CN114283396A

CN114283396A - Method, apparatus, and computer-readable storage medium for autonomous driving

Info

Publication number: CN114283396A
Application number: CN202010987392.0A
Authority: CN
Inventors: 李飞; 范时伟; 李向旭
Original assignee: Huawei Technologies Co Ltd
Current assignee: Huawei Technologies Co Ltd
Priority date: 2020-09-18
Filing date: 2020-09-18
Publication date: 2022-04-05

Abstract

Embodiments of the present disclosure relate to methods, apparatuses, and computer-readable storage media for autonomous driving. The method comprises the following steps: obtaining a plurality of sets of exit features for a plurality of candidate exits associated with a vehicle, each candidate exit being a portion of a road associated with a drivable route of the vehicle; determining a plurality of sets of vehicle characteristics of the vehicle relative to the plurality of candidate exits, each set of vehicle characteristics including at least a state of motion of the vehicle relative to a respective candidate exit of the plurality of candidate exits; and determining an exit matching the travel route of the vehicle from the plurality of candidate exits based on the plurality of sets of exit features and the plurality of sets of vehicle features. The method abstracts the target high-order intention prediction into a target and exit matching problem, can be adapted to various intersection scenes, and has good generalization performance.

Description

Method, apparatus, and computer-readable storage medium for autonomous driving

Technical Field

Embodiments of the present disclosure relate generally to autonomous driving, and more particularly, to a method, apparatus, electronic device, autonomous vehicle, and computer-readable storage medium for autonomous driving.

Background

In autonomous driving, vehicle intention prediction refers to prediction of a behavior intention of a target vehicle according to a target state at a current time and a historical time, for example, a vehicle intention may include lane keeping, lane changing right, and the like in an on-road scene, and a vehicle intention may include straight running, left turning, right turning, turning around, and the like in an intersection scene. The method can predict the intention of the vehicle accurately and reliably in real time, help the vehicle to predict the traffic condition ahead, establish the traffic situation around the vehicle, help to judge the importance of other vehicle targets around, screen interactive key targets, facilitate the vehicle to plan the path in advance and safely pass through complex intersection scenes.

For the urban traffic intersection scene, as the road structure is complex and changeable and no clear entity lane line constraint exists, the behavior difference of the vehicles is more obvious and more difficult to predict; due to the fact that different intersection structures have large differences, simple semantic level intents such as straight driving, left turning, right turning and turning around are difficult to accurately describe the behavior intents of the target vehicle.

Disclosure of Invention

Embodiments of the present disclosure may provide an autonomous driving solution that at least partially solves the technical problem in the conventional solution.

In a first aspect of the disclosure, a method for autonomous driving is provided. The method comprises the following steps: obtaining a plurality of sets of exit features for a plurality of candidate exits associated with a vehicle, each candidate exit being a portion of a road associated with a drivable route of the vehicle; determining a plurality of sets of vehicle characteristics of the vehicle relative to the plurality of candidate exits, each set of vehicle characteristics including at least a state of motion of the vehicle relative to a respective candidate exit of the plurality of candidate exits; and determining an exit matching the travel route of the vehicle from the plurality of candidate exits based on the plurality of sets of exit features and the plurality of sets of vehicle features.

In the conventional scheme, the road area is divided by taking a vehicle as a reference, so that it is difficult to adapt to a large number of exits in an intersection scene and to adapt to various intersection scenes. The method abstracts the target high-order intention prediction into a target and exit matching problem, and the intersection does not need to be divided into sector areas, so that the method can be adapted to various intersection scenes and has good generalization performance. In addition, when the target intention is not obvious, a plurality of intentions are output, so that the multi-modal problem is solved.

In some embodiments, each set of exit features includes a state of motion of the vehicle relative to the candidate exits in a cartesian coordinate system.

In some embodiments, the exit characteristics of the respective candidate exits include at least one of: a size, a travelable direction, and a boundary of the respective candidate exit.

In some embodiments, each set of vehicle characteristics includes at least one of: the coordinates of the vehicle, the distance between the vehicle and the origin of the Cartesian coordinate system, the orientation of the vehicle, the rate of change of the coordinates of the vehicle, the rate of change of the distance between the vehicle and the origin, and the rate of change of the orientation of the vehicle. The coordinate system is established with the exit as a reference, and the characteristics of the vehicle can be conveniently expressed.

In some embodiments, each set of vehicle features further comprises: and the motion state of the vehicle under a curve coordinate system taking a reference lane line as a coordinate axis, wherein the lane line closest to the vehicle is selected from the lane lines connected with the corresponding candidate outlets as the reference lane line. In the case of considering the lane line, the moving state of the vehicle with respect to the lane line can be conveniently represented according to a curvilinear coordinate system with the exit as a reference.

In some embodiments, the motion state of the vehicle in the curvilinear coordinate system comprises at least one of: the coordinates of the vehicle, an angular difference of an orientation of the vehicle from a reference traveling direction, the reference traveling direction being a traveling direction of a lane line at a projection point of the vehicle to the reference lane line, a rate of change in the coordinates of the vehicle, and a rate of change in the angular difference.

In some embodiments, the method further comprises: determining, by a first multi-layer perceptron, a feature representation for each of the plurality of sets of outlet features; determining, by a second multi-layered perceptron and recurrent neural network, a feature representation for each of the plurality of sets of vehicle features; and wherein determining the outlet further comprises: determining the egress based on the feature representation for each of the sets of egress features and the feature representation for each of the sets of vehicle features. By using the multilayer perceptron and the recurrent neural network, the convolutional neural network can be avoided, thereby saving the system computing resources.

In some embodiments, determining the outlet comprises: determining, by a third multi-tier perceptron, a matching coefficient between each of the plurality of sets of egress features and a respective vehicle feature of the plurality of sets of vehicle features; and determining the outlet based on the matching coefficient. By using a multi-layer perceptron and a recurrent neural network, the use of a convolutional neural network can be avoided, thereby improving the computing performance of the system.

In some embodiments, determining the exit based on the matching coefficients comprises: obtaining a historical matching coefficient relating to a historical location of the vehicle, the historical matching coefficient being a matching coefficient between each of the plurality of sets of exit features and a respective vehicle feature of the plurality of sets of vehicle features when the vehicle is at the historical location; calculating an updated matching coefficient based on the historical matching coefficient and the determined matching coefficient; and determining the exit based on the updated matching coefficient. By calculating the historical fusion probability, the influence of the state change of the vehicle can be further considered, and the prediction accuracy is improved.

In some embodiments, the plurality of candidate exits are exits of the intersection region. The technical scheme disclosed by the invention is particularly suitable for complex and changeable intersection scenes.

In a second aspect of the present disclosure, an apparatus for autonomous driving is provided. The device comprises: a first obtaining module configured to obtain a plurality of sets of exit features of a plurality of candidate exits associated with a vehicle, each candidate exit being a portion of a road associated with a drivable route of the vehicle; a first determination module configured to determine a plurality of sets of vehicle characteristics of the vehicle relative to the plurality of candidate exits, each set of vehicle characteristics including at least a state of motion of the vehicle relative to a respective candidate exit of the plurality of candidate exits; and a second determination module configured to determine an exit matching the travel route of the vehicle from the plurality of candidate exits based on the plurality of sets of exit features and the plurality of sets of vehicle features.

In a third aspect of the disclosure, an electronic device is provided. The electronic device includes: one or more processors; and memory for storing one or more programs that, when executed by the one or more processors, cause the electronic device to implement the method according to the first aspect.

In a fourth aspect of the present disclosure, an autonomous vehicle is provided. The autonomous vehicle comprises an electronic device according to the second aspect of the present disclosure.

In a fifth aspect of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored, which program, when executed by a processor, implements the method according to the first aspect of the present disclosure.

This summary is provided to introduce a selection of concepts in a simplified form that are further described below in the detailed description. This summary is not intended to identify key features or essential features of the disclosure, nor is it intended to be used to limit the scope of the disclosure.

Drawings

The above and other objects, features and advantages of the present disclosure will become more apparent by describing in more detail exemplary embodiments thereof with reference to the attached drawings, in which like reference numerals generally represent like parts throughout.

FIG. 1 illustrates a block diagram of an environment in which embodiments of the present disclosure can be implemented;

FIG. 2 illustrates a block diagram of an intent prediction system in accordance with some embodiments of the present disclosure;

FIG. 3 illustrates a block diagram of an intent prediction system in accordance with some embodiments of the present disclosure;

FIG. 4 illustrates a schematic diagram of a curvilinear coordinate system, in accordance with some embodiments of the present disclosure;

FIG. 5 illustrates a block diagram of an intent prediction system in accordance with some embodiments of the present disclosure;

FIG. 6 illustrates a flow chart of a method for autonomous driving according to some embodiments of the present disclosure;

FIG. 7 illustrates a block diagram of an apparatus for autonomous driving, according to some embodiments of the present disclosure; and

FIG. 8 illustrates a block diagram of an electronic device capable of implementing embodiments of the present disclosure.

Detailed Description

Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While certain embodiments of the present disclosure are shown in the drawings, it is to be understood that the present disclosure may be embodied in various forms and should not be construed as limited to the embodiments set forth herein, but rather are provided for a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the disclosure are for illustration purposes only and are not intended to limit the scope of the disclosure.

The term "include" and variations thereof as used herein is meant to be inclusive in an open-ended manner, i.e., "including but not limited to". Unless specifically stated otherwise, the term "or" means "and/or". The term "based on" means "based at least in part on". The terms "one example embodiment" and "one embodiment" mean "at least one example embodiment". The term "another embodiment" means "at least one additional embodiment". The terms "first," "second," and the like may refer to different or the same object. Other explicit and implicit definitions are also possible below.

As used herein, "vehicle" refers to any type of implement capable of carrying people and/or things and being mobile, examples of which include, but are not limited to, cars, trucks, buses, electric vehicles, and the like. As used herein, "autonomous driving" means having a certain level of autonomous driving capability, including assisted driving, smart driving, unmanned driving, and the like. The following will be described and illustrated in connection with smart driving, however it should be understood that embodiments of the present disclosure may also be applied to any other suitable driving scenarios and driving functions, such as assisted driving, adaptive cruise, and the like.

FIG. 1 illustrates a block diagram of an example environment 100 in which embodiments of the present disclosure can be implemented. It should be understood that the example environment 100 is described for exemplary purposes only and is not intended to suggest any limitation as to the scope of the disclosure. The present disclosure may be implemented in different environments. The example environment 100 may be applied in smart driving, autonomous driving, assisted driving, unmanned driving, etc. scenarios. The example environment 100 illustrates an intersection scenario in which an intersection is an intersection region of two or more roads. In an intersection scenario, multiple exits are typically involved, especially in view of lane lines, even virtual lane lines. For example, one road at an intersection may have a plurality of virtual lane lines thereon, corresponding to a plurality of exits.

As shown in fig. 1, a vehicle 101 (also referred to as "own vehicle") travels on a road, and a vehicle 102 (also referred to as "target vehicle" or "other vehicle") also travels on the road. The own vehicle 101 may be an autonomous vehicle, while the other vehicle 102 may be an autonomous vehicle or a non-autonomous vehicle. The own vehicle 101 and the other vehicle 102 may travel on the same lane line or different lane lines. During the driving of the own vehicle 101, the own vehicle 101 can determine the intention of his vehicle 102 at the intersection, for example, which exit at the intersection the his vehicle 102 selects. For example, the exit may be

exit

103, 104, 105, or 106, which are all located on the travelable route of his car 102, also referred to as candidate exits of his car 102. The exit may not only indicate which road but may also include which lane line or other map elements. The own vehicle 101 can make a better plan of the route according to the intention of the other vehicle 102. For example, the host vehicle 101 may be scheduled to exit the exit 103. If it is determined that the intention of the other vehicle 102 is the exit 103, the own vehicle 101 may not need to travel at a reduced speed. However, if it is determined that the intent of his car 102 is the exit 104, then the own car 101 may need to be slowed down to prevent snagging with his car 102.

To determine the intent of his vehicle 102, the own vehicle 101 may be equipped with a sensing system to sense the environment around the own vehicle 101 and output environmental information around the own vehicle 101. For example, the perception system may identify lane lines, passable areas of roads, vehicles, pedestrians, traffic signs, etc. around the host vehicle 101 in real time. The sensing system may include various sensors, such as cameras, lidar, millimeter wave radar, ultrasonic radar, and the like.

The host vehicle 101 may combine the environmental information acquired by the sensing system with a map, and map each target in the environmental information onto the map. For example, the map may be a conventional electronic map, or may be a high-precision map (also referred to as a lane-level map). The high-precision map is accurate to centimeter level, and an auxiliary line (namely a virtual lane) can be added in the area without the real lane line at the intersection so as to be convenient for restricting the driving of the automatic driving vehicle. High-precision maps can be used as auxiliary tools or prior information in a high-level automatic driving system to help an automatic driving vehicle to better predict the behavior intention of other target vehicles.

The own vehicle 101 can determine information about the target, such as the position, moving speed, moving direction, and the like of the target, from the sensing system and the map. The own vehicle 101 can determine the intention of the target, for example, which exit the target is to select, from the information about the target. The intention prediction can help the self-vehicle 101 predict the intentions of other surrounding vehicles (such as other vehicles 102), and is beneficial to planning control to take emergency safety measures to guarantee the safety of the vehicle and avoid collision in dangerous scenes.

The actual road environment, particularly the intersection scenario, is often complex and may be more complex than the road environment shown in fig. 1. Conventional common modes such as straight running, left turning, right turning and turning around are difficult to accurately describe the movement intention of the target vehicle in such a complex scene. Embodiments of the present disclosure may solve the problem that the movement intention is difficult to predict in such a complex scene, and will be described in detail below with reference to fig. 2 to 5.

Fig. 2 illustrates a block diagram of an intent prediction system 200 in accordance with some embodiments of the present disclosure. The intent prediction system 200 as illustrated in FIG. 2 will be described below in conjunction with the example environment of FIG. 1. For example, the intent prediction system 200 may be implemented on the host vehicle 101, perform data collection via sensors at the host vehicle 101, and perform calculations based on the collected data to achieve intent prediction. In addition, the intention prediction system 200 may also collect data through sensors at the host vehicle 101 and transmit the collected data to the cloud. The cloud calculates from the data received from the host vehicle 101 to implement intent prediction, and passes the intent prediction results back to the host vehicle 101.

As shown in FIG. 2, the intent prediction system 200 includes a feature extractor 202 that may be used to extract features related to exits and vehicles. For convenience, features related to the exit will be referred to as exit features and features related to the vehicle will be referred to as vehicle features hereinafter.

Still described in conjunction with the example environment 100 of FIG. 1, a Cartesian coordinate system corresponding to the candidate exits 105 is shown. In this example cartesian coordinate system, O is used as an origin, the travelable direction is used as the x-axis, and the vertical direction of the x-axis is used as the y-axis direction. It should be appreciated that this cartesian coordinate system is provided as an example only, and the origin may also be provided at any other suitable location of the candidate exit 105, for example, translating O a number of distances forward along the direction of travelable. The feature extractor 202 may extract a set of exit features for the candidate exits 105 in the cartesian coordinate system, which may include the size, travelable direction, and/or boundaries of the candidate exits 105, and the like. For example, the size of the candidate exit 105 may include the width of the candidate exit 105, etc., the travelable direction of the candidate exit 105 is the x-axis direction, and the boundary of the candidate exit 105 may include the boundary position or coordinates on both sides of the candidate exit, etc. For the other candidate exits 103, 104, and 106, similar cartesian coordinate systems may be defined, respectively, to extract corresponding candidate exit features by the feature extractor 202.

In some embodiments, the vehicle features may also be represented with reference to candidate exits, e.g., other vehicles 102 may have one set of vehicle features with respect to one candidate exit and multiple sets of vehicle features with respect to multiple candidate exits. Each set of vehicle characteristics may include at least a state of motion of the other vehicle 102 relative to the respective candidate exit. For example, for a candidate exit 105, the vehicle characteristics of his vehicle 102 may include at least a state of motion of his vehicle 102 relative to the candidate exit 105, e.g., a direction of travel of his vehicle 102 relative to a direction of travelable of the candidate exit 105. The motion state of the other vehicle 102 may be represented in a cartesian coordinate system with the origin at the respective candidate exit, thereby indicating the motion state of the other vehicle 102 with respect to the respective candidate exit. For example, each set of vehicle characteristics of the other vehicle 102 may include coordinates of the other vehicle 102, a distance between the other vehicle 102 and an origin, an orientation of the other vehicle 102, a rate of change of coordinates of the other vehicle 102, a rate of change of distance between the other vehicle 102 and the origin, and/or a rate of change of orientation of the other vehicle 102. In conjunction with the cartesian coordinate system of the candidate exit 105 of fig. 1, the set of vehicle characteristics of the other vehicle 102 relative to the candidate exit 105 may include coordinates of the other vehicle 102, a distance between the other vehicle 102 and the origin O, an orientation of the other vehicle 102, a rate of change of the coordinates of the other vehicle 102, a rate of change of the distance between the other vehicle 102 and the origin O, and/or a rate of change of the orientation of the other vehicle 102.

In some embodiments, each set of vehicle features may further include a motion state of the other vehicle 102 in a curved coordinate system with an origin at an end point of the reference lane line and the reference lane line as one coordinate axis, wherein a lane line closest to the vehicle is selected as the reference lane line from the lane lines connected to the corresponding candidate exits. The above-mentioned curvilinear coordinate system is also called Frenet coordinate system, wherein the direction along the reference line is called longitudinal direction S and the normal direction of the reference lane line is transverse direction D. Based on the position of the reference lane line, the longitudinal distance and the lateral distance can be used to describe an arbitrary position, and it is convenient to calculate the longitudinal and lateral velocities, acceleration, jerk, and the like. For example, the reference lane line may include a physical lane line and may also include a virtual lane line.

In some embodiments, the motion state of the other vehicle 102 in the above-described curved coordinate system may include the coordinates of the other vehicle 102, an angular difference between the orientation of the other vehicle 102 and a reference traveling direction, which is the traveling direction of the lane line at the projection point of the other vehicle 102 to the reference lane line, a rate of change in the coordinates of the other vehicle 102, and a rate of change in the angular difference.

By determining the vehicle characteristics with the exits as a reference, the intent prediction system 200 can predict from which exit of the intersection area the target vehicle will exit, or the probability of exiting from each exit. In this way, the high-order intent prediction problem of the target vehicle can be converted into a matching problem of the vehicle state with map elements (i.e., exits).

The feature extractor 202, after extracting the vehicle features and the egress features, may provide the vehicle features and the egress features to a vehicle feature learning network 204 and an egress feature learning network 206, respectively. The vehicle feature learning network 204 may obtain the vehicle features from the feature extractor 202 to determine a feature representation of the vehicle features, e.g., the vehicle features represented in a vector form. For example, the vehicle feature learning network 204 may be implemented by a neural network. Since the motion state of the vehicle is dynamic, the vehicle feature learning Network 204 may be implemented by a Recurrent Neural Network (RNN) in combination with a Multi Layer Perceptron (MLP) to track the state of the vehicle. The RNN has internal memory to record its hidden state and performs the same function for each input data, and the input at the current time includes the output at the previous time. After the output is computed, the output is copied and sent back to the RNN. In other words, the RNN takes into account the input at the current time and the output learned from the input at the previous time, and updates its hidden state at each time. Examples of RNNs include Gated Round Units (GRUs), Long Short-Term Memory units (LSTM), and the like. The MLP is a fully-connected neural network and can realize the function of a nonlinear function. The MLP includes an input layer, an output layer, and at least one hidden layer connecting the input layer and the output layer.

The exit feature learning network 206 may obtain the exit features from the feature extractor 202 to determine a feature representation of the exit features, e.g., vehicle features in a vector representation. For example, the exit feature learning network 206 may be implemented by a neural network, such as a multi-layer perceptron (MLP).

The vehicle feature learning network 204 and the egress feature learning network 206 may provide feature representations of the vehicle features and the egress features, respectively, to the metric learning network 208. Metric learning (Metric learning), also known as similarity learning, is used to learn or construct metrics or similarities between different objects. Metric learning network 208 may obtain a feature representation of the vehicle features from vehicle feature learning network 204 and a feature representation of the exit features from exit feature learning network 206, and determine a degree or probability of match between his vehicle 102 and the exit based on the vehicle features and the feature representations of the exit features. The metric learning network 208 may be implemented using a variety of different neural networks, such as a multi-layer perceptron MLP.

The metric learning network 208 may provide the degree of match or the probability of match to the intent tracker 210. The intent tracker 210 can track the intent of his car 102, e.g., the exit of an intersection, based on the degree or probability of match at different times. The vehicle intent may be tracked temporally in a sliding window fashion. For example, a fusion probability (e.g., a weighted average) for the current time may be calculated based on the state probabilities within a window length. In some embodiments, the intent tracker 210 may output the one of the exits with the highest probability as the final output result. In other embodiments, the intent tracker 210 may output a plurality of exits with a higher probability as the final output. For example, when the probabilities of a plurality of exits having higher probabilities are closer, it may be difficult to determine which exit is the intention of the target vehicle. Multiple output results may be output at this time to prevent erroneous determination.

In the intention prediction system 200, features of the target vehicle with respect to map elements (i.e., exits) are extracted, the features including two parts: the first part is the vehicle motion state (namely, the vehicle position, the orientation and the position variation and the orientation variation of the adjacent frames) in a Cartesian coordinate system relative to the exit, and the vehicle motion state in a curvilinear coordinate system relative to the exit lane line; the second part is the exit feature, including the width of each exit, the direction of travel, and the two side boundary points of the exit. And respectively acquiring the expression vectors of the target vehicle and the exit through a vehicle characteristic learning network and an exit characteristic learning network, and then calculating the matching probability between the vehicle expression vector and the vehicle expression vector by utilizing a metric learning network. The matching probabilities of different outlets can be normalized, and the target intention probability is tracked in a sliding window weighted average mode.

The intention prediction system 200 abstracts the target high-order intention prediction into a target and exit matching problem, and the method can be adapted to various intersection scenes and has good generalization performance. When the intention of the target is not obvious, for example, the probability corresponding to the intention with the highest probability is still low and is relatively close to the probability corresponding to the intention with the second highest probability, in this case, a plurality of intentions can be output, thereby solving the multi-modal problem. In addition, the intention prediction system does not need to use CNN to extract features from the rasterized image, and only needs simple MLP and RNN network, so that the complexity of the model can be reduced, the calculation efficiency, the real-time performance and the timeliness can be improved, and the intention prediction system can react in time and is more adaptive to the scene of automatic driving.

Fig. 3 illustrates a schematic diagram of an intent prediction system 300, in accordance with some embodiments of the present disclosure. As shown in FIG. 3, the vehicle feature extracted at time t-1 is represented as a vehicle feature vector F_i，t-1And S_i，t-1And the vehicle feature extracted at time t is represented as a vehicle feature vector F_i，tAnd S_i，tWhere i denotes the number of exits, a vehicle feature vector F_iRepresenting the characteristics of the vehicle relative to the exit i, and a vehicle characteristic vector S_iWhich characterizes the lane line of the vehicle relative to exit i. In addition, the extracted exit features are represented as an exit feature vector E_i. In the context of the present disclosure, the subscripts indicating time may be omitted for convenience, without confusion.

In some embodiments, the vehicle feature vector F_iRepresenting the state of motion of the vehicle relative to the exit i in the exit coordinate system. For example, the vehicle feature vector F may be extracted in a cartesian coordinate system or a reference coordinate system with the exit i as the origin and the traveling direction of the exit i as one coordinate axis direction (e.g., x axis)_i. For example, the vehicle feature vector F_iMay include the position P of the vehicle in the reference coordinate system_iDistance d between vehicle and origin_iDirection of vehicle θ_iAnd position P of the vehicle_iRate of change of (d), distance d between vehicle and origin_iRate of change of (2), orientation of vehicle theta_iThe rate of change of (c), etc. The change rate of these feature amounts can be represented by the change amount of the feature amount from the previous time. For example, the position P of the vehicle_iCan be determined from the position P of the vehicle_iAnd the amount of change at the previous time. Although the vehicle feature vector F is given here_iA plurality of features of (a), it being understood that the vehicle feature vector F_iMore or fewer features may be included.

In some embodiments, the vehicle feature vector S_iRepresenting the state of motion of the vehicle relative to the exit i in the reference lane line coordinate system. For example, a lane line closest to the vehicle may be selected from the lane lines connected to the exit i as a reference lane line, thereby establishing a curved coordinate system with the end point of the reference lane line as the origin and the reference lane line as one coordinate axis (e.g., vertical axis), as shown in fig. 4. The reference lane line includes a lane line end 401, a lane line start 407, and lane line sampling points 402-. The point P represents the position of the target vehicle, and the projected point on the reference lane line is R, i.e., the projected point corresponding to the minimum projected distance from the target point P to the lane line. The coordinates of the vehicle, the angular difference of the orientation of the vehicle from a reference traveling direction, which is the traveling direction of the lane line at the projected point of the vehicle to the reference lane line, the rate of change of the coordinates of the vehicle, and the rate of change of the angular difference are calculated under the curved coordinate system. The change rate of these feature amounts can be represented by the change amount of the feature amount from the previous time. For example, vehiclesThe change rate of the coordinates of (a) may be represented by the amount of change of the coordinates of the vehicle from the previous time. Although the vehicle feature vector S is given here_iA plurality of features of (a), it being understood that the vehicle feature vector S_iMore or fewer features may be included.

As shown in fig. 4, taking the lane line as an example of the vertical axis, the position of the target point P in the curved coordinate system can be represented by equations (1) and (2):

x＝|PR| (2)

where x denotes the x coordinate of the target point P, y denotes the y coordinate of the target point P, L_jIs the length of the lane line segment, j is 0, 1, 2, 3, 4; l is₅The distance between the end point of the previous line segment and the target projection point, namely the distance between the lane line 405 and the R; | is the vector length. It should be understood that the "lane lines" described herein may include both physical lane lines and virtual lane lines. In some embodiments, the "lane lines" described herein may also include only physical lane lines if the map does not support virtual lane lines.

As shown in FIG. 3, at time t-1, the vehicle feature vector F_i，t-1Input to a multi-level perceptron (MLP) 302. The multi-layered perceptron is a fully-connected neural network that includes an input layer, an output layer, and at least one hidden layer connecting the input layer and the output layer. For example, the MLP 302 may have a hidden layer to improve computational efficiency. The output of the multi-layered perceptron 302 is provided to a Recurrent Neural Network (RNN) 303. In addition, at time t-1, the vehicle feature vector S_i，t-1Input to a multi-layered perceptron (MLP)304, the output of the multi-layered perceptron 304 is provided to a recurrent neural network 305. For example, the MLP 304 may have a hidden layer to improve computational efficiency. For example,

RNNs

303 and 305 may be implemented by recurrent neural network elements such as Gated Recurrent Units (GRUs), long and short term memory units (LSTM), and the like.

The recurrent neural network has internal memory that records its hidden state and performs the same function for each input data, and the input at the current time depends on the output at the previous time. After the output is computed, the output is replicated and sent back to the recurrent neural network. In other words, the recurrent neural network considers the input at the present time and the output learned from the input at the previous time, and updates its hidden state at each time. It should be understood that fig. 3 shows the recurrent neural network in expanded form, i.e., RNN303 and RNN313 are the same RNN at different times, and RNN 305 and RNN 315 are the same RNN at different times. In other words, only two RNNs are included in the feature learning network layer of fig. 3. Accordingly,

MLPs

302 and 312 represent the same MLP,

MLPs

304 and 314 represent the same MLP, and matching

functions

306 and 316 also represent the same matching function.

At time t-1, the RNN303 receives an input from the MLP 302 at time t-1 and takes the output of the RNN303 at time t-2 as another input at time t-1. For clarity, FIG. 3 does not show the RNN303 receiving as input the output of the RNN at time t-2. By calculation, RNN303 may update its hidden state and may provide an output at time t-1 as an input to the RNN at time t + 1. For example, at time t-1, the RNN303 hidden state is represented by vector g_ei，t-1And the output of RNN303 at time t-1 may be provided as an input to RNN at time t, i.e., RNN 313.

At time t-1, the RNN 305 receives an input from the MLP 304 at time t-1 and takes the output of the RNN 305 at time t-2 as another input at time t-1. For clarity, FIG. 3 does not show RNN 305 receiving as input the output of the RNN at time t-2. By calculation, the RNN 305 may update its hidden state and may provide an output at time t. For example, at time t-1, the RNN 305 hidden state is represented by vector h_li，t-1And the output of RNN 305 at time t-1 may be provided as an input to the RNN at time t, namely RNN 315. At time t-1, a representation vector h may be generated_ei，t-1And h_li，t-1Spliced together to obtain a final representation vector h_i，t-1。

As shown in FIG. 3, at time t, the RNN313 receives input from the MLP 312 at time t and acquires the RNN303 at time tThe output at time-1 serves as the other input at time-t. By calculation, RNN313 may update its hidden state and may provide an output at time t to RNN at time t +1 (not shown) as an input. For clarity, FIG. 3 does not show that the output of RNN313 is provided to RNN at time t +1 as an input to RNN at time t + 1. For example, at time t, the hidden state of RNN313 is represented by vector h_ei，tAnd the output of RNN313 at time t may be provided as an input to RNN at time t +1 (not shown).

At time t, RNN 315 receives an input from MLP 314 at time t and takes the output of RNN 305 at time t-1 as another input at time t. By calculation, RNN 315 may update its hidden state and may provide an output at time t as an input to the RNN at time t +1 (not shown). For clarity, FIG. 3 does not show that the output of RNN 315 is provided to RNN at time t +1 as an input to RNN at time t + 1. For example, RNN 315 may be hidden from view at time t as a vector h_li，tAnd the output of RNN 305 at time t may be provided as an input to the RNN at time t +1 (not shown). At time t, a representation vector h may be represented_ei，tAnd h_li，tSpliced together to obtain a final representation vector h_i，t。

Returning now to FIG. 3, Outlet i feature E_iMay include the width w of the outlet i_iDirection of travel alpha of exit i_iBoundary point position P on left and right sides of outlet_li、P_ri. As shown in fig. 3, an exit feature vector E_iIs provided to the multi-tier perceptron 301, and the multi-tier perceptron 301 maps the egress feature vector to an egress representation vector e_i。

At time t-1, the exit represents vector e_iAnd a vehicle representation vector h_i，t-1Provided to a matching function 306, the matching function 306 calculates a vehicle representation vector h_i，t-1And the exit representation vector e_iThe degree of matching between the two groups can be represented by similarity or matching probability. For example, the vehicle represents a vector h_i，t-1And the exit representation vector e_iMay be spliced together and provided to the matching function 306, and the matching function 306 may be rootDetermining a match probability P from the stitched representation vector_i，t-1. The matching probabilities can then be normalized by the activation function to obtain the relative probabilities between the different exits, i.e. the probabilities of the different intents. For example, the activation function may be a softmax activation function. For example, the matching function 306 may include one or more multi-tier perceptrons.

At time t, the exit represents vector e_iAnd a vehicle representation vector h_i，tProvided to a matching function 316, the matching function 316 calculates a vehicle representation vector h_i，tAnd the exit representation vector e_iThe degree of matching between the two groups can be represented by similarity or matching probability. For example, the vehicle represents a vector h_i，t-1And the exit representation vector e_iCan be stitched together and provided to the matching function 316, and the matching function 306 can determine the matching probability P from the stitched representation vector_i，t. The matching probabilities can then be normalized by the activation function to obtain the relative probabilities between the different exits, i.e. the probabilities of the different intents. For example, the activation function may be a softmax activation function. For example, the matching function 316 may include one or more multi-tier perceptrons.

In some embodiments, the target vehicle intent may be tracked temporally by a sliding window approach. For example, the fusion probability at the current time (weighted average) may be calculated based on the state probabilities within a certain window length, for example, as shown in equation (3):

wherein w_jAnd M is the weight corresponding to the probability at the moment t-j, and is the window length. The weight can adopt average weighting or exponential decreasing weighting, and K is the total number of the exits of the intersection where the target vehicle is located. By calculating the fusion probability, the influence of the change in state of the vehicle can be further taken into account.

One specific embodiment of the intent prediction system 300 is described below, it being understood that this specific embodiment is provided by way of example only. For example, the MLPs 302 and 304 (i.e., MLPs 312 and 314) in the vehicle feature learning network layer may be single-layer MLPs, i.e., include only one hidden layer. Single-layer MLP can achieve feature fusion and correlation with low computational overhead. The hidden unit dimension of the single-layer MLP is 64, and the activation function is a linear rectification function (ReLU). It should be understood that any other suitable hidden unit dimensions and activation functions may be used.

RNNs 303 and 305 (i.e., RNNs 313 and 315) in a vehicle feature learning network may be implemented by GRUs, with hidden unit dimensions of 128. It should be understood that any other suitable RNN, e.g., LSTM, may be used, as well as any other suitable hidden cell dimension.

The MLP 301 in the egress feature learning network may be implemented by a single layer MLP, i.e., containing only one hidden layer. The hidden unit dimension of the single-layer MLP is 16, and the activation function is ReLU. It should be understood that any other suitable hidden unit dimensions and activation functions may be used.

The metric learning network can be implemented by three layers of MLPs, where the first layer of hidden unit dimensions is 128, and the activation function is ReLU; the dimension of a second layer of hidden units is 64, and the activation function is ReLU; the third level hidden unit dimension is 1, and the output of the three levels MLP is normalized by a normalization function (e.g., softmax function) to obtain a normalized target intention probability. It should be understood that any other suitable MLP, hidden cell dimensions, and normalization functions may be used.

After the target intention probability passes through a classifier (e.g., sigmoid function), an absolute matching probability of the vehicle with each exit is obtained. It should be understood that any other suitable classifier may be used in addition to the sigmoid function.

In the training process, weighted two-class cross entropy or multi-class cross entropy may be used as the loss function. It should be understood that any other suitable loss function may be used instead. In one embodiment, the batch size of the training data is 128, the initial learning rate is 0.001, the learning rate can be changed in an exponential decreasing form, the attenuation step size is set to 10 rounds, the total number of training rounds is 100 rounds, and one round of training refers to one traversal of all the training data. It should be appreciated that any other suitable training parameter may be used instead.

In the intent prediction system 300 shown in fig. 3, it is not necessary to use CNN to extract features from a high-dimensional rasterized image, so that the model complexity can be reduced, and the calculation efficiency, real-time performance and timeliness can be improved, thereby enabling timely reaction and being more suitable for the scene of automatic driving. In addition, the intention prediction system 300 represents the state of the target vehicle with the exit as a reference or benchmark, and thus, the intersection does not need to be divided into sector areas, so that the intention prediction system can be adapted to various intersection scenes and has good generalization performance.

Fig. 5 illustrates a schematic diagram of an intent prediction system 500, in accordance with some embodiments of the present disclosure. Unlike the intention prediction system 300 in FIG. 3, the intention prediction system 500 does not extract the vehicle feature vector S_iOnly the vehicle feature vector F is extracted_i. Vehicle feature vector S is extracted in the intention prediction system 500_iAnd an exit feature vector E_i。

In some casesIn an embodiment, the exit feature vector E_iMay include the width w of the outlet i_iDirection of travel alpha of exit i_iBoundary point position P on left and right sides of outlet_li、P_ri。

As shown in FIG. 5, at time t-1, the vehicle feature vector F_i，t-1Input to a multilayer perceptron (MLP) 502. For example, the MLP502 may have a hidden layer to improve computational efficiency. The output of the multi-layered perceptron 502 is provided to a Recurrent Neural Network (RNN) 503. For example, the RNN 503 may be implemented by a recurrent neural network unit such as a Gated Recurrent Unit (GRU), a long-short term memory unit (LSTM), or the like. It should be understood that fig. 5 shows the recurrent neural network in expanded form, i.e., RNN 503 and RNN513 are the same RNN at different times. In other words, only one RNN is included in the feature learning network layer of fig. 5.

At time t-1, RNN 503 receives an input from MLP502 at time t-1 and obtains the output of RNN 503 at time t-2 as another input at time t-1. For clarity, FIG. 5 does not show RNN 503 receiving as input the output of the RNN at time t-2. By calculation, RNN 503 may update its hidden state and may provide an output at time t-1 as an input to the RNN at time t + 1. For example, at time t-1, the hidden state of RNN 503 is represented by vector h_i，t-1And the output of RNN 503 at time t-1 may be provided as an input to RNN at time t, RNN 513.

As shown in FIG. 5, at time t, RNN513 receives an input from MLP 512 at time t and takes the output of RNN 503 at time t-1 as another input at time t. By calculation, RNN513 may update its hidden state and may provide an output at time t as an input to the RNN at time t +1 (not shown). For clarity, FIG. 5 does not show that the output of RNN513 is provided to RNN at time t +1 as an input to RNN at time t + 1. For example, at time t, the RNN513 hidden state is represented by vector h_i，tAnd the output of time t RNN513 may be provided as an input to RNN at time t +1 (not shown).

As shown in fig. 5, an exit feature vector E_iIs provided to a multi-layered perceptron 501, the multi-layered perceptron 501The exit feature vector E_iMapping to an egress representation vector e_i。

At time t-1, the exit represents vector e_iAnd a vehicle representation vector h_i，t-1Provided to a matching function 506, which matching function 506 calculates a vehicle representation vector h_i，t-1And the exit representation vector e_iThe degree of matching between the two groups can be represented by similarity or matching probability. For example, the vehicle represents a vector h_i，t-1And the exit representation vector e_iCan be stitched together and provided to a matching function 506, and the matching function 506 can determine a matching probability P from the stitched representation vector_i，t-1. The matching probabilities can then be normalized by the activation function to obtain the relative probabilities between the different exits, i.e. the probabilities of the different intents. For example, the activation function may be a softmax activation function. For example, the matching function 506 may include one or more multi-tier perceptrons.

At time t, the exit represents vector e_iAnd a vehicle representation vector h_i，tProvided to a matching function 516, the matching function 516 calculates a vehicle representation vector h_i，tAnd the exit representation vector e_iThe degree of matching between the two groups can be represented by similarity or matching probability. For example, the vehicle represents a vector h_i，tAnd the exit representation vector e_iCan be stitched together and provided to the matching function 516, and the matching function 516 can determine the matching probability P according to the stitched representation vector_i，t. The matching probabilities can then be normalized by the activation function to obtain the relative probabilities between the different exits, i.e. the probabilities of the different intents. For example, the activation function may be a softmax activation function. For example, the matching function 516 may include one or more multi-tier perceptrons.

In some embodiments, the target vehicle intent may be tracked temporally by a sliding window approach. For example, the fusion probability at the current time (weighted average) may be calculated based on the state probabilities within a certain window length, for example, as shown in equation (4).

Wherein w_jAnd M is the weight corresponding to the probability at the moment t-j, and is the window length. The weight can adopt average weighting or exponential decreasing weighting, and K is the total number of the exits of the intersection where the target vehicle is located.

One specific embodiment of intent prediction system 500 is described below, it being understood that this specific embodiment is provided by way of example only. For example, the MLP502 (i.e., MLP 512) in the vehicle feature learning network layer may be a single-layer MLP, i.e., include only one hidden layer. Single-layer MLP can achieve feature fusion and correlation with low computational overhead. The hidden unit dimension of the single-layer MLP is 64, and the activation function is a linear rectification function (ReLU). It should be understood that any other suitable hidden unit dimensions and activation functions may be used.

The RNN 503 (i.e., RNN 513) in the vehicle feature learning network may be implemented by a GRU with a hidden unit dimension of 128. It should be understood that any other suitable RNN, e.g., LSTM, may be used, as well as any other suitable hidden cell dimension.

The MLP 501 in the egress feature learning network may be implemented by a single-layer MLP, i.e., containing only one hidden layer. The hidden unit dimension of the single-layer MLP is 16, and the activation function is ReLU. It should be understood that any other suitable hidden unit dimensions and activation functions may be used.

After the target intention probability passes through a classifier (e.g., s igmoid function), an absolute matching probability of the vehicle with each exit is obtained. It should be understood that any other suitable classifier may be used in addition to the sigmoid function.

In the intention prediction system 500 shown in fig. 5, it is not necessary to use CNN to extract features from a high-dimensional rasterized image, so that the model complexity can be reduced, and the calculation efficiency, real-time performance and timeliness can be improved, thereby enabling timely reaction and being more suitable for the scene of automatic driving. In addition, the intention prediction system 500 represents the state of the target vehicle with the exit as a reference or benchmark, and thus, the intersection does not need to be divided into sector areas, so that the intention prediction system can be adapted to various intersection scenes and has good generalization performance. Compared with the intention prediction system 300, the intention prediction system does not use the vehicle feature vector S related to the lane line_iThe intent prediction system 500 does not require a high precision map, and thus may still achieve better performance when the virtual lane line is missing or inaccurate.

Fig. 6 illustrates a flow chart of a method 600 for autonomous driving, according to some embodiments of the present disclosure. The method 600 may be performed at the host vehicle 101, or may be performed partially at the host vehicle 101 and partially in the cloud.

At block 602, sets of exit characteristics for a plurality of candidate exits associated with a vehicle (e.g., other vehicle 102, also referred to as a target vehicle) are obtained, each candidate exit being a portion of a road associated with a drivable path of the vehicle.

At block 604, a plurality of sets of vehicle characteristics of the vehicle relative to the plurality of candidate exits are determined, each set of vehicle characteristics including at least a state of motion of the vehicle relative to a respective candidate exit of the plurality of candidate exits.

At block 606, an exit matching the travel route of the vehicle is determined from a plurality of candidate exits based on the plurality of sets of exit features and the plurality of sets of vehicle features.

In some embodiments, each set of outlet features comprises: and under a Cartesian coordinate system with the corresponding candidate exit as an origin, the vehicle moves relative to the candidate exit, wherein the Cartesian coordinate system takes the road running direction at the corresponding candidate exit as a coordinate axis direction.

In some embodiments, the exit characteristics of the respective candidate exits include at least one of: size, travelable direction, and boundary of the respective candidate exit.

In some embodiments, each set of vehicle characteristics includes at least one of: the coordinates of the vehicle, the distance between the vehicle and the origin, the orientation of the vehicle, the rate of change of the coordinates of the vehicle, the rate of change of the distance between the vehicle and the origin, and the rate of change of the orientation of the vehicle.

In some embodiments, each set of vehicle features further comprises: and the vehicle is in a motion state under a curve coordinate system taking the terminal point of the reference lane line as the origin and the reference lane line as one coordinate axis, wherein the lane line closest to the vehicle is selected from the lane lines connected with the corresponding candidate outlets as the reference lane line.

In some embodiments, the motion state of the vehicle in the curvilinear coordinate system comprises at least one of: the coordinates of the vehicle, the angular difference of the orientation of the vehicle from a reference travel direction, which is the travel direction of the lane line at the projection point of the vehicle to the reference lane line, the rate of change of the coordinates of the vehicle, and the rate of change of the angular difference.

Additionally, in some embodiments, a feature representation for each of the plurality of sets of exit features may also be determined by the first multi-layered perceptron, and a feature representation for each of the plurality of sets of vehicle features may be determined by the second multi-layered perceptron and the recurrent neural network.

For example, the exit may be determined based on the feature representation of each of the plurality of sets of exit features and the feature representation of each of the plurality of sets of vehicle features. Alternatively, in some embodiments, a matching coefficient between each of the sets of egress features and a respective vehicle feature of the sets of vehicle features may be determined by a third multi-tier perceptron, and the egress determined based on this matching coefficient.

In some embodiments, determining the exit based on the matching coefficients comprises: obtaining a historical matching coefficient relating to a historical location of the vehicle, the historical matching coefficient being a matching coefficient between each of the plurality of sets of exit features and a respective one of the plurality of sets of vehicle features when the vehicle is at the historical location; calculating an updated matching coefficient based on the historical matching coefficient and the determined matching coefficient; and determining an exit based on the updated matching coefficients.

Fig. 7 illustrates a block diagram of an apparatus 700 for autopilot according to some embodiments of the present disclosure. The apparatus 700 may be implemented at the vehicle 101, or may be partially implemented at the vehicle 101, and partially implemented at the cloud.

As shown in fig. 7, the apparatus 700 includes a first obtaining module 702 configured to obtain sets of exit characteristics for a plurality of candidate exits associated with the vehicle, each candidate exit being a portion of a road associated with a drivable path of the vehicle.

The apparatus 700 further includes a first determination module 704 configured to determine a plurality of sets of vehicle characteristics of the vehicle relative to a plurality of candidate exits, each set of vehicle characteristics including at least a state of motion of the vehicle relative to a respective candidate exit of the plurality of candidate exits.

The apparatus 700 further includes a second determination module 706 configured to determine an exit from the plurality of candidate exits that matches the travel route of the vehicle based on the plurality of sets of exit characteristics and the plurality of sets of vehicle characteristics.

In some embodiments, each set of outlet features comprises: and under the Cartesian coordinate system, the vehicle moves relative to the candidate exit.

In some embodiments, each set of vehicle characteristics includes at least one of: the coordinates of the vehicle, the distance between the vehicle and the origin of the cartesian coordinate system, the orientation of the vehicle, the rate of change of the coordinates of the vehicle, the rate of change of the distance between the vehicle and the origin, and the rate of change of the orientation of the vehicle.

In some embodiments, the motion state of the vehicle in the curvilinear coordinate system comprises at least one of: the coordinates of the vehicle, the angular difference of the orientation of the vehicle from a reference travel direction, which is the travel direction of the lane line at the projection point of the vehicle to the reference lane line, the rate of change of the coordinates of the vehicle, the rate of change of the angular difference.

In some embodiments, the apparatus 700 further includes a third determination module configured to determine, by the first multi-tier perceptron, a feature representation for each of the plurality of sets of outlet features; a fourth determination module configured to determine a feature representation for each of the plurality of sets of vehicle features through the second multi-layered perceptron and the recurrent neural network; and wherein the second determination module 706 is configured to: the exit is determined based on the feature representation of each of the sets of exit features and the feature representation of each of the sets of vehicle features.

In some embodiments, the second determination module 706 comprises: a fifth determination module configured to determine, by a third multi-tier perceptron, matching coefficients between each of the sets of egress features and respective ones of the sets of vehicle features; and a sixth determination module configured to determine an exit based on the matching coefficient.

In some embodiments, the sixth determining module comprises: a second obtaining module configured to obtain a history matching coefficient relating to a history position of the vehicle, the history matching coefficient being a matching coefficient between each of the plurality of sets of exit features and a corresponding vehicle feature of the plurality of sets of vehicle features when the vehicle is at the history position; a calculation module configured to calculate an updated matching coefficient based on the historical matching coefficient and the determined matching coefficient; and a seventh determination module configured to determine an exit based on the updated matching coefficient.

FIG. 8 shows a schematic block diagram of an electronic device 800 that may be used to implement embodiments of the present disclosure. Device 800 may be used to implement apparatus 700 of fig. 7. As shown, device 800 includes a Central Processing Unit (CPU)801 that may perform various appropriate actions and processes in accordance with computer program instructions stored in a Read Only Memory (ROM)802 or loaded from a storage unit 808 into a Random Access Memory (RAM) 803. In the RAM 803, various programs and data required for the operation of the device 800 can also be stored. The CPU 801, ROM 802, and RAM 803 are connected to each other via a bus 804. An input/output (I/O) interface 805 is also connected to bus 804.

A number of components in the device 800 are connected to the I/O interface 805, including: an input unit 806, such as a keyboard, a mouse, or the like; an output unit 807 such as various types of displays, speakers, and the like; a storage unit 808, such as a magnetic disk, optical disk, or the like; and a communication unit 809 such as a network card, modem, wireless communication transceiver, etc. The communication unit 809 allows the device 800 to exchange information/data with other devices via a computer network such as the internet and/or various telecommunication networks.

The various processes and processes described above, such as method 600, may be performed by processing unit 801. For example, in some embodiments, the method 600 may be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 808. In some embodiments, part or all of the computer program can be loaded and/or installed onto device 800 via ROM 802 and/or communications unit 809. When the computer program is loaded into RAM 803 and executed by CPU 801, one or more steps of method 600 described above may be performed. Alternatively, in other embodiments, the CPU 801 may be configured to perform the method 600 by any other suitable means (e.g., by way of firmware).

The present disclosure may be methods, apparatus, systems, and/or computer program products. The computer program product may include a computer-readable storage medium having computer-readable program instructions embodied thereon for carrying out various aspects of the present disclosure.

The computer readable storage medium may be a tangible device that can hold and store the instructions for use by the instruction execution device. The computer readable storage medium may be, for example, but not limited to, an electronic memory device, a magnetic memory device, an optical memory device, an electromagnetic memory device, a semiconductor memory device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium would include the following: a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a Static Random Access Memory (SRAM), a portable compact disc read-only memory (CD-ROM), a Digital Versatile Disc (DVD), a memory stick, a floppy disk, a mechanical coding device, such as punch cards or in-groove projection structures having instructions stored thereon, and any suitable combination of the foregoing. Computer-readable storage media as used herein is not to be construed as transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission medium (e.g., optical pulses through a fiber optic cable), or electrical signals transmitted through electrical wires.

The computer-readable program instructions described herein may be downloaded from a computer-readable storage medium to a respective computing/processing device, or to an external computer or external storage device via a network, such as the internet, a local area network, a wide area network, and/or a wireless network. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers and/or edge servers. The network adapter card or network interface in each computing/processing device receives computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in the respective computing/processing device.

The computer program instructions for carrying out operations of the present disclosure may be assembler instructions, Instruction Set Architecture (ISA) instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smal ltalk, C + + or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet service provider). In some embodiments, the electronic circuitry that can execute the computer-readable program instructions implements aspects of the present disclosure by utilizing the state information of the computer-readable program instructions to personalize the electronic circuitry, such as a programmable logic circuit, a Field Programmable Gate Array (FPGA), or a Programmable Logic Array (PLA).

Various aspects of the present disclosure are described herein with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the disclosure. It will be understood that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer-readable program instructions.

These computer-readable program instructions may be provided to a processing unit of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processing unit of the computer or other programmable data processing apparatus, create means for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks. These computer-readable program instructions may also be stored in a computer-readable storage medium that can direct a computer, programmable data processing apparatus, and/or other devices to function in a particular manner, such that the computer-readable medium storing the instructions comprises an article of manufacture including instructions which implement the function/act specified in the flowchart and/or block diagram block or blocks.

The computer readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other devices to produce a computer implemented process such that the instructions which execute on the computer, other programmable apparatus or other devices implement the functions/acts specified in the flowchart and/or block diagram block or blocks.

The flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.

The foregoing description of the embodiments of the present disclosure has been presented for purposes of illustration and description, and is not intended to be exhaustive or limited to the embodiments disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen in order to best explain the principles of the embodiments, the practical application, or improvements made to the technology in the marketplace, or to enable others of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A method for autonomous driving, comprising:

obtaining a plurality of sets of exit features for a plurality of candidate exits associated with a vehicle, each candidate exit being a portion of a road associated with a drivable route of the vehicle;

determining a plurality of sets of vehicle characteristics of the vehicle relative to the plurality of candidate exits, each set of vehicle characteristics including at least a state of motion of the vehicle relative to a respective candidate exit of the plurality of candidate exits; and

determining an exit from the plurality of candidate exits that matches the travel route of the vehicle based on the plurality of sets of exit features and the plurality of sets of vehicle features.

2. The method of claim 1, wherein each set of exit features includes a state of motion of the vehicle relative to the candidate exits in a Cartesian coordinate system.

3. The method of claim 2, wherein the exit characteristics of the respective candidate exits comprise at least one of:

a size, a travelable direction, and a boundary of the respective candidate exit.

4. The method of claim 2, wherein each set of vehicle characteristics includes at least one of:

the coordinates of the vehicle, the distance between the vehicle and the origin of the Cartesian coordinate system, the orientation of the vehicle, the rate of change of the coordinates of the vehicle, the rate of change of the distance between the vehicle and the origin, and the rate of change of the orientation of the vehicle.

5. The method of claim 1, wherein each set of vehicle features further comprises:

and the motion state of the vehicle under a curve coordinate system taking a reference lane line as a coordinate axis, wherein the lane line closest to the vehicle is selected from the lane lines connected with the corresponding candidate outlets as the reference lane line.

6. The method of claim 5, wherein the state of motion of the vehicle in the curvilinear coordinate system comprises at least one of:

the coordinates of the vehicle, an angular difference of an orientation of the vehicle from a reference traveling direction, the reference traveling direction being a traveling direction of a lane line at a projection point of the vehicle to the reference lane line, a rate of change in the coordinates of the vehicle, and a rate of change in the angular difference.

7. The method of claim 1, further comprising:

determining, by a first multi-layer perceptron, a feature representation for each of the plurality of sets of outlet features;

determining, by a second multi-layered perceptron and recurrent neural network, a feature representation for each of the plurality of sets of vehicle features; and

wherein determining the outlet comprises:

determining the egress based on the feature representation for each of the sets of egress features and the feature representation for each of the sets of vehicle features.

8. The method of claim 1, wherein determining the outlet comprises:

determining, by a third multi-tier perceptron, a matching coefficient between each of the plurality of sets of egress features and a respective vehicle feature of the plurality of sets of vehicle features; and

determining the outlet based on the matching coefficient.

9. The method of claim 8, wherein determining the exit based on the matching coefficient comprises:

obtaining a historical matching coefficient relating to a historical location of the vehicle, the historical matching coefficient being a matching coefficient between each of the plurality of sets of exit features and a respective vehicle feature of the plurality of sets of vehicle features when the vehicle is at the historical location;

calculating an updated matching coefficient based on the historical matching coefficient and the determined matching coefficient; and

determining the exit based on the updated matching coefficient.

10. The method of claim 1, wherein the plurality of candidate exits are exits of an intersection area.

11. An apparatus for autonomous driving, comprising:

a first obtaining module configured to obtain a plurality of sets of exit features of a plurality of candidate exits associated with a vehicle, each candidate exit being a portion of a road associated with a drivable route of the vehicle;

a first determination module configured to determine a plurality of sets of vehicle characteristics of the vehicle relative to the plurality of candidate exits, each set of vehicle characteristics including at least a state of motion of the vehicle relative to a respective candidate exit of the plurality of candidate exits; and

a second determination module configured to determine an exit from the plurality of candidate exits that matches the travel route of the vehicle based on the plurality of sets of exit features and the plurality of sets of vehicle features.

12. An electronic device, the electronic device comprising:

one or more processors; and

memory storing one or more programs that, when executed by the one or more processors, cause the electronic device to implement the method of any of claims 1-10.

13. An autonomous vehicle comprising the electronic device of claim 12.

14. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the method according to any one of claims 1-10.