WO2020000192A1 - Method for providing vehicle trajectory prediction - Google Patents
Method for providing vehicle trajectory prediction Download PDFInfo
- Publication number
- WO2020000192A1 WO2020000192A1 PCT/CN2018/092905 CN2018092905W WO2020000192A1 WO 2020000192 A1 WO2020000192 A1 WO 2020000192A1 CN 2018092905 W CN2018092905 W CN 2018092905W WO 2020000192 A1 WO2020000192 A1 WO 2020000192A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- trajectory
- vehicle
- cost function
- lane
- trajectories
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims abstract description 51
- 206010048669 Terminal state Diseases 0.000 claims abstract description 5
- 230000001133 acceleration Effects 0.000 claims description 14
- 230000007613 environmental effect Effects 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 2
- 230000006870 function Effects 0.000 description 36
- 238000002474 experimental method Methods 0.000 description 13
- 230000006399 behavior Effects 0.000 description 5
- 238000009826 distribution Methods 0.000 description 5
- 238000010200 validation analysis Methods 0.000 description 5
- 238000012549 training Methods 0.000 description 4
- 230000036461 convulsion Effects 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 238000005070 sampling Methods 0.000 description 3
- 239000013598 vector Substances 0.000 description 3
- 238000013461 design Methods 0.000 description 2
- 238000006073 displacement reaction Methods 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 230000001360 synchronised effect Effects 0.000 description 2
- 230000008094 contradictory effect Effects 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 230000001955 cumulated effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000009472 formulation Methods 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- BULVZWIRKLYCBC-UHFFFAOYSA-N phorate Chemical compound CCOP(=S)(OCC)SCSCC BULVZWIRKLYCBC-UHFFFAOYSA-N 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000011524 similarity measure Methods 0.000 description 1
- 238000005303 weighing Methods 0.000 description 1
Images
Classifications
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B60—VEHICLES IN GENERAL
- B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
- B60W50/00—Details of control systems for road vehicle drive control not related to the control of a particular sub-unit, e.g. process diagnostic or vehicle driver interfaces
- B60W50/0097—Predicting future conditions
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B60—VEHICLES IN GENERAL
- B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
- B60W2554/00—Input parameters relating to objects
- B60W2554/80—Spatial relation or speed relative to objects
- B60W2554/801—Lateral distance
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B60—VEHICLES IN GENERAL
- B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
- B60W2554/00—Input parameters relating to objects
- B60W2554/80—Spatial relation or speed relative to objects
- B60W2554/802—Longitudinal distance
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B60—VEHICLES IN GENERAL
- B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
- B60W2554/00—Input parameters relating to objects
- B60W2554/80—Spatial relation or speed relative to objects
- B60W2554/804—Relative longitudinal speed
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B60—VEHICLES IN GENERAL
- B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
- B60W30/00—Purposes of road vehicle drive control systems not related to the control of a particular sub-unit, e.g. of systems using conjoint control of vehicle sub-units
- B60W30/14—Adaptive cruise control
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B60—VEHICLES IN GENERAL
- B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
- B60W60/00—Drive control systems specially adapted for autonomous road vehicles
- B60W60/001—Planning or execution of driving tasks
- B60W60/0011—Planning or execution of driving tasks involving control alternatives for a single driving scenario, e.g. planning several paths to avoid obstacles
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B60—VEHICLES IN GENERAL
- B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
- B60W60/00—Drive control systems specially adapted for autonomous road vehicles
- B60W60/001—Planning or execution of driving tasks
- B60W60/0015—Planning or execution of driving tasks specially adapted for safety
Definitions
- an autonomous vehicle needs not only fulfill its mission (e.g. achieving a destination efficiently and safely) , but also follow human drivers behavior so as to let the passengers of a driverless car feel comfortable and convinced, and other traffic participants of the road keep their usual way in risk prediction and decision making.
- a cost function that encodes human drivers preference in decision making and operation.
- Machine learning methods have been used to learn cost functions or parameter settings from the data of human demonstration.
- Evaluations are conducted at both the initial and terminal time of the maneuver, and denoted by subscripts s and e respectively.
- Each potential target lane is evaluated individually, including the left, right and current lanes, and denoted by subscript tar. Velocities of the preceding and back vehicles at the initial time are given in the driving situation S i , while those at the terminal time are predicted on a linear motion model. The predicted values are denoted by a superscript . It would be noted that the greater the differences, the more attraction from the target lane.
Landscapes
- Engineering & Computer Science (AREA)
- Automation & Control Theory (AREA)
- Human Computer Interaction (AREA)
- Transportation (AREA)
- Mechanical Engineering (AREA)
- Traffic Control Systems (AREA)
Abstract
A method for providing a vehicle trajectory prediction is provided, the method comprising: recording, at a processor, a human driven sample H i consisting of driving situation S i at a certain time and human driven trajectory T GT i from the certain time to a real terminal time or in a fixed duration; generating a set of sampled trajectories T i at the driving situation S i using a tessellation of terminal states to describe potential candidates of future motion sequences; calculating a distance measure between a sampled trajectory and the human driven trajectory at the driving situation; calculating a cost function f of the sampled trajectory using a set of coefficients; converting the cost function f to probability with a Softmax function; and calibrating the cost function weight for minimizing the total distance of all trajectories combined with the probability of the trajectory adopted with gradient descent. Such method increases acceptability of autonomous driving by surrounding drivers and is essential to predict surrounding human driven vehicles trajectories.
Description
The present invention is directed to the domain of automobile vehicles, and more particularly, to a method for providing vehicle trajectory prediction, which would be applied to human-like motion planning and surrounding human driven vehicle trajectory prediction.
In the last few decades, there are tremendous development and progress in Advanced Driving Assistant Systems (ADAS) and autonomous vehicles. Advanced intelligent vehicles have great potential to improve the performance and safety of the transportation system and free people from the tasks of driving. While an autonomous driving system needs not only to keep safety, but also follow human drivers behavior, so as to let the passengers of a driverless car feel comfortable and convinced, and other traffic participants of the road keep their usual way in risk prediction and decision making. It is thus important to develop motion planning methods that achieve human-like autonomous driving.
Motion planning has generally been framed as finding the lowest cost trajectory from a set of trajectory candidates, and trajectory optimization has also been conducted to minimize the cost of an initial estimate. The motion planner for on-road autonomous driving usually takes the following steps:
1. Giving an initial state of the vehicle, sampling a set of end points;
2. Generating trajectories by linking the initial state to each end point with consideration to the vehicles kinematic and dynamic constraints;
3. Evaluating the trajectories by a cost function, and select the optimal one for the vehicle to execute.
However, finding a proper cost function to evaluate candidate trajectories is highly non-trivial. Cost functions are carefully designed by incorporating the terms on efficiency, comfort and safety, while tuning parameters usually require significant amount of hand-engineering by experts, which needs to balance the contributions (e.g. weights) of many components that could potentially be correlated or even contradictory, and it is even harder to design a correct setting to generalize enough at various conditions. Such motion planning methods have technological feasibility of achieving vehicle autonomy, while humanization has been ignored, more importantly the resultant behaviors may be much different with those of human drivers.
Furthermore, an autonomous vehicle needs not only fulfill its mission (e.g. achieving a destination efficiently and safely) , but also follow human drivers behavior so as to let the passengers of a driverless car feel comfortable and convinced, and other traffic participants of the road keep their usual way in risk prediction and decision making. Hence, it is great important to design a cost function that encodes human drivers preference in decision making and operation. Machine learning methods have been used to learn cost functions or parameter settings from the data of human demonstration.
SUMMARY OF THE INVENTION
The present invention aims to challenge of human-like autonomous driving on crowded highway scenes, and proposes a human-like motion planning method by learning from naturalistic driving data. At a certain driving condition, a set of trajectory samples is first generated containing both longitudinal and lateral movements, a desired one is then selected with a cost function. A cost function is formulated by incorporating not only the factors on such as comfort, efficiency and safety as an existing autonomous driving system does, but also concerning lane incentive by referring to a human driver’s lane change decisions. Coefficients of the cost components are learnt by correlating the probability of a trajectory being selected with its distance (i.e. similarity) to the human driven one at the same driving situation.
To this end, according to one aspect of the present invention, a method for providing a vehicle trajectory prediction is provided, the method comprising:
recording, at a processor, a human driven sample H
i consisting of driving situation S
i at a certain time and human driven trajectory T
GT
i from the certain time to a real terminal time or in a fixed duration;
generating, at the processor, a set of sampled trajectories T
i at the driving situation S
i using a tessellation of terminal states to describe potential candidates of future motion sequences;
calculating, at the processor, a distance measure between a sampled trajectory and the human driven trajectory at the driving situation;
calculating, at the processor, a cost function f of the sampled trajectory using a set of coefficients;
converting, at the processor, the cost function f to probability with a Softmax function; and
calibrating, at the processor, the cost function weight for minimizing the total distance of all trajectories combined with the probability of the trajectory adopted with gradient descent.
In accordance with the foregoing technical concept, the present invention may further include any one or more of the following alternative forms.
In some alternative forms, the human driven sample is recorded in Frenet frame with the origin at the ego vehicle’s location at the initial time, in which the driving situation is defined based on the states of the ego and environmental vehicles in local surroundings.
In some alternative forms, the human driven trajectory T
GT
i is a time series of trajectory points in view of a variable duration in case of lane change and a fixed duration in case of car following.
In some alternative forms, the distance between each pair of the sampled trajectory and the human driven trajectory are estimated on both the location and velocity components.
In some alternative forms, the set of coefficients of the cost function f are associated with comfort, efficiency, lane incentive and safety of a trajectory.
In some alternative forms, the comfort of a trajectory is evaluated in view of acceleration and acceleration change ratio on longitudinal and lateral dimensions.
In some alternative forms, the efficiency of a trajectory is evaluated in view of average speed of the trajectory.
In some alternative forms, the lane incentive of a trajectory is evaluated in view of repulsion from the current lane and attraction of the target lane, wherein the repulsion from the current lane is evaluated using the velocity difference of the ego vehicle with its leading vehicle, and the attraction of the target lane is evaluated using the velocity difference of the ego vehicle with the preceding and back vehicles on the target lane.
In some alternative forms, the safety of a trajectory is evaluated in view of distances between the ego vehicle and the surrounding vehicles at each discrete time during the course of the trajectory.
In accordance with another aspect of the invention, a use of the above method for human-like trajectory planning of autonomous driving vehicle is provided, wherein a trajectory of the vehicle is selected from a probability which is transformed from the cost function with the cost function weight generated at calibration phase.
In some alternative forms, the trajectory of the vehicle is obtained in view of situation defined by the states of the ego and surrounding vehicles.
In accordance with another aspect of the invention, a use of the above method for surrounding human driven vehicles trajectories prediction is provided, wherein a trajectory of a vehicle to be predicted is selected from a probability which is transformed from the cost function with the cost function weight generated at calibration phase.
In some alternative forms, the trajectory of the vehicle is obtained in view of situation defined by the states of the vehicle for which trajectory is to be predicted and surrounding vehicles.
In accordance with another aspect of the invention, a vehicle trajectory prediction system arranged and operable to carry out the above method or the above use is provided.
In accordance with another aspect of the invention, a processing means programmed and operable to execute instructions for carrying out the above method or the above use is provided.
Whereas human factors are taken into account in usual traditional methods for trajectory planning through criteria based on comfort and safety, they do not explicitly tend towards human generated trajectories.
The method proposed in this invention enables selection of a human-like trajectory and can therefore be used
· for human-like trajectory planning of an ego autonomous driving vehicle; or
· for surrounding human driven vehicles trajectories prediction.
Such method increases acceptability of autonomous driving by surrounding drivers and is essential to predict surrounding human driven vehicles trajectories.
These and other aspects of the present invention will be apparent from and elucidated with reference to the embodiments described hereinafter.
Throughout the course of the following detailed description, reference will be made to the drawings, and in which:
Fig. 1 schematically illustrates a driving situation concerned in the present invention;
Fig. 2 schematically illustrates the results of experiment 1, showing the distributions of the trajectories over distance;
Fig. 3 schematically illustrates the results of experiment 2, showing the distributions of the trajectories over distance; and
Fig. 4 schematically illustrates the results of experiment 3, showing the distributions of the trajectories over distance.
DESCRIPTION OF EMBODIMENTS
Although the invention may be susceptible to embodiment in different forms, there is shown in the drawings, and herein will be described in detail, specific embodiments with the understanding that the present disclosure is to be considered an exemplification of the principles of the invention, and is not intended to limit the invention to that as illustrated and described hereinafter. Therefore, unless otherwise noted, features disclosed herein may be combined together to form additional combinations that were not otherwise shown for purposes of brevity.
As mentioned, whereas human factors are taken into account in usual traditional methods for trajectory planning through criteria based on comfort and safety, they do not explicitly tend towards human generated trajectories. The method proposed in this invention is based on Machine Learning and needs recorded human trajectories first. Such method increases acceptability of autonomous driving by surrounding drivers and is essential to predict surrounding human driven vehicles trajectories.
At a certain driving situation S, a traditional motion planning method would first generate a set of synthesized trajectories {T
1, T
2, ..., T
n} , then select an optimal one for the autonomous driving system’s execution, e.g. T
opt:
based on a cost function f.
However, the cost function would be hand crafted to balance the contributions of different components, and the selected trajectory may not follow human driver’s behavior. The present method exploits the same workflow, but learns a cost function f that has preference to the trajectories T
j that are similar with the human driven ones.
In the concept of this invention, a human driven sample H
i= (S
i, T
GT
i) is firstly recorded, which consists of a S
i, describing the driving situation at timet
s and a T
GT
i, of a human driven trajectory from t
s to t
e, where t
e is the real terminal time for lane changes, while a fixed duration τ=t
e-t
s is considered for car followings.
Let T
i= {T
1
i, T
2
i, ..., T
n
i} be a set of trajectories that are generated at the driving situation S
i using a tessellation of terminal states (ending position and ending time) . A quantic polynomial is optimal regarding the criterion
in which p denotes the vehicle position and which tends to minimize the jerk L
2-norm, then showed that ending position and ending time were sufficient parameters to define such an optimal trajectory using quantic polynomials.
Here, T
i thus serves as a proposal for all candidate motion sequences in future seconds.
Let d
j
i=d (T
j
i, T
GT
i) be a measure evaluating distance about a synthesized trajectory T
j
i with the human driven one T
GT
i at the driving situation.
Let f be a cost function that associates with comfort, efficiency, lane incentive and safety of a trajectory using a set of coefficients Ω. Considering that lower cost prompting higher probability of the trajectory being selected, it would be better to convert cost function to probability prob (f (T
j
i) ) with a Softmax transform function prob. Therefore, it would be appreciated that the more similarity (i.e. less distance) a trajectory with human driven one, the higher probability it would be selected.
Given a set of human driven samples {H
i} , the problem is therefore formulated as finding a cost function f (i.e. its coefficients Ω) to minimize the following objective function:
where m is the number of different measured situations and n is the number of generated trajectories.
Learning method
In this invention, pseudo code of the learning method is described as Algorithm 1. The goal is to find a set of coefficients Ω= {Ω
1, Ω
2, ..., Ω
K} that minimizes the total distance of all trajectories combined with probability of the trajectory adopted with gradient descent methods.
Algorithm 1: Learning Method
end for
end for
It would be noted that this function enables selection of a human-like trajectory and can therefore be used for trajectory planning of an ego autonomous driving vehicle or for surrounding vehicles trajectory prediction.
Human Driven Sample and Road Situation
Each data sample is described in a Frenet frame with the origin at the ego vehicle’s location at the initial time.
Let p= (s, d) be a location at the Frenet frame, where s and d are displacements from the origin on longitudinal and lateral road directions, and
as well as
are velocity and acceleration vectors in the same frame.
As presented on Fig. 1, a driving situation S= {S
ego, S
env, S
road} is described with three components, where the first two are the states (position, velocity and acceleration) of the ego vehicle and of environmental vehicles in local surroundings, such that
where j is the index of a particular surrounding vehicle.
In this invention, the preceding and back vehicles on the left, right and current lanes are concerned. S
road represents the road situation, indicating that the ego vehicle is driving at the farthest left lane, a middle lane or the farthest right lane, respectively S
road=-1, S
road=0 and S
road=+1.
On the other hand, a trajectory T
GT is a time series of trajectory points {P
k/t
s≤t
s+k×Δ
t≤t
e} , where
t
s and t
e respectively are the initial and terminal times of the trajectory, and Δ
t is the time interval of data sampling. In case of a lane change sample, t
e is the time when the maneuver finished and τ=t
e-t
s is the duration. At real world driving, τ varies within a range of [τ
min=6s; τ
max=10s] . In case of car following samples, a fixed duration (τ=8s) can be used in the present method.
Trajectory Sampling
At a certain driving situation S
i, a set of trajectories T
i= {T
1
i, T
2
i, ..., T
n
i} are generated describing potential candidates of future motion sequences.
In the trajectory generation method, a trajectory is represented as using two quintic polynomials for lateral and longitudinal axes, respectively
d (t) =a
0+a
1t+a
2t
2+a
3t
3+a
4t
4+a
5t
5
and
s (t) =b
0+b
1t+b
2t
2+b
3t
3+b
4t
4+b
5t
5.
For simplicity, let set t
s=0 and t
e=τ. Given the value of τ and two trajectory points
at the initial and terminal times t=0 and t=τ respectively, six equations can be derived for both d (t) and s (t) consequently the coefficients {a
0, ..., a
5} and {b
0, ..., b
5} are estimated.
Assuming that velocity on lateral axis, and acceleration on both longitudinal and lateral axes are zero at both the initial and end times and since the trajectory point at the initial time is given by S
i,
Highway driving has some properties that can be used in trajectory generation:
· lane width (D
lane) and speed limits (V
max) at a certain road are known;
· lateral vehicle positions at the initial and terminal time of each maneuver are usually at the middle of lanes;
· longitudinal displacement could vary largely with respect to different velocity at the initial time, while longitudinal speed may vary only slightly (Δ
V) to keep constant and smooth driving etc.
The following states can thus be tessellated for trajectory definition:
Distance measure
Given a pair of synthesized trajectory T
j and human driven one T
GT, discretization is first conducted to convert continuous trajectories to sequences of synchronized points at an equal interval Δ
T.
where p
j (t) = (s
j (t) , d
j (t) ) and
are the location and velocity at the Frenet frame of the synthesized trajectory T
j at time t, while
and
are those of the human driven one.
Distance between each pair of synchronized trajectory points are estimated on both the location
and velocity
components.
Weighing by a parameter λ
d and taking an average of these distances during the course, a distance measure between T
j and T
GT is formulated below
Cost Function
This proposal addresses the motion planning problem when the target lane has not been decided, e.g. a lane change could be made to either the left or right lane, the vehicle could perform either longitudinal driving or lane change, etc. Lane incentive is incorporated in addition to the well-used components on comfort, efficiency and safety. Below, the formulations on each component, followed by a summary of the total cost will be described in detail.
Comfort
High acceleration and high acceleration change ratio (i.e. jerk) could be the major reasons of lack of comfort. In Frenet frame, vehicle motion is decomposed into longitudinal and lateral movements. Comfort of a trajectory is evaluated on its acceleration and jerk on two individual dimensions as formulated below.
Efficiency
It would be appreciated that the higher speed, the higher efficiency. Let
be the average speed of the trajectory, which is estimated in this research by
efficiency is formulated as:
Lane Incentive
Lane incentive evaluates the potential of a lane to be selected as the target.
In this invention, it would not consider the case where the vehicle must leave the current lane for its destination, which is addressed at the strategic level by path planning. A different lane is selected when the current driving condition is not satisfying (e.g. due to a slow preceding vehicle) and the new target lane has the potential of improving driving condition. Therefore, lane incentive is evaluated on two aspects:
· repulsion from the current lane; and
· attraction of the target lane.
In case the current and the target lanes are the same, longitudinal driving is kept and car following is executed.
Traffic situation on each lane is crucial in lane selection. The current sensing system measures environmental vehicles location, hence their second derivative (acceleration) could be highly noisy. On the other hand, vehicles instantaneous locations change dramatically, yielding an algorithm sensitive to estimation timing. Comparing to instantaneous locations, vehicle velocities are influential in a driver’s decision making and providing more stable estimation. To this end, repulsion from the current lane is evaluated using the signed velocity difference of the ego with its leading vehicle as formulated below. It would be noted that the greater the difference, the more repulsion from the current lane.
Attraction of the target lane is evaluated using the signed velocity difference of the ego vehicle with the preceding and back vehicles on the target lane:
Evaluations are conducted at both the initial and terminal time of the maneuver, and denoted by subscripts s and e respectively. Each potential target lane is evaluated individually, including the left, right and current lanes, and denoted by subscript tar. Velocities of the preceding and back vehicles at the initial time are given in the driving situation S
i, while those at the terminal time are predicted on a linear motion model. The predicted values are denoted by a superscript
. It would be noted that the greater the differences, the more attraction from the target lane.
Safety
Distances between the ego vehicle and the surrounding vehicles at each discrete time during the course of the trajectory are examined to evaluate its safety.
Motion sequences of surrounding vehicles are predicted on a linear motion model based on the initial states that are described in S
i, and discretized to trajectory points T
env at interval Δ
t:
Here subscript q denotes the id of a surrounding vehicle, while the preceding and back vehicles on the left, right and current lanes are only concerned.
As longitudinal and lateral distances may contribute to different safety costs, they are estimated individually, respectively:
and
and weighed with a parameter λ
s. An exponential average is then taken on these distances to formulate safety cost:
Total Cost
Summarizing all above components in a vector
the total cost f (T
j) of a trajectory T
j∈T is formulated by weighting C
j with a set of coefficients:
Here Ω
k is a vector of coefficients weighting on each component, and K is the maximum exponential coefficient of the cost function.
Probability Estimation
A Softmax function prob is used to transform cost f (T
j) to probability prob (f (T
j) ) , where the lower the cost, the higher the probability of a trajectory of being selected:
Examples
The method proposed here has been tested on real life data. Over the drives have been cumulated 135 Left Lane Changes (LLC) , 135 Right Lane Changes (RLC) and 143 Car Following (CF) situations. Car Followings have been extracted for fixed durations of 8s and drives at speed lower than 8 m/shave been discarded to avoid any jammed situations that may exhibit other behavior.
Three experiments have been conducted to seek answers whether the learnt planner could propose autonomously a trajectory that is close to that of human driven one at the following prepositions:
1. Assuming the algorithm knows a lane change which has been decided and the target lane which the driver aims at (i.e. lane change motion planning to a target lane) , evaluate whether the algorithm predicts a trajectory that is at lowest distance than the real recorded human one.
2. Assuming the algorithm only knows a lane change which has been decided but not on what lane (the left lane or the right lane) the driver aims at (i.e. lane change motion planning with simultaneous decision of the target lane) , evaluate whether the algorithm predicts a trajectory that is at lowest distance than the real recorded human one.
3. Assuming the algorithm only knows a lane change or a car following which has been decided, subsequent maneuver on either left/right lane change or car following has not been decided (i.e. motion planning with simultaneous decision of maneuver) , evaluate whether the algorithm predicts a trajectory that is at lowest distance than the real recorded human one.
Performance of the proposed method has been demonstrated by experimental results.
For trajectory generation, tessellation of the terminal state was over the range τ∈ [6s, 10s] at resolution 1s, and Δ
V=4m/s at resolution 1 m/s.
For each experiment, the method has been applied using Algorithm 1 for learning over 90 cases of each situation. Validation was performed over the remaining cases. Table 1 presents the number of cases per situation and the split between training and validation data. Note that the CF cases were only used for experiment 2.
Table 1 -Number of cases per situation (Left Lane Change LLC, Right Lane Change RLC and Car Following CF) and split between training and validation data.
Number of training situations | Number of validation situations | |
CF | 90 | 53 |
LLC | 90 | 45 |
RLC | 90 | 45 |
Fig. 2 presents the distributions obtained for experiment 1 of
· MinDis: the minimal distance between the real human driven trajectory and the generated trajectories (best possible case with this method) .
· MaxPro: the distance between the real human driven trajectory and the selected ones (Result obtained with the method) .
· AllDis: the distance between the real human driven trajectory and all the generated ones (Worst possible case with this method) .
The distances used are as presented in part Distance measure with λ=1.
It can be seen that the MaxPro is at lower distances, meaning that the obtained selected trajectories are close to the human driven ones.
Tables 2 and 3 present the prediction results for experiments 2 and 3 respectively, over the training and the validation data. It can be noted that the results are promising with rates higher than 90%on target lane prediction when lane change occurs and around 70 % (min is 64.2%) on decision prediction (LC or CF) .
Table 2 –Results obtained with experiment 2
Table 3 –Results obtained with experiment 3
Figs. 3 and 4 present the distributions obtained for experiments 2 and 3 respectively. While results of experiment 2 are clustered at lower distances, meaning that the obtained selected trajectories are close to the human driven ones, results of experiment 3 are more sparse which denotes lower prediction rates, as also observed from table 3.
Implementation
As mentioned above, the learning method is applied for determination of a cost function f that can be used for two main cases by selection of a human like trajectory:
· for human-like trajectory planning of autonomous driving vehicle; or
· for surrounding human driven vehicles trajectories prediction.
Human-like trajectory planning of autonomous driving vehicle
The trajectory planning module of autonomous driving software can use the method presented here for human-like trajectory selection. Such an algorithm would then appear like the following.
Algorithm 2: Human-like trajectory planning of autonomous driving vehicle
Require:
A “situation” S, requiring measurement of position, velocity and acceleration of self and surrounding vehicles
Ensure: Human-like trajectory planning T
HL
Generate sampled trajectories set T= {T
1, ..., T
j, ...}
For each T
j∈T do
Transform the cost to the probability p
j=prob (f (T
j) )
end for
end for
Surrounding human driven vehicles trajectories prediction
Prediction of the surrounding vehicles trajectories is essential for performance and safety of autonomous driving vehicles. One solution would be to integrate its velocity but this would lead to only short term prediction. A way to improve the prediction quality over a longer term is to take the driving situation into account such that the most probable trajectory of a surrounding vehicle is evaluated. This would lead to improve anticipation of surrounding traffic.
Note that the information required, i.e. position, velocity and acceleration of any vehicle in the surrounding of the vehicle that the prediction is made, may not be fully available.
Such an algorithm would appear like the following.
Algorithm 3: Surrounding human driven vehicles trajectories prediction
Require:
A vehicle V
i for which trajectory is to be predicted
“situation” S associated with vehicle V
i, requiring measurement of position, velocity and acceleration of vehicle V
i and its surrounding vehicles
Ensure: Prediction of trajectory T
HL
Generate sampled trajectories set T= {T
1, ..., T
j, ...}
For each T
j∈T do
Transform the cost to the probability p
j=prob (f (T
j) )
end for
end for
The proposed method has at least the following advantages:
1. Similarity of trajectories are measured on both path and velocity profiles (not only on features) , and learns a cost function that favors on the trajectories similar to the real one.
2. The method generates and analyzes trajectories with respect to each driving situation, and the feature correlations are considered (i.e. trajectories cannot be isolated from its driving situation) .
It will be understood that there are numerous modifications of the illustrated embodiments described above which will be readily apparent to one skilled in the art, such as many variations and modifications as for the structure of the device for guiding and the assistant component. These modifications and variations fall within the scope of the claims, which follow.
Claims (15)
- A method for providing a vehicle trajectory prediction, comprising:recording, at a processor, a human driven sample H i consisting of driving situation S i at a certain time and human driven trajectory T GT i from the certain time to a real terminal time or in a fixed duration;generating, at the processor, a set of sampled trajectories T i at the driving situation S i using a tessellation of terminal states to describe potential candidates of future motion sequences;calculating, at the processor, a distance measure between a sampled trajectory and the human driven trajectory at the driving situation;calculating, at the processor, a cost function f of the sampled trajectory using a set of coefficients;converting, at the processor, the cost function f to probability with a Softmax function; andcalibrating, at the processor, the cost function weight for minimizing the total distance of all trajectories combined with the probability of the trajectory adopted with gradient descent.
- The method according to claim 1, characterized in that the human driven sample is recorded in Frenet frame with the origin at the ego vehicle’s location at the initial time, in which the driving situation is defined based on the states of the ego and environmental vehicles in local surroundings.
- The method according to claim 1 or 2, characterized in that the human driven trajectory T GT i is a time series of trajectory points in view of a variable duration in case of lane change and a fixed duration in case of car following.
- The method according to anyone of the preceding claims, characterized in that the distance between each pair of the sampled trajectory and the human driven trajectory are estimated on both the location and velocity components.
- The method according to anyone of the preceding claims, characterized in that the set of coefficients of the cost function f are associated with comfort, efficiency, lane incentive and safety of a trajectory.
- The method according to claim 5, characterized in that the comfort of a trajectory is evaluated in view of acceleration and acceleration change ratio on longitudinal and lateral dimensions.
- The method according to claim 5, characterized in that the efficiency of a trajectory is evaluated in view of average speed of the trajectory.
- The method according to claim 5, characterized in that the lane incentive of a trajectory is evaluated in view of repulsion from the current lane and attraction of the target lane, wherein the repulsion from the current lane is evaluated using the velocity difference of the ego vehicle with its leading vehicle, and the attraction of the target lane is evaluated using the velocity difference of the ego vehicle with the preceding and back vehicles on the target lane.
- The method according to claim 5, characterized in that the safety of a trajectory is evaluated in view of distances between the ego vehicle and the surrounding vehicles at each discrete time during the course of the trajectory.
- A use of the method according to anyone of claims 1 to 9 for human-like trajectory planning of autonomous driving vehicle, characterized in that a trajectory of the vehicle is selected from a probability which is transformed from the cost function with the cost function weight generated at calibration phase.
- The use according to claim 10, characterized in that the trajectory of the vehicle is obtained in view of situation defined by the states of the ego and surroundings vehicles.
- A use of the method according to anyone of claims 1 to 9 for surrounding human driven vehicles trajectories prediction, characterized in that a trajectory of a vehicle to be predicted is selected from a probability which is transformed from the cost function with the cost function weight generated at calibration phase.
- The use according to claim 12, characterized in that the trajectory of the vehicle is obtained in view of situation defined by the states of the vehicle for which trajectory is to be predicted and surroundings vehicles.
- A vehicle trajectory prediction system arranged and operable to carry out the method according to anyone of claims 1 to 9 or the use according to anyone of claims 10 to 13.
- A processing means programmed and operable to execute instructions for carrying out the method according to anyone of claims 1 to 9 or the use according to anyone of claims 10 to 13.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/CN2018/092905 WO2020000192A1 (en) | 2018-06-26 | 2018-06-26 | Method for providing vehicle trajectory prediction |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/CN2018/092905 WO2020000192A1 (en) | 2018-06-26 | 2018-06-26 | Method for providing vehicle trajectory prediction |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2020000192A1 true WO2020000192A1 (en) | 2020-01-02 |
Family
ID=68985414
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2018/092905 WO2020000192A1 (en) | 2018-06-26 | 2018-06-26 | Method for providing vehicle trajectory prediction |
Country Status (1)
Country | Link |
---|---|
WO (1) | WO2020000192A1 (en) |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111267846A (en) * | 2020-02-11 | 2020-06-12 | 南京航空航天大学 | Game theory-based peripheral vehicle interaction behavior prediction method |
CN112286197A (en) * | 2020-11-03 | 2021-01-29 | 交通运输部公路科学研究所 | Vehicle longitudinal movement speed optimization method based on discrete time and space |
CN112631295A (en) * | 2020-12-17 | 2021-04-09 | 北京汽车研究总院有限公司 | Path planning method and device and vehicle |
CN112697146A (en) * | 2020-11-19 | 2021-04-23 | 北京电子工程总体研究所 | Steady regression-based track prediction method |
CN112721929A (en) * | 2021-01-11 | 2021-04-30 | 成都语动未来科技有限公司 | Decision-making method for lane changing behavior of automatic driving vehicle based on search technology |
CN112810630A (en) * | 2021-02-05 | 2021-05-18 | 山东大学 | Method and system for planning track of automatic driving vehicle |
CN113468698A (en) * | 2021-06-30 | 2021-10-01 | 上海理工大学 | Traffic flow model parameter calibration method facing intersection traffic trajectory discreteness |
WO2022005576A1 (en) * | 2020-06-29 | 2022-01-06 | Woven Planet North America, Inc. | Systems and methods for optimizing trajectory planner based on human driving behaviors |
US11753006B2 (en) | 2020-06-08 | 2023-09-12 | Robert Bosch Gmbh | Representing objects in a surrounding environment of a vehicle using a frenet coordinate system |
CN117775078A (en) * | 2024-02-28 | 2024-03-29 | 山西阳光三极科技股份有限公司 | Method for judging running direction of freight train in mine based on deep learning |
WO2024081742A1 (en) * | 2022-10-14 | 2024-04-18 | Motional Ad Llc | Systems and methods for autonomous driving based on human-driven data |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101118500A (en) * | 2007-08-02 | 2008-02-06 | 上海交通大学 | Software emulation method of self-determination driving vehicle running process |
CN103218489A (en) * | 2013-04-11 | 2013-07-24 | 浙江大学 | Method for simulating vehicle personalized driving characteristic based on video sample |
US20150073663A1 (en) * | 2013-09-12 | 2015-03-12 | Volvo Car Corporation | Manoeuver generation for automated driving |
US20170139423A1 (en) * | 2015-11-12 | 2017-05-18 | King Fahd University Of Petroleum And Minerals | Control system and method for multi-vehicle systems |
CN107544516A (en) * | 2017-10-11 | 2018-01-05 | 苏州大学 | Automated driving system and method based on relative entropy depth against intensified learning |
-
2018
- 2018-06-26 WO PCT/CN2018/092905 patent/WO2020000192A1/en active Application Filing
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101118500A (en) * | 2007-08-02 | 2008-02-06 | 上海交通大学 | Software emulation method of self-determination driving vehicle running process |
CN103218489A (en) * | 2013-04-11 | 2013-07-24 | 浙江大学 | Method for simulating vehicle personalized driving characteristic based on video sample |
US20150073663A1 (en) * | 2013-09-12 | 2015-03-12 | Volvo Car Corporation | Manoeuver generation for automated driving |
US20170139423A1 (en) * | 2015-11-12 | 2017-05-18 | King Fahd University Of Petroleum And Minerals | Control system and method for multi-vehicle systems |
CN107544516A (en) * | 2017-10-11 | 2018-01-05 | 苏州大学 | Automated driving system and method based on relative entropy depth against intensified learning |
Non-Patent Citations (1)
Title |
---|
KUDERER, MARKUS ET AL.: "Learning Driving Styles for Autonomous Vehicles from Demonstration", 2015 IEEE INTERNATIONAL CONFERENCE ON ROBOTICS AND AUTOMATION (ICRA, 30 May 2015 (2015-05-30), pages 2641, XP033168769, DOI: 10.1109/ICRA.2015.7139555 * |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111267846A (en) * | 2020-02-11 | 2020-06-12 | 南京航空航天大学 | Game theory-based peripheral vehicle interaction behavior prediction method |
US11753006B2 (en) | 2020-06-08 | 2023-09-12 | Robert Bosch Gmbh | Representing objects in a surrounding environment of a vehicle using a frenet coordinate system |
US11458991B2 (en) | 2020-06-29 | 2022-10-04 | Woven Planet North America, Inc | Systems and methods for optimizing trajectory planner based on human driving behaviors |
WO2022005576A1 (en) * | 2020-06-29 | 2022-01-06 | Woven Planet North America, Inc. | Systems and methods for optimizing trajectory planner based on human driving behaviors |
CN112286197A (en) * | 2020-11-03 | 2021-01-29 | 交通运输部公路科学研究所 | Vehicle longitudinal movement speed optimization method based on discrete time and space |
CN112286197B (en) * | 2020-11-03 | 2024-02-13 | 交通运输部公路科学研究所 | Vehicle longitudinal movement speed optimization method based on discrete time and space |
CN112697146A (en) * | 2020-11-19 | 2021-04-23 | 北京电子工程总体研究所 | Steady regression-based track prediction method |
CN112697146B (en) * | 2020-11-19 | 2022-11-22 | 北京电子工程总体研究所 | Steady regression-based track prediction method |
CN112631295A (en) * | 2020-12-17 | 2021-04-09 | 北京汽车研究总院有限公司 | Path planning method and device and vehicle |
CN112721929A (en) * | 2021-01-11 | 2021-04-30 | 成都语动未来科技有限公司 | Decision-making method for lane changing behavior of automatic driving vehicle based on search technology |
CN112810630B (en) * | 2021-02-05 | 2022-03-04 | 山东大学 | Method and system for planning track of automatic driving vehicle |
CN112810630A (en) * | 2021-02-05 | 2021-05-18 | 山东大学 | Method and system for planning track of automatic driving vehicle |
CN113468698A (en) * | 2021-06-30 | 2021-10-01 | 上海理工大学 | Traffic flow model parameter calibration method facing intersection traffic trajectory discreteness |
WO2024081742A1 (en) * | 2022-10-14 | 2024-04-18 | Motional Ad Llc | Systems and methods for autonomous driving based on human-driven data |
CN117775078A (en) * | 2024-02-28 | 2024-03-29 | 山西阳光三极科技股份有限公司 | Method for judging running direction of freight train in mine based on deep learning |
CN117775078B (en) * | 2024-02-28 | 2024-05-07 | 山西阳光三极科技股份有限公司 | Method for judging running direction of freight train in mine based on deep learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2020000192A1 (en) | Method for providing vehicle trajectory prediction | |
JP6931937B2 (en) | A learning method and learning device that uses human driving data as training data to perform customized route planning by supporting reinforcement learning. | |
CN112347567B (en) | Vehicle intention and track prediction method | |
CN111098852B (en) | Parking path planning method based on reinforcement learning | |
CN110386145B (en) | Real-time prediction system for driving behavior of target driver | |
CN110850854A (en) | Autonomous driver agent and policy server for providing policies to autonomous driver agents | |
US7107107B2 (en) | Predictive action decision device and action decision method | |
CN110850861A (en) | Attention-based hierarchical lane change depth reinforcement learning | |
Grigorescu et al. | Neurotrajectory: A neuroevolutionary approach to local state trajectory learning for autonomous vehicles | |
CN111145552B (en) | Planning method for vehicle dynamic lane changing track based on 5G network | |
He et al. | A human-like trajectory planning method by learning from naturalistic driving data | |
JP2022506404A (en) | Methods and devices for determining vehicle speed | |
JP2016053846A (en) | Automatic driving support system, automatic driving support method and automatic driving device | |
CN113095558B (en) | Iterative optimization multi-scale fusion vehicle speed prediction method for intelligent networked automobile | |
Kuutti et al. | End-to-end reinforcement learning for autonomous longitudinal control using advantage actor critic with temporal context | |
CN110490275A (en) | A kind of driving behavior prediction technique based on transfer learning | |
CN106708044A (en) | Full-hovering hovercraft course control method based on grey prediction hybrid genetic algorithm-PID | |
US20220203997A1 (en) | Apparatus and method for determining optimal velocity of vehicle | |
WO2020164089A1 (en) | Trajectory prediction using deep learning multiple predictor fusion and bayesian optimization | |
CN117104310A (en) | Virtual marshalling control method and system based on data-driven predictive control | |
EP4171080A1 (en) | Intelligent messaging framework for vehicle ecosystem communication | |
CN116448134A (en) | Vehicle path planning method and device based on risk field and uncertain analysis | |
Okuda et al. | Model predictive driver assistance control for cooperative cruise based on hybrid system driver model | |
Phuong et al. | Building a mathematics model for lane‐change technology of autonomous vehicles | |
Schröter et al. | A particle filter for the dynamic window approach to mobile robot control |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 18924261 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
32PN | Ep: public notification in the ep bulletin as address of the adressee cannot be established |
Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 15/03/2021) |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 18924261 Country of ref document: EP Kind code of ref document: A1 |