CN115817455A

CN115817455A - Automatic parking path planning method oriented to multiple parking space scenes

Info

Publication number: CN115817455A
Application number: CN202211398938.4A
Authority: CN
Inventors: 陈慧; 张书恺; 孙宏伟
Original assignee: Tongji University
Current assignee: Tongji University
Priority date: 2022-11-09
Filing date: 2022-11-09
Publication date: 2023-03-21

Abstract

The invention relates to an automatic parking path planning method for multiple parking space scenes, which comprises the following steps: acquiring a library position category, library position corner point coordinates and an initial pose; coordinate conversion is carried out by combining with the type of the parking space, and coordinate systems of the vehicles under parallel, oblique and vertical parking are unified; under a unified coordinate system, acquiring a rough planning path scatter sequence through a neural network; and acquiring a planned path through post-processing including simulation tracking, DWA and end smoothing on the rough planned path scatter sequence. Compared with the prior art, the method finishes planning under different library positions and different initial poses by utilizing a unified algorithm, and has the advantages of zero error of a path relative to a target pose, self-updating potential of the algorithm, low complexity of a neural network and the like.

Description

Automatic parking path planning method oriented to multiple parking space scenes

Technical Field

The invention relates to the technical field of driving assistance, in particular to an automatic parking path planning method for multiple parking space scenes.

Background

The automatic parking technology has received a high degree of attention from colleges and enterprises as a representative technology of automobile intellectualization. The automatic parking path planning algorithm with high precision performance can not only enable the vehicle to park without overlarge safety margin, is beneficial to improving the storehouse planning and increasing the land utilization efficiency, but also can be organically combined with the automatic charging technology and the like which need the vehicle to park accurately in the future.

Compared to a general automatic driving scenario, path planning for automatic parking has the following requirements: (1) the method can plan a sectional path to deal with the warehouse kneading process of multi-section parking and parallel parking; (2) the planning should achieve high precision as far as possible on the basis of ensuring the safety, so that the space is fully utilized and the planning is matched with the future charging pile technology; (3) the planning is completed by a uniform planning algorithm as much as possible, so that the algorithm is simplified; (4) with the further intellectualization of future vehicles, the algorithm itself should have the potential of self-updating, so that the algorithm can be continuously self-optimized as people do. The existing mixed A-x algorithm, RRT algorithm, lattice algorithm and artificial potential field method can not be independently applied to realize the planning of multi-section paths.

Chinese patent CN 113830079A proposes a parking path planning algorithm combining a hybrid a-star algorithm with CC curve splicing, and although the planning function can be implemented in different parking spaces, the design of the cost function of the hybrid a-star algorithm requires a large amount of early-stage calculation work, and different CC curve splicing forms still need to be designed for different parking spaces, and the unification of the planning process in different parking spaces is not completely implemented.

The chinese patent CN 114906128A performs parallel parking planning by using reinforcement learning, although the subsequent self-updating capability is retained, since the neural network is limited to be trained in a parallel parking scene, the universality of the network in other scenes is limited, and meanwhile, the training data used by the network takes 0.04m as an extension step length, which may cause the increase of the number of initial poses, the data volume may increase sharply, and the neural network is forced to be more complex to accept more data, while the more complex neural network corresponds to a longer training speed and a larger calculation amount, which may also increase the calculation burden on the vehicle-mounted system.

In summary, the current algorithm can not completely meet the above 4 requirements for a while, and it is significant to develop a set of algorithms that can meet the above requirements.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provide an automatic parking path planning method for various parking space scenes, so that the requirements are met.

The purpose of the invention can be realized by the following technical scheme:

an automatic parking path planning method oriented to multiple parking space scenes comprises the following steps:

acquiring a library position category, library position corner point coordinates and an initial pose;

coordinate conversion is carried out by combining with the type of the parking space, and coordinate systems of the vehicles under parallel, oblique and vertical parking are unified;

under a unified coordinate system, acquiring a rough planning path scatter sequence through a neural network;

and acquiring a planned path through post-processing including simulation tracking, DWA and end smoothing on the scattered point sequence of the rough planned path.

As a preferred technical solution, the method for performing coordinate transformation and unifying coordinate systems specifically comprises the steps of:

the initial pose of the vehicle under the original coordinate system is

Object pose is

Parallel parking takes a warehousing pose as a target pose; the final pose is taken as a target pose for oblique and vertical parking;

the target position posture of the vehicle in a new coordinate system is (0, pi/2) for coordinate conversion, and the new coordinate system deviates from the original coordinate system in the directions of x and y

Rotate

Obtaining an initial pose and a library position angular point under a new coordinate system according to the coordinate system conversion relation;

initial pose (x) in the new coordinate system ₀ ,y ₀ ,θ ₀ ) Initial pose under original coordinate system

And object position and posture

The coordinate conversion formula of (a) is:

in the formula (I), the compound is shown in the specification,

and

respectively is the horizontal and vertical coordinates and the angle of the initial pose in the original coordinate system;

and

respectively is the horizontal and vertical coordinates and the angle of the target pose in the original coordinate system; x is the number of ₀ 、y ₀ And theta ₀ Respectively the horizontal and vertical coordinates and the angle of the initial pose in the new coordinate system.

As a preferred technical solution, the step of obtaining the rough planning path scatter sequence through the neural network includes:

taking the current pose of the vehicle and the library position angular point after coordinate conversion as neural network input, and outputting an action combination of an expansion direction pi and an expansion curvature k;

based on the action combination output by the neural network, the vehicle carries out large-step expansion to the next state in a geometric calculation mode; until the vehicle expands from the starting pose to the target pose.

As a preferred technical solution, the training step of the neural network includes:

performing simulated learning on the neural network by using path data spliced based on RS curves;

and optimizing and exploring by using a neural network through setting different library position types and different initial poses for multiple times, acquiring data with higher return value, updating the network again, and finishing reinforcement learning.

As a preferred technical solution, the path data based on RS curve splicing is generated data, and the generation process is as follows:

defining multiple RS curve splicing forms based on the parking experience of a human driver in parallel, oblique and vertical garage positions;

obtaining the length and curvature of the RS curve through geometric calculation by utilizing the position relation between different initial poses and target poses at different library positions;

and generating position information and corresponding action classes required by training the neural network based on the RS curve.

As a preferred technical scheme, the return value R _total The function of (d) is:

R _total ＝R _{change_dir} +R _space +R _lengt +R _reach +R _colli +R _{change_cur}

in the formula:

R _{chang_dir} for the gear shift return value, the more the gear shift times, the lower the gear shift return value;

R _space the more the space is utilized, the lower the path space utilization return value;

R _lengt the path length return value is the longer the path length is, the lower the path length return value is;

R _reac if the final pose evaluation value represents whether the target pose is reached or not, a return value can be obtained within an allowable range, and if the final pose evaluation value can be fitted to the target pose with smaller error on the basis again, the return value can be additionally obtained;

R _colli the collision report value is a collision report value, and the report value is greatly reduced if the collision occurs in the expansion process;

R _{chang_cur} to report values for curvature changes, the larger the curvature change the lower the value is reported during expansion.

As a preferred technical solution, the analog tracking method comprises:

establishing a transverse control state equation by using a vehicle kinematics model, designing a front wheel corner through feedback and feedforward, considering vehicle dynamics constraint, simulating the tracking process of a vehicle to a rough planning path scatter point, and simulating a path scatter point sequence obtained by tracking through the kinematics model by small step length expansion to serve as planning data;

and the extension of the kinematic model is started from the target pose and reversely extended to the starting pose.

As a preferred solution, the DWA includes:

sampling a front wheel steering angle of the vehicle within an allowable range;

setting a fixed extension length, and comparing the return values of extension paths corresponding to different sampling results; selection value r _total The maximum value is taken as the ideal front wheel rotation angle value at the moment;

wherein the value r _total The function of (d) is defined as:

r _total ＝r _{follow_pat} +r _{avoid_obstacle}

in the formula:

r _{follow_path} representing a conformity degree return value of the rotation angle value calculated by the analog tracking, wherein the smaller the deviation between the sampling value and the rotation angle value calculated by the analog tracking is, the higher the conformity degree return value is;

r _{avoid_obstacle} and the collision return value of the vehicle after the expansion is finished is represented, and the smaller the distance is, the lower the collision return value is.

As a preferred technical scheme, the end smoothing is to perform two-stage simulation tracking by taking a centerline of an exploration termination position and a target pose as a boundary:

the first section keeps the direction of the reverse expansion and continues to expand a certain distance in a simulation tracking mode, so that the transverse deviation between the vehicle and the center line is eliminated; the second section is opposite to the first section, and the transverse deviation between the vehicle and the target pose is eliminated.

As a preferred technical scheme, when parallel parking path planning is carried out, a database kneading path scatter sequence based on a geometric circular arc curve is additionally added after a planning path is obtained through post-processing.

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

1) Uniformity: the invention can complete the planning under different library positions and different initial poses by utilizing a unified algorithm through coordinate conversion and planning under a new coordinate system.

2) Precision: according to the invention, the pose of the target is reversely expanded during path post-processing, so that zero error of the path relative to the pose of the target is realized.

3) Self-renewability: the invention uses the neural network, and the neural network has exploration capability and has an opportunity to explore more excellent data, which means that the algorithm has the potential of self-updating.

4) Low-cost realizability: the invention provides that the expansion step length of the neural network is 1m, so that more working conditions can be covered by using less data volume during training without increasing the complexity of the neural network.

Drawings

FIG. 1 is an overall schematic block diagram of the present invention;

FIG. 2 is a diagram illustrating the effect of coordinate transformation at different bin positions;

FIG. 3 is a diagram of the summary effect of different bin processes after coordinate transformation;

FIG. 4 is a schematic diagram of a parallel parking garage pose calculation method;

FIG. 5 is a schematic diagram of four RS curve splicing forms;

FIG. 6 is a schematic diagram of the calculation of the size of a straight line-arc-straight line curve;

FIG. 7 is a data diagram of a data set for mock learning based on RS curves;

FIG. 8 is a schematic view of corner points of an obstacle vehicle at different garage positions;

FIG. 9 is a schematic view of a rolling exploration process for reinforcement learning;

FIG. 10 is a schematic diagram of a vehicle performing simulated tracking of a neural network planning anchor point;

FIG. 11 is a schematic diagram of lateral deviation during simulated tracking of a vehicle;

FIG. 12 is a schematic view of the vehicle being "flush" with the anchor point;

FIG. 13 is a schematic diagram of DWA (dynamic windowing) principle;

FIG. 14 is a schematic view of a vehicle outline covering an octagon with an envelope circle;

FIG. 15 is a schematic illustration of the principle of tip smoothing;

FIG. 16 is a schematic diagram of a parallel parking plan result;

FIG. 17 is a diagram illustrating a result of an oblique parking plan;

fig. 18 is a diagram showing the results of the vertical parking plan.

Detailed Description

The invention is described in detail below with reference to the figures and specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation manner and a specific operation process are given, but the scope of the present invention is not limited to the following embodiments.

The invention provides an automatic parking path planning method for various parking space scenes, and as shown in fig. 1, the whole method is divided into a coordinate unification module and a path planning module.

It has been mentioned previously that the future automatic parking path planning algorithm should satisfy four requirements: (1) the method can plan a sectional path to deal with the warehouse kneading process of multi-section parking and parallel parking; (2) the planning should achieve high precision as far as possible on the basis of ensuring the safety so as to fully utilize the space and match with the future charging pile technology; (3) the planning is completed by a uniform planning algorithm as much as possible, so that the algorithm is simplified; (4) with the further intellectualization of future vehicles, the algorithm itself should have the potential of self-updating, so that the algorithm can be continuously self-optimized like a person. According to the requirement of point (1), the method which is inconvenient to use alone for planning multiple segments of paths, such as the mixed a-star algorithm, the RRT algorithm, the lattice algorithm, the artificial potential field method, etc., is not used, but the conceivable methods are only the traditional curve splicing proposed in chinese patent CN 113830079A, such as the planning method of CC curve splicing, and the machine learning method proposed in chinese patent CN 114906128A. According to the requirement of point (4), the traditional curve splicing method has no self-updating capability, so that the method of machine learning is most suitable. However, the algorithm limits a parallel parking scene, the network can only be applied to parallel parking, and secondly, the training of the network adopts trace point data with 5ms as an interval, so that a single working condition corresponds to a large amount of data, and a limited network structure cannot accept training data of multiple working conditions. Moreover, the trace points are obtained by network exploration completely, and in order to ensure that the exploration can be ended, a wider termination condition needs to be set, and the accuracy of the planning result cannot be ensured by the wider termination condition.

Based on the above analysis, the use of neural network can satisfy the requirements of points (1) and (4), and reinforcement learning is an effective means for obtaining excellent performance of the network. In reinforcement learning, the value of the data is determined by a reward function, and in the reward function, the degree of fit between the planning termination pose and the target pose is an extremely important item. Because a single neural network can only be trained based on one return function, and the termination condition is unique, the target poses of parallel parking, oblique parking and vertical parking are required to be uniform, so that the neural network can search data under different library positions according to a uniform standard and search according to a uniform termination condition, can be generally used for different library positions, and the algorithm uniformity mentioned in the requirement of point (3) is realized. Because of the randomness of the neural network exploration, in order to ensure that the exploration process can be smoothly ended, a tolerant termination condition needs to be set, namely the exploration is ended when the pose error with the target pose is lower than a certain range, and once the tolerant range exists, the neural network stops the exploration when the error exists, so that the precision requirement mentioned in the point (2) cannot be ensured. Since this requirement cannot be reliably completed by only relying on the neural network, the assistance of other algorithms needs to be added on the basis of the neural network.

Therefore, it can be concluded that before the neural network is applied, the target poses at different library positions need to be unified through coordinate transformation for training. Although the training can realize the unification of algorithms under different library positions, the training also means that the training object of the neural network is any initial pose under three library positions, and the data volume is inevitably huge, so that countermeasures need to be made, the data volume is reduced on the premise of ensuring the training effectiveness, and the problem of network complexity increase caused by the overlarge data volume is solved:

A. the first way is to reduce the network input dimension as much as possible, thereby reducing the number of corresponding scenes to reduce the data amount. The network input used in the chinese patent CN 114906128A is the coordinates and heading angle (x, y, θ) of the vehicle and the current steering wheel angle of the vehicle

And vehicle speed v, output as steering wheel angle variation

And acceleration a, which ensures that the planning results comply with the dynamic constraints, which can make a scenario be determined by five quantities, doubling the number of scenarios that need to be trained. Thus, the new neural network will use the vehicle coordinates and heading angles (x, y, θ), as well as the bin-corner points (x) representing the bin _k1 ,y _k1 )、(x _k2 ,y _k2 ) The outputs are the heading pi and the expansion curvature k. The above inputs are the simplest case to be able to represent different poses in different scenes.

B. The second way is to increase the exploration step size. The data used by the chinese patent CN 114906128A is track data with an interval of 5ms, which makes a single working condition correspond to hundreds of groups of data, and once the working condition increases, the data volume will increase rapidly. Therefore, the new neural network is explored by taking 1m as a step length, so that a single working condition only has about 20 groups of data at most, and the network can accept more data of different working conditions without increasing complexity.

The above two measures can relieve the training pressure of the neural network, but the defects are also obvious: firstly, a neural network is planned to be a path point which is not provided with a real-time speed and a front wheel corner track point but simply represents a pose, and a speed command and a corner command cannot be designed simultaneously; secondly, the exploration result does not consider the dynamic constraint at all, and the curvature is discontinuous; thirdly, the distance between the path points is 1m, and the path points lack information and cannot be directly used for tracking. The three points determine that the exploration of the neural network can only be a rough plan, and post-processing needs to be added to realize fine planning.

The fine programming step size can be set to 0.04m, which is approximately one code wheel grid length of the vehicle. And in order to ensure that the path can reach the target pose with zero error, the finely planned extension direction is extended from the target pose to the starting pose. Because the planning result of the neural network is not convenient to be directly modified, the method of simulation tracking is adopted. Under the condition of considering vehicle dynamics constraint, a lateral control state equation of the vehicle to a neural network planning path is established by utilizing a kinematic model, an actual path of the vehicle when the vehicle follows a path with discontinuous curvature is simulated, and a simulated path result can be used as a planning result. Meanwhile, in order to ensure that the vehicle does not collide with front and rear obstacle vehicles and boundaries in the planning process, a Dynamic Window Analysis (DWA) method is used as an aid, and the vehicle extension direction is made to coincide with the extension direction obtained by the simulation tracking calculation as much as possible on the premise that collision can be avoided. Finally, as the vehicle expands from the target pose to the initial pose, deviation possibly exists between the vehicle and the initial pose when the expansion stops, and a mechanism of end smoothing is added. So far, path post-processing is also designed, and can be well matched with the rough planning of the neural network.

Based on the above analysis, the two modules involved in the present invention will now be described in detail.

1. Coordinate unification module

The unification of the coordinates aims to enable parking target poses in various storage positions to be unified, a single neural network is convenient to train and is universal for all the storage positions, and the selection of the target poses needs to be considered. During planning, the parallel warehouse positions, the vertical warehouse positions and the oblique warehouse positions have warehouse entering processes, and the warehouse kneading part is a special process under the parallel warehouse positions, so that the warehouse kneading planning of parallel parking is not considered when the algorithm is uniformly designed, and the uniform algorithm frame is only planned to the warehouse entering pose when the algorithm frame is in the parallel warehouse positions. This means that the target pose of parallel parking should actually be the warehousing pose, not the final pose; the target poses of vertical and diagonal parking can be set as the final poses.

The coordinate transformation effects of the parking paths in the three parking spaces are shown in fig. 2. The parking coordinate systems under the three storerooms are different initially, but once the target pose is determined, coordinate conversion is carried out, and the target poses can be unified. As can be seen from fig. 2, the parking process of the vehicle at different parking spaces is started from different positions, avoids obstacles with different distribution conditions, and finally reaches the same position. This also means that the parking process of the vehicle in different parking spaces can be summarized as an obstacle avoidance problem. Fig. 3 is a summary of the above conclusions:

a. the vertical parking is sent out from the point A, and the obstacle avoidance vehicles (1) and (2) are planned to the point O;

b. the oblique parking is sent out from the point B, and the obstacle avoidance vehicles (3) and (4) are planned to the point O;

c. parallel parking is sent out from the point C, and obstacle vehicles (5) and (6) are avoided to plan to the point O;

in addition, when the vehicle is planned to the O point under different garage positions, collision with a planned boundary is avoided when the vehicle avoids an obstacle.

If the initial pose of the vehicle in the original coordinate system is

Object pose is

Then, in order to make the target pose of the vehicle under the new coordinate system be (0, pi/2), the new coordinate system should be shifted in the x and y directions compared to the original coordinate system

Rotate

Then can be constructed according to the coordinate conversion formulaInitial pose (x) standing under new coordinate system ₀ ,y ₀ ,θ ₀ ) And

and

the relationship of (1):

library corner point (x) under new coordinate system _k1 ,y _k1 )、(x _k2 ,y _k2 ) Can be found in a similar manner.

As mentioned above, parallel parking requires the calculation of the parking pose before coordinate transformation. The calculation of the parallel parking garage pose is shown in fig. 4. And the garage kneading path is simply designed into an arc curve, the vehicle rubs the garage from a point O at the termination position in the garage position to a point O' through the arc curve, and if the vehicle moves forwards from the current position with the minimum curvature radius at the moment, the point K at the most dangerous point can not collide with the front obstacle car, the current pose of the vehicle is considered as the garage entering pose. During the garage kneading process, the vehicle is ensured not to collide with the front and rear obstacle vehicles and the right boundary, namely, the delta l in the figure 4 is ensured ₁ 、Δl ₂ 、Δl ₃ Always greater than a threshold value, which may be set at 15cm to 20cm.

So far, coordinate unification is completely finished, and the important basis of subsequent algorithm implementation is.

2. Path planning module

According to the foregoing design, the path planning module includes two parts: coarse planning of the neural network and fine planning of path post-processing. Each will be described in detail below.

(1) Coarse neural network planning

The neural network needs to be trained before being used to ensure good performance, and in order to reduce the training process, the neural network can firstly obtain basic performance through simulation learning, and then further improve the performance through reinforcement learning. Therefore, there is a need to produce a data set that can be applied to mock learning.

a. Data set production

Although the CC curve splicing method used in chinese patent CN 113830079A does not have self-updating capability, the method relies on strict geometric calculation, and its performance is relatively stable. Therefore, the production of the data set can also be based on this method. Because the calculation of the CC curve is more complex and the previous design also mentions that the neural network does not need to give data points with continuous curvature, the production of the data set can reduce the requirements and the RS curve can be used to meet the requirements.

Before splicing by using the RS curve, a plurality of splicing forms are defined to perform geometric calculation. Based on the experience of a human driver parking in parallel, oblique and vertical garage, four curve forms shown in fig. 5 are defined in total, namely a straight line-arc-straight line, a straight line-arc-straight line, an arc-straight line-arc-straight line, and a straight line-arc-straight line. The crosses in fig. 5 represent the boundary points of the different shape segments.

The calculation of the path is based on a strict geometric relationship, and the calculation process is described by taking the simplest straight line-circular arc-straight line as an example, as shown in FIG. 6, the vehicle starts from the point S, and the initial coordinate (x) is _s ,y _s ,θ _s ) Parking to the end point O, the coordinate of which is (x) _o ,y _o ,θ _o ) The most collision-prone point in the whole process is C (x) _c ,t _c ). The vehicle plans a path in a straight line-circular arc-straight line manner, and an inflection point A (x) appears _a ,y _a ,θ _a ) And B (x) _b ,y _b ,θ _b ) And according to the geometric relationship, the position relationship between the points A and B and the points O and S is

Here l _BS 、l _OA The lengths of the line segments BS and OA. When course angles A and B are determined and the arc angle delta theta is determined, once the arc radius r is given, the position relation between the A position and the B position can be determined based on the arc shape

Therefore, if the radius r of the arc can ensure that the vehicle does not collide with C, l can be obtained by r _BS 、l _OA 。

Other forms of curve calculation are similar. When the arc radius r, namely the arc curvature k, is selected, in order to ensure that the curvature k can be matched with the subsequent neural network training, the value of the given curvature k is between-0.2 and 0.2, and the interval is 0.05, and the number of the curvature k is 9. The RS curve obtained by calculation according to the rule is in an implicit form as shown in a formula (2.3)

In the formula I ₁ 、l ₂ 、l ₃ Is a segment boundary along the arc length; rho ₁ 、ρ ₂ 、ρ ₃ Is the curvature value in this segment; q. q of ₁ 、q ₂ 、q ₃ The running direction of the vehicle in the path is (forward is 1, reverse is-1); pi ₁ 、π ₂ 、π ₃ The steering direction of the steering wheel of the path is shown as (the left turn is 1, and the right turn is-1); s _d Is the length along the path. The collected data points may be shaped and expanded according to equation (2.3), taking one data every 1 m.

The data should include planning data of multiple initial poses in parallel, oblique and vertical library positions, and all planned target poses are (0, pi/2), and finally the data result shown in fig. 7 can be extracted. In fig. 7, only the x and y coordinate positions of the data points are shown, but actually, the heading angle, the moving direction, the curvature and the coordinates of the angular points of the obstacle vehicle also need to be recorded together. FIG. 8 shows the corner points K of the obstacle vehicles to be recorded at different garage positions ₁ 、K ₂ . All the data recorded above should be data after coordinate conversion.

b. Imitation learning/reinforcement learning

As described aboveWill be used to model learning. Here again, each set of data set includes the position information (x, y, theta) of the vehicle and the coordinates (x) of the corner points of the obstacle vehicle at the garage position where the vehicle is located _k1 ,y _k1 ) And (x) _k2 ,y _k2 ) The vehicle motion direction pi and the vehicle running curvature kappa. The 7 inputs of the neural network are vehicle position information and garage position obstacle corner points, and the 18 classified outputs are formed by combining different vehicle motion directions and different running curvatures in pairs, wherein the motion directions are 2 forward and backward, the running curvatures are-0.2 to 0.2, and the intervals are 0.05.

Since data is extracted by a large step, the amount of data is moderate even if the operating conditions are numerous. In the invention, 10000-15000 groups of data under about 500 different working conditions are generated by changing the initial pose under three kinds of library positions. This makes the neural network architecture unnecessarily complex. The network used in the invention comprises two hidden layers except the input layer and the classification layer, and the number of the neurons is 25. 70% of the above data are training set, 15% are validation set, and 15% are test set.

The neural network has preliminary planning capability after the simulation learning, and the performance of the network is further improved by the reinforcement learning. The goal of reinforcement learning is to obtain more excellent data sequence(s) by exploring the existing strategy _i ,a _i ) I =0,1.. N-1} and stores, and finally, the network is updated again by using the data sequences stored under different working conditions, so that the network performance is further improved. Wherein s is _i Representing the ith state, including 7 data of the coordinates, the course angle and the obstacle corner point of the vehicle under a certain working condition, a _i Is represented in state s _i The following optimal motion set comprising the motion direction and the running curvature. S when i =0 ₀ Representing an initial state with the vehicle at s _i At state, take action group a _i Transition to the next state s _i+1 When i = n-1, the vehicle takes action group a _n-1 Will transition to the final state s _n 。

The vehicle is extended following the kinematic model, but due to the large step (1 m) extension, the extension result of the vehicle needs to be calculated by means of geometry:

(1) when curvature kappa _i When =0

(2) When curvature kappa _i >At 0 time

(3) When curvature kappa _i <At 0 time

Return function R for reinforcement learning _total Is designed as

R _total ＝R _{change_dir} +R _space +R _length +R _reach +R _colli +R _{chang_cur} (2.7)

Each part is specifically defined as

In the above formulas:

①R _{change_dir} a representative shift return value is calculated according to the shift frequency dir _ change;

②R _space representing a space utilization reward value, x, according to the leftmost x coordinate value of the path _left The rightmost x coordinate value x _right The uppermost y coordinate value y _up The lowest coordinate value y _down Calculating;

③R _leng representing a path length report value, and calculating according to the total path length _ total;

④R _reach representing the final pose estimate, by the base return value R _{reac_basic} And additional return valueR _{reach_extra} Two parts are formed. Because the neural network adopts large-step exploration and the neural network still has a post-processing process after planning, the exploration of the network does not need to reach the final pose very accurately, and only needs to reach the pose convenient for post-processing. If the search result (x, y, θ) satisfies

Then R can be obtained _{rea_basic} =40000, if the search result can further approach the target pose on the basis, an additional return value R can be obtained _{reach_extra} Can be calculated as

R _{rea_extr} ＝20*(0.05-|x|)/0.05+100*(0.25-|θ-pi/2|)/0.25(2.10)

⑤R _colli Representing a collision return value, if the path collides with the obstacle vehicle, is _ colli is 1, otherwise, is 0;

⑥R _{change_cur} the representative curvature change return value is calculated by two consecutive steps of the maximum curvature change amount max _ curve _ change.

The optimization process is shown in fig. 9. After the library position and the initial pose are selected, a single-step MCTS is used for exploration: before the termination condition is not reached, the current state is expanded, all 18 types of action combinations are traversed to obtain 18 corresponding sub-states, then rolling exploration is carried out by combining a neural network with a roulette rule from the 18 sub-states to obtain 18 path point sequences, an action group corresponding to a sequence with the highest return value is selected as an action group of the current expansion, the next state is entered, the sequence is compared with a sequence corresponding to the maximum return value under the current library position and the initial pose, and a sequence with a larger return value is reserved; repeating the above process after reaching the next state; and after the termination condition is reached, storing the maximum return value sequence under the initial pose into a data pool. The direct effect of the above exploration is that an attempt is continuously made to update the existing sequence by using a more optimal path point sequence in the expansion process, so that the path point sequence in the current library position and the initial pose is optimal. If the process is carried out for multiple times under different library positions and different initial poses, a large amount of optimized data is stored in the data pool, and the neural network is trained again by the data, so that the network effect can be enhanced.

(2) Path post-processing fine planning

As the planning result of the neural network is discrete path points with the distance of 1m, the problem of discontinuous expansion curvature of adjacent points exists, and the discrete path points cannot be directly used for tracking, so that post-processing is needed, and the step length can be set to be 0.04m of the length of one code disc unit. Moreover, the planning of the neural network only reaches the pose convenient for post-processing and does not accurately reach the target pose, and in order to ensure the planning accuracy, the extension of the post-processing is reversely extended from the target pose to the initial pose.

a. Analog tracking

The simulation tracking is realized by assuming that a vehicle is tracking a certain path as the name implies, although the vehicle cannot completely track the path due to dynamic constraints, the path actually taken by the vehicle to track the path necessarily conforms to the dynamic constraints of the vehicle and can be really tracked, so that the actual path taken by the vehicle during the simulation tracking can be used as a final path planning result.

The path points planned by the neural network can be actually regarded as a plurality of anchor points, the vehicle only needs to track along the extension direction marked by the anchor points during analog tracking, and whether the anchor points can be accurately reached is not necessarily required. The schematic diagram of the simulated tracking is fig. 10.

Because the anchor point of the neural network comprises the coordinates and the course angle of the anchor point

Extended direction and extended curvature at this position

This makes the path between any two points planned by the neural network to be a simple circular arc, while the simulation tracking is an attempt to follow the path with discontinuous curvature, which is connected end to end by a plurality of circular arcs, under the dynamic constraint. Since the path is a circleAnd arc, so that the transverse offset and the angle error between the vehicle and the path are easy to calculate, and a transverse controller can be designed by adopting simple feedforward and feedback. The process of the vehicle following the circular arc curve is shown in fig. 11, and with the vehicle reference point at the midpoint of the rear axle of the vehicle, the controller is designed as follows:

(1) transverse deviation e ₁ Rate of change of

Deviation from vehicle speed V and angle e ₂ Correlation, combined with small angle assumptions, of

(2) Angular deviation e ₂ Rate of change of

Related to vehicle speed V, current front wheel steering angle psi of the vehicle, vehicle wheelbase L, path curvature kappa, and combined with small angle assumptions, there are

Setting the control variable u = tan ψ, there is a state equation

Establishing a control variable u and a deviation e ₁ 、e ₂ And adding a feed forward amount of

u＝-k ₁ e ₁ -k ₂ e ₂ -Lκ (2.14)

The lateral controller is thus set up for analog tracking, in which case the range of variation of the front wheel angle ψ is set in conjunction with the dynamic constraints.

It was previously mentioned that simulation of tracking does not require accurate vehicle arrival at anchor pointsAs a requirement, the expansion is performed only according to the direction of the anchor point guide. In practice, when the vehicle reaches a position "level" with the anchor point in the simulation tracking, the vehicle is considered to be capable of tracking the next arc. The "flush" condition of the vehicle with the anchor point is shown in FIG. 12, when the vehicle condition is (x, y, θ), the anchor point is

And the action group stored at the anchor point is

It is considered "flush" when the vector of the anchor point to the vehicle is at an angle of less than pi/2 to the direction of the anchor point itself.

DWA (dynamic windowing)

It should be noted that the neural network does not consider the boundary of the parking environment during planning, and in order to further ensure that the vehicle does not collide with the front and rear obstacle vehicles and other obstacles which may appear in the environment during the expansion process, an auxiliary method capable of flexibly combining the current position for obstacle avoidance is required.

The present invention adopts DWA (dynamic windowing). The principle is that possible actions are sampled, the sampling action is used for advancing for a certain distance, the sampling action is evaluated according to whether the sampling action collides with an obstacle in the advancing process, whether the sampling action accords with the expected design advancing direction and the like, and the action with the highest evaluation is selected as the action selected by advancing at this time. DWA is widely used for path planning of robots, while the present invention is directed to path planning of vehicles, so the following modifications are made when using DWA:

(1) instead of evaluating with a full forward path, only the end positions reached after a single sampling action are used for evaluation. This is because the size of the vehicle is much larger than that of a general robot, and the distance of the expansion of the sampling motion is less than half of the total length of the vehicle body, and if the vehicle collides with an obstacle during the expansion process, the end position will actually collide with the obstacle with a high probability, and the state of the end position is sufficient to represent the state of the whole route. Meanwhile, only the calculation of the termination position saves more calculation, and excessive pressure on the vehicle-mounted system is avoided.

(2) When the front wheel steering angle is sampled, only the maximum range which can be reached by the front wheel steering angle is considered, and whether the sampling action meets the rotating speed requirement under the current steering angle is not considered. The reason is that when the vehicle is parked, the environment where the vehicle is located is narrow, the vehicle always keeps a short distance from the obstacle, if the rotation speed is considered, the search range is reduced, the situation that all the vehicle collides with the obstacle when all sampling actions are expanded is probably caused, and at the moment, the vehicle needs to know whether the vehicle has a corner in which one direction can avoid the obstacle instead of only meeting the rotation speed requirement.

The schematic diagram of the DWA is shown in fig. 13. In the sampling process, the vehicle front wheel rotation angle range is +/-30 degrees, in order to avoid excessive increase of the calculated amount of the sampling action, the sampling process takes 3 degrees as an interval, 21 sampling actions are totally performed, in combination with the actual effect, the expansion of 40 x 0.04=1.6m is selected, 21 end positions are obtained, the 21 end positions are evaluated, and the action corresponding to the highest evaluation is selected as the next action.

The merit function is defined as

r _total ＝r _{follow_path} +r _{avoid_obstacle} (2.15)

Wherein:

①r _{follow_path} representing sampled values psi _c Angle phi calculated from analog tracking _p The smaller the deviation of the sampling value from the angle value calculated by the analog tracking is, the higher the value is, which is defined as

r _{follow_path} ＝-|ψ _c -ψ _p | (2.16)

②r _{avoid_obstacle} The value of the collision report value of the vehicle after the expansion is finished is calculated by using the distance dis between the vehicle and the obstacle, and the smaller the distance is, the lower the value is, which is specifically defined as

It should be noted here that the actual calculation of the distance dis between the vehicle and the obstacle is troublesome, the vehicle is originally an octagonal contour, and the calculation of the distance to other obstacles is actually the calculation of the distance between the polygon and the other polygons, which is troublesome. The invention covers the outline of the vehicle as shown in figure 14 by using 4 enveloping circles, so that the calculation of the distance between the vehicle and other obstacles can be simplified into the calculation of the distance between a point and a polygon although the calculation takes some space.

c. End smoothing

As mentioned previously, post-processing is backward explored from the target pose to the start pose. However, because the curvature of the path point of the neural network is discontinuous, the simulation tracking can not completely follow the path point; second, the use of a DWA allows the vehicle to make slight adjustments when approaching an obstacle. The two points can cause that the vehicle cannot accurately return to the initial pose when the vehicle is expanded reversely. If the difference between the reverse exploration result and the initial pose is too large, the vehicle has a large error at the initial time, so that a large turning angle value can be calculated by a follow-up path tracking module, the turning angle of a steering wheel is strongly shaken when the vehicle initially runs, and the comfort level of passengers is reduced. Therefore, an end smoothing strategy needs to be supplemented to avoid the initial error from being too large.

The steering wheel angle jitter is caused by the fact that a large transverse error exists between the end position of the reverse exploration and the target pose, and therefore the aim of the smooth end is to eliminate the transverse deviation through extra forward and backward movement by increasing the running process of the vehicle. Since in most cases the passenger will not deliberately activate the auto park function near an obstacle, the end smoothing does not take into account obstacle avoidance for the purpose of simplifying the design.

As shown in fig. 15, when the vehicle has a lateral deviation between the reverse exploration termination position and the target pose, two segments of simulation tracking can be performed by taking the central line of the vehicle and the target pose as a boundary: the first section keeps the direction of reverse exploration and eliminates the transverse deviation between the vehicle and the center line; the second section is opposite to the first section, and the transverse deviation between the vehicle and the target pose is eliminated.

Through the process, the unified frame can be utilized to complete the planning of the vehicle at different storage positions and different initial poses. Particularly, for parallel parking, the framework can only plan the warehousing-in path, and the path in the warehouse location needs to be merged with the warehousing path into the final planned path based on the geometric circular arc line. Fig. 16, 17 and 18 are diagrams of effects of vehicle planning under parallel, oblique and vertical garage positions respectively, and simulation results also show that the planned path can be completely tracked.

The foregoing detailed description of the preferred embodiments of the invention has been presented. It should be understood that numerous modifications and variations could be devised by those skilled in the art in light of the present teachings without departing from the inventive concepts. Therefore, the technical solutions available to those skilled in the art through logic analysis, reasoning and limited experiments based on the prior art according to the concept of the present invention should be within the scope of protection defined by the claims.

Claims

1. An automatic parking path planning method oriented to multiple parking space scenes is characterized by comprising the following steps:

2. The automatic parking path planning method oriented to multiple parking space scenes as claimed in claim 1, wherein the method for performing coordinate transformation and unifying the coordinate system comprises the following specific steps:

the initial pose of the vehicle under the original coordinate system is

Object pose is

Parallel parking takes a warehousing pose as a target pose; oblique and vertical parking takes the final pose as a target pose;

Rotate

And object position and posture

The coordinate conversion formula of (a) is:

in the formula (I), the compound is shown in the specification,

and

respectively representing the horizontal and vertical coordinates and the angle of the initial pose in the original coordinate system;

and

respectively representing the horizontal and vertical coordinates and the angle of the target pose in the original coordinate system; x is the number of ₀ 、y ₀ And theta ₀ Respectively the horizontal and vertical coordinates and the angle of the initial pose in the new coordinate system.

3. The automatic parking path planning method for multiple parking space scenes as claimed in claim 1, wherein the step of obtaining the rough planning path scatter sequence through the neural network comprises:

4. The automatic parking path planning method for multiple parking space scenes according to claim 3, wherein the training step of the neural network comprises:

5. The automatic parking path planning method oriented to multiple parking space scenes as claimed in claim 4, wherein the path data based on RS curve splicing is generated data, and the generation process is as follows:

6. The automatic parking path planning method for multiple parking space scenes as claimed in claim 4, wherein the report back value R is _total The function of (d) is:

R _total ＝R _{change_dir} +R _space +R _leng +R _reach +R _colli +R _{chang_cur}

in the formula:

R _{chang_dir} for the shift return value, the more the number of shifts, the lower the shift return value;

R _reach if the final pose evaluation value represents whether the target pose is reached or not, a return value can be obtained within an allowable range, and if the target pose can be fitted with smaller errors on the basis again, the return value can be additionally obtained again;

7. The automatic parking path planning method for multiple parking space scenes as claimed in claim 1, wherein the simulation tracking method comprises:

establishing a transverse control state equation by using a vehicle kinematics model, designing a front wheel corner by feedback and feedforward, considering vehicle dynamics constraint, simulating the tracking process of a vehicle to a rough planning path scatter point, and simulating a path scatter point sequence obtained by tracking by using the kinematics model in small step length expansion to serve as planning data;

8. The method for automatic parking path planning for multiple parking space scenes according to claim 1, wherein the DWA step comprises:

sampling a front wheel steering angle of the vehicle within an allowable range;

wherein the value r _total The function of (d) is defined as:

r _total ＝r _{follow_path} +r _{avoid_obstacle}

in the formula:

r _{avoid_obstacle} and the collision report value of the vehicle after the expansion is finished is represented, and the smaller the distance is, the lower the collision report value is.

9. The automatic parking path planning method for multiple parking space scenes as claimed in claim 1, wherein the end smoothing is to perform two-segment simulation tracking at the boundary of the exploration termination position and the central line of the target pose:

the first section keeps the direction of the reverse expansion and continues to expand a certain distance in a simulation tracking mode, and the transverse deviation between the vehicle and the center line is eliminated; the second section is opposite to the first section, and the transverse deviation between the vehicle and the target pose is eliminated.

10. The automatic parking path planning method oriented to multiple parking space scenes according to claim 1, characterized in that during parallel parking path planning, a garage kneading path scatter-point sequence based on a geometric circular arc curve is additionally added after a planning path is obtained through post-processing.