Lane level positioning method and device

Lane-level positioning is solved through the decision tree model based on random forests, and the problem of high cost of deep learning methods is achieved, efficient and accurate lane positioning is achieved, and hardware requirements and training time are reduced.

CN118980388AInactive Publication Date: 2024-11-19HOZON NEW ENERGY AUTOMOBILE CO LTD

Patent Information

Application Number
CN202411460562.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-18
Publication Date
2024-11-19
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

In the existing lane-level positioning technology, deep learning methods require a large amount of labeling data, have a long training time and high hardware configuration requirements, resulting in high cost.

Method used

The decision tree model based on random forest is used for lane-level positioning, and multiple decision trees are used for voting prediction, reducing the variance of a single decision tree and improving the robustness and reliability of the model.

Benefits of technology

Efficient and accurate lane-level positioning at lower costs is achieved, which reduces hardware configuration requirements and training time, and improves the robustness and reliability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118980388A_ABST
    Figure CN118980388A_ABST
Patent Text Reader

Abstract

The invention discloses a lane level positioning method and device, and relates to the technical field of vehicle automatic driving. According to the scheme, after automatic driving operation is started on a target vehicle, a sensor on the target vehicle is used for collecting road condition data information around driving, and local map information, corresponding to the current position where the target vehicle is located, on a navigation map is obtained; the method comprises the following steps: firstly, processing acquired road condition data information by using a preset sensing model, and outputting obstacle tracking information and lane line tracking information through sensing processing; processing the obstacle tracking information, the lane line tracking information and the local map information by using a preset lane level positioning model; the preset lane level positioning model is a decision tree model based on random forest training, so that the lane where the target vehicle is located can be finally predicted and positioned by using decision results obtained by a plurality of decision trees in the model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of vehicle autonomous driving technology, and in particular to a lane-level positioning method and device. Background Art

[0002] "Image-free" technology generally refers to achieving similar or better functions and effects through other technical means without relying on or significantly reducing the reliance on specific types of images or maps. This technology can be applied to many fields, such as autonomous driving.

[0003] In the field of autonomous driving, the "mapless" technology used means that autonomous driving can be achieved without relying on previously established high-precision maps. In the application of this technology, the existing low-cost maps in actual use can be used (that is, there is no need to spend processing costs to build high-precision maps). By sensing and identifying road signs, road boundary lines and other road conditions, the specific lane position of the vehicle on the road can be determined, thus realizing the lane-level positioning of the vehicle itself.

[0004] At present, in the implementation process of lane-level positioning, in order to obtain high-quality positioning accuracy, deep learning is usually used in lane-level positioning algorithms. However, due to the large amount of data that needs to be annotated and the long training time in deep learning, this not only requires high model computing power, but also requires relatively high hardware configuration support, which leads to a high cost for lane-level positioning in general. Summary of the invention

[0005] The present application provides a lane-level positioning method and device, the main purpose of which is to use random forests to predict and locate lanes. The training process based on decision trees is simpler and more efficient, and the trained random forest model is more robust and reliable, thereby making the lane-level positioning implemented by the present application more efficient and accurate at a lower cost.

[0006] In order to achieve the above objectives, this application mainly provides the following technical solutions: A first aspect of the present application provides a lane-level positioning method, the method comprising: Acquire first data information collected by a sensor on the target vehicle; Acquire current local map information corresponding to the current position of the target vehicle as second data information; Processing the first data information using a preset perception model to output obstacle tracking information and lane line tracking information corresponding to the target vehicle; The obstacle tracking information, the lane line tracking information and the second data information are processed using a preset lane-level positioning model to output lane positioning information corresponding to the target vehicle. The preset lane-level positioning model is a decision tree model based on random forest training, which is used to use multiple decision trees to predict and locate the lane where the target vehicle is located.

[0007] A second aspect of the present application provides a lane-level positioning device, the device comprising: A first acquisition unit, used to acquire first data information collected by a sensor on the target vehicle; A second acquisition unit is used to acquire current local map information corresponding to the current position of the target vehicle as second data information; A first processing unit, configured to process the first data information using a preset perception model, and output obstacle tracking information and lane line tracking information corresponding to the target vehicle; The second processing unit is used to process the obstacle tracking information, the lane line tracking information and the second data information using a preset lane-level positioning model, and output the lane positioning information corresponding to the target vehicle. The preset lane-level positioning model is a decision tree model based on random forest training, which is used to use multiple decision trees to predict and locate the lane where the target vehicle is located.

[0008] A third aspect of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the lane-level positioning method as described above is implemented.

[0009] A fourth aspect of the present application provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the above lane-level positioning method is implemented when the computer program is executed by the processor.

[0010] By means of the above technical solution, the technical solution provided by this application has at least the following advantages: The present application provides a lane-level positioning method and device. After starting the automatic driving operation on the target vehicle, the sensors on the target vehicle are used to collect road condition data information around the vehicle, and obtain the local map information on the navigation map corresponding to the current position of the target vehicle. First, the preset perception model is used to process the collected road condition data information, and the obstacle tracking information and lane line tracking information are output after perception processing; then the preset lane-level positioning model is used to process the obstacle tracking information, lane line tracking information, and the local map information; since the preset lane-level positioning model is a decision tree model based on random forest training, the present application can use the decision results obtained by each of the multiple decision trees in the model to finally predict the lane where the target vehicle is located.

[0011] Compared with the existing technology, this application is a solution based on random forest to predict and locate lanes. Because random forest has the characteristics of bootstrap sampling and random feature selection when training the model, the variance of a single decision tree is reduced, making the model more robust and reliable as a whole. Random forest belongs to the field of machine learning, but not deep learning. The training process of decision trees is simpler and more efficient, solving the existing technical problem of high cost caused by deep learning. This application uses random forest to implement the lane positioning solution, which is not only efficient and low-cost, but also based on the robustness and reliability of the model, making the prediction more accurate.

[0012] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings: Figure 1 A flow chart of a lane-level positioning method provided in an embodiment of the present application; Figure 2 A flow chart of another lane-level positioning method provided in an embodiment of the present application; Figure 3 A schematic diagram of an enhanced lane change detection method is provided for an embodiment of the present application; Figure 4 A simple schematic diagram of building a decision tree provided in an embodiment of the present application; Figure 5 A block diagram of a lane-level positioning device provided in an embodiment of the present application; Figure 6 A block diagram of another lane-level positioning device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0014] The exemplary embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art.

[0015] Now, for lane-level positioning of autonomous driving, we can make use of the existing low-cost maps in actual use (that is, there is no need to spend processing costs to build high-precision maps), and determine the specific lane position of the vehicle on the road by sensing and identifying information such as road signs and road boundaries, thereby achieving the vehicle's own lane-level positioning.

[0016] In the specific implementation process, in order to obtain high-quality positioning accuracy, deep learning is usually used in lane-level positioning algorithms, such as convolutional neural networks (CNN) and deep learning model architectures based on self-attention mechanisms (Transformer), to perform end-to-end learning of sensor data and construct the required positioning map.

[0017] Because deep learning algorithms require a large amount of labeled data, long training time, high model computing power, and relatively high hardware configuration support, the overall cost of lane-level positioning is also very high.

[0018] Deep learning is a branch of machine learning. Machine learning has many advantages, such as powerful data processing capabilities, accurate prediction and classification, and auxiliary decision support. When applied to the field of autonomous driving, it helps to improve the quality of lane-level positioning. Therefore, on the basis of machine learning, in order to reduce costs, if another branch learning method is sought, it is possible to achieve both efficiency and accuracy in lane-level positioning.

[0019] Based on the above considerations, the embodiment of the present disclosure provides a lane-level positioning method, which uses random forest to achieve predictive lane-level positioning, such as Figure 1 As shown, the present application embodiment provides the following specific steps: 101. Acquire first data information collected by a sensor on a target vehicle.

[0020] In the embodiment of the present application, a variety of sensors will be pre-installed around the vehicle body, such as at least: cameras, lidars and millimeter-wave radars. The embodiment of the present application does not specifically limit the number and installation positions of different types of sensors, but it needs to be ensured that during driving, the road condition information around the vehicle that needs to be collected by these sensors includes at least: data information on dynamic / static objects around the vehicle body; data information on traffic signs on the road surface and around the vehicle.

[0021] It should be noted that, in order to distinguish it from the numerous vehicles traveling on the road, the embodiment of the present application uses the term "target vehicle" to refer to the object in the driving lane to be positioned in the lane-level positioning; in addition, the embodiment of the present application uses the terms "first" and "second" as identifiers to distinguish and refer to the data information obtained in steps 101 and 102 respectively.

[0022] 102. Obtain current local map information corresponding to the current position of the target vehicle as second data information.

[0023] For a moving vehicle, the road condition information within a certain distance coverage range around the vehicle, with the vehicle itself as the origin, will affect the lane-level positioning. And as the vehicle moves continuously, this "certain distance coverage range around the vehicle" is also constantly changing.

[0024] The embodiment of the present application can make use of the existing low-cost maps in actual use, for example, but not limited to, including the following: Standard Definition (SD) maps have relatively low accuracy, usually around the meter level, and are mainly used to meet daily navigation needs.

[0025] "SD+" map: Compared with SD map, it is a navigation map with some high-precision information added. This type of map usually has detailed attribute information such as the number of lanes and lane topology, but the accuracy is still lower than that of high-precision map, and it does not provide high-precision lane-level geometric shape data. It is more inclined to enhance the navigation function.

[0026] "SD Pro" map: It is a type of map between navigation maps and high-precision maps. It can supplement the deficiencies of real-time perception and is mainly used to assist autonomous driving operations.

[0027] In an embodiment of the present application, any one of the above lower-cost maps can be selected and applied to the present application scheme. After the vehicle position is located in real time while the vehicle is driving, a certain coverage area is delineated around the vehicle position as the origin on the lower-cost map, and the current local map information is obtained based on the delineated coverage area, and the current local map information includes lane semantic information.

[0028] It should be noted that, compared to not relying on a high-precision map (High Definition Map), the purpose of the embodiment of the present application is to only use a lower-cost map to be able to provide comprehensive and rich lane semantic information. These lane semantic information will be input into the preset lane-level positioning model of 104 for processing to help improve the accuracy of lane-level positioning.

[0029] For example, the lane semantic information that can be obtained includes but is not limited to: lane type (such as straight lane, turning lane, emergency lane, etc.), lane markings (such as solid line, dashed line, yellow line, white line, etc.), the association between lanes and road signs (such as speed limit signs, traffic lights, etc.), the relationship between lanes and surrounding obstacles, etc.

[0030] 103. Use a preset perception model to process the first data information, and output obstacle tracking information and lane line tracking information corresponding to the target vehicle.

[0031] The embodiment of the present application can pre-train one or more perception models to perceive and identify lane-related visual, geometric, and semantic information, which are the basic data for autonomous driving decision-making and planning.

[0032] Exemplarily, the preset perception model may be, but is not limited to: a perception model based on vision, such as an image segmentation model, which separates lane lines from the background through image segmentation technology to form a lane line segmentation mask. This model can distinguish different lane lines (such as left and right lines, solid lines and dashed lines) and perform instance segmentation; a perception model based on radar and laser radar, which can provide distance and three-dimensional spatial information of the vehicle's surroundings. By processing the data of these sensors, a point cloud map of the vehicle's surroundings can be constructed, and the position and shape information of the lane lines can be extracted from it.

[0033] Therefore, the embodiment of the present application utilizes a preset perception model to process the sensor data obtained by 101, with the purpose of obtaining information including: information on obstacles existing around the vehicle, which may be, but is not limited to, vehicles traveling around the vehicle, which may be referred to as "obstacle tracking information"; information on lane lines that the vehicle passes through, such as taking into account situations such as normal lane driving and overtaking and lane changing, which may be referred to as "lane line tracking information".

[0034] 104. The obstacle tracking information, lane line tracking information and second data information are processed using a preset lane-level positioning model to output lane positioning information corresponding to the target vehicle. The preset lane-level positioning model is a decision tree model based on random forest training, which is used to use multiple decision trees to predict the lane where the target vehicle is located.

[0035] The "current local map information corresponding to the current position of the target vehicle" obtained as 102 and the "obstacle tracking information and lane line tracking information corresponding to the target vehicle" obtained as 103, that is, the "current local map information" obtained based on the lower cost map and the "obstacle tracking information and lane line tracking information" obtained based on the preset perception model, are comprehensively processed by the embodiment of the present application using a pre-trained lane-level positioning model.

[0036] In addition to the above two aspects of data processing, in the embodiment of the present application, the preset lane-level positioning model is a decision tree model based on random forest training. During the model training phase, the random forest reduces the variance of a single decision tree through bootstrap sampling and random feature selection, making the overall model more robust and reliable. In the model application phase, when predicting, the random forest votes on the prediction results of all trees, which improves the generalization ability and anti-overfitting ability of the model. All of the above factors will ultimately help improve the accuracy of the positioning results output by the preset lane-level positioning model.

[0037] In addition, for the preset lane-level positioning model based on random forest, during the model training phase, the training of multiple decision trees can be carried out in parallel, and during the model application phase, multiple decision trees can be applied simultaneously to make decisions, and finally vote based on the decision results of multiple decision trees to obtain the final prediction result. The above is more suitable for processing large-scale data in both the model training and application phases. At the same time, compared with deep neural networks, it does not require the annotation of large amounts of data and complex computing power, and is simpler and more efficient in overall operation.

[0038] As mentioned above, the embodiment of the present application provides a lane-level positioning method. After the automatic driving operation is started on the target vehicle, the sensors on the target vehicle are used to collect road condition data information around the vehicle, and obtain the local map information on the navigation map corresponding to the current position of the target vehicle. First, the preset perception model is used to process the collected road condition data information as above, and the obstacle tracking information and lane line tracking information are output after perception processing; then the preset lane-level positioning model is used to process the obstacle tracking information, lane line tracking information, and the local map information as above; since the preset lane-level positioning model is a decision tree model based on random forest training, the present application can use the decision results obtained by each of the multiple decision trees in the model to finally predict the lane where the target vehicle is located.

[0039] Compared with the prior art, the embodiment of the present application is a solution based on random forest to predict and locate lanes, and because random forest has the characteristics of bootstrap sampling and random feature selection when training the model, the variance of a single decision tree is reduced, making the model as a whole more robust and reliable. Random forest belongs to the field of machine learning, but not deep learning. The training process of decision trees is simpler and more efficient, solving the existing technical problem of high cost caused by deep learning. The embodiment of the present application uses random forest to implement the lane positioning solution, which is not only efficient and low-cost, but also based on the robustness and reliability of the model, making the prediction more accurate.

[0040] In some specific embodiments, in order to explain in more detail, the embodiments of the present application also provide another method for analyzing the via quantity of a battery pack installed on a vehicle body, such as Figure 2 As shown, the present application embodiment provides the following specific steps: 201. Acquire first data information collected by a sensor on a target vehicle.

[0041] 202. Obtain current local map information corresponding to the current position of the target vehicle as second data information.

[0042] This step can be broken down into the following: (1)-(4); (1) Obtaining integrated navigation information corresponding to the target vehicle, where the integrated navigation information at least includes latitude, longitude and altitude information corresponding to the target vehicle.

[0043] Integrated navigation refers to a navigation system that integrates various navigation devices, such as inertial navigation system (INS), global satellite navigation system (such as GPS, Beidou), wheel speed sensor, vehicle environment perception sensor (such as radar, lidar, camera), etc., and is controlled by monitors and computers. These systems each have different advantages and limitations, and by using them in combination, they can complement each other and improve the overall positioning accuracy and reliability.

[0044] In the specific application of lane-level positioning of autonomous driving, taking the current time t as the reference, the combined navigation is used to obtain at least the latitude and longitude information of the vehicle, for example, expressed as: LLH t ={X, Y, Z}, where the first L stands for "longitude", the second L stands for "latitude", and H stands for "height".

[0045] (2) Based on the latitude, longitude and altitude information, determine the current location of the target vehicle on the vehicle navigation map.

[0046] The latitude, longitude and altitude information is used to determine the current location of the vehicle on a vehicle navigation map (such as, but not limited to, an SD map, an "SD+" map, or an "SDPro" map).

[0047] (3) Taking the current location as the origin, select the preset driving range on the vehicle navigation map.

[0048] (4) On the vehicle navigation map, the lane semantic information is obtained from the preset driving range as the current local map information corresponding to the current position of the target vehicle.

[0049] In order to improve the safety of lane-level automatic positioning, in actual driving scenarios, a preset driving range can be defined around the vehicle with the vehicle's own position as the origin. The embodiment of the present application needs to sense and identify road conditions within the preset driving range to assist in achieving automatic lane positioning while ensuring a safe driving distance.

[0050] Accordingly, after locating and obtaining the latitude, longitude and altitude information of the vehicle in 202, in order to still ensure the existence of the preset driving range, on the vehicle navigation map (for example, it can be but not limited to SD map, "SD+" map, "SD Pro" map), the embodiment of the present application takes the current position as the origin and selects this preset driving range on the vehicle navigation map. Since the vehicle navigation map can provide comprehensive and rich lane semantic information, by identifying the lane semantic information within the preset driving range, it will be possible to obtain but not limited to: such as road center, road topology, road speed limit, interval speed limit, presence or absence of traffic lights, road grade, functional grade, number of lanes, lane width, main and auxiliary roads, bridges, tunnels, roundabouts, elevated roads, slopes, zone rates, etc. The above uses the various lane semantic information obtained by identification to form the current local map information corresponding to the current position of the vehicle. For example, the current local map information obtained at the current time t can be expressed as MAP t .

[0051] 203. Process the first data information using a preset perception model, and output obstacle tracking information and lane line tracking information corresponding to the target vehicle.

[0052] The embodiment of the present application collects sensor data information at the current time t, which includes road condition information around the vehicle, including at least: data information of dynamic / static objects around the vehicle body; data information of traffic signs on the road surface and around the vehicle.

[0053] Assume that at the current time t, in the collected sensor data, there are n obstacles (pedestrians, vehicles, etc.) around the vehicle body in the current frame, and m sensing lane lines. Using a local coordinate system in the navigation system, such as the ENU (East, North, Up) coordinate system (which is used to describe the three-dimensional spatial position of an object near the earth's surface), this step can be refined to include the following: (1) and (2); (1) Performing environmental perception processing on the first data information, and outputting the position information, speed information, and posture information of the obstacle as obstacle tracking information corresponding to the target vehicle.

[0054] Among them, the environmental perception processing operation is a processing operation possessed by the preset perception model. By using this processing operation, other obstacle information in the vehicle's environment can be identified (such as data information about dynamic / static objects around the vehicle body). During the vehicle's driving, the surrounding environment is constantly changing, so the information about other obstacles is also constantly changing. Therefore, by using this processing operation, the continuous updating and tracking of the obstacle information encountered around the vehicle is realized.

[0055] During the driving process of the vehicle, the obstacles around the vehicle include at least pedestrians or other vehicles. Based on this, the specific content of the obstacle tracking information obtained by performing environmental perception processing in the preset perception model may include but is not limited to: obtaining the location information, speed information and posture information of the obstacle. An example is shown as follows: The position detection box of n obstacles is expressed as: ; Among these n obstacles, the position detection frame of obstacle 1 is taken as an example: The position vector of the first corner point of obstacle 1 is: ; Based on the same representation method in three-dimensional space, it is also necessary to obtain the position vectors of the other seven corner points of obstacle 1. And based on the same principle, it is also necessary to obtain the vectors of the other n-1 obstacles and represent them as in the representation method of obstacle 1 in the example.

[0056] The speed of n obstacle targets is expressed as follows: Among these n obstacles, for example, the velocity vector of obstacle 1 is: ; As in the representation of obstacle 1, based on the same principle, it is also necessary to obtain the vectors of the other n-1 obstacles and represent them.

[0057] The orientation of n obstacle targets is expressed as follows: Among these n obstacles, for example, the posture vector of obstacle 1 is: ; As in the representation of obstacle 1, based on the same principle, it is also necessary to obtain the vectors of the other n-1 obstacles and represent them.

[0058] As described above, the embodiment of the present application uses the position information, speed information and posture information of n obstacles as the obstacle tracking information corresponding to the target vehicle.

[0059] (2) Perform lane line detection processing on the first data information, and output geometric information and semantic information of the lane line as lane line tracking information corresponding to the target vehicle.

[0060] Among them, the lane line detection operation is another processing operation possessed by the preset perception model. By using this operation, the geometric and semantic information of the lane line can be identified from the traffic sign data information on the road surface and around the vehicle.

[0061] Lane lines are mainly composed of straight lines, curves and their combination, which are used to guide and control vehicle driving. The geometric information of lane lines mainly involves its shape, size, position and relative relationship with the road and vehicles; the semantic information of lane lines refers to the meaning and function of lane lines in traffic scenes.

[0062] In the embodiment of the present application, for example, by performing lane line detection processing on the first data information, m perceived lane lines are output, including geometric information and semantic information, which can be expressed as: , where the geometric equation of the first sensing lane line can be expressed as: ,Semantic information can be expressed as type, such as solid line, dotted line, double yellow line, curb, etc.

[0063] In some specific embodiments, for the data information output by the preset perception model, due to the characteristics of missed detection and false detection in perception detection, these data information can usually be further processed after perception, for example, using Kalman filtering to process the obstacle tracking information obtained as above (1) and the lane line tracking information obtained as above (2).

[0064] 204. Input the obstacle tracking information, the lane line tracking information and the second data information into the input layer of each decision tree in the preset lane-level positioning model, and each decision tree corresponds to a unique input-to-output mapping rule.

[0065] The preset lane-level positioning model provided in the embodiment of the present application is a decision tree model based on random forest training, which includes multiple pre-trained decision trees. The decision results of multiple decision trees are combined for voting to ultimately predict and locate the lane in which the vehicle is located.

[0066] In the training phase of the preset lane-level positioning model, some historical data can be obtained as sample data to train multiple decision trees, so that each decision tree corresponds to a unique input-to-output mapping rule, so that in the application phase of the model, after inputting data into the decision tree, the data result can be output according to the mapping rule. Here, the model training phase is explained in the embodiment of the present application, including the following: Exemplarily, a historical moment t' can be determined first, and the same working principle (such as 201-203 above) can be used to obtain some historical data at the historical moment t', such as at least including: historical first data information collected by the sensor on the target vehicle; current local map information corresponding to the current position of the target vehicle, as historical second data information; the first data information is processed using a preset perception model to output historical obstacle tracking information and historical lane line tracking information corresponding to the target vehicle. Therefore, based on the data information in the three dimensions of "historical obstacle tracking information", "historical lane line tracking information" and "current local map information" at the historical moment t', as sample data, it is used to train the model based on random forest.

[0067] Secondly, during the model training phase, the label information in the decision tree of the random forest is constructed. For example, starting from the leftmost lane in the direction of the vehicle's travel, the labels are constructed as 1, 2, 3, 4, etc. This label information is used to characterize the numbering of the lanes. During the model training phase, the label information is introduced to assist in model training, so that during the model application phase, when the model is used to process data information, it is convenient to use the specified label information output by the model to indicate which numbered lane the vehicle is in. Since the decision tree is actually a classification process, the label is represented as a "category" in the decision tree. For example, "label is 1" corresponds to "category is 1" in the decision tree.

[0068] Furthermore, in order to adapt to decision-making in lane-level positioning application scenarios, during the model training phase, the embodiment of the present application will pre-construct some preset features related to the application scenario, and will use the following three key preset features ((1)-(3)) as an example for explanation. However, it should be noted that these are only illustrative examples, and the embodiment of the present application does not limit the content and quantity of the preset features.

[0069] (1) The preset feature 1 constructed is the lane line semantic and quantity features, as follows: The number of perceived lane lines on the left is L_lane, the number of perceived lane lines on the right is R_lane, the number of perceived lane edges on the left is L_roadedge, the number of perceived lane edges on the right is R_roadedge, the number of opposite lane lines on the left is L_oppo_lane, and the number of opposite lane lines on the right is R_oppo_lane.

[0070] The preset feature 1 input is: L_lane, R_lane, L_roadedge, R_roadedge, L_oppo_lane, R_oppo_lane.

[0071] (2) The preset feature 2 constructed is the lane-changing feature. The lane-changing feature algorithm is as follows: Obtain lane change status information from the control algorithm. When a lane change is identified, use an enhanced perception lane change detection method, such as Figure 3 In the lane recognition thread, the vehicle is always checked for lateral acceleration. When the lateral acceleration exceeds acc max When, usually acc max Please refer to 0.1m / s 2 , turn on lateral acceleration sliding window detection and lane line C0 sliding window detection.

[0072] The lateral acceleration sliding window detection is as follows: the lateral acceleration in the Macc frame exceeds the Nacc frame by more than acc max , it indicates the detection is successful, otherwise it indicates the detection fails. Usually, Macc can refer to 100 and Nacc can refer to 50.

[0073] The sensing lane line C0 sliding window detection is as follows: The sensing lane line equation is usually expressed as y = C3x3 + C2x2 + C1x + C0, where C0, C1, C2, and C3 are the lane line distance, slope, curvature, and curvature derivative, respectively, and x and y are the coordinates in the vehicle body coordinate system, respectively. With the center of the rear axle of the vehicle as the origin of the coordinate system, in practice, the C0l value on the left side of the lane line is greater than 0, and the C0r value on the right side is less than 0. When changing lanes to the left without crossing the lane line, the C0l value gradually decreases and approaches 0, and instantly increases and approaches the width of a lane when just crossing the lane line. The same is true for changing lanes to the right. Therefore, by setting a sliding window and dynamically recording the C0 value within the time TC0 (the specific value is adjusted according to the vehicle speed) to determine whether to change lanes, only when it is determined that the vehicle begins to cross the lane and move to the adjacent lanes on the left and right sides, can it be determined that the vehicle is changing lanes, otherwise it may be driving in the curve of the lane.

[0074] When both the lateral acceleration sliding window detection and the lane line C0 sliding window detection detect a lane change, the current state is set to the lane change state, otherwise it is not a lane change.

[0075] Through the above algorithm, the preset feature 2 can finally be represented as a successful lane change or a failed lane change.

[0076] (3) Construct preset feature 3 as the position, posture and semantic features of the target, as follows: Calculate each obstacle to get the lateral distance from the vehicle under the vehicle system. The orientation information of the target can be obtained through the posture, that is, the semantic label can be added to determine whether the target is in the opposite lane. Taking obstacle 1 at time t as an example, the model can be obtained. ,in It has positive and negative values. OPP=1 means that the obstacle target is in the same direction as the vehicle, and OPP=0 means that the obstacle target is in the opposite direction to the vehicle.

[0077] The calculation is as follows: The position detection frame of obstacle 1: First calculate the center point of the vehicle system horizontal axis position of obstacle 1 POS_CEN = The yaw angle of the vehicle is , the yaw angle of obstacle 1 is , when >C degrees, the threshold C can refer to 145 degrees, indicating that obstacle 1 is an oncoming vehicle, OPP=0, otherwise OPP=1.

[0078] Therefore, the preset feature 3 can finally be expressed as the model .

[0079] In the embodiment of the present application, the above preset features 1-3 are only illustrative examples. Since lane-level positioning involves various aspects, a variety of preset features can be pre-constructed. After the preset features are constructed, the next step is to further extract key information based on each preset feature as a refined feature (that is, referred to as a "target feature" in the embodiment of the present application) based on the data information involved in each preset feature, and then construct constraints based on each target feature. When the constraints composed of the target features are placed on a node in a decision tree, the sample data flowing into the node can be screened according to the constraints to make a decision on how to allocate the sample data to the next node.

[0080] Based on the sample data, label information, preset features, and constraint conditions composed of target features constructed above, the present application embodiment provides, for example, Figure 4 A simple diagram of building a decision tree is shown in Figure . Figure 4 The root node, internal node and leaf node in the tree structure are shown. Assume that there are six labels on the vehicle navigation map, each label identifies a lane (that is, it is equivalent to numbering the lane). Since the decision tree is actually a prediction and classification process, Figure 4 As shown in , "category" is used to represent "label"; and it is assumed that target features 1-7 are target features constructed based on preset feature 1; Figure 4 The "a, b, c, d, e, f, g" shown in the figure are combined with corresponding target features to express corresponding constraints. It should be noted that these letters are not limited to representing numbers.

[0081] Taking preset feature 1 as an example, since the key information (as target feature) included in preset feature 1 is: "L_lane", "R_lane", "L_roadedge", "R_roadedge", "L_oppo_lane", "R_oppo_lane", constraints can be added based on each refined feature to construct diverse target features. For example, "L_lane<=3" (as target feature 1<=a) means that the number of left perceived lane lines is less than or equal to 3.

[0082] Moreover, under preset feature 1, based on the above target features, a variety of constraints can be constructed to be arranged on each internal node of the tree structure respectively. It should be noted that in order to assist in achieving the purpose of predicting lane-level positioning, the arrangement operation on each internal node should be implemented according to a certain prediction logic.

[0083] In the actual model training process, Figure 4 It is shown that after the sample data is input on the root node, relevant operations are performed based on the constraints on the node to distribute the sample data to the next adjacent node, and then the operations of operation and distribution to the next adjacent node are repeated until the leaf node is reached, completing the training of a decision tree.

[0084] In some specific embodiments, during the model training phase, hundreds or thousands of decision trees will be obtained in the random forest. Although the purpose of the random forest is to train the independence and randomness of each decision tree to enhance the generalization ability of the entire model, some design rules may also be added during the specific implementation process so that there may still be some "tendencies" or "preferences" between different decision trees. For example, it may be preferred to randomly select target features and combine constraints from a large number of target features split by two or more associated preset features to train the same decision tree. Therefore, under the premise of ensuring randomness and independence, some "tendencies" or "preferences" are adopted to increase the weight of the decision results output by a certain decision tree, so as to ultimately affect the prediction results output by the model and adapt to the needs of different driving scenarios.

[0085] 205. In each decision tree, the data information inputted in the input layer is processed by using the mapping rules corresponding to the decision tree, and the prediction result corresponding to the decision tree is outputted.

[0086] In the model training phase, the random forest will obtain hundreds or thousands of decision trees, and after model training, each decision tree will obtain the data processing rules, that is, the data is input from the root node to the leaf node. The data processing rules in the entire process can be called the mapping rules corresponding to each decision tree. Therefore, this step can be refined to include the following: In the tree structure corresponding to the decision tree, first, determine the root node, internal nodes and leaf nodes and the traversal path between the three; second, determine the constraints composed of different preset features on the root node and internal nodes; finally, for the data information input on the input layer of the decision tree, starting from the root node, process it according to the traversal path and the constraints on each node until it reaches the leaf node, output the prediction result on the leaf node, and obtain the corresponding prediction result at each leaf node as the corresponding prediction result of the decision tree.

[0087] 206. Process the prediction results corresponding to each decision tree based on the preset voting rule, and output the target prediction result as the lane positioning information corresponding to the target vehicle.

[0088] The voting rule is mainly used for classification tasks. In a random forest, each decision tree will classify the sample according to the input features and give a classification result (i.e., category label). There are many ways to implement the voting rule, including but not limited to: Majority Voting: If a category receives more than half of the votes, then that category will be used as the final prediction result. If no category receives more than half of the votes, you can choose to reject the prediction or adopt other strategies based on the specific situation.

[0089] Plurality Voting: Also known as simple voting or majority voting. In this method, the category with the most votes is selected as the final prediction result. If there are multiple categories with the same number of votes, one of them may be randomly selected as the final result (depending on the specific implementation).

[0090] Weighted Voting: Similar to the weighted average method, but in the classification task, the prediction result (i.e., category label) of each decision tree is assigned a weight. Then, the categories are weighted voted according to the weights, and finally the category with the largest weight and the largest value is selected as the prediction result.

[0091] As mentioned above, by integrating the prediction results of multiple decision trees, the lane-level positioning result of the target vehicle at the current moment is obtained, that is, the real-time prediction of which lane the vehicle is in is made. The embodiment of the present application uses random forests to reduce the error rate of a single decision tree and improve the overall prediction accuracy and robustness.

[0092] As the above Figure 1 , 2In order to realize the method shown in the figure, the embodiment of the present application provides a lane-level positioning device. The device embodiment corresponds to the aforementioned method embodiment. For ease of reading, the device embodiment will not repeat the details of the aforementioned method embodiment one by one, but it should be clear that the device in this embodiment can correspond to all the contents of the aforementioned method embodiment. The device is used to efficiently and accurately realize lane-level positioning, such as Figure 5 , the device comprises: A first acquisition unit 31 is used to acquire first data information collected by a sensor on the target vehicle; A second acquisition unit 32 is used to acquire current local map information corresponding to the current position of the target vehicle as second data information; A first processing unit 33 is used to process the first data information using a preset perception model, and output obstacle tracking information and lane line tracking information corresponding to the target vehicle; The second processing unit 34 is used to process the obstacle tracking information, the lane line tracking information and the second data information using a preset lane-level positioning model, and output the lane positioning information corresponding to the target vehicle. The preset lane-level positioning model is a decision tree model based on random forest training, which is used to use multiple decision trees to predict and locate the lane where the target vehicle is located.

[0093] In some specific embodiments, Figure 6 As shown, the preset lane-level positioning model includes multiple decision trees, and the second processing unit 34 includes: An input module 341, used to input the obstacle tracking information, the lane line tracking information and the second data information into the input layer of each decision tree, each decision tree corresponding to a unique input-to-output mapping rule; A first processing module 342 is used to process the data information input on the input layer in each decision tree using the mapping rule corresponding to the decision tree, and output a prediction result corresponding to the decision tree; The second processing module 343 is used to process the prediction results corresponding to each of the decision trees based on a preset voting rule, and output a target prediction result as lane positioning information corresponding to the target vehicle.

[0094] In some specific embodiments, the first processing module 342 is specifically used to: In the tree structure corresponding to the decision tree, determining the root node, internal nodes and leaf nodes and the traversal paths existing between the three; Determine the constraint conditions consisting of different preset features existing on the root node and the internal node respectively; For the data information inputted on the input layer of the decision tree, starting from the root node, the data information is processed according to the traversal path and the constraint conditions existing on each node until the leaf node is reached, and the prediction result on the leaf node is outputted; A corresponding prediction result is obtained based on each leaf node as the prediction result corresponding to the decision tree.

[0095] In some specific embodiments, Figure 6 As shown, the second acquisition unit 32 includes: An acquisition module 321 is used to acquire the combined navigation information corresponding to the target vehicle, wherein the combined navigation information at least includes the latitude, longitude and height information corresponding to the target vehicle; A determination module 322, configured to determine the current position of the target vehicle on a vehicle navigation map based on the latitude, longitude and height information; A selection module 323 is used to select a preset driving range on the vehicle navigation map with the current position as the origin; The identification module 324 is used to identify lane semantic information from the preset driving range on the vehicle navigation map as the current local map information corresponding to the current position of the target vehicle.

[0096] In some specific embodiments, Figure 6 As shown, the first processing unit 33 includes: The third processing module 331 is used to perform environment perception processing on the first data information, and output the position information, speed information and posture information of the obstacle as the obstacle tracking information corresponding to the target vehicle; The fourth processing module 332 is used to perform lane line detection processing on the first data information, and output the geometric information and semantic information of the lane line as the lane line tracking information corresponding to the target vehicle.

[0097] In some specific embodiments, the device further comprises: The third processing unit 35 is used to perform post-perception processing on the data information output by the perception model, and the post-perception processing at least includes: using Kalman filtering processing.

[0098] In summary, the embodiments of the present application provide a lane-level positioning method and device. The embodiments of the present application are based on random forests to achieve a solution for predicting and locating lanes. Since random forests have the characteristics of bootstrap sampling and random feature selection when training models, the variance of a single decision tree is reduced, making the model as a whole more robust and reliable. Random forests belong to the field of machine learning, but not deep learning. The training process of decision trees is simpler and more efficient, solving the existing technical problem of high cost caused by the use of deep learning. The embodiments of the present application use random forests to implement a lane positioning solution, which is not only efficient and low-cost, but also based on the robustness and reliability of the model, making the prediction more accurate.

[0099] The lane-level positioning device includes a processor and a memory. The first acquisition unit, the second acquisition unit, the first processing unit, the second processing unit, etc. are all stored in the memory as program units, and the processor executes the program units stored in the memory to implement corresponding functions.

[0100] The processor includes a kernel, which retrieves the corresponding program unit from the memory. One or more kernels can be set, and the random forest is used to predict and locate the lane by adjusting the kernel parameters. The training process based on the decision tree is simpler and more efficient, and the trained random forest model is more robust and reliable, so that the lane-level positioning implemented by the present application is more efficient and accurate at a lower cost.

[0101] An embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the lane-level positioning method as described above is implemented.

[0102] An embodiment of the present application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the above lane-level positioning method is implemented when the computer program is executed by the processor.

[0103] The embodiment of the present application also provides a computer program product, which, when executed on a data processing device, is suitable for executing a program that initializes the steps of the upper lane level positioning method.

[0104] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0105] In a typical configuration, the device includes one or more processors (CPU), memory and bus. The device may also include input / output interface, network interface and the like.

[0106] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip. The memory is an example of a computer-readable medium.

[0107] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0108] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0109] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0110] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A lane-level positioning method, characterized in that: The method comprises: Acquire first data information collected by a sensor on a target vehicle; Acquire current local map information corresponding to the current position of the target vehicle as second data information; Processing the first data information using a preset perception model to output obstacle tracking information and lane line tracking information corresponding to the target vehicle; The obstacle tracking information, the lane line tracking information and the second data information are processed using a preset lane-level positioning model to output lane positioning information corresponding to the target vehicle. The preset lane-level positioning model is a decision tree model based on random forest training, which is used to use multiple decision trees to predict and locate the lane where the target vehicle is located.

2. The lane-level positioning method according to claim 1, characterized in that: The preset lane-level positioning model includes a plurality of decision trees. The preset lane-level positioning model is used to process the obstacle tracking information, the lane line tracking information and the second data information to output the lane positioning information corresponding to the target vehicle, including: Inputting the obstacle tracking information, the lane line tracking information and the second data information into the input layer of each decision tree, each decision tree corresponding to a unique input-to-output mapping rule; On each of the decision trees, the data information input on the input layer is processed using the mapping rule corresponding to the decision tree, and a prediction result corresponding to the decision tree is output; The prediction results corresponding to each of the decision trees are processed based on a preset voting rule, and a target prediction result is output as lane positioning information corresponding to the target vehicle.

3. The lane-level positioning method according to claim 2, characterized in that: The step of processing the data information inputted on the input layer by using the mapping rule corresponding to the decision tree and outputting the prediction result corresponding to the decision tree comprises: In the tree structure corresponding to the decision tree, determining the root node, internal nodes and leaf nodes and the traversal paths existing between the three; Determine the constraint conditions consisting of different preset features existing on the root node and the internal node respectively; For the data information inputted on the input layer of the decision tree, starting from the root node, the data information is processed according to the traversal path and the constraint conditions existing on each node until the leaf node is reached, and the prediction result on the leaf node is outputted; A corresponding prediction result is obtained based on each leaf node as the prediction result corresponding to the decision tree.

4. The lane-level positioning method according to any one of claims 1 to 3, characterized in that: The obtaining of the current local map information corresponding to the current position of the target vehicle as the second data information includes: Acquire the combined navigation information corresponding to the target vehicle, wherein the combined navigation information at least includes the latitude, longitude and altitude information corresponding to the target vehicle; Based on the latitude, longitude and height information, determining the current position of the target vehicle on a vehicle navigation map; Taking the current position as the origin, selecting a preset driving range on the vehicle navigation map; On the vehicle navigation map, lane semantic information is identified from the preset driving range as current local map information corresponding to the current position of the target vehicle.

5. The lane-level positioning method according to any one of claims 1 to 3, characterized in that: The using of the preset perception model to process the first data information and output obstacle tracking information and lane line tracking information corresponding to the target vehicle includes: Performing environmental perception processing on the first data information, and outputting position information, speed information, and posture information of the obstacle as obstacle tracking information corresponding to the target vehicle; The first data information is subjected to lane line detection processing, and geometric information and semantic information of the lane line are output as lane line tracking information corresponding to the target vehicle.

6. The lane-level positioning method according to claim 5, characterized in that: The method further comprises: The data information output by the preset perception model is subjected to post-perception processing, wherein the post-perception processing at least includes: using Kalman filtering processing.

7. A lane-level positioning device, characterized in that: The device comprises: A first acquisition unit, used to acquire first data information collected by a sensor on a target vehicle; A second acquisition unit is used to acquire current local map information corresponding to the current position of the target vehicle as second data information; A first processing unit, configured to process the first data information using a preset perception model, and output obstacle tracking information and lane line tracking information corresponding to the target vehicle; The second processing unit is used to process the obstacle tracking information, the lane line tracking information and the second data information using a preset lane-level positioning model, and output the lane positioning information corresponding to the target vehicle. The preset lane-level positioning model is a decision tree model based on random forest training, which is used to use multiple decision trees to predict and locate the lane where the target vehicle is located.

8. The lane level positioning device according to claim 7, characterized in that: The preset lane-level positioning model includes a plurality of decision trees, and the second processing unit includes: An input module, used to input the obstacle tracking information, the lane line tracking information and the second data information into an input layer of each decision tree, each decision tree corresponding to a unique input-to-output mapping rule; A first processing module is used to process the data information input on the input layer in each decision tree by using the mapping rule corresponding to the decision tree, and output a prediction result corresponding to the decision tree; The second processing module is used to process the prediction results corresponding to each of the decision trees based on a preset voting rule, and output a target prediction result as lane positioning information corresponding to the target vehicle.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the lane-level positioning method according to any one of claims 1 to 6 is implemented.

10. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the lane-level positioning method as claimed in any one of claims 1 to 6 is implemented.

Citation Information

Patent Citations

  • Lane positioning method and vehicle positioning system

    CN110556012A

  • Lane positioning method and device, storage medium and electronic equipment

    CN113916242A

  • Automobile domain controller and automatic mapping and positioning method for vehicles on commuting road

    CN115994985A

  • Automatic driving behavior decision-making method and system, electronic equipment and storage medium

    CN116461555A

Cited By

  • Method for realizing strip mine vehicle trajectory prediction through combination of random forest and Kalman

    CN120145808A

  • Method for realizing vehicle trajectory prediction in open-pit mine by combining random forest and kalman

    CN120145808B

  • Systems and methods for vehicle localization using adaptive filtering and a standard map

    US20260274287A1