A Path Planning Learning Method for Autonomous Vehicles

Abstract

The application provides a path planning learning method of an automatic driving vehicle, comprising the following steps: collecting images of targets around the vehicle through a camera and a radar of the vehicle, and classifying the targets around the vehicle into left, middle and right regions according to the images; establishing a safety degree calculation model, and calculating the driving safety degrees of the targets in the left, middle and right regions according to the speed and position of the vehicle; obtaining lane information of the vehicle, and selecting the lanes and planning the path according to the driving safety degrees of the targets in the left, middle and right regions from high to low. The application can solve the inconvenient and inaccurate problems of the path planning of the existing automatic driving vehicle, improve the accuracy and efficiency of the path planning of the automatic driving vehicle, and increase the safety of the automatic driving vehicle.

Description

Technical Field

[0001] This invention relates to the technical field of autonomous driving, specifically to a path planning learning method for autonomous vehicles. Background Technology

[0002] Existing autonomous vehicles often employ machine learning algorithms for path planning, such as A* and RRT algorithms. This algorithmic path planning requires learning the driver's lane-changing maneuvers and recording multiple parameters of the lane-changing process. In an autonomous driving scenario, if the identified road condition information corresponds to a learned road condition, the autonomous vehicle will perform a lane-changing operation according to the lane-changing parameters of that learned road condition. This is a supervised learning process, requiring the labeling of different road condition information (i.e., mapping road condition information to lane-changing parameters). This requires significant time and road condition information, placing high demands on the training dataset. If road condition information not present in the training dataset appears during autonomous driving, lane-changing operations will be difficult to execute, affecting the accuracy of path planning. Simultaneously, autonomous vehicles also use high-precision map modeling for path planning, applying path planning algorithms to the modeled map to find feasible paths between the start and end points. However, this method is mainly suitable for autonomous driving scenarios with relatively fixed geographical environments, and its use is inconvenient. Therefore, finding a more convenient and accurate path planning method for autonomous vehicles is of great significance. Summary of the Invention

[0003] This invention provides a path planning learning method for autonomous vehicles, which solves the problems of inconvenience and inaccuracy in existing path planning for autonomous vehicles, improves the accuracy and efficiency of path planning for autonomous vehicles, and increases the safety of autonomous vehicles.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A path planning learning method for autonomous vehicles includes:

[0006] Images of targets around the vehicle are collected using the vehicle's cameras and radar, and the targets around the vehicle are classified into left, center, and right regions based on the images.

[0007] Establish a safety calculation model and calculate the driving safety of each target in the left, middle and right areas based on the vehicle's speed and position;

[0008] The system acquires information about the vehicle's lane and selects lanes and plans routes in descending order of driving safety for each target in the left, middle, and right regions.

[0009] Preferred options also include:

[0010] An offline simulation environment is constructed, and offline simulation data for targets around the vehicle is provided based on the images.

[0011] A deep learning neural network for predicting safety level was constructed, and self-learning was performed based on the offline simulation data to optimize the safety level calculation model.

[0012] Preferred options also include:

[0013] Perceptual information collected in the real environment is transformed into a 2D bird's-eye view through perspective transformation, and traffic participants are categorized and dynamically labeled.

[0014] Extract the features of the bird's-eye view, and perform clustering processing on the bird's-eye view based on the features and the driving safety level to determine whether the bird's-eye view corresponds to the left lane change category, the right lane change category, or the straight-ahead category, and then perform lane change operation according to the corresponding category.

[0015] Preferred options also include:

[0016] Path planning is learned based on the test simulation results of the vehicle in the offline simulation environment, and the bird's-eye view of the real environment of the actual lane change operation is added to the dataset according to the results of the actual lane change operation, so as to gradually replace the data generated by the offline simulation with real environmental data.

[0017] Preferred options also include:

[0018] In a real-world scenario, after a vehicle completes a lane change, the driving safety before and after the lane change is compared. If the driving safety does not decrease after the lane change, then the lane change was successful.

[0019] Preferably, the construction of the deep learning neural network for security prediction and the optimization of the security calculation model include:

[0020] The security prediction quantity is self-learned through a deep learning neural network to predict the security change of each target.

[0021] The sum of the calculated driving safety level and the predicted safety level change for each target is used as the optimized safety level prediction for each target.

[0022] Preferably, the self-learning through the deep learning neural network for the security prediction includes:

[0023] Construct a deep learning neural network, and use the relative distance, relative speed, and type of the vehicle in front of the vehicle corresponding to the lane where the vehicle plans to change lanes, as well as the relative distance, relative speed, and type of the vehicle behind the vehicle, as input parameters of the neural network. The output parameter of the neural network is the safety prediction quantity.

[0024] The offline simulation data is used as input samples for training the neural network. Each set of offline simulation data includes the relative distance to the vehicle in front, the relative speed of the vehicle in front, the type of the vehicle in front, the relative distance to the vehicle behind, the relative speed of the vehicle behind, the type of the vehicle behind, and the change in safety before and after the lane change.

[0025] By iteratively optimizing using gradient descent, the neural network converges to obtain the mapping relationship between input and output.

[0026] Preferably, the step of transforming the 3D scene into a 2D bird's-eye view by perspective transformation of the perceived information collected in the real environment, and classifying and dynamically labeling traffic participants, includes:

[0027] In a real-world environment, real-scene images captured by cameras and radar are transformed into a 2D bird's-eye view through perspective transformation. The bird's-eye view is a 100m×100m perception plane. The category labeling of all traffic participants is completed by machine vision algorithms of vehicle-mounted camera images, and the dynamic characteristic labeling is completed by vehicle-mounted radar and perception algorithms.

[0028] Preferably, the step of determining whether the bird's-eye view corresponds to a left lane change category, a right lane change category, or a straight-ahead category, and then performing a lane change operation according to the corresponding category, includes:

[0029] If the bird's-eye view converted in the real environment is similar to the bird's-eye view corresponding to the left lane change, right lane change, or straight-ahead lane change, then the lane change operation in the real environment is determined to be: left lane change, right lane change, or straight-ahead lane change.

[0030] This invention provides a path planning learning method for autonomous vehicles. Based on the classification of left, center, and right regions of targets around the vehicle and the calculation of safety level, lane selection and path planning are performed. This solves the problems of inconvenience and inaccuracy in the path planning of existing autonomous vehicles, and can improve the accuracy and efficiency of path planning for autonomous vehicles, thereby increasing the safety of autonomous vehicles. Attached Figure Description

[0031] To more clearly illustrate the specific embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below.

[0032] Figure 1 This is a schematic diagram of a path planning learning method for autonomous vehicles provided by the present invention.

[0033] Figure 2 This is a flowchart of an offline reinforcement learning method provided by the present invention. Detailed Implementation

[0034] To enable those skilled in the art to better understand the embodiments of the present invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and implementation methods.

[0035] To address the inconvenience and inaccuracy in current path planning for autonomous vehicles, this invention provides a path planning learning method for autonomous vehicles. This method solves the problems of inconvenience and inaccuracy in existing path planning for autonomous vehicles, improves the accuracy and efficiency of path planning, and enhances the safety of autonomous vehicles.

[0036] like Figure 1 As shown, a path planning learning method for autonomous vehicles includes:

[0037] S1: Collect images of targets around the vehicle using the vehicle's cameras and radar, and classify the targets around the vehicle into left, center, and right regions based on the images.

[0038] S2: Establish a safety calculation model and calculate the driving safety of each target in the left, middle and right areas based on the vehicle's speed and position.

[0039] S3: Obtain the lane information of the vehicle and select lanes and plan routes in descending order of driving safety level for each target in the left, middle and right areas.

[0040] Specifically, driving safety is calculated by combining each frame of image captured by the camera with radar ranging information. When driving in a lane with heavy traffic, for each frame of image captured by the camera and radar, the driving safety score for the left, center, and right regions is calculated using a specific algorithm based on target category information, target distance information, and target relative speed information collected within the image range. The basic principle followed by the algorithm is: a low driving safety score is indicated by the presence of pedestrians, large tractor-trailers, etc., at close range; a high driving safety score is indicated by an empty lane and nearby stable-moving vehicles. For example, if there are vehicles ahead in the current lane, but the relative distance and relative speed are controllable, then a higher driving safety score is defined as M. Security =90%, there are no vehicles in the left lane, but changing lanes poses a safety risk, therefore the driving safety level is L. Security =80%. Similarly, if there are vehicles in the right lane and they are close together, changing lanes poses a risk, resulting in a driving safety rating of R. Security =50%.

[0041] In one embodiment, when a dependent variable is defined, such as relative vehicle speed, relative distance, traffic density, vehicle category information, etc., the driving safety level of the lane ahead can be calculated: Security = Φ(x1, x2, x3...x iAmong them, (x1, x2, x3...x i Let be the dependent variable and Φ be the calculated driving safety factor. Provided lane-changing conditions permit, autonomous vehicles should prioritize lane-changing maneuvers in directions with higher safety factors.

[0042] The method also includes:

[0043] An offline simulation environment is constructed, and offline simulation data for targets around the vehicle is provided based on the images.

[0044] A deep learning neural network for predicting safety level was constructed, and self-learning was performed based on the offline simulation data to optimize the safety level calculation model.

[0045] In practical applications, offline simulation environments can define traffic participants (Actors) and plan their behavior. Some leading autonomous driving companies have their own dedicated simulation test cases, which define lanes, traffic participants, and their travel paths. Ego Car, on the other hand, needs to test autonomous vehicles that run in a simulation test environment according to pre-defined autonomous driving control logic. The test environment observes whether collisions or other safety incidents occur during the operation, in order to train a feasible lane selection and path planning method in an offline simulation environment.

[0046] Furthermore, the construction of the deep learning neural network for security prediction and the optimization of the security calculation model include:

[0047] The safety prediction quantity is self-learned through a deep learning neural network to predict the safety change of each target; the sum of the calculated driving safety of each target and the predicted safety change is used as the optimized safety prediction quantity of each target.

[0048] Specifically, to ensure high accuracy in safety calculations, the computational model is optimized in an offline environment. In addition to calculating based on the dependent variables mentioned above, the predicted change in safety is also added. These factors together form the final safety calculation result, guiding lane-change planning for autonomous vehicles.

[0049] The improved formula for calculating driving safety is as follows: Security = Φ(x1, x2, x3...x i )+ΔS; where ΔS is the added safety prediction quantity, and ΔS is optimized based on the offline simulation learning process.

[0050] In one embodiment, a safety prediction calculation model can be implemented using a neural network. A deep learning neural network is constructed, with input parameters including the relative distance, relative speed, and vehicle type of the vehicle ahead in the lane the vehicle plans to change lanes from; and the relative distance, relative speed, and vehicle type of the vehicle behind. To facilitate neural network calculation, the vehicle types are quantified numerically, with common vehicle categories such as light trucks, heavy trucks, cars, two-wheeled vehicles, three-wheeled vehicles, and agricultural vehicles represented by numbers. The output parameter of the neural network is the safety prediction. Fully connected layers, hidden layers, and dropout layers are added between the output and input layers of the neural network to improve its performance. Sufficient offline simulation data is input into the neural network. Each set of offline simulation data includes the relative distance, relative speed, and vehicle type of the vehicle ahead in the lane the vehicle plans to change lanes from; the relative distance, relative speed, and vehicle type of the vehicle behind; and the change in safety before and after the lane change. Through gradient descent iterative optimization, the neural network converges, learning the mapping relationship between input and output.

[0051] The method also includes:

[0052] Perceptual information collected in the real environment is transformed into a 2D bird's-eye view through perspective transformation, and traffic participants are categorized and dynamically labeled.

[0053] Extract the features of the bird's-eye view, and perform clustering processing on the bird's-eye view based on the features and the driving safety level to determine whether the bird's-eye view corresponds to the left lane change category, the right lane change category, or the straight-ahead category, and then perform lane change operation according to the corresponding category.

[0054] Path planning is learned based on the test simulation results of the vehicle in the offline simulation environment, and the bird's-eye view of the real environment of the actual lane change operation is added to the dataset according to the results of the actual lane change operation, so as to gradually replace the data generated by the offline simulation with real environmental data.

[0055] In a real-world scenario, after a vehicle completes a lane change, the driving safety before and after the lane change is compared. If the driving safety does not decrease after the lane change, then the lane change was successful.

[0056] Furthermore, the process of transforming the 3D scene into a 2D bird's-eye view by perspective transformation of the perceived information collected in the real environment, and classifying and dynamically labeling traffic participants, includes:

[0057] In a real-world environment, real-scene images captured by cameras and radar are transformed into a 2D bird's-eye view through perspective transformation. The bird's-eye view is a 100m×100m perception plane. The category labeling of all traffic participants is completed by machine vision algorithms of vehicle-mounted camera images, and the dynamic characteristic labeling is completed by vehicle-mounted radar and perception algorithms.

[0058] Furthermore, the step of determining whether the bird's-eye view corresponds to a left lane change, a right lane change, or a straight-ahead lane change, and then performing a lane change operation according to the corresponding category, includes:

[0059] If the bird's-eye view converted in the real environment is similar to the bird's-eye view corresponding to the left lane change, right lane change, or straight-ahead lane change, then the lane change operation in the real environment is determined to be: left lane change, right lane change, or straight-ahead lane change.

[0060] In practical applications, such as Figure 2 As shown, a typical offline reinforcement learning method is as follows:

[0061] Build a simulation scenario, set up driving roads, add traffic participants, and plan the driving paths of traffic participants.

[0062] Perspective transformation converts a 3D scene into a 2D scene viewed from above, resulting in a planar image. Perspective transformation can transform the relative positions of objects in three-dimensional space onto a two-dimensional plane. Since we can ignore the height information of traffic participants during lane-changing decisions, this dimensionality reduction operation does not affect our action decision-making outcome. Furthermore, the converted 2D image can be processed using many computer vision methods, making it much easier to process than the complex perceptual information of 3D.

[0063] A sufficient number of labeled bird's-eye view images and their corresponding lane-change operations are input into the reinforcement learning module. The lane-change operations are determined through extensive offline simulation data, using a sufficiently accurate driving safety calculation model, to identify the optimal lane-change operation for a given bird's-eye view. Through convolutional neural networks and self-attention mechanisms, the reinforcement learning module learns methods for classifying bird's-eye views. When given an unfamiliar bird's-eye view in the future, the module will quickly determine, based on its prior knowledge, whether the view corresponds to a left lane change, right lane change, or straight-ahead lane.

[0064] Feature extraction is performed on the bird's-eye view images used in offline simulations. Feature extraction of images is a routine operation in the field of data mining, and this invention does not limit it. The bird's-eye view images are then clustered according to the extracted features. The clustering operation can use k-means or other clustering methods, which are not limited here. All offline simulation bird's-eye view images are classified into one of the following categories: left lane change, right lane change, or straight ahead.

[0065] In real-world autonomous driving, the real-world images captured by cameras and radar are also transformed into 2D bird's-eye view through perspective transformation. This bird's-eye view is also a 100m×100m perception plane, and all traffic participants have completed category labeling and dynamic characteristic labeling. Category labeling is completed by machine vision algorithms of vehicle-mounted camera images, and dynamic characteristic labeling is completed by vehicle-mounted radar and perception algorithms.

[0066] To determine if the features extracted from the bird's-eye view transformed in the real-world environment are similar to any category of the bird's-eye view learned offline, the criterion for similarity is the similarity of the extracted features. If the bird's-eye view transformed in the real-world environment is similar to a bird's-eye view of the left lane change category, then the optimal operation in the real-world environment is determined to be a left lane change.

[0067] After a vehicle completes a lane change maneuver in a real-world environment, the driving safety before and after the change is compared. If the driving safety does not decrease after the lane change, the lane change is considered successful. The bird's-eye view of this lane change maneuver in the real-world environment is added to our dataset. Real-world environmental data is gradually used to replace the offline simulation-generated data, making the dataset more realistic. If the perceived information from the real-world environment cannot be categorized based on the extracted features when converted into a bird's-eye view, experienced drivers classify the perceived information into one of three categories: changing lanes to the left, changing lanes to the right, or going straight.

[0068] Therefore, this invention provides a path planning learning method for autonomous vehicles. Based on the classification of left, middle and right regions of targets around the vehicle and the calculation of safety, lane selection and path planning are performed. This solves the problems of inconvenience and inaccuracy in the path planning of existing autonomous vehicles, improves the accuracy and efficiency of path planning for autonomous vehicles, and increases the safety of autonomous vehicles.

[0069] The structure, features, and effects of the present invention have been described in detail above with reference to the embodiments shown in the figures. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, shall be within the protection scope of the present invention as long as they do not exceed the spirit covered by the specification and figures.

Claims

1. A path planning learning method for autonomous vehicles, characterized in that, The method comprises the following steps: Collecting images of targets around the vehicle through the camera and radar of the vehicle, and classifying the targets around the vehicle into left, middle and right regions according to the images; A safety degree calculation model is established, and the safety degree of each target in the left, middle and right areas is calculated according to the speed and position of the vehicle; the driving safety degree is calculated for each frame of image perceived by the camera and fused with the radar ranging information to generate, and the driving safety degree of the front lane is calculated: Security=Φ(x1,x2,x3...x i ), wherein (x1,x2,x3...x i ) is a dependent variable, the dependent variable includes relative speed, relative distance, traffic density and vehicle category information, and Φ(x1,x2,x3...x i ) is the calculated driving safety degree; Obtaining lane information of the vehicle, and selecting a lane and planning a path according to the driving safety degrees of the targets in the left, middle and right regions from high to low; Building an offline simulation environment, and providing offline simulation data for the targets around the vehicle according to the images; Building a deep learning neural network for safety degree prediction, and performing self-learning according to the offline simulation data to optimize the safety degree calculation model; The building of the deep learning neural network for safety degree prediction and the optimization of the safety degree calculation model comprises the following steps: Performing self-learning through the deep learning neural network for safety degree prediction to predict the change amount of the safety degree of each target; Taking the sum of the driving safety degree of each target calculated and the predicted change amount of the safety degree as the safety degree prediction of each target after optimization.

2. The path planning learning method for an autonomous vehicle according to claim 1, wherein The method further comprises the following steps: Converting a 3D scene into a 2D bird's eye view through perspective transformation of the perception information collected in a real environment, and labeling the traffic participants according to categories and dynamic features; Extracting bird's eye view features, and performing clustering processing on the bird's eye view according to the bird's eye view features and the driving safety degrees to determine whether the bird's eye view corresponds to a left lane changing category, a right lane changing category or a straight driving category, and then performing a lane changing operation according to the corresponding category.

3. The path planning learning method for an autonomous vehicle according to claim 2, wherein The method further comprises the following steps: Learning path planning according to the test simulation results of the vehicle in the offline simulation environment, and adding the bird's eye view of the real environment of this lane changing operation to the data set according to the results of the actual lane changing operation to gradually replace the data generated by the offline simulation with real environment data.

4. The path planning learning method for an autonomous vehicle according to claim 3, wherein The method further comprises the following steps: After the vehicle completes the lane changing operation in the real environment, comparing the driving safety degrees before and after the lane changing operation, and if the driving safety degree after the lane changing operation does not decrease, the lane changing operation is successful.

5. The path planning learning method for an autonomous vehicle according to claim 4, wherein The self-learning through the deep learning neural network for safety degree prediction comprises the following steps: Building a deep learning neural network, taking the relative distance, relative speed and category of the front vehicle of the lane where the vehicle plans to change lanes, the relative distance, relative speed and category of the rear vehicle as input parameters of the neural network, and taking the safety degree prediction as output parameters of the neural network; Taking the offline simulation data as input samples of the neural network for training, and each group of offline simulation data comprises the relative distance, relative speed and category of the front vehicle of the lane where the vehicle plans to change lanes, the relative distance, relative speed and category of the rear vehicle, and the change value of the safety degree before and after the lane changing operation; Optimizing the neural network through gradient descent iteration to obtain the mapping relationship between the input and the output.

6. The path planning learning method for an autonomous vehicle according to claim 5, wherein The conversion of the 3D scene into the 2D bird's eye view through perspective transformation of the perception information collected in the real environment, and the labeling of the traffic participants according to categories and dynamic features comprise the following steps: In a real environment, real scene images collected by a camera and a radar are converted into a 2D bird's eye view through perspective change, the bird's eye view is a 100m*100m perception plane, and the category labels of all traffic participants are completed through a machine vision algorithm of a vehicle-mounted camera, and dynamic characteristic labels are completed through a vehicle-mounted radar and a perception algorithm.

7. The path planning learning method for an autonomous vehicle according to claim 6, wherein The judgment of the bird's eye view corresponding to the left lane change class, the right lane change class and the straight driving class, and the execution of the lane change operation according to the corresponding class, comprises: If the bird's eye view converted in the real environment is similar to the bird's eye view corresponding to the left lane change class, the right lane change class or the straight driving class, it is determined that the lane change operation in the real environment corresponds to the left lane change, the right lane change or the straight driving.

Images