A method for realizing emergency navigation and positioning of an autonomous driving vehicle

Through reinforcement learning algorithms and vehicle state estimation, the problem of failure of navigation and positioning equipment of autonomous vehicles is solved, and the safe navigation and positioning of vehicles is realized without relying on redundant equipment and external communications is achieved, reducing costs and improving Lu Awesome and accuracy.

CN115158342BActive Publication Date: 2024-11-08YANGZHOU UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210913249.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-07-29
Publication Date
2024-11-08
Estimated Expiration
2042-07-29

AI Technical Summary

Technical Problem

When the navigation and positioning equipment fails, autonomous vehicles cannot ensure safe driving. The existing technology relies on redundant sensors, roadside equipment and networked communication equipment, resulting in high cost, complexity and robustness.

Method used

Using reinforcement learning algorithms and vehicle state estimation, road information is obtained through front-end visual sensors, Kalman filtering and long-term memory network algorithms are used to estimate vehicle state, and navigation mapping functions and reward functions are designed to realize vehicle autonomous navigation and obstacle avoidance, avoid redundancy. Device and external communication.

Benefits of technology

When the navigation and positioning equipment fails, ensure that the vehicle can drive safely, reduce costs, and improve the robustness and accuracy of navigation and positioning, which is suitable for complex road environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115158342B_ABST
    Figure CN115158342B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for realizing emergency navigation and positioning of an autonomous driving vehicle, and obtains t i The velocity set, posture angle set, and position information set at the moment; according to t i The velocity set, posture angle set, and position information set estimate at time t j The velocity set, posture angle set, and position information set at the moment; get t j The obstacle set at the moment, t j The speed set, posture angle set, position information set, and obstacle set at each moment are used as the state space of reinforcement learning, and the steering angle of the vehicle steering wheel is used as the action space of reinforcement learning. The navigation mapping function is optimized by reinforcement learning to obtain a new action space and state space for the interaction between the vehicle and the environment. The new action space and state space are used for vehicle control. When the state of the vehicle cannot meet its estimated range value, reinforcement learning is stopped, the vehicle automatically stops, and the navigation positioning function stops. The present invention ensures that the vehicle can drive safely by relying on its own vehicle algorithm characteristics when all navigation positioning sensors fail.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method for realizing emergency navigation and positioning of an automatic driving vehicle, and belongs to the technical field of intelligent vehicles. Background Art

[0002] The navigation and positioning technology of autonomous vehicles is the key technology to realize the real-time autonomous position and posture of the vehicle. Today's navigation and positioning technology is generally based on the GPS / IMU (inertial measurement unit) inertial navigation and positioning mode, assisted by visual SLAM (State Estimation in Robotics) and lidar SLAM to realize the navigation and positioning function of the vehicle. However, GPS mainly relies on satellite and radio communication technology, and its receiving end is easily degraded by high buildings in the city, and is also prone to signal loss due to tunnel scenes. Even when the weather is rainy, snowy, and thick clouds, the signal is degraded and lost, resulting in loss of navigation and positioning function; the inertial navigation element IMU leads to a large cumulative error as the driving distance increases, causing the vehicle to deviate far from the standard position.

[0003] LiDAR SLAM and visual SLAM technologies are relatively complex, and the components are relatively sensitive; when GPS / IMU, LiDAR SLAM and visual SLAM sensors fail at the same time, the vehicle's navigation and positioning performance disappears. There are also some solutions to solve the problem of sensor failure, such as adding redundant sensor equipment. This technology can offset the failure of other sensors to some extent. However, redundant sensor equipment not only increases the functions of electronic control equipment, but also makes the electronic and electrical architecture extremely complex. The system cannot ensure robustness and the cost is high. Even if additional redundant equipment is added, the safety of the system cannot be ensured. If the additional sensor equipment fails again, the system cannot support the redundant sensor equipment. Therefore, adding redundant sensor equipment cannot ensure the safety and robustness of the system.

[0004] Another positioning method is achieved through roadside equipment. The inability to ensure accurate positioning technology on the roadside equipment side has become a major difficulty restricting the industrialization of roadside equipment. In particular, in severe weather conditions, the instability and uncertainty of roadside equipment are increased. The maintenance time after a failure will affect the real-time positioning and safe driving of the vehicle. The price cost is high and it is difficult to promote.

[0005] The third auxiliary positioning method is through the networked communication mode, but the networked communication equipment cannot ensure the robustness, real-time and security in various environments, so it is not the best alternative.

[0006] Therefore, navigating an autonomous vehicle in the event of sensor failure is a technical problem that technical personnel in this field urgently need to solve. Summary of the invention

[0007] Purpose: In order to overcome the deficiencies in the prior art, the present invention provides an emergency navigation and positioning implementation method for an autonomous driving vehicle. Based on the situation of sensor failure, the vehicle body is used as the main research method to carry out research on positioning. The system can realize the navigation and positioning function of the vehicle without relying on any redundant sensor equipment, roadside equipment and networked communication equipment, so as to ensure safe driving in a relatively short period of time in the future and finally stop in a safe road environment.

[0008] Technical solution: To solve the above technical problems, the technical solution adopted by the present invention is:

[0009] A method for realizing emergency navigation and positioning of an autonomous driving vehicle comprises the following steps:

[0010] Get i The velocity set, posture angle set, and position information set at each moment, i=0,…,n.

[0011] According to t i The velocity set, posture angle set, and position information set estimate at time t j The velocity set, posture angle set, and position information set at each moment, j=n+1,…,n+N.

[0012] Get i The obstacle set at time t i The velocity set, posture angle set, position information set, and obstacle set at each moment are used as the state space of reinforcement learning. The steering angle of the vehicle steering wheel is used as the action space of reinforcement learning. Reinforcement learning is used to optimize the navigation mapping function to obtain a new action space and state space for the interaction between the vehicle and the environment. The new action space and state space are used for vehicle control.

[0013] When the vehicle's state cannot meet its estimated range value, reinforcement learning is stopped, the vehicle stops automatically, and the navigation and positioning function stops.

[0014] As a preferred solution, according to t n The speed state at the moment is v n , according to the Kalman filter algorithm to estimate t j The speed set at the moment is calculated as follows:

[0015]

[0016] In the formula, v n , v n+1 , …, v n+N is the estimated velocity, u n+1 ,u n+2 ,…,u n+Nis the input value of the system; A is the system state matrix, and B is the control matrix.

[0017] As a preferred solution, according to t i The pose angle set at time t and the position information set are estimated by the long short-term memory network algorithm i The pose angle set and position information set at the moment.

[0018] As a preferred solution, the navigation mapping function calculation formula is as follows:

[0019] δ=f(v β p(xy) obs)

[0020] Among them, p(xy) is the position information, obs is the obstacle, v is the speed information, β is the posture angle information, δ is the steering angle of the vehicle steering wheel, and f( ) is the exploration process of reinforcement learning from the vehicle state space to the action space.

[0021] As a preferred solution, the reward function of reinforcement learning includes: a reward value for tracking the vehicle position point. If there is a reference point in front of the vehicle within the lane line range, a positive reward value r is given to the vehicle. tracking , other reward values ​​are 0; if the vehicle collides with an obstacle, it will get a negative reward value, if it safely avoids the obstacle, it will get a positive reward value; if the estimated speed state satisfies the lane line range or the obstacle avoidance state estimation, a positive reward value is given, otherwise 0 is given.

[0022] As a preferred solution, when the state of the vehicle cannot meet its estimated range value, reinforcement learning is stopped and the vehicle automatically stops, including:

[0023] The front-end visual sensor recognizes the lane line to ensure that the vehicle is driving on the lane line. When the vehicle's posture angle deviates from its standard route angle and exceeds the posture angle limit, the estimation of the vehicle speed state is stopped and the vehicle is stopped through wire control braking.

[0024] As a preferred solution, the attitude angle limit values ​​include:

[0025] The horizontal driving route of the vehicle is taken as the standard reference line and set as the standard posture angle, which is defined as β s If the estimated posture angle of the vehicle continues to deviate from the upward or downward angle of the reference line for a period of time in the future, the limit angle of the vehicle is set to the angle at which the vehicle and the lane lines on both sides of the road contact each other instantaneously. At this time, the posture angle of the vehicle is set to the posture angle limit value.

[0026] As a preferred solution, when the state of the vehicle cannot meet its estimated range value, reinforcement learning is stopped and the vehicle automatically stops, including:

[0027] The front-end visual sensor recognizes the lane lines to ensure that the vehicle is driving on the lane lines. When the vehicle's position deviates from its standard route and exceeds the limit, the estimation of the vehicle speed state is stopped and the vehicle is stopped through wire control braking.

[0028] As a preferred solution, the extreme positions include:

[0029] The horizontal driving route of the vehicle is used as the standard reference line and the center line of the starting road is set as the center line. If the estimated position of the vehicle continues to deviate from the upward or downward position of the reference line for a period of time in the future, the limit position of the vehicle is set to the position where it instantly contacts the lane lines on both sides of the road.

[0030] As a preferred solution, t j The velocity set, posture angle set, and position information set at each moment also include:

[0031] In t j The speed set, posture angle set, and position information set at the moment are respectively increased with the vehicle speed state correction value, the posture angle state correction value, and the vehicle position correction value.

[0032] As a preferred solution, it also includes: training the neural network of reinforcement learning, the steps are as follows:

[0033] Get i The velocity set, posture angle set, position information set, and obstacle set at each moment are stored in the experience pool as samples.

[0034] Randomly extract batch samples from the experience pool to optimize the neural network parameters for reinforcement learning.

[0035] When the state of X deviates from the actual state by less than a threshold, the training is stopped and a trained reinforcement learning neural network is obtained.

[0036] Said

[0037] Among them, x, y, δ are the lateral coordinates, longitudinal coordinates and steering angle of the vehicle; v is the vehicle speed; represents the first derivative of the transverse coordinate, Represents the first derivative of the longitudinal coordinate.

[0038] Beneficial effects: The present invention provides a method for implementing emergency navigation and positioning of an autonomous driving vehicle. When the navigation and positioning equipment of the vehicle fails, without adding any redundant navigation and positioning equipment and without relying on any roadside navigation and positioning equipment and communication equipment, the method can ensure that the vehicle can still perform accurate navigation and positioning tasks on the road by adopting vehicle state estimation and reinforcement learning algorithms, while taking into account the vehicle's kinematic characteristics to better serve the accuracy of the estimated vehicle state for navigation and positioning.

[0039] The present invention can be used for emergency navigation and positioning tasks of vehicles when navigation and positioning fail instantaneously, ensuring that the vehicle can drive safely by relying on its own vehicle algorithm characteristics when all navigation and positioning sensors fail. The present invention not only saves costs, but also helps to better use it for navigation and positioning tasks in many complex road environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Overall schematic diagram of the method for implementing emergency navigation and positioning for autonomous vehicles.

[0041] Figure 2 Schematic diagram of vehicle speed state stop estimation.

[0042] Figure 3 Schematic diagram of vehicle speed state, posture angle and position state estimation and correction. DETAILED DESCRIPTION

[0043] The present invention will be further described below in conjunction with specific embodiments.

[0044] A method for realizing emergency navigation and positioning of an autonomous driving vehicle, wherein the specific implementation steps are as follows:

[0045] Data preprocessing:

[0046] Defined in t n At all times, the vehicle relies on GPS / IMU, lidar SLAM and visual SLAM sensors to obtain vehicle speed, posture angle information, road characteristics, vehicle characteristics and environmental semantic information. The road environment on which the vehicle is traveling does not install any roadside sensor communication equipment, and does not rely on any external communication equipment. All positioning systems are completed by their own on-board sensors.

[0047] Data normalization. In order to eliminate the dimensionality problem between vehicle status data, the minimax normalization is used to normalize the posture angle information state, limiting it to a certain range and increasing the data standardization of the calculation; the vehicle speed state does not need to be normalized, and the specific vehicle speed plan is set according to the actual road environment.

[0048] Define the vehicle at t n The state set at the previous moment is:

[0049]

[0050] In the formula, Yes n and the velocity set at previous times, Yes nand the pose angle sets at previous moments. They mainly use GPS / IMU sensors to complete data acquisition; LiDAR SLAM and visual SLAM sensors do not intervene in the acquisition of vehicle speed and pose angles. Their main task is to realize the map creation task of road features, vehicle features and environmental semantic information, and match the fused data of the two to update the environmental information in real time.

[0051] Estimation of vehicle speed state. Assume that at t n When the vehicle navigation and positioning sensor equipment such as GPS / IMU inertial navigation and positioning mode, LiDAR SLAM and visual SLAM fails, t n+1 At this moment, the vehicle navigation positioning sensor device cannot output positioning data.

[0052] First, the visual sensor is used to perceive the front-end road environment, such as 150 meters ahead. The perception layer sends the perceived front-end road data information to the server of the operating system. After the navigation positioning layer receives the information about the road environment from the perception layer, it uses the Kalman filter algorithm to perform t n+1 Vehicle speed status at time v n+1 The estimate of v n+1 is the initial speed state, and t n+2 Vehicle speed status at time v n+2 The estimated speed state and the estimated speed state are marked as The sum of the vehicle speeds is denoted by v:

[0053]

[0054] The specific estimation process is as follows:

[0055] Set t n The vehicle speed state at the moment is v n , then according to the Kalman filter algorithm formula, we can get:

[0056]

[0057] In the formula, v n , v n+1 , …, v n+N is the estimated vehicle speed, u n+1 ,u n+2 ,…,u n+N is the input value of the system; A is the system state matrix, and B is the control matrix.

[0058] When the vehicle speed estimation state is disturbed by external factors, the estimated state equation can be changed and set to the following formula:

[0059]

[0060] In the formula, χ n It is external interference factors.

[0061] When the vehicle is in v j When the vehicle is disturbed by external factors, the state v is estimated according to the vehicle speed before the disturbance. j+1 , and the estimated speed state v after the vehicle is disturbed j+1 +χ n When the vehicle is disturbed by external factors and has no time to make an estimated response, in order to better set the accurate estimation of the vehicle speed, a threshold range is added each time the vehicle speed is estimated to determine the elasticity of the estimated vehicle speed, which can be set to v j+1 +σ, σ is the threshold range, usually set to a minimum value.

[0062] The vehicle's pose angle and vehicle position are estimated. n The vehicle’s posture angle information and vehicle position information are recorded at all times and input into the long short-term memory network algorithm (LSTM) to predict the continuous time t n+1 The posture angle and position state of the adjacent continuous moments are predicted by analogy, and the accumulated position and posture angle state of the vehicle in the future is used as the reference path point and azimuth. According to the estimated speed state and the reference path point and azimuth, a new reference path in the future is estimated. The specific steps are described as follows:

[0063] For the predicted vehicle pose angle, set t n At time t and before, the vehicle's posture angle is β, then its historical posture angle can be expressed as:

[0064] β={(β1 t1) (β2 t2) … (β n t n )}

[0065] The encoded LSTM historical data is expressed as:

[0066] e t =FC(LSTM(β n ))

[0067] FC() represents the encoding function.

[0068] The historical data of the decoded LSTM is expressed as:

[0069] e t =LSTM(e t )

[0070] The output predicted pose angle is β pre ={(β n+1 t n+1 ) (βn+2 t n+2 ) … (β n+N t n+N )}, and solve the loss function:

[0071] L a =10·(β pre -β) 2

[0072] For predicting the vehicle’s position information, set t n At and before time, the vehicle's location information is:

[0073] p={(x1 y1),t1 (x2 y2),t2 … (x n y n ),t n},

[0074] Its time series historical position information can be expressed as:

[0075] p={(x1 y1),t1 (x2 y2),t2 … (x n y n ),t n}

[0076] Similarly, encoding LSTM and decoding LSTM are used for historical location information to output the predicted vehicle location information:

[0077] p pre ={(x n+1 y n+1 ),t n+1 (x n+2 y n+2 ),t n+2 … (x n+N y n+N ),t n+N}

[0078] Solve the loss function:

[0079] L p =10·(p pre -p) 2

[0080] In the formula, e t Indicates the information encoded by the output pose angle, L a Represents the loss function of pose angle solution, L p Represents the loss function for solving the vehicle position.

[0081] Vehicle speed state stop estimation. The stop estimation of the vehicle speed state is derived from the degree of deviation of the posture angle and position state. The vehicle's driving path can first be obtained by the front-end visual sensor's recognition of the lane line to ensure that the vehicle is driving on the lane line. When the vehicle's posture angle deviates from its standard route angle to a certain extent, it means that the vehicle's heading angle deviates from the safe range of the lane line. At this time, the system automatically stops estimating the vehicle speed state and stops the vehicle through the wire control brake system.

[0082] When the vehicle position state deviates from its standard position limit, the estimation of the vehicle speed state is immediately stopped and the vehicle is stopped by wire control, such as Figure 2 shown.

[0083] Standardized description of vehicle posture angle and position state. The posture angle and position state of the vehicle determine the heading angle and position state of the autonomous vehicle. Their limit states should be set to prove that the vehicle's posture angle estimation, position angle estimation, and speed estimation are all within a specific limit angle range.

[0084] First, the setting of the attitude angle limit is as follows:

[0085] Taking the horizontal driving route of the vehicle as the standard reference, the standard posture angle is set, which is defined as β s If the estimated posture angle of the vehicle continues to deviate from the upward or downward angle of the reference line for a period of time in the future, then the limit angle of the vehicle is set to the angle at which the vehicle and the lane lines on both sides of the road contact each other instantaneously. At this time, the posture angle of the vehicle is set to the limit angle, and the estimation of the posture angle is stopped when the limit angle is exceeded.

[0086] Assume t n+1 The pose angle at the moment is β n+1 , t n+2 The pose angle at the moment is β n+2 , t max The moment the posture angle reaches the limit angle is recorded as β max , then the degree to which the vehicle's posture angle deviates from its standard reference value can be mathematically expressed as:

[0087]

[0088] In the formula, is the impact factor, β error is the posture angle deviation.

[0089] When β max ≥β error , which means that the vehicle's heading angle deviates from the standard reference value, and the estimation of the vehicle speed is automatically stopped at this time.

[0090] The safe posture angle of the vehicle is:

[0091] 0≤β n+i ≤|β max |(i=1,…,N)

[0092] Similarly, the vehicle's position prediction is also based on the vehicle's horizontal driving route as the standard reference line, which is set as the center line of the starting road. If the vehicle's estimated position state continues to deviate from the upward or downward position of the reference line for a period of time in the future, then the vehicle's limit position is set to the predicted vehicle limit position and the lane lines on both sides of the road are in instantaneous contact with each other. At this time, the vehicle's position is set to the limit position, and the estimation of the position state is stopped when the vehicle exceeds the limit position.

[0093] Assume t n+1 The vehicle position at the moment is p(x n+1 y n+1 ), t n+2 The position at the moment is p(x n+2 y n+2 ), t max The vehicle position reaches the limit position at the moment recorded as p(x max y max ), then the vehicle's position status can be expressed as:

[0094] p(x n+i y n+i )≤|p(x max y max )|(i=1,…,N)

[0095] y n+i ≤|y max |

[0096] Estimation and correction of vehicle speed state, posture angle and position state. Assume that the vehicle and the vehicle ahead on the road are both traveling at normal speeds. The vehicle senses the vehicle ahead on the road through a visual sensor and obtains the speed, posture angle and position information of the vehicle ahead as a reference for recording the standardization of its own vehicle. When the estimated speed, posture angle and position path of the own vehicle exceeds or is far lower than the speed, posture angle and position of the vehicle ahead, the speed needs to be corrected to compensate for the error in the estimated speed state, such as Figure 3 shown.

[0097] Assume that at t n At the moment, the speed, posture angle and vehicle position information of the vehicle in front obtained by the visual sensor are: f = {v fn β fn (x fn y fn )}.

[0098] The vehicle uses the visual sensor to take the state of the vehicle in front as a standard reference for the driving of its own vehicle, that is, to ensure that both the front and rear vehicles are driving normally and are within the same lane line at the same time.

[0099] The vehicle is at t n The estimated state at each moment is expressed as:

[0100] s={v n β n (x n y n )}

[0101] Solve for s and s f The error between the two states is expressed as:

[0102] L err =(ss f )

[0103] When L err >0&L err <0, indicating that the estimated vehicle state deviates from the standard reference value. At this time, the vehicle speed state needs to be corrected to compensate for the state estimation error of the vehicle speed. Then, the correction value is added to the estimated vehicle state, and the result is:

[0104] s f-x = {v n +τ fn β n +μ fn (x n y n )+ρ fn}

[0105] 0≤τ fn <|v n -v fn |,0≤μ fn <|β n -β fn |,0≤ρ fn <|(x n -x fn y n -y fn )|

[0106] In the formula, τ fn is the vehicle speed state correction value, μ fn is the posture angle state correction value, ρ fn is the vehicle position correction value, s f-x is the corrected vehicle state estimate, L err is the vehicle state error value; usually, τ fn , μ fn, fn The numerical value takes the minimum value and the minimum value can be selected within the fine-tunable range.

[0107] Navigation Algorithm:

[0108] Construct a navigation mapping function from vehicle status and environmental information to vehicle decision planning to ensure that the decision planning layer can obtain effective data from the navigation layer to improve the real-time transmission and utilization of data channels. Considering the failure of sensor equipment, in order to improve the same-frequency interaction between the vehicle and the dynamic environment, a reinforcement learning algorithm is used to achieve the interactivity of the navigation algorithm. The constructed navigation mapping function is:

[0109] δ=f(v β p(xy) obs)

[0110] p(xy) is the position state of the vehicle, obs is the obstacle set, v is the estimated vehicle speed state, β is the estimated pose angle information, and δ is the steering wheel angle. The formula f( ) expresses the exploration process of reinforcement learning from the vehicle state space to the action space, where the state space contains the vehicle's position information, speed, pose angle, and obstacle information, and the action space contains the vehicle's steering wheel angle.

[0111] Design a reward function. Design a reward value for tracking the vehicle position point. If there is a reference point in front of the vehicle within the lane line, give the vehicle a positive reward value r tracking , other reward values ​​are 0; if the vehicle collides with an obstacle, it will get a negative reward value, if it safely avoids the obstacle, it will get a positive reward value; if the estimated vehicle speed state meets the above state estimation, a positive reward value is given, otherwise 0 is given.

[0112] The obtained samples are stored in the experience pool, and batch samples are randomly extracted from the experience pool to optimize the parameters of the reinforcement learning neural network to ensure that the vehicle can interact with the environment to obtain new state and action pairs. When the vehicle state cannot meet its estimated range value, the training is stopped. At this time, the autonomous driving vehicle automatically stops and the vehicle navigation and positioning function stops. The network model used in the training is a fully connected layer with four sub-hidden layers, and the neuron activation function of the steering angle action is the Tanh function.

[0113] In order to verify the navigation and positioning accuracy of the reinforcement learning algorithm output, a vehicle kinematic model based on posture angle and position state is constructed, and the output action value δ steering wheel angle is substituted into the vehicle kinematic model to verify whether the δ value meets the requirements of the actual vehicle kinematic characteristics.

[0114] The specific vehicle kinematic model can be expressed as:

[0115]

[0116] Where: x, y, δ are the lateral coordinates, longitudinal coordinates and steering angle of the vehicle; v is the vehicle speed; represents the first derivative of the transverse coordinate, Represents the first derivative of the longitudinal coordinate.

[0117] When the desired state of X deviates greatly from the actual state, the network needs to be retrained until the appropriate vehicle state is solved.

[0118] When there are many obstacles in the road environment, it may affect the position of the vehicle. First, the vehicle's front-end visual sensor is used to obtain the obstacle position and attitude angle state of the road environment. Then, the state estimation of the vehicle itself is adjusted based on the obtained obstacle position and attitude angle state. The position information of the vehicle itself is continuously evolved, and reinforcement learning is used to continuously interact with the surrounding dynamic environment and obstacles to achieve the vehicle's avoidance task, which can ensure that the vehicle's navigation and positioning can operate safely.

[0119] The unevenness of the road environment will affect the accuracy of the vehicle's position, thereby affecting the accuracy of the reinforcement learning algorithm. First, a visual sensor is used to obtain the unevenness information and obstacle information of the road environment. In order to prevent the vehicle from positionally shifting when passing through uneven roads, some standard virtual vehicle position points are set on the uneven roads. If the vehicle deviates from the established track, the standard virtual vehicle position points can be used as the vehicle's final position state and input into the reinforcement learning model for training, thereby obtaining a more accurate navigation and positioning function under reinforcement learning.

[0120] Based on the failure of the navigation and positioning sensor of the vehicle, the present invention proposes to use vehicle state estimation and reinforcement learning algorithm to realize the emergency braking mode of navigation and positioning, and use the long short-term memory network algorithm to predict the vehicle's posture and position information to determine the driving state of the vehicle in the future, and by designing a variety of constraint limit states and vehicle kinematic characteristics, determine the accuracy of vehicle state estimation, which is used to realize the task of emergency navigation and positioning of the vehicle. This technical invention can not only solve the situation of failure of the navigation and positioning equipment of the vehicle in a complex road environment, but also help save costs and make it more convenient to operate.

[0121] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for realizing emergency navigation and positioning of an autonomous driving vehicle, characterized in that: The steps include: Get i The velocity set, posture angle set, and position information set at each moment, i=0,…,n; According to t i The velocity set, posture angle set, and position information set estimate at time t j The velocity set, posture angle set, and position information set at the moment, j = n+1,…,n+N; Get j The obstacle set at time t j The speed set, posture angle set, position information set, and obstacle set at each moment are used as the state space of reinforcement learning. The steering angle of the vehicle steering wheel is used as the action space of reinforcement learning. The navigation mapping function is optimized using reinforcement learning to obtain a new action space and state space for the interaction between the vehicle and the environment. The new action space and state space are used for vehicle control. When the vehicle's state cannot meet its estimated range value, reinforcement learning stops, the vehicle automatically stops, and the navigation and positioning function stops; According to t n The speed state at the moment is v n , according to the Kalman filter algorithm, t is estimated j The speed set at the moment is calculated as follows: In the formula, v n+1 ,…,v n+N is the estimated velocity, u n+1 ,u n+2 ,…,u n+N is the input value of the system; A is the system state matrix, B is the control matrix, v n Yes n The speed of the moment; According to t i The pose angle set at time t and the position information set are estimated by the long short-term memory network algorithm j The pose angle set and position information set at the moment; The calculation formula of the navigation mapping function is as follows: δ=f(v β p(xy) obs) Among them, p(xy) is the position information, obs is the obstacle, v is the speed information, β is the pose angle information, δ is the steering angle of the vehicle steering wheel, and f() is the mapping function of the exploration process from the vehicle state space to the action space in reinforcement learning; The reward function of reinforcement learning includes: the reward value for tracking the vehicle position point, if there is a reference point in front of the vehicle within the lane line range, then give the vehicle a positive reward value r tracking , other reward values ​​are 0; if the vehicle collides with an obstacle, it will get a negative reward value, if it safely avoids the obstacle, it will get a positive reward value; if the estimated speed state satisfies the lane line range or the obstacle avoidance state estimation, a positive reward value is given, otherwise 0 is given.

2. The method for realizing emergency navigation and positioning of an autonomous driving vehicle according to claim 1, characterized in that: When the vehicle's state cannot meet its estimated range value, reinforcement learning is stopped and the vehicle stops automatically, including: The front-end visual sensor recognizes the lane line to ensure that the vehicle is driving on the lane line. When the vehicle's posture angle deviates from its standard route angle and exceeds the posture angle limit, the estimation of the vehicle speed state is stopped and the vehicle is stopped through wire control braking.

3. The method for realizing emergency navigation and positioning of an autonomous driving vehicle according to claim 2, characterized in that: The pose angle limits include: The horizontal driving route of the vehicle is taken as the standard reference line and set as the standard posture angle, which is defined as β s If the estimated posture angle of the vehicle continues to deviate from the upward or downward angle of the reference line for a period of time in the future, the limit angle of the vehicle is set to the angle at which the vehicle and the lane lines on both sides of the road contact each other instantaneously. At this time, the posture angle of the vehicle is set to the posture angle limit value.

4. The method for realizing emergency navigation and positioning of an autonomous driving vehicle according to claim 1, characterized in that: When the vehicle's state cannot meet its estimated range value, reinforcement learning is stopped and the vehicle stops automatically, including: The front-end visual sensor recognizes the lane lines to ensure that the vehicle is driving on the lane lines. When the vehicle's position deviates from its standard route and exceeds the limit, the estimation of the vehicle speed state is stopped and the vehicle is stopped through wire control braking.

5. The method for realizing emergency navigation and positioning of an autonomous driving vehicle according to claim 1, characterized in that: The extreme positions include: The horizontal driving route of the vehicle is used as the standard reference line and the center line of the starting road is set as the center line. If the estimated position of the vehicle continues to deviate from the upward or downward position of the reference line for a period of time in the future, the limit position of the vehicle is set to the position where it instantly contacts the lane lines on both sides of the road.

6. The method for realizing emergency navigation and positioning of an autonomous driving vehicle according to claim 1, characterized in that: t j The velocity set, posture angle set, and position information set at the moment also include: j The speed set, posture angle set, and position information set at the moment are respectively increased with the vehicle speed state correction value, the posture angle state correction value, and the vehicle position correction value.