CN117094459A - Motion planning method based on safe track tree network - Google Patents

Abstract

The invention discloses a motion planning method based on a safe track tree network, which comprises the following steps: generating a track tree capable of representing discretized feasible spaces of all traffic participants through a TTNET network; obtaining depth feature codes through multi-layer perceptron MLP codes, and obtaining depth feature codes of boundary frame features and local reference lines of automatic driving automobiles and surrounding traffic participants; the depth feature codes are sent into a space transform coder after being cascaded, and the integral interactive codes are output after full feature fusion; and inputting the integral interactive codes into a track prediction auxiliary and safe track tree network model, and outputting a final motion planning track. Compared with the existing method, the driving score of the TTNET network on the CARLA Town05 closed loop test standard is improved by 8.3%, and the route completion degree is improved by 39.2%; on the calla Longest6 benchmark, the score of TTNET networks for compliance with traffic rules is improved by 10.6%. Meanwhile, the TTNET network also has an inference speed of at least 28 Hz, which is 1.5 times faster than the best method.

Description

Motion planning method based on safe track tree network

Technical Field

The invention belongs to the field of neural networks and path planning, and particularly relates to a motion planning method based on a safe track tree network.

Background

The automatic driving automobile technology has the advantages of reducing traffic accidents, improving traffic efficiency, reducing environmental pollution and providing mobility for disabled people. With rapid developments in the last decade, this technology has begun to be incorporated into people's daily lives. When the current automatic driving automobile performs point-to-point navigation tasks in different traffic scenes, the whole framework still follows the basic design of a mobile robot system 'perception-planning-control'. After the global route is determined, the motion planning section is responsible for generating short-term collision-free trajectories required for real-time operation of the vehicle. Exercise planning often plays a critical role in the overall framework after being coupled with behavioral decisions. The problem of motion planning is solved by considering various constraints at the same time, and the method mainly comprises feasibility constraints in terms of vehicle dynamics, real-time interaction constraints of surrounding dynamic environments, traffic rule constraints such as traffic lights and the like, time efficiency constraints required by completing tasks and the like. The superposition of these different constraints makes the motion planning task very complex and makes it difficult for the autonomous vehicle to make a reasonable decision. Thus, there are still many open challenges to be addressed for the motion planning task at present.

In addition, motion planning methods based on imitation learning still cannot solve the "inertia problem", i.e. the phenomenon that an autonomous car remains stationary for a long period of time in a complex scene.

Disclosure of Invention

The invention focuses on the safety of a planning module and the completeness of an automatic driving overall task, provides a motion planning method based on a safe track tree network, and solves the problems by using the safe track tree network (safe Trajectory Tree Network, TTNET).

The invention is realized by adopting the following technical scheme:

the motion planning method based on the safe track tree network is used for completing the planning task of an automatic driving automobile and the multi-mode track prediction task of surrounding traffic participants through the combination of the track tree and the neural network, and specifically comprises the following steps:

1) The track tree set of the automatic driving automobile and surrounding traffic participants is generated through a TTNET network in an offline data processing stage; the track tree set can represent the discretization feasible space of all traffic participants, and the track tree of the automatic driving automobile is encoded by the multi-layer perceptron MLP to obtain the depth feature code of the track tree;

2) The depth feature codes are cascaded and then sent to a transducer encoder of a TTNET network, and after the three types of information of the vehicle, surrounding traffic participants and a local reference line are subjected to full feature fusion in a space transducer encoder, the overall interactive codes are output;

3) The TTNET network applies two auxiliary tasks to enhance the processing capacity of the network for related information; the first auxiliary task is a multi-mode track prediction task, and the second auxiliary task is to map depth feature code input after vehicle fusion into the current speed of the vehicle by using MLP.

The invention is further improved in that in step 1), the discretized feasible space of all traffic participants is represented by a TTNET network, which selects a discretized representation mode, and the representation consists of 5 parts: a) an agent observation feature set, b) traffic light status, c) target points, d) local reference lines, e) track trees.

In the step 1), the track tree set is represented by TTNET, the track tree generation process is divided into two stages, namely an expansion stage and a refining stage, and the whole process is finished offline before the training stage.

In the development stage, the TTNET network uses multiple time intervals, in practice, the TTNET network uses two time steps as intervals, and then each intermediate node develops by adopting the same process until a final leaf node is obtained, so that a preliminary simplified track tree is obtained.

The invention is further improved in that the TTNET network introduces a standard Euler spiral line to optimize the track tree structure and improve the quality of the track tree in the refining stage, namely after the expansion process is finished.

A further improvement of the present invention is that in step 2), the depth feature codes are output as global interaction codes by a transducer encoder, which is used to extract spatial interaction features between the vehicle and surrounding traffic participants, and also to extract spatial interaction features between the local reference line Orf and the agent observation feature set Oob.

The invention is further improved in that the space Transformer encoder is composed of 8 groups of identical network layers, each network layer is divided into two network sublayers, the first network sublayer is a multi-head attention sublayer, the second network sublayer is a multi-layer feedforward network sublayer, and residual connection and layer standardization are sequentially carried out among each network sublayer.

In the step 3), the TTNET network adopts track measurement operation to calculate L2 distances between all leaf tracks in the track tree and the true value of future tracks; after the calculation is completed, the She Guiji index number closest to the future real track is set to be q, and the score vector p is subjected to supervised learning by the index number; meanwhile, the TTNET network introduces a focus loss function into the training process:

wherein: k represents the index of the leaf track; alpha _k For adjusting the loss weights of different tags; γ=2 is used to balance the distribution of the tags.

Compared with the prior art, the invention has at least the following beneficial technical effects:

the invention provides a motion planning method based on a safe track tree network, wherein a key module of a TTNET network is a predefined track tree. Specifically, TTNET networks introduce interpretable input expressions in traditional motion planning methods into a learning-based motion planning method framework. The interpretable input representation includes a bounding box of perceived results and local reference lines obtained from the global route. The expression can be easily obtained from the auxiliary task of the end-to-end method, and can also be obtained from the perception module of the traditional automatic driving integral frame. Thus, the TTNET network may be seamlessly spliced into the autopilot overall framework as part of an end-to-end approach or as an alternative to a traditional motion planning approach. The track tree not only can meet the kinematic constraint of the vehicle and explicitly express different behavior intentions of the automatic driving automobile, but also can act on the core motion planning task and the track prediction auxiliary task at the same time, so that the TTNET network has the capability of exploring interaction between track prediction and motion planning.

Furthermore, the invention provides a simulation learning-based multitasking planning framework, which utilizes a trunk network based on a Transformer to accurately capture interaction information between different traffic participants and local reference lines of an automatic driving automobile.

Further, a trajectory tree with curvature continuity and kinematic feasibility is constructed, which can be used not only as a planning output, but also for trajectory prediction, thereby forming a predictive planning multitasking framework.

Further, an interpretable input representation is introduced, thereby enabling the TTNET network to be easily accessed into the overall task of automatic driving. Meanwhile, a focus loss function is designed, and the capability of the TTNET network for processing complex scenes is effectively enhanced.

Drawings

Fig. 1 is a simplified schematic diagram of a trajectory tree generation process, wherein fig. 1 (a) is an expansion phase and fig. 1 (b) is a refining phase.

Fig. 2 is a schematic diagram of a TTNET network.

Detailed Description

Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art. It should be noted that, without conflict, the embodiments of the present invention and features of the embodiments may be combined with each other. The invention will be described in detail below with reference to the drawings in connection with embodiments.

The invention provides a motion planning method based on a safe track tree network, which comprises the following steps:

1. input output expression

Generating a track tree set of the automatic driving automobile and surrounding traffic participants in an offline data processing stage through a TTNET network, wherein the track tree set can represent discretized feasible spaces of all the traffic participants;

TTNET networks employ a two-dimensional coordinate system from top to bottom. The origin is the position of the lidar sensor from the roof of the vehicle. This coordinate system ensures that the host vehicle is always at the origin at the current time. At the same time, the top-down two-dimensional view provides a simple and easy to interpret way of representing the spatial positional relationship between traffic participants. This characteristic is critical to the generalization performance of TTNET networks.

End-to-end methods typically convert raw sensor data into depth characterization representations as input to a planning module. These depth feature characterization approaches often lack interpretability, impeding the effectiveness of the planning module. The TTNET network selects a simpler discretized representation. The characterization consisted of 5 parts:

a) Intelligent observation feature set (O) _ob ): representing the position features and the boundary frame feature sets of different types of traffic participants at the current moment;

wherein:representing four corner positions and a center point set of a traffic participant boundary box in a self-vehicle or perception range; />Complete observation information representing an individual 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 own vehicle is fixedly set to i=0, the index number of the traffic participants in the perception range is set to i epsilon {1, …, N }, and the total number N of the traffic participants can change along with scene and time change;

b) Traffic light statusThe traffic light signal Boolean value detected by the vehicle sensor at the current moment;

c) Target pointA short term target location in the global path offer that is in front of the host vehicle;

d) Local reference line: a spline curve interpolation is used for the path points in a set distance range from the current position of the vehicle on the global path to obtain a path point set; in order to help the self-vehicle advance along the local reference line, manually generating a corner point on the left side and the right side of each reference line path point; the distance between the corner point and the reference line path point is half of the width of the vehicle; the entire local reference line is denoted as:

wherein: o (o) _rf Feature set representing single reference line path point and corresponding left and right corner points thereof, including type of reference line (alpha is forcedly set to 9), position of reference line path point and left and right corner pointsLane width w; n (N) ^′ Representing the number of reference line path points.

e) Track tree: track trees are quickly generated according to the current state information of the own vehicles and surrounding traffic participants; a single track tree is represented asWhere K is the number of leaf trajectories contained throughout the tree, t=4 refers to the predicted time length;

the TTNET network has three parts including planning track of own vehicleCurrent speed of own vehicle->Score vector set P for evaluating corresponding track tree of surrounding traffic participants _prod The method comprises the steps of carrying out a first treatment on the surface of the The planned trajectory of the vehicle can be transferred to a controller, generating steering, throttle and braking commands required for driving the vehicle.

2. Generation of track tree

The track tree generation process is divided into two stages, namely an expansion stage and a refining stage, and the whole process is finished offline before the training stage. The method comprises an expansion stage and a refining stage 2, wherein in the step 1), the generation process of the track tree is divided into two stages, namely the expansion stage and the refining stage, and the whole process is finished offline before the training stage.

The following describes an example of a process for generating a vehicle track tree. See FIG. 1

1) Expansion stage

Variation { delta } of steering wheel angle discrete value according to possible self-vehicle ₁ ,…,δ _J Sum of velocity discrete valuesGenerating a set of speed-steering pairs +.>Self-settingCurrent position of vehicle->Speed->And heading angle->The current state of the composition is the root state; then, all the speed-steering pairs are sequentially brought into a vehicle kinematic model to obtain an intermediate node of the expansion of the root node to the designated speed and the corner; the single intermediate node expansion process is expressed as:

wherein: dt is the time interval between the intermediate node and the root node.

2) Refining stage

After the expansion process is finished, introducing a standard Euler spiral line into the TTNET network to optimize the track tree structure and improve the quality of the track tree; giving the states of the root node and the leaf node, and re-representing each track in the track tree as a standard Euler spiral line;

wherein: (x, y) is arc length l ⁴ⁿ⁺¹ A location of the site; a represents the rate of change of curvature;

the process of refining the track tree by using the Euler spiral line can ensure the curvature continuity of the track; after a track tree consisting of a group of standard Euler spiral lines is obtained, mapping discrete time step sizes onto the Euler spiral lines to obtain a final discrete point track tree; the own vehicle track tree can obtain the track tree of surrounding traffic participants through rotation and translation operations.

3. Space transducer encoder

The depth feature codes are sent to a space transducer encoder after being cascaded, and the three types of information of the own vehicle, surrounding traffic participants and the local reference line are output to the whole interactive code after the space transducer encoder is fully subjected to feature fusion;

the space transducer encoder consists of 8 groups of identical network layers, and each layer of network is divided into two network sublayers; the first sub-layer is a multi-head attention sub-layer, and the second sub-layer is a multi-layer feedforward network sub-layer; each sub-layer is sequentially subjected to residual connection and layer standardization;

(1) Input processing

First, local reference line O _rf And agent observation feature set O _ob Is processed by two independent embedded layers;

wherein:and->Respectively is O _rf And O _ab Is a depth feature encoding of (2); d, d _m =512 is the feature dimension; />And->Is two full connection layer networks;

then E _rf And E is _ab Performing cascading operation to obtain the input of a space Transformer encoder;

wherein:representing a cascading operation; e (E) _ab+rf Three different types of information can be provided, namely, a vehicle, surrounding traffic participants and a local reference line;

(2) Multi-headed attention sub-layer

Multi-head attention sub-layer uses self-attention mechanism to process input characteristic E of current time space information _ab+ Self-attention mechanisms are understood to be messaging mechanisms; when E is _ab Depth feature encoding of arbitrary object iWhen it corresponds to inquiry->Key->Sum->Can be obtained independently through different full connection layer networks; wherein->And-> d _q ,d _k And d _v Is three feature dimensions; n (N) _h =8 denotes the number of heads; the h-th version of the query, key, and value are expressed as:

wherein:and->Is three full connection layer networks;

the message from object j to object i is defined as:

at this time, the attention head of the object i is expressed as:

wherein:for improving gradient stability;

cascade i all N _h Multiple attention multi head (Q) by linear mapping after each attention head ⁱ ,K ⁱ ,V ^k )；

Wherein: concat represents a cascading operation; d, d _q ＝d _k ＝d _v ＝d _m /N _h ；

At this point, object i has collected messages delivered by all traffic participants and local reference lines, including itself; the message transmission process is applied to all objects in the same scene, and depth characteristic information interaction among the objects is completed; in the second network sub-layer, all depth feature information learns higher-dimensional features in the multi-layer feed-forward network sub-layer; the space transform encoder outputs an updated whole interactive coding set;

the collection provides depth profile information for the vehicle and surrounding traffic participants required for the multitasking branch.

4. Trajectory prediction auxiliary and safety trajectory tree network model

The invention provides a safe track tree network based on a simulation learning multitasking motion planning framework neural network, which is mainly characterized in that the simulation learning-based multitasking planning framework and a track tree with curvature continuity and kinematic feasibility are constructed:

(1) Planning a primary task

First, the depth feature code of the own vehicle of the whole interactive code setAnd performing compression dimension reduction through a full connection layer network. Then, the target point information O _tp Traffic light status information O _tl And the self-vehicle coding feature cascade after dimension reduction, which generates a self-vehicle depth feature code for planning a main task>

The vehicle code can provide the basic depth features required to plan the primary task.

On the other hand, the own vehicle track treeA set of tree encodings is generated by the multi-layer perceptron that contains all leaf-track depth feature encodings.

Wherein:is the depth feature encoding of the kth leaf track.

For efficient evaluation of each leaf track, TTNET network codes in own vehicleAnd tree coding->Applying an attention mechanism, generating a score vector p of the vehicle code about the tree code.

Wherein:and->Is two full connection layer networks; t represents the matrix transpose.

In the track tree, the leaf track closest to the last real track provides the most reasonable interpretation of the behavior of the vehicle, for which the highest score needs to be given. Therefore, the TTNET network adopts track measurement operation to calculate L between all leaf tracks and final track true values in the track tree ₂ Distance. After the calculation is completed, the She Guiji index number closest to the last real track is set to q, and is used for supervised learning of the score vector p.

The imitation learning method is severely affected by the "inertia problem". The inertia problem refers to the problem that the own vehicle remains stationary in the training data for a long period of time, thereby allowing the own vehicle to learn a suboptimal stationary strategy. To address this problem, TTNET networks introduce a focus loss function into the training process.

Wherein k represents an index of a leaf track; alpha _k For adjusting the loss weights of different tags; γ=2 is used to balance the distribution of the tags.

In training, α of the true rest trajectory is set to 0.25 (α _k＝1 =0.25), the true trajectory α of the other index values is set to 0.75 (α _k！＝1 =0.75). The static track data is changed into background data, and other tracks are changed into foreground data, so that the loss contribution of the static real track is effectively reduced, and the inertia problem caused by a large number of static tracks in the training data is relieved.

During training phase, bicycle codeAnd She Guiji coding of the authentic tag->Cascading and then generating a final planned trajectory by using a GRU autoregressive formula.

She Guiji coding of genuine tags in the inference phaseIs replaced by She Guiji code with highest scoreThe same procedure is then used to generate the planned trajectory.

(2) Trajectory prediction auxiliary task

Except forIn addition to completing the primary planning task, TTNET networks also introduce an auxiliary task to predict future multimodal trajectories of surrounding traffic participants. Having the ability to predict the trajectories of surrounding traffic participants is critical to making reasonable decisions for the own vehicle. The trajectory prediction assistance task may provide the spatial transducer encoder backbone network with dynamic interaction information between the host and surrounding traffic participants that is needed for the host to make a reasonable decision. The input of the multi-modal trajectory prediction auxiliary task is the coding part of surrounding traffic participants in the whole interactive coding set, called predictive codingUnlike the complex design of planning tasks, trajectory prediction tasks output scoring vectors for surrounding traffic participants' corresponding trajectory trees using only classifiers that include multiple layers of perceptrons and Softmax layers.

Leaf track index set Q nearest to future real track of surrounding traffic participants _pred Also obtained by a trajectory measurement operation. The multi-modal trajectory prediction task also uses the focus loss function in training.

Wherein,a focal function representing trajectory prediction.

(3) Integral loss function

The overall loss function of TTNET networks consists of four parts: planning a focus loss functionPredictive focus loss function->Planning regression loss function->And the current speed regression loss function->

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

wherein:representing smoothL ₁ Loss function [188 ]]。

To provide the ability of the host to estimate the current state, TTNET networks introduce an auxiliary task for current host speed prediction. The task predicts the current speed of the vehicle by inputting the vehicle code into the multi-layer perceptron

The prediction of the current speed of the own vehicle also uses Smooth L ₁ A loss function.

TTNET networks train a multitasking model using a weighted combination of multiple loss functions. The overall loss function is expressed as follows:

in which lambda is _scave 、λ _pred 、λ _plan And lambda (lambda) _spd Weights for each loss term.

While the invention has been described in detail in the foregoing general description and with reference to specific embodiments thereof, it will be apparent to one skilled in the art that modifications and improvements can be made thereto. Accordingly, such modifications or improvements may be made without departing from the spirit of the invention and are intended to be within the scope of the invention as claimed.

Claims (8)

1. The motion planning method based on the safe track tree network is characterized by completing the planning task of an automatic driving automobile and the multi-mode track prediction task of surrounding traffic participants through the combination of the track tree and the neural network, and specifically comprises the following steps:

1) The track tree set of the automatic driving automobile and surrounding traffic participants is generated through a TTNET network in an offline data processing stage; the track tree set can represent the discretization feasible space of all traffic participants, and the track tree of the automatic driving automobile is encoded by the multi-layer perceptron MLP to obtain the depth feature code of the track tree;

2) The depth feature codes are cascaded and then sent to a transducer encoder of a TTNET network, and after the three types of information of the vehicle, surrounding traffic participants and a local reference line are subjected to full feature fusion in a space transducer encoder, the overall interactive codes are output;

3) The TTNET network applies two auxiliary tasks to enhance the processing capacity of the network for related information; the first auxiliary task is a multi-mode track prediction task, and the second auxiliary task is to map depth feature code input after vehicle fusion into the current speed of the vehicle by using MLP.

2. The method for motion planning based on a safe trajectory tree network according to claim 1, wherein in step 1), the discretized feasible space of all traffic participants is represented by a TTNET network, and the TTNET network selects a discretized representation mode, and the representation is composed of 5 parts: a) an agent observation feature set, b) traffic light status, c) target points, d) local reference lines, e) track trees.

3. The method of claim 1, wherein in step 1), the track tree set is represented by TTNET, the track tree generation process is divided into two stages, namely an expansion stage and a refining stage, and the whole process is finished offline before the training stage.

4. A method of motion planning based on a secure trajectory tree network as claimed in claim 3, characterized in that in the expansion phase the TTNET network uses multiple time intervals, in practice the TTNET network uses two time steps as intervals, and then each intermediate node is expanded by the same process until the final leaf node is obtained, whereby a preliminary simplified trajectory tree is obtained.

5. A method of motion planning based on a secure trajectory tree network as claimed in claim 3, wherein the TTNET network introduces a standard euler spiral to optimize the trajectory tree structure and improve the trajectory tree quality, i.e. after the end of the expansion process in the refining stage.

6. The method of claim 1, wherein in step 2), the depth feature codes are output as global interactive codes by a transducer encoder, and the spatial transducer encoder is used for extracting spatial interactive features between the vehicle and surrounding traffic participants, and is also used for extracting spatial interactive features between the local reference line Orf and the agent observation feature set Oob.

7. The method of claim 6, wherein the spatial transform encoder is comprised of 8 identical network layers, each of the network layers being divided into two network sub-layers, the first network sub-layer being a multi-headed attention sub-layer and the second network sub-layer being a multi-layered feed forward network sub-layer, each of the network sub-layers being sequentially subjected to residual connection and layer normalization.

8. The method for motion planning based on a safe trajectory tree network according to claim 1, wherein in step 3), the TTNET network calculates L2 distances between all leaf trajectories and future trajectory truth values in the trajectory tree by using a trajectory measurement operation; after the calculation is completed, the She Guiji index number closest to the future real track is set to be q, and the score vector p is subjected to supervised learning by the index number; meanwhile, the TTNET network introduces a focus loss function into the training process:

wherein: k represents the index of the leaf track; alpha _k For adjusting the loss weights of different tags; γ=2 is used to balance the distribution of the tags.