CN118270013A

CN118270013A - Method and device for predicting future running track of motor vehicle and unmanned equipment

Info

Publication number: CN118270013A
Application number: CN202211737534.3A
Authority: CN
Inventors: 安向京; 朱亚东; 罗辉武; 胡庭波
Original assignee: Changsha Xingshen Intelligent Technology Co Ltd
Current assignee: Changsha Xingshen Intelligent Technology Co Ltd
Filing date: 2022-12-30
Publication date: 2024-07-02

Abstract

The invention relates to the technical field of unmanned vehicles and discloses a method and a device for predicting a future running track of a motor vehicle and unmanned equipment. According to the invention, the historical running track of the motor vehicle at the intersection is combined with the constraint of the lane line nodes, so that the track prediction model can be finely modeled, the problem of insufficient future track prediction precision of the motor vehicle caused by the fact that the vehicle has large running randomness in the driving scene of the intersection without the light control and the lane line connection relationship is complex can be better solved, and the accuracy and the stability of the prediction result are integrally improved.

Description

Method and device for predicting future running track of motor vehicle and unmanned equipment

Technical Field

The invention relates to the technical field of unmanned aerial vehicles, in particular to a method and a device for predicting future running tracks of motor vehicles and unmanned equipment.

Background

With the gradual landing of unmanned technologies, autopilot automobiles have gradually emerged in real life. In a traffic scene without a light control intersection, subjective driving will of a vehicle is strong, and the subjective driving will of the vehicle brings great challenges to unmanned safety. The unmanned automobile needs to accurately pre-judge the future running track of other vehicles, so that pre-judgment is made in advance, a reasonable and smooth path is planned, the safety of the running process is effectively ensured, and the intelligence of the unmanned automobile is improved. Therefore, modeling is necessary to model the future track of the motor vehicle without the light control intersection, and the future running track is predicted, so that additional reference is provided for downstream planning and decision making, and the safety and fluency of the unmanned vehicle are improved.

In the prior art, when predicting the track of a motor vehicle running on a road, feature information such as speed, position and the like of the motor vehicle is often obtained, and the prediction is performed based on a rule or model method by combining the position and semantic features of a lane line in a high-precision map. In addition, a bird's eye view can be established by imaging the running motor vehicle, and corresponding analysis can be performed by combining the features extracted from the bird's eye view. In the prediction technology, the high-precision map is used as important information to effectively improve the accuracy of prediction. Under the driving scene of the intersection without the light control, the driving behavior of the social vehicle is mainly judged by the subjective judgment of a driver, and the high-precision map is generally only capable of modeling more standard traffic behavior, so that how to refer to the past driving behavior track of the intersection, and the high-precision map is fully utilized to infer the future driving track of the motor vehicle which is driving, and the method is a problem which needs to be solved urgently.

Patent publication number CN115320623a entitled "vehicle trajectory prediction method, apparatus, mobile device, and storage medium", the disclosed method includes: when a vehicle enters an intersection, determining a first layer of exit candidate points for normal running of the vehicle and a second layer of exit candidate points for non-normal running of the vehicle based on map information of the intersection, and predicting a plurality of predicted tracks from the vehicle to the first layer of exit candidate points and the second layer of exit candidate points; determining the turning amplitude cost of each of a plurality of predicted tracks based on a preset cost function; and selecting the predicted track corresponding to the turning amplitude cost with the lowest value from the turning amplitude cost of each of the plurality of predicted tracks as a track prediction result of the vehicle. It is easy to see that the invention determines the final predicted result by generating a candidate predicted trajectory using the lane lines and performing predicted trajectory selection based on the vehicle history trajectory. However, in this method, the cost calculation is intuitively performed using the history track and the predicted track for the standard running, and the variability of the running of the vehicle cannot be considered. Meanwhile, in the prediction process, the vehicle can be accurately judged only when being positioned at a specific position, and the intelligent can better cope with complex traffic scenes under the intersection without the light control only when the intelligence is further improved.

The invention with the publication number of CN115009275A and the name of "method, system and storage medium for predicting vehicle track in urban-oriented scene" discloses a method comprising: searching a lane sequence in a certain range around a predicted target, and extracting a candidate lane sequence of the predicted target from the lane sequence to obtain a candidate target point set; the method comprises the steps of representing a track sequence and a lane sequence of an intelligent agent around a predicted target by vectors, and extracting spatial interaction characteristics of the predicted target; fusing the space interaction characteristics and the track time sequence characteristics to obtain scene context characteristics; and decoding the candidate point target set and the scene context characteristics to output the future track of the predicted target. The invention uses the lane topological structure to restrict the vehicle driving by combining the high-precision map, but when the lane topological structure is used, the transmission among the characteristics is carried out by taking the lane lines as units, the complex lane line topological relationship under the intersection without the light control cannot be well utilized, and the topological relationship restriction between the history track of the vehicle and the lane is not considered, so that the final prediction precision is limited.

In summary, the future track prediction of the motor vehicle without the light control intersection has a great challenge, and the existing method cannot be well adapted to the complexity of the future track prediction, so that a relatively accurate prediction result is obtained.

Disclosure of Invention

Aiming at the defects in the prior art, the invention provides a method and a device for predicting the future running track of a motor vehicle at a non-light-control intersection and unmanned equipment, and the accuracy and the stability of a predicted result are integrally improved by restraining the historical running track of the motor vehicle at the intersection and combining lane line nodes.

In a first aspect, an embodiment of the present application provides a method for predicting a future travel track of a motor vehicle, where the method includes the following steps:

Step 1, acquiring vehicle information and scene information in a target vehicle position selection range, wherein the vehicle information is each vehicle information aggregate comprising a target vehicle in the target vehicle position selection range, and the scene information is each lane line information aggregate in the target vehicle position selection range;

step 2, extracting historical track information in each vehicle information aggregate, and extracting lane line information in each lane line information aggregate;

Step 3, constructing track node characteristics corresponding to each vehicle information aggregate based on the historical track information, and constructing lane line node characteristics corresponding to each lane line information aggregate based on the lane line information;

Step 4, extracting the track node characteristics to output track node dimension-increasing characteristics, and extracting the lane line node characteristics to output lane line node dimension-increasing characteristics;

Step 5, transmitting the lane line node dimension-increasing characteristics to the track node dimension-increasing characteristics connected with the lane line node dimension-increasing characteristics to perform characteristic fusion, so as to obtain fused track node fusion characteristics;

and 6, obtaining a future running track prediction result of the target vehicle based on the track node fusion characteristics through neural network regression.

In one embodiment, prior to step 1, the method further comprises:

Collecting the type and distance between the vehicle and the vehicle of the obstacle, and selecting a target vehicle needing to be predicted; the target vehicle is a vehicle which is interacted with the own vehicle or is within a preset distance;

And acquiring track coordinates in a period T before the predicted time of each target vehicle.

In one embodiment, the vehicle information is defined as:

S＝{S₀,S₁,…,S_i,…,S_R}

wherein S ₀ is a vehicle information aggregate of the target vehicle, S _i is a vehicle information aggregate of the ith vehicle, and R is the number of other vehicles except the target vehicle in the target vehicle position selection range;

the vehicle information aggregate is defined as:

S_i＝{S_i1…,S_ij,…,S_iT}

wherein S _i1～S_iT represents the vehicle state in the period T before the i-th vehicle prediction time;

The vehicle state is defined as:

In the method, in the process of the invention, The world coordinates of the ith vehicle at the jth time.

In one embodiment, the scene information is defined as:

C＝{C₁,…,C_i,…,C_M}

Wherein C _i is a lane line information aggregate formed by the ith lane line, M is the number of the lane line information aggregates obtained by dividing the acquired high-precision map data according to the ID in the target vehicle position selection range;

The lane line information aggregate is defined as:

C_i＝{C_i1,…,C_ij,…,C_iN}

wherein C _ij is the j coordinate point line segments in the i-th lane line information aggregate, and N is the number of coordinate point line segments in the lane line information aggregate;

the coordinate point line segment is defined as:

In the method, in the process of the invention, Is the world coordinates of the lane line coordinate points.

In one embodiment, in step 3, the historical track information includes historical track point coordinates, and the constructing track node features corresponding to each vehicle information aggregate based on the historical track information includes:

for any vehicle information aggregate, the world coordinate of the previous moment of the adjacent historical track point coordinate is taken as a starting point, the world coordinate of the next moment is taken as an end point, and the adjacent moment historical track point coordinates are subtracted to form a first node characteristic representation with direction information Simultaneously taking the center of the adjacent historical track coordinates as the position information of the first node features, and jointly forming track node features S' corresponding to the vehicle information aggregate by all the first node features;

the lane line information comprises lane line coordinates, and the construction of lane line node characteristics corresponding to each lane line information aggregate based on the lane line information comprises the following steps:

for any lane line information aggregate, the world coordinate of the previous moment of the adjacent lane line coordinate is taken as a starting point, the world coordinate of the next moment is taken as an ending point, and the adjacent lane line coordinates are subtracted to form a second node characteristic representation with direction information And simultaneously taking the coordinate centers of the adjacent lane lines as the position information of the second node features, wherein all the second node features jointly form lane line node features C' corresponding to the lane line information aggregate.

In one embodiment, in step 4, the feature extracting the trace node feature to output a trace node dimension-increasing feature includes:

step 411, mutually splicing the track node features to form three-dimensional input features;

step 412, inputting the three-dimensional input features into a neural network and extracting features;

step 413, extracting and obtaining a track node dimension-increasing feature by using a maximum value pooling method;

The feature extraction of the lane line node features to output lane line node dimension-increasing features comprises:

Step 421, obtaining the front-rear connection relation of the lane lines according to the front-rear connection relation of the lane lines;

step 422, thinning the front-rear connection relationship of the lane lines to obtain the connection relationship between the lane line nodes;

Step 423, searching left and right adjacent lane lines according to the lane line to which the target lane line node belongs, and searching the nearest left and right nodes on the left and right lane lines so as to obtain the topological relation of the lane line nodes;

Step 424, determining a weighted value of the connection relationship between the lane line nodes and a weighted topological relationship of the lane line nodes;

In step 425, the lane line node obtains the last extracted dimension-increasing feature of the lane line node through multiple information transfer.

In one embodiment, in step 425, the lane line node obtains the last extracted lane line node dimension-increasing feature through multiple information transfer, including:

For any node i, the lane line node characteristic C _i 'receives the lane line node characteristic C _j' with a connection relation, and carries out information transfer for a plurality of times after being weighted by the weight of the neural network, and the characteristics of the node i after the information transfer are expressed as follows:

Wherein alpha _j is a weighted value in a topological relation, C _i 'and C _j' are adjacent lane line node characteristics, and W _i、W_j is a neural network weight acting on nodes i and j respectively;

Therefore, the lane line node characteristics are subjected to information transfer for a plurality of times according to the mode, and the finally extracted lane line node dimension-increasing characteristic C) is obtained.

In a second aspect, an embodiment of the present application further provides a device for predicting a future travel track of a motor vehicle, where the device includes:

the information acquisition module is used for acquiring vehicle information and scene information in a target vehicle position selection range, wherein the vehicle information is each vehicle information aggregate comprising a target vehicle in the target vehicle position selection range, and the scene information is each lane line information aggregate in the target vehicle position selection range;

The information extraction module is used for extracting historical track information in each vehicle information aggregate and extracting lane line information in each lane line information aggregate;

the node characteristic construction module is used for constructing track node characteristics corresponding to each vehicle information aggregate based on the historical track information and constructing lane line node characteristics corresponding to each lane line information aggregate based on the lane line information;

The node feature extraction module is used for carrying out feature extraction on the track node features to output track node dimension-increasing features, and carrying out feature extraction on the lane line node features to output lane line node dimension-increasing features;

The feature fusion module is used for transmitting the lane line node dimension-increasing features to track node dimension-increasing features connected with the lane line node dimension-increasing features to perform feature fusion, so as to obtain fused track node fusion features;

And the track prediction module is used for obtaining a future running track prediction result of the target vehicle based on the track node fusion characteristics through neural network regression.

In a third aspect, an embodiment of the present application further provides an unmanned apparatus, where the unmanned apparatus is provided with:

A memory for storing a program;

A processor for executing the program stored in the memory, the processor being for executing part or all of the steps of the future travel track prediction method of the motor vehicle as described above when the program is executed

In a fourth aspect, an embodiment of the present application further provides a computer readable storage medium, where computer executable instructions are stored in the computer readable storage medium; the computer-executable instructions, when executed by the processor, are adapted to carry out some or all of the steps of a method for predicting a future travel path of a motor vehicle as described hereinbefore.

The beneficial technical effects of the invention are as follows:

According to the future running track prediction method of the motor vehicle, vehicle information and scene information in a running scene are converted into similar node representation, and information mining is carried out under the node scale, so that the lane line node dimension-increasing characteristics are effectively transferred into the vehicle history track node dimension-increasing characteristics, and track prediction is completed under the fine granularity of node level; in addition, through restricting the historical running track of the motor vehicle at the intersection and combining with the lane line nodes, the track prediction model can be finely modeled, the problem of insufficient future track prediction precision of the motor vehicle caused by the fact that the random running of the vehicle is large in a driving scene of the intersection without a light control and the connection relation of the lane lines is complex can be better solved, the accuracy and the stability of the prediction result are integrally improved, and the method has very practical use significance for future track prediction of the motor vehicle.

It will be appreciated that the future travel track prediction apparatus, the unmanned device, and the computer-readable storage medium for a motor vehicle according to the second to fourth aspects also have the above advantages, and are not described herein.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings may be obtained according to the structures shown in these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a flowchart of a method for predicting a future travel track of a motor vehicle according to an embodiment of the present invention;

FIG. 2 is a diagram illustrating a vehicle history track and a high-precision map in accordance with an embodiment of the present invention;

FIG. 3 is an example of a vehicle history track and lane line construction node feature in an embodiment of the present invention;

FIG. 4 is a flowchart of extracting trace node feature in an embodiment of the invention;

FIG. 5 is an exemplary diagram of a trace node dimension-increasing feature extraction process in accordance with an embodiment of the present invention;

FIG. 6 is a flow chart of feature extraction of lane line nodes in an embodiment of the invention;

FIG. 7 is an exemplary diagram of a process for constructing a topological relation of lane line nodes in an embodiment of the invention;

FIG. 8 is a diagram illustrating an example of a process for extracting ascending dimension features of lane line nodes in an embodiment of the present invention;

FIG. 9 is a diagram illustrating an example construction of a joint topology of track nodes and lane line nodes in an embodiment of the present invention;

FIG. 10 is a process of receiving lane line node dimension increasing feature information transfer by a track node dimension increasing feature in an embodiment of the present invention;

FIG. 11 is a flowchart of a method for predicting a future travel path of a motor vehicle according to another embodiment of the present invention;

FIG. 12 is a block diagram showing a future travel path predicting apparatus of a motor vehicle according to an embodiment of the present invention;

fig. 13 is a partial block diagram of the unmanned aerial vehicle in the embodiment of the invention.

The achievement of the objects, functional features and advantages of the present invention will be further described with reference to the accompanying drawings, in conjunction with the embodiments.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are only some, but not all embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

It should be noted that all directional indicators (such as up, down, left, right, front, and rear … …) in the embodiments of the present invention are merely used to explain the relative positional relationship, movement, etc. between the components in a particular posture (as shown in the drawings), and if the particular posture is changed, the directional indicator is changed accordingly.

In addition, the technical solutions of the embodiments of the present invention may be combined with each other, but it is necessary to be based on the fact that those skilled in the art can implement the technical solutions, and when the technical solutions are contradictory or cannot be implemented, the combination of the technical solutions should be considered as not existing, and not falling within the scope of protection claimed by the present invention.

In order to cope with the challenges of large random driving of vehicles at the intersection without light control and complex connection relation of lane lines, the embodiment develops a future driving track prediction method of the motor vehicle, which is mainly used for processing future driving track predictions under different scenes of the motor vehicle and assisting unmanned equipment in comprehensive decision-making, and is different from other schemes in which lane lines and tracks are used in line units, the scheme of the embodiment fully excavates effective information for future track prediction by carrying out refined construction on the topological structures of the vehicle tracks and lane line nodes, thereby improving the prediction precision of the future tracks and ensuring that the whole set of prediction flow is well applied to track prediction of the motor vehicle at the intersection without light control.

As shown in fig. 1, a method for predicting a future driving track of a motor vehicle is provided, which is mainly applied to unmanned equipment, and specifically includes the following steps 1 to 6.

And step 1, acquiring vehicle information and scene information in a target vehicle position selection range.

In one embodiment, the vehicle information is each vehicle information aggregate including the target vehicle in the target vehicle position selection range, and the acquiring process includes:

Firstly, selecting a vehicle needing prediction, and defining the vehicle as a target vehicle; then selecting vehicles with the number R in the distance range as surrounding vehicles according to the current coordinates of the target vehicle, and forming vehicle information together with the vehicles; as shown in fig. 2, a trajectory line 101 is a trajectory of a target vehicle, a trajectory line 102 is a trajectory of surrounding vehicles, and a broken line portion is a high-precision map.

Therefore, the vehicle information includes a number r+1 of vehicle information aggregates, that is, the vehicle information may be defined as s= { S ₀,S₁,…,S_i,…,S_R }, where S ₀ is the vehicle information aggregate of the target vehicle and S _i is the vehicle information aggregate of the i-th vehicle;

each vehicle information aggregate is composed of vehicle states in a period T before the current moment of the corresponding vehicle, and each moment of vehicle state comprises world coordinates x and y of the current moment of the information;

Each vehicle information aggregate S _i＝{S_i1,…,S_ij,…,S_iT, where S _i1～S_iT represents a vehicle state within a period T before the i-th vehicle current time;

Vehicle state Wherein,The world coordinates of the ith vehicle at the jth time.

In one embodiment, the scene information is an aggregate of lane line information of each lane line in the target vehicle position selection range, and the acquiring process includes:

Selecting a high-precision map lane line in a range according to the current coordinates of a target vehicle, and dividing the high-precision map data into a plurality of lane line information aggregates according to the ID, wherein the number is M, namely scene information C= { C ₁,…,C_i,…,C_M }, and C _i is the lane line information aggregate of the ith lane line;

each lane line information aggregate comprises a plurality of lane line coordinate points, and for convenience of use, the plurality of lane line coordinate points are uniformly sampled into N, namely each lane line information aggregate comprises a fixed number N of coordinate point line segments, namely the lane line information aggregate C _i＝{C_i1,…,C_ij,…,C_iN }, wherein C _ij is a j-th line coordinate point line segment in the i-th lane line information aggregate;

the coordinate point line segment is as follows: Wherein, Is the world coordinates of the lane line coordinate points.

Meanwhile, for each coordinate point, a semantic attribute value is given according to the semantic attribute of the lane line to which the coordinate point belongs.

In one embodiment, prior to step 1, the method further comprises:

According to the sensing result acquired by the sensing system on the unmanned equipment, determining the type of the vehicle of the obstacle and the distance between the vehicle and the vehicle, and selecting a target vehicle needing to be predicted; and selecting vehicles which can interact with the own vehicle or are close to the own vehicle, and defining the vehicles as target vehicles to be predicted. For each target vehicle to be predicted, it is necessary to acquire the track coordinates within a period T before the predicted time of each of the target vehicles.

Step 2, extracting historical track information in each vehicle information aggregate, and extracting lane line information in each lane line information aggregate; the history track information comprises history track point coordinates, and the lane line information comprises lane line coordinates.

In one embodiment, the historical track information of all vehicles constitutes vehicle information, and the lane line information of all scenes constitutes scene information.

And 3, constructing track node characteristics corresponding to each vehicle information aggregate based on the historical track information, and constructing lane line node characteristics corresponding to each lane line information aggregate based on the lane line information.

As shown in fig. 3, the specific operation procedure of step 3 in this embodiment is:

(1) And constructing a historical track. Node feature construction is performed by using track point coordinates: for any vehicle information aggregate, the world coordinate of the previous moment of the adjacent historical track point coordinate is taken as a starting point, the world coordinate of the next moment is taken as an end point, and the adjacent moment historical track point coordinates are subtracted to form a first node characteristic representation with direction information And simultaneously taking the center of the adjacent historical track coordinates as the position information of the first node features, and forming track node features S' corresponding to the vehicle information aggregate by all the first node features.

(2) And constructing a lane line. Similarly, for any lane line information aggregate, the world coordinates of the adjacent lane line coordinates at the previous time are used as the starting point, the world coordinates of the next time are used as the end point, and the adjacent lane line coordinates are subtracted to form a second node characteristic representation with direction informationAnd simultaneously taking the coordinate centers of the adjacent lane lines as the position information of the second node features, and forming the lane line node feature C' of the motor vehicle by all the second node features.

And 4, extracting the track node characteristics to output track node dimension-increasing characteristics, and extracting the lane line node characteristics to output lane line node dimension-increasing characteristics.

More specifically, as shown in fig. 4, the feature extraction of the trace node feature to output a trace node dimension-increasing feature specifically includes:

Step 411, firstly, mutually splicing the track node features to form a three-dimensional input feature;

step 412, then inputting the three-dimensional input feature into a neural network as shown in fig. 5; the neural network is constituted by VGG16 of 1D in the present embodiment;

step 413, finally, using maximum value pooling to obtain a feature node of one dimension per track; thus, the input features pass through the neural network to form a two-dimensional output feature expressed as a vehicle track node dimension-increasing feature.

As shown in fig. 6 to 7, the feature extraction of the lane line node feature to output a lane line node dimension-increasing feature specifically includes:

Step 421, determining the connection relationship between the lane lines: obtaining the front-rear connection relation of the lane lines according to the front-rear sequence of the lane lines in the high-precision map;

Step 422, determining the connection relationship between the lane line node characteristics: the topological relation among the lane line node characteristics is obtained after the topological relation of the lane lines is thinned; the front-back relation of the node characteristics of the lane lines is used for obtaining the left-right relation of the final node according to the front-back relation of the lane lines;

Step 423, determining a topological relation of the lane line node characteristics: searching left and right adjacent lane lines according to the lane lines to which the target lane line nodes belong, and searching nearest left and right nodes on the left and right lane lines so as to obtain the topological relation of the lane line node characteristics;

Step 424, determining the weight of the connection relationship: the topological connection relation only considers whether the node characteristics have connection relation or not, and different lane line types are different in vehicle driving influence, so that semantic attributes of the nodes are used as relation weights in the embodiment; therefore, the weighted topological relation of the lane line node characteristics can be obtained;

in step 425, the lane line node feature is transmitted multiple times to obtain the last extracted lane line node dimension-increasing feature.

In one embodiment, in step 425, the lane-line node obtains the last extracted lane-line node dimension-increasing feature through multiple information transfer, and further includes:

The information transmission is completed by using a neural network, for any node i, the lane line node characteristic C _i 'of the node i receives the lane line node characteristic C _j' with a connection relation with the node i, and the node i is weighted by the weight of the neural network and then is subjected to multiple information transmission, and the characteristics of the node i after the information transmission are expressed as follows:

thus, as shown in fig. 8, the lane line node feature performs information transfer multiple times in the above manner, and the finally extracted lane line node dimension-increasing feature C) is obtained.

And 5, transmitting the lane line node dimension-increasing characteristics to the connected track node dimension-increasing characteristics to perform characteristic fusion, so as to obtain the fused track node fusion characteristics.

In a specific application process, a connection relation between the track node and the lane line node needs to be constructed, and the connection relation determines whether information transfer is carried out between the track node characteristics and the lane line node characteristics. The construction process uses the position information of the node characteristics in construction. Thus, for each track node, a lane line coordinate point having a connection relationship with the track node can be found to form a joint topological relationship, as shown in fig. 9. Likewise, the present embodiment utilizes semantic attribute values of lane line node features to weight joint topology.

The feature fusion is mainly to enable the track node dimension-increasing feature to receive the adjacent lane line node dimension-increasing feature, so that richer features are provided for subsequent track prediction. The implementation process is as follows: as shown in fig. 10, the lane line node dimension-increasing feature is transferred to the connected track node dimension-increasing feature, so as to complete feature fusion. The feature delivery is accomplished using a neural network. And for any track node i, receiving the ascending dimension characteristic of the lane line node with a connection relation with the track node i, and weighting the ascending dimension characteristic by a learnable neural network weight. Therefore, the track node dimension-increasing characteristics are subjected to information transfer for a plurality of times according to the mode, and the fused track node fusion characteristics are obtained.

And outputting a future running track prediction result, namely carrying out a regression process of the future track by using the obtained track node fusion characteristics. And (3) giving track node fusion characteristics, and obtaining future track prediction results of the track by means of neural network regression for nodes belonging to the same historical track. The number of output nodes of the neural network is twice the number of preset regression coordinates, and each two neural network nodes corresponds to X and Y coordinate values of each regression coordinate; and then, applying the obtained future running track prediction result to a subsequent flow of the self-driving business.

In the prediction process, a plurality of targets to be predicted are formed at each moment. And the targets to be predicted circularly enter the future running track prediction result generation flow to form the future track of each target to be predicted. The generated future track is used as a position to be avoided in subsequent self-vehicle path planning, so that the self-vehicle has prediction capability, and the rationality of the path planning is improved.

According to the embodiment, vehicle information and scene information in a driving scene are converted into node representations in similar forms, information mining is carried out under the node scale, the lane line node dimension-increasing characteristics are effectively transferred into the vehicle history track node dimension-increasing characteristics through the topological connection relation, and track prediction is completed under the fine granularity of the node level. Compared with other schemes for completing prediction at the lane line/track level, the method can better solve the problem of insufficient prediction precision of future tracks of motor vehicles caused by the fact that the random driving of vehicles at the non-light-control intersection is large and the connection relationship of the lane lines is complex by establishing the weighted topological relationship between the tracks of the vehicles and the nodes of the lane lines. In addition, through the establishment of the topological connection relationship, the embodiment can accurately carry out information transfer among nodes, avoid the consumption of computing resources caused by invalid searching, and has advantages in feasibility, identification effect and computing power consumption. As described, the embodiment can be effectively applied to a driving scene without light control, and has very practical use significance for future track prediction of a vehicle.

As shown in fig. 11, the present application further provides another method for predicting a future driving track of a motor vehicle, including the following steps:

1. And a driving data acquisition system. Vehicle information of the motor vehicle in a driving scene is collected, wherein the vehicle information comprises historical driving data and lane line information. The sampling may be performed at regular time intervals, recording the track of the vehicle at several subsequent moments in time and recording for network training.

2. A data preprocessing system. And extracting the historical track of the target vehicle and extracting lane line information in the high-definition map. And normalizing the lane line information to the position of the target vehicle at the latest moment.

3. Training the system. In the training process, training data is firstly divided into a training set and a verification set. The training data is used as data to be input after being constructed through node characteristics. In the process, the four steps of track node feature extraction, lane line node feature extraction, feature fusion and track output jointly form an end-to-end network structure.

4. An inference system. After the trained network model is obtained, deployment of the network model is applied to the reasoning speed of the corresponding platform acceleration model so as to meet the condition of real-time output track prediction in a driving scene. Therefore, vehicle information is collected on a driving platform in real time, corresponding to high-precision map information, and node characteristics are built by using the vehicle information and the high-precision map information and sent to an inference model to obtain a regression prediction track.

As shown in fig. 12, corresponding to the method for predicting a future driving track of a motor vehicle according to the above embodiment, the embodiment of the present application further provides a device for predicting a future driving track of a motor vehicle, where the device includes an information acquisition module 201, an information extraction module 202, a node feature construction module 203, a node feature extraction module 204, a feature fusion module 205, and a track prediction module 206, where

An information obtaining module 201, configured to obtain vehicle information and scene information in a target vehicle position selection range, where the vehicle information is each vehicle information aggregate including a target vehicle in the target vehicle position selection range, and the scene information is each lane line information aggregate in the target vehicle position selection range;

the information extraction module 202 is configured to extract historical track information in each vehicle information aggregate, and extract lane line information in each lane line information aggregate;

The node feature construction module 203 is configured to construct track node features corresponding to each vehicle information aggregate based on the historical track information, and construct lane line node features corresponding to each lane line information aggregate based on the lane line information;

the node feature extraction module 204 is configured to perform feature extraction on the track node feature, output a track node dimension-increasing feature, and perform feature extraction on the lane line node feature, and output a lane line node dimension-increasing feature;

The feature fusion module 205 is configured to transfer the lane line node dimension-increasing feature to a track node dimension-increasing feature connected with the lane line node dimension-increasing feature to perform feature fusion, so as to obtain a fused track node fusion feature;

The track prediction module 206 is configured to obtain a future driving track prediction result of the target vehicle based on the track node fusion feature through neural network regression.

It should be noted that, because the content of information interaction and execution process between the above devices/modules is based on the same concept as the method embodiment of the present application, specific functions and technical effects thereof may be referred to in the method embodiment section, and will not be described herein.

As shown in fig. 13, an unmanned apparatus disclosed in this embodiment may be an automatic or semi-automatic unmanned apparatus such as an unmanned vehicle, a mobile robot, or the like, on which a transmitter 302, a receiver 301, a memory 304, and a processor 303 are disposed. Wherein the transmitter 302 is configured to transmit instructions and data, the receiver 301 is configured to receive instructions and data, the memory 304 is configured to store computer-executable instructions, and the processor 303 is configured to execute the computer-executable instructions stored in the memory 304, so as to implement some or all of the steps performed by the future travel track prediction method of the motor vehicle. The specific implementation process is the same as the method for predicting the future running track of the motor vehicle.

It should be noted that the memory 304 may be separate or integrated with the processor 303. When the memory is provided separately, the unmanned device apparatus further comprises a bus for connecting the memory 304 and the processor 303.

The embodiment also discloses a computer readable storage medium, wherein computer executable instructions are stored in the computer readable storage medium, and when the processor executes the computer executable instructions, part or all of the steps executed by the method for predicting the future running track of the motor vehicle are realized.

The foregoing description is only of the preferred embodiments of the present invention and is not intended to limit the scope of the invention, and all equivalent structural changes made by the description of the present invention and the accompanying drawings or direct/indirect application in other related technical fields are included in the scope of the invention.

Claims

1. A method for predicting a future travel path of a motor vehicle, the method comprising the steps of:

Step 4, outputting track node dimension-increasing characteristics after extracting the characteristics of the track node, and outputting lane line node dimension-increasing characteristics after extracting the characteristics of the lane line node;

2. The method for predicting future travel path of a motor vehicle of claim 1, wherein prior to step 1, the method further comprises:

3. The method for predicting future travel path of motor vehicle according to claim 2, wherein the vehicle information is defined as:

S＝{S₀,S₁,…,S_i,…,S_R}

the vehicle information aggregate is defined as:

S_i＝{S_i1…,S_ij,…,S_iT}

The vehicle state is defined as:

4. The method for predicting future travel path of motor vehicle according to claim 1, wherein the scene information is defined as:

C＝{C₁,…,C_i,…,C_M}

The lane line information aggregate is defined as:

C_i＝{C_i1,…,C_ij,…,C_iN}

the coordinate point line segment is defined as:

5. The method according to claim 1, wherein in step 3, the history track information includes history track point coordinates, and the constructing track node features corresponding to each vehicle information aggregate based on the history track information includes:

for any vehicle information aggregate, the world coordinate of the previous moment of the adjacent historical track point coordinate is taken as a starting point, the world coordinate of the next moment is taken as an end point, and the adjacent moment historical track point coordinates are subtracted to form a first node characteristic representation with direction information Simultaneously taking the center of the adjacent historical track coordinates as the position information of the first node features, and forming track node features S' corresponding to the vehicle information aggregate by all the first node features together;

6. The method for predicting a future travel path of a motor vehicle according to claim 5, wherein in step 4, the feature extracting the path node feature to output a path node dimension-increasing feature comprises:

step 412, inputting the three-dimensional input feature into a neural network;

step 421, obtaining the front-rear connection relationship of the lane lines according to the front-rear sequence of the lane lines;

step 422, thinning the front-rear connection relationship of the lane lines to obtain the connection relationship between the node characteristics of the lane lines;

step 423, searching left and right adjacent lane lines according to the lane lines to which the node characteristics of the target lane line belong, and searching the nearest left and right nodes on the left and right lane lines so as to obtain the topological relation of the node characteristics of the lane lines;

Step 424, determining a weighted value of the connection relationship between the lane line node features and a weighted topological relationship of the lane line node features;

7. The method according to claim 6, wherein in step 425, the lane-line node obtains the last extracted lane-line node dimension-increasing feature through multiple information transfer, and the method comprises:

For any node i, the lane line node characteristic C _i 'receives the lane line node characteristic C _j' with a connection relation, and performs multiple information transfer after weighting by the weight of the neural network; the characteristics of the node i after information transmission are expressed as follows:

8. A future travel path prediction apparatus for a motor vehicle, the apparatus comprising:

9. An unmanned aerial vehicle, characterized in that, be equipped with on the unmanned aerial vehicle:

A memory for storing a program;

a processor for executing the program stored in the memory, the processor being configured to execute part or all of the steps of the future travel path prediction method of an automobile as claimed in any one of claims 1 to 7 when the program is executed.

10. A computer-readable storage medium having stored therein computer-executable instructions; the computer-executable instructions, when executed by a processor, for performing part or all of the steps of the method for predicting a future travel path of a motor vehicle as claimed in any one of claims 1 to 7.