the vehicle, surrounding traffic participants and local reference lines in the spatial Transformer encoder, the overall interaction code is output;

[0009] 3) The TTNET network applies two auxiliary tasks to enhance the network's ability to process relevant information; the first auxiliary task is the multimodal trajectory prediction task, and the second auxiliary task is to use MLP to map the deep feature encoding input after ego-vehicle fusion into the current speed of the ego-vehicle.

[0010] A further improvement of the present invention is that in step 1), the discretized feasible space of all traffic participants is represented by the TTNET network, and the TTNET network selects a discretized representation method, which consists of five parts: a) an intelligent agent observation feature set, b) a traffic light state, c) a target point, d) a local reference line, and e) a trajectory tree.

[0011] A further improvement of the present invention is that in step 1), the trajectory tree set is represented by TTNET, and the generation process of the trajectory tree is divided into two stages, namely the expansion stage and the refinement stage, and the whole process is completed offline before the training stage.

[0012] A further improvement of the present invention is that in the expansion phase, the TTNET network utilizes multiple time intervals. In practice, the TTNET network uses two time steps as intervals, and then each intermediate node is expanded using the same process until the final leaf node is obtained, thereby obtaining a preliminary simplified trajectory tree.

[0013] A further improvement of the present invention is that, in the refinement stage, that is, after the expansion process is completed, the TTNET network introduces a standard Euler spiral to optimize the trajectory tree structure and improve the quality of the trajectory tree.

[0014] A further improvement of the present invention is that in step 2), the deep feature encoding is output as an overall interaction encoding through a Transformer encoder, and the spatial Transformer encoder is used to extract the spatial interaction features between the vehicle and the surrounding traffic participants, and is also used to extract the spatial interaction features between the local reference line Orf and the intelligent agent observation feature set Oob.

[0015] A further improvement of the present invention is that the spatial Transformer encoder is composed of 8 groups of identical network layers, each network layer is divided into two network sub-layers, the first network sub-layer is a multi-head attention sub-layer, and the second network sub-layer is a multi-layer feedforward network sub-layer, and each network sub-layer is sequentially connected through residual connections and layer normalization.

[0016] A further improvement of the present invention is that in step 3), the TTNET network uses a trajectory measurement operation to calculate the L2 distance between all leaf trajectories in the trajectory tree and the true value of the future trajectory; after the calculation is completed, the leaf trajectory index number closest to the future true trajectory is set to q, and it is used to supervise the learning of the score vector p; at the same time, the TTNET network introduces the focal loss function into the training process:

[0017]

[0018] Where: k represents the index of the leaf trajectory; α k Used to adjust the loss weights of different labels; γ = 2 is used to balance the distribution of labels.

[0019] Compared with the prior art, the present invention has at least the following beneficial technical effects:

[0020] The present invention provides a motion planning method based on a safe trajectory tree network, in which the key module of the TTNET network is a predefined trajectory tree. Specifically, the TTNET network introduces the interpretable input expression in the traditional motion planning method into the learning-based motion planning method framework. The interpretable input expression includes the bounding box of the perception result and the local reference line obtained from the global route. This expression can be easily obtained from the auxiliary tasks of the end-to-end method, or from the perception module of the traditional autonomous driving overall framework. Therefore, the TTNET network can be seamlessly connected to the overall autonomous driving framework as part of the end-to-end method or as an alternative to the traditional motion planning method. The trajectory tree can not only meet the vehicle kinematic constraints and explicitly express the different behavioral intentions of the autonomous driving vehicle, but also act on the core motion planning task and the trajectory prediction auxiliary task at the same time, so that the TTNET network has the ability to explore the interaction between trajectory prediction and motion planning.

[0021] Furthermore, the present invention proposes a multi-task planning framework based on imitation learning, which uses a Transformer-based backbone network to accurately capture the interaction information between different traffic participants and the local reference lines of the autonomous driving vehicle.

[0022] Furthermore, a trajectory tree with curvature continuity and kinematic feasibility is constructed, which can not only be used as planning output but also for trajectory prediction, thus forming a predictive planning multi-task framework.

[0023] Furthermore, an interpretable input expression is introduced, so that the TTNET network can be easily integrated into the overall task of autonomous driving. At the same time, a focal loss function is designed to effectively enhance the ability of the TTNET network to handle complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A simplified diagram of the trajectory tree generation process is shown in Figure 2. Figure 1 (a) is the expansion stage, Figure 1 (b) is the refining stage.

[0025] Figure 2 This is a schematic diagram of the TTNET network. DETAILED DESCRIPTION

[0026] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to be able to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features described in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0027] The present invention provides a motion planning method based on a safety trajectory tree network, comprising the following steps:

[0028] 1. Input and output expressions

[0029] The trajectory tree set of the autonomous vehicle and surrounding traffic participants is generated through the TTNET network in the offline data processing stage. The trajectory tree set can represent the discretized feasible space of all traffic participants.

[0030] TTNET uses a top-down two-dimensional coordinate system. The origin is the position of the lidar sensor on the roof of the ego vehicle. This coordinate system ensures that the ego vehicle is always at the origin at the current moment. At the same time, the top-down two-dimensional view provides a simple and easy-to-interpret way to represent the spatial position relationship between traffic participants. This feature is crucial to the generalization performance of the TTNET network.

[0031] End-to-end methods usually convert raw sensor data into deep feature representations as input to the planning module. These deep feature representations often lack interpretability, hindering the effectiveness of the planning module. TTNET network chooses a simpler discretization representation. The representation consists of 5 parts:

[0032] a) Agent observation feature set (O ob ): represents the location features and bounding box feature sets of different types of traffic participants at the current moment;

[0033]

[0034] Where: Represents the four corner points and the center point of the bounding box of the vehicle or traffic participants within the perception range; Represents the complete observation information of a single traffic participant i, including category a, bounding box location point set box, speed v, heading angle θ, width w and length h; the index number of the ego vehicle is fixed to i = 0, and the index number of the traffic participant within the perception range is set to i∈{1,…,N}, and the total number of traffic participants N will change with the scene and time;

[0035] b) Traffic light status The Boolean value of the traffic light signal detected by the vehicle sensor at the current moment;

[0036] c) Target point The short-term target position in front of the ego vehicle in the global path provision;

[0037] d) Local reference line: A set of path points obtained by using spline curve interpolation on the path points within a set distance range from the current position of the ego vehicle on the global path; in order to help the ego vehicle move along the local reference line, a corner point is manually generated on the left and right sides of each reference line path point; the distance between the corner point and the reference line path point is half the width of the ego vehicle; the entire local reference line is represented as:

[0038]

[0039] Where: rf Represents a feature set of a single reference line path point and its corresponding left and right corner points, including the type of reference line (α is forced to be 9), the position of the reference line path point and the left and right corner points and lane width w; N ′ Indicates the number of reference line path points.

[0040] e) Trajectory tree: A trajectory tree quickly generated based on the current status information of the vehicle and surrounding traffic participants; a single trajectory tree is represented as Where K is the number of leaf trajectories contained in the entire tree, and T = 4 refers to the prediction time length;

[0041] The output of the TTNET network consists of three parts, including the planned trajectory of the vehicle The current speed of the vehicle And the score vector set P that evaluates the trajectory tree of surrounding traffic participants prod The trajectory planned by the ego vehicle can be passed to the controller, which generates the steering, throttle, and braking commands required to drive the vehicle.

[0042] 2. Generation of trajectory tree

[0043] The generation process of the trajectory tree is divided into two stages, namely the expansion stage and the refinement stage, and the whole process is completed offline before the training stage. It is divided into two stages, namely the expansion stage and the refinement stage. In the step 1), the generation process of the trajectory tree is divided into two stages, namely the expansion stage and the refinement stage, and the whole process is completed offline before the training stage.

[0044] The following takes the generation process of the vehicle trajectory tree as an example to illustrate. Figure 1

[0045] 1) Expansion stage

[0046] According to the possible discrete value changes of the steering wheel angle of the vehicle {δ1,…,δ J} and speed discrete values Generate a set of speed-steering pairs Set the current position of the vehicle speed and heading angle The current state of the composition is the root state; then, all speed-turn pairs are sequentially brought into the vehicle kinematic model to obtain the intermediate nodes that extend from the root node to the specified speed and angle; the expansion process of a single intermediate node is expressed as:

[0047]

[0048]

[0049]

[0050] Where: dt is the time interval between the intermediate node and the root node.

[0051] 2) Refining stage

[0052] After the expansion process, the TTNET network introduces the standard Euler spiral to optimize the trajectory tree structure and improve the quality of the trajectory tree; given the state of the root node and the leaf node, each trajectory in the trajectory tree is re-expressed as a standard Euler spiral;

[0053]

[0054]

[0055] Where: (x, y) is the arc length l 4n+1 The position of the curve; a represents the rate of change of curvature;

[0056] The process of using Euler spirals to refine the trajectory tree can ensure the curvature continuity of the trajectory; after obtaining a trajectory tree composed of a set of standard Euler spirals, the discrete time steps are mapped to the Euler spiral to obtain the final discrete point trajectory tree; the trajectory tree of the vehicle can be used to obtain the trajectory trees of surrounding traffic participants through rotation and translation operations.

[0057] 3. Spatial Transformer Encoder

[0058] These deep feature codes are cascaded and sent to the spatial Transformer encoder. After sufficient feature fusion between the three types of information, namely the vehicle, surrounding traffic participants and local reference lines, the spatial Transformer encoder outputs the overall interaction code.

[0059] The spatial Transformer encoder consists of 8 groups of identical network layers, each of which is divided into two network sublayers; the first sublayer is a multi-head attention sublayer, and the second sublayer is a multi-layer feedforward network sublayer; each sublayer is sequentially connected with residual connections and layer normalization;

[0060] (1) Input processing

[0061] First, the local reference line O rf and the agent observation feature set O ob is processed by two separate embedding layers;

[0062]

[0063] Where: and O rf and O ab Deep feature encoding; d m =512 is the feature dimension; and It is a network with two fully connected layers;

[0064] Then, E rf and E ab Perform cascade operations between them to obtain the input of the spatial Transformer encoder;

[0065]

[0066] Where: Indicates cascade operation; E ab+rf It can provide three different types of information: the vehicle itself, surrounding traffic participants, and local reference lines;

[0067] (2) Multi-head attention sub-layer

[0068] The multi-head attention sublayer uses the self-attention mechanism to process the input feature E of the spatial information at the current moment ab+ , the self-attention mechanism is understood as a message passing mechanism; when E ab The deep feature encoding of any object i in When key Sum can be obtained independently through different fully connected layer networks; and d q ,d k and d v are three feature dimensions; N h = 8 indicates the number of headers; the query, key and value of the hth version are expressed as:

[0069]

[0070]

[0071] Where: and It is a network of three fully connected layers;

[0072] A message from object j to object i is defined as:

[0073]

[0074] At this time, the attention head of object i is expressed as:

[0075]

[0076] Where: Used to improve gradient stability;

[0077] Cascade i all N h After the attention heads are linearly mapped, the multi-head attention MultiHead (Q i ,K i ,V k );

[0078]

[0079] Where: Represents a fully connected layer network: Concat represents a cascade operation; d q =d k =d v =d m / N h ;

[0080] At this point, object i has collected the messages transmitted by all traffic participants including itself and local reference lines; the message transmission process is applied to all objects in the same scene to complete the deep feature information interaction between them; in the second network sublayer, all deep feature information learns higher-dimensional features in the multi-layer feedforward network sublayer; the spatial Transformer encoder outputs the updated overall interaction encoding set;

[0081]

[0082] This set provides the deep feature information of the ego vehicle and surrounding traffic participants required by the multi-task branches.

[0083] 4. Trajectory prediction assistance and safety trajectory tree network model

[0084] The safety trajectory tree network of the multi-task motion planning framework neural network based on imitation learning provided by the present invention has two main innovations: a multi-task planning framework based on imitation learning and a trajectory tree with curvature continuity and kinematic feasibility.

[0085] (1) Planning the main tasks

[0086] First, the vehicle deep feature encoding of the overall interaction encoding set The fully connected layer network is used for compression and dimensionality reduction. Then, the target point information O tp , Traffic light status information tl The ego vehicle encoding feature is concatenated with the dimension-reduced ego vehicle encoding feature to generate the ego vehicle deep feature encoding for planning the main task.

[0087]

[0088] This ego-vehicle encoding is able to provide the basic deep features required for the main planning task.

[0089] On the other hand, the vehicle trajectory tree A tree encoding set containing deep feature encodings of all leaf trajectories is generated through a multi-layer perceptron.

[0090]

[0091] Where: is the deep feature encoding of the kth leaf track.

[0092] In order to effectively evaluate each leaf trajectory, the TTNET network is used in the vehicle encoding and tree encoding The attention mechanism is applied between them to generate a score vector p of the ego-car encoding with respect to the tree encoding.

[0093]

[0094] Where: and are two fully connected layer networks; T represents matrix transpose.

[0095] In the trajectory tree, the leaf trajectory closest to the future true trajectory provides the most reasonable explanation of the vehicle's behavior and needs to be given the highest score. Therefore, the TTNET network uses trajectory measurement operations to calculate the L2 distance between all leaf trajectories in the trajectory tree and the future trajectory truth value. After the calculation is completed, the leaf trajectory index number closest to the future true trajectory is set to q, and it is used to supervise the score vector p.

[0096] Imitation learning methods are severely affected by the "inertia problem". The inertia problem refers to the problem that the ego vehicle remains stationary for a long time in the training data, causing the ego vehicle to learn a suboptimal stationary strategy. To solve this problem, the TTNET network introduces the focal loss function into the training process.

[0097]

[0098] Where: k represents the index of the leaf trajectory; α k Used to adjust the loss weights of different labels; γ = 2 is used to balance the distribution of labels.

[0099] During training, α of the ground-truth stationary trajectory was set to 0.25 (α k＝1 =0.25), the true trajectory α of other index values ​​is set to 0.75 (α k！＝1 =0.75). This setting makes the static trajectory data become the "background data" and other trajectories become the "foreground data", which effectively reduces the loss contribution of static real trajectories and alleviates the inertia problem caused by the presence of a large number of static trajectories in the training data.

[0100] During the training phase, the car encodes and leaf trajectory encoding of the true label Cascade, and then use GRU autoregression to generate the final planning trajectory.

[0101]

[0102] During the inference phase, the leaf trajectory encoding of the true label is replaced by the leaf trajectory encoding with the highest score The same process is then used to generate the planning trajectory.

[0103] (2) Trajectory prediction auxiliary task

[0104] In addition to completing the main planning task, the TTNET network also introduces an auxiliary task of predicting the future multimodal trajectories of surrounding traffic participants. The ability to predict the trajectories of surrounding traffic participants is crucial for the ego vehicle to make reasonable decisions. The trajectory prediction auxiliary task can provide the spatial Transformer encoder backbone network with the dynamic interaction information between the ego vehicle and the surrounding traffic participants required for the ego vehicle to make reasonable decisions. The input of the multimodal trajectory prediction auxiliary task is the encoding part of the surrounding traffic participants in the overall interaction encoding set, called the prediction encoding. Different from the complex design of the planning task, the trajectory prediction task only uses a classifier consisting of a multi-layer perceptron and a Softmax layer to output the score vector of the trajectory tree corresponding to the surrounding traffic participants.

[0105]

[0106] The leaf trajectory index set Q that is closest to the future true trajectory of surrounding traffic participants pred It is also obtained through the trajectory measurement operation. The multimodal trajectory prediction task also uses the focal loss function during training.

[0107]

[0108] in, Represents the focal function for trajectory prediction.

[0109] (3) Overall loss function

[0110] The overall loss function of the TTNET network consists of four parts: planning focus loss function Prediction focal loss function Planning regression loss function And the current speed regression loss function

[0111] In the planning task, the regression loss function can be expressed as:

[0112]

[0113] Where: represents the SmoothL1 loss function

[188] .

[0114] In order to provide the vehicle with the ability to estimate its current state, the TTNET network introduces an auxiliary task of predicting the current speed of the vehicle. This task predicts the current speed of the vehicle by inputting the vehicle code into a multilayer perceptron.

[0115]

[0116] The prediction of the current speed of the vehicle also uses the Smooth L1 loss function.

[0117]

[0118] TTNET network uses a weighted combination of multiple loss functions to train a multi-task model. The overall loss function is expressed as follows:

[0119]

[0120] Where: scave , pred , plan and λ spd is the weight of each loss item.

[0121] Although the present invention has been described in detail above with general descriptions and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements may be made thereto based on the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all fall within the scope of protection claimed by the present invention.

Claims

1. A motion planning method based on a safe trajectory tree network, characterized in that: This method uses a combination of trajectory trees and neural networks to complete the planning task of the autonomous vehicle and the multimodal trajectory prediction task of surrounding traffic participants. Specifically, it includes the following steps: 1) The trajectory tree set of the autonomous vehicle and surrounding traffic participants is generated through the TTNET network during the offline data processing stage; the trajectory tree set can represent the discretized feasible space of all traffic participants, and the trajectory tree of the autonomous vehicle is encoded by the multi-layer perceptron MLP to obtain the deep feature encoding of the trajectory tree; 2) These deep feature codes are cascaded and sent to the Transformer encoder of the TTNET network. After sufficient feature fusion between the three types of information of the vehicle, surrounding traffic participants and local reference lines in the spatial Transformer encoder, the overall interaction code is output; 3) The TTNET network applies two auxiliary tasks to enhance the network's ability to process relevant information; the first auxiliary task is the multimodal trajectory prediction task, and the second auxiliary task is to use MLP to map the deep feature encoding input after ego-vehicle fusion into the current speed of the ego-vehicle.

2. A motion planning method based on a safe trajectory tree network according to claim 1, characterized in that: In step 1), the discretized feasible space of all traffic participants is represented by the TTNET network. The TTNET network selects a discretized representation method, which consists of five parts: a) the agent observation feature set, b) the traffic light state, c) the target point, d) the local reference line, and e) the trajectory tree.

3. A motion planning method based on a safe trajectory tree network according to claim 1, characterized in that: In step 1), the trajectory tree set is represented by TTNET. The generation process of the trajectory tree is divided into two stages, namely the expansion stage and the refinement stage. The whole process is completed offline before the training stage.

4. A motion planning method based on a safety trajectory tree network according to claim 3, characterized in that: In the expansion phase, the TTNET network utilizes multiple time intervals. In practice, the TTNET network uses two time steps as the interval. Subsequently, each intermediate node is expanded using the same process until the final leaf node is obtained, thereby obtaining a preliminary simplified trajectory tree.

5. The motion planning method based on a safety trajectory tree network according to claim 3 is characterized in that: In the refinement phase, that is, after the expansion process, the TTNET network introduces the standard Euler spiral to optimize the trajectory tree structure and improve the quality of the trajectory tree.

6. The motion planning method based on a safe trajectory tree network according to claim 1, characterized in that: In step 2), the deep feature encoding is output as the overall interaction encoding through the Transformer encoder, and the spatial Transformer encoder is used to extract the spatial interaction features between the vehicle and the surrounding traffic participants, and also to extract the spatial interaction features between the local reference line Orf and the agent observation feature set Oob.

7. A motion planning method based on a safety trajectory tree network according to claim 6, characterized in that: The spatial Transformer encoder consists of 8 groups of identical network layers. Each network layer is divided into two network sub-layers. The first network sub-layer is a multi-head attention sub-layer, and the second network sub-layer is a multi-layer feed-forward network sub-layer. Each network sub-layer is sequentially connected with residual connections and layer normalization.

8. The motion planning method based on a safety trajectory tree network according to claim 1, characterized in that: In step 3), the TTNET network uses the trajectory measurement operation to calculate the L2 distance between all leaf trajectories in the trajectory tree and the true value of the future trajectory. After the calculation is completed, the leaf trajectory index number closest to the future true trajectory is set to q, and it is used to supervise the learning of the score vector p. At the same time, the TTNET network introduces the focal loss function into the training process: Where: k represents the index of the leaf trajectory; α k Used to adjust the loss weights of different labels; γ = 2 is used to balance the distribution of labels.

Images