Complex scene road-level positioning method and system suitable for automatic driving system

Abstract

The invention relates to the technical field of vehicle automatic driving, in particular to a complex scene road-level positioning method and system suitable for an automatic driving system. According to the invention, road and environment information around the vehicle can be obtained through perception extension of the automatic driving system; probability allocation is carried out by integrating information around a vehicle and a global road network, so that the road positioning effect of a complex environment can be remarkably improved (a road environment detection head is added based on a backbone network of a current automatic driving system BEV sensing network (certain marking and training need to be carried out from end to end, and the ability is achieved after pre-training is carried out on a VLM / VLA system); and the environment (high-rise semantic information of a main road, an auxiliary road, a ramp connecting road, a highway, a tunnel, a bridge, a garage, a viaduct and the like) where the current vehicle is located can be obtained with relatively high accuracy. Through verification of an own data set, the method can obviously improve the road-level positioning accuracy in a road environment.

Description

A road-level localization method and system for complex scenarios applicable to autonomous driving systems Technical Field

[0001] This invention relates to the field of autonomous driving technology, specifically to a road-level positioning method and system for complex scenarios applicable to autonomous driving systems. Background Technology

[0002] With the development of intelligent driving technology, the adoption rate of driver assistance systems in intelligent connected vehicles is increasing, especially with the significant improvement in L2-level highway and urban NOA coverage over the past 25 years. Simultaneously, with the development of large-scale models in the field of artificial intelligence, high-precision maps have become less common in highway and urban NOA. The basic technical paradigm is an end-to-end, lightweight map model. In this model, the end-to-end autonomous driving model constructs a perception environment (online high-precision map and surrounding dynamic obstacles) based on the vehicle's perception (camera and other sensors) and the current road location (based on route positioning sent from the vehicle's navigation system or positioning results corrected by the autonomous driving system). This information is then explicitly or implicitly passed to the PNC module for planning and control to achieve overall autonomous driving functionality. If the road matching deviates, it can lead to autonomous driving takeover or abnormalities, even resulting in safety accidents.

[0003] One current technical solution involves directly using the route and its location sent to the autonomous driving system from the vehicle's navigation system. Currently, vehicle navigation positioning and matching are based on filtered inertial navigation. Additionally, based on the location and navigation route, a Hidden Markov Model (HMM)-based road network matching and voting process is used to find the optimal road and location. However, due to errors in the positioning data source and the complexity of road structures, vehicle navigation's road-level matching can encounter problems in complex scenarios, such as errors in main / secondary road positioning, layered road structure, and deviations in longitudinal positioning. Another solution involves the autonomous driving system receiving the route and location and then applying an additional layer of strategy within the system. This includes maintaining a global position (based on filtered navigation) to obtain global coordinates, and combining this with semantic information from vision, such as lane numbers, to further filter the route and road network and correct the longitudinal position. Generally, autonomous driving integrated navigation performs better than vehicle navigation, and the addition of visual semantic information significantly improves positioning accuracy compared to the first solution. However, due to the complexity of the road environment, road positioning errors still occur in complex scenarios (main road / auxiliary road error, layered road error, deviation in longitudinal positioning).

[0004] Therefore, there is an urgent need for a road-level positioning method and system suitable for autonomous driving systems in complex scenarios to address the technical deficiencies of the existing technical solutions. Summary of the Invention

[0005] The purpose of this invention is to design a road-level positioning method for complex scenarios suitable for autonomous driving systems, compared with existing technical solutions, to improve the road positioning problem of autonomous driving in complex road environments, such as overpass areas, main and auxiliary roads, and reduce the problems of autonomous driving takeover or abnormality caused by inaccurate road-level positioning.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a road-level localization method for complex scenarios applicable to autonomous driving systems, comprising the following steps: S1, acquiring navigation route information and vehicle current position information transmitted from the cockpit domain to obtain vehicle navigation route position information; S2, obtaining the vehicle's global coordinate position by performing ESKF filtering on the acquired GNSS information, IMU information, and vehicle speed information, and acquiring global road-level network data within a preset range around the vehicle based on the global coordinate position; S3, recording the vehicle's historical data; S4, performing scene recognition and selecting candidate links based on the vehicle's current localization stage; wherein, the vehicle's current localization stage includes the initial localization stage and the non-initial localization stage; when the vehicle is in the initial localization stage... In the first stage, candidate links are determined by matching global coordinate positions and global road network data, and the target link is determined by a geometric projection distance and direction weighting method. When the vehicle is not in the initial localization stage, candidate links with similar directions are searched based on the vehicle's historical data to determine whether the vehicle is in a complex scene. When the vehicle is in a non-complex scene, the global coordinate position is located using a Hidden Markov Model to determine the target link. When the vehicle is in a complex scene, the target link is located using a Hidden Markov Model based on the road semantics and global coordinate position within the vehicle's field of vision output by the perception model or end-to-end model. In the second stage, the global coordinate position is projected onto the target link, and the link ID, projection coordinates, and offset of the target link are output.

[0007] The above technical solution produces the following technical effects: This application can obtain road and environmental information around the vehicle by extending the perception of the autonomous driving system. By integrating the information around the vehicle and the global road network for probability allocation, the road localization effect in complex environments can be significantly improved (by adding a road environment detection head to the backbone network of the current autonomous driving system BEV perception network (end-to-end annotation and training are required, and this capability is available after pre-training for VLM / VLA systems), the current environment of the vehicle (main roads, auxiliary roads, ramp connecting roads, highways, tunnels, bridges, garages, and high-level semantic information such as above / below elevated bridges) can be obtained with high accuracy. Validation of this method using proprietary datasets shows that it can significantly improve the road-level localization accuracy in road environments.

[0008] As a further improvement to the road-level localization method for complex scenarios applicable to autonomous driving systems in this application, in step S4, the road semantics include one or more of the following: road type, whether it is elevated or ground level, number of lanes, and ramp attributes.

[0009] As a further improvement to the road-level localization method for complex scenarios applicable to autonomous driving systems proposed in this application, the process of determining the target link using a Hidden Markov Model includes the following steps: S41, Observation probability calculation: The observation probability of the Hidden Markov Model includes distance observation probability, heading observation probability, and road semantic observation probability when the vehicle is in a complex scenario; S42, Transition probability calculation: The transition probability of the Hidden Markov Model is determined based on road topology. When there is a topological connection between two candidate links, the state is allowed to transition from the previous candidate link to the next candidate link; when there is no topological connection between two candidate links, their transition probability is set to zero, and state transition is prohibited; S43, Probability calculation and filtering: A variable sliding window is used on the time axis, and forward probability calculation is performed only on the observation sequence within a preset historical length. The driving distance corresponding to the historical length is not less than a preset threshold. The forward cumulative probability of the candidate links within the window is normalized and sorted, and the maximum probability after normalization is compared with a preset confidence threshold. When the maximum probability is greater than the confidence threshold, the corresponding candidate link is output as the target link.

[0010] As a further improvement to the road-level localization method for complex scenarios applicable to autonomous driving systems in this application, in step S41, the formula for calculating the distance observation probability is as follows:

[0011] Where dis is the geometric projection of the vehicle's global coordinate position onto the candidate link; The probability of distance observation is given by the formula; the probability of heading observation is calculated as follows:

[0012] in, For the probability of heading observation, The heading is the angle between the vehicle's direction of travel and due north.

[0013] As a further improvement to the road-level localization method for complex scenes applicable to autonomous driving systems proposed in this application, in step S41, the formula for calculating the semantic observation probability is as follows:

[0014] in, Similarity of road types; Similarity of road attributes; To ensure consistency in the number of lanes, the observation probability output by the Hidden Markov Model is calculated as follows: ;in, This represents the observation probability of a candidate link output by the Hidden Markov Model; when X=1, the vehicle is in a complex scene. 0.5; When X=0, the vehicle is in a non-complex scenario. 0.

[0015] As a further improvement to the road-level localization method for complex scenarios applicable to autonomous driving systems in this application, in step S42, candidate links are numbered according to sequence and an adjacency matrix based on road topology is constructed, and the transition probability of two candidate links is calculated through the adjacency matrix.

[0016] As a further improvement to the road-level localization method for complex scenarios applicable to autonomous driving systems in this application, the size of the variable sliding window is adjusted according to the vehicle speed to ensure that the historical driving distance covered by the window is greater than or equal to 300 meters; when the vehicle speed increases, the time length of the time axis is shortened, and when the vehicle speed decreases, the time length of the time axis is extended.

[0017] As a further improvement to the road-level localization method for complex scenarios applicable to autonomous driving systems in this application, in step S4, the method for determining the target link using a geometric projection distance and orientation weighted method is to obtain the best-matching target link by finding the minimum value after linear normalization weighting; the calculation formula for linear normalization weighting is: Where X is the normalized geometric projection distance, and Y is the normalized angular error. and The weights are set as segmented parameters based on vehicle speed.

[0018] As a further improvement to the road-level localization method for complex scenarios applicable to autonomous driving systems proposed in this application, the historical data records historical operational data including global coordinates, road network matching coordinates, and perception semantic information.

[0019] A road-level localization system for complex scenarios suitable for autonomous driving systems includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute any of the road-level localization methods for complex scenarios suitable for autonomous driving systems described above. Attached Figure Description

[0020] Figure 1 is a flowchart of the present invention; Detailed Description of Embodiments The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] To facilitate a precise understanding of the solutions provided in the following embodiments of the present invention, the following explanations are given regarding the terms involved in the present invention before describing the technical solutions provided: Global Position: refers to the position and attitude information of the vehicle in the global coordinate system obtained by fusion filtering algorithms (such as Extended Kalman Filter ESKF) using GNSS (Global Navigation Satellite System), IMU (Inertial Measurement Unit), and vehicle speed information. This coordinate system is typically updated at high frequency and is used to represent the continuous trajectory of the vehicle in the world coordinate system.

[0022] Road Network Data: Refers to road-level map information composed of road links as basic units, including road geometry, topology (upstream and downstream relationships, connectivity), number of lanes, road type, elevation information, and other road attributes. Vehicles construct candidate road sets and road topology relationships based on this data during the implementation of this invention.

[0023] Link: In road network data, the smallest logical unit used to describe the road structure is typically a road segment with a unique ID, including a start point, end point, geometric shape, and road attributes. This application uses candidate links as the localization target for road-level localization.

[0024] Projection distance: This refers to the perpendicular distance from the vehicle's current global position point, as projected onto a candidate link, to the vehicle's current position. This distance measures the spatial proximity of the vehicle relative to a link and is an important component of the observation probability.

[0025] Candidate links: refer to the set of links near the current location that meet preset spatial or semantic conditions and may be used as the actual road where the vehicle is located.

[0026] Heading angle: refers to the angle between the vehicle's forward direction and the road's geometric direction, reflecting the consistency between the vehicle's direction of motion and the road's direction. In this application, its calculation error (deltaYaw) is used as an input factor for the observation probability.

[0027] Road Semantic Information (DSI) refers to the high-level semantic description of roads output by autonomous driving perception models (including BEV models or end-to-end perception models) within the vehicle's field of vision. This includes, but is not limited to, road type (main road, auxiliary road, elevated ramps, etc.), road attributes (whether it is an elevated road, bridge, tunnel, etc.), and the number of available lanes. This invention utilizes this semantic information to improve positioning accuracy in complex scenarios.

[0028] The BEV (Bird's Eye View Model) is a deep neural network model that transforms vehicle-surrounding sensor data into a bird's-eye view coordinate system and extracts road structure, road type, and road attributes within this coordinate system. This application utilizes the road semantic information output by the BEV model as observations in complex scenarios to improve the accuracy and stability of road-level positioning. This model typically includes: 1) a basic feature extraction network: used to extract spatial geometric and semantic features from raw sensor data and map them uniformly to the BEV space.

[0029] 2) Detection Head or Semantic Head: Used to output road-related elements in the BEV space, including road geometry, road type, road attributes, number of lanes, lane direction, and high-level semantics such as bifurcation / merging.

[0030] Historical Trajectory: refers to the continuous sequence of historical locations recorded by a vehicle during its journey, including the original global coordinate historical trajectory and the road-level trajectory obtained based on road network matching. This application constructs a sliding window based on this historical trajectory for probability recursion in a Hidden Markov Model.

[0031] Hidden Markov Model (HMM): A probabilistic graphical model consisting of a sequence of hidden states and an observation sequence, where road links are the hidden states and vehicle observation information (projected distance, heading error, semantic information, etc.) are the observations. This application constructs a probabilistic update and filtering mechanism for road-level positioning based on HMM.

[0032] Transition Probability: In a Hidden Markov Model (HMM), the probability of transitioning from one candidate link state to another is determined based on the road network topology. If two links are topologically connected, the transition probability is 1; otherwise, it is 0. This invention constructs candidate links as an adjacency matrix to calculate the transition probability.

[0033] Observation Probability: This refers to the probability that the vehicle's current observation information is generated by a candidate link. This application establishes a multi-dimensional observation probability model based on the projected distance between the vehicle and the candidate link, heading deviation, semantic matching degree, etc., and adaptively adjusts the weights according to the complexity of the road scene.

[0034] Complex scenarios refer to situations where the road structure involves multiple layers, intersections, and merging, resulting in high uncertainty. These include elevated multi-level roads, ramp merging / diversion zones, parallel roads under bridges, and urban canyons. In such scenarios, GNSS and traditional geometric matching are unreliable. This application employs an enhanced semantic fusion strategy for localization.

[0035] Sliding window: A time window used in Hidden Markov Model (HMM) forward probability calculation, consisting of a fixed-length historical trajectory and a historical semantic sequence. This application dynamically adjusts the window length based on vehicle speed and ensures it covers more than 300 meters of historical data to improve localization robustness.

[0036] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this invention is for describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0037] As shown in Figure 1, in order to address the technical deficiencies in the existing technical solutions, this application designs a road-level positioning method suitable for complex scenarios in autonomous driving systems, specifically including the following steps: S1, obtaining navigation route information and vehicle current location information transmitted from the cockpit domain to obtain vehicle navigation route location information; wherein, when the user performs route navigation, deviates, or updates the route during travel, the cockpit will send the navigation route information and current location to the autonomous driving system, and the road-level algorithm module of the autonomous driving system needs to receive and parse the information and store it in the internal data cache.

[0038] S2. By performing ESKF filtering on the acquired GNSS information, IMU information, and vehicle speed information, the global coordinate position of the vehicle is obtained. Based on the global coordinate position, global road-level road network data within a preset range around the vehicle is obtained. Specifically, the autonomous driving system's integrated navigation system will fuse GNSS, IMU, and vehicle speed information through filtering to obtain a high-frequency (50-100 Hz) global position coordinate information.

[0039] Furthermore, based on the global coordinates of the autonomous driving system, information on the global road network (geometric, topological, and road attribute information) within a certain distance range (e.g., 1 km) around the vehicle can be obtained. Based on the global coordinates, the road network information around the vehicle is loaded. Considering actual system performance issues, a designed caching mechanism is used to load, update, and use the data in real time. During cache initialization, road network data within a 500m radius (configurable) around the vehicle is loaded. As the vehicle moves, the coverage area of ​​the road network around the vehicle is calculated periodically every 2 seconds (configurable). When the minimum distance around the vehicle is less than 200m (configurable), the road network within a 500m radius centered on the vehicle's global coordinates is automatically updated and loaded. This avoids performance issues caused by frequent loading.

[0040] S3. Record the vehicle's historical data; by recording the historical coordinate position (global coordinate position and road network matching position) of the autonomous driving vehicle and the real-time perceived environmental information of the historical frames, the original historical trajectory and the real-time perceived historical information and road network matching information can be obtained. Specifically, the historical operation information of the vehicle is continuously recorded during the localization process, including: (1) the global coordinate position sequence obtained by GNSS / IMU fusion, which is used to form the vehicle's original driving trajectory; (2) the road network matching position sequence obtained by projecting the global coordinates onto the road candidate link, which is used to form the road-level trajectory; (3) the real-time environmental semantic information such as road type, road attributes, and number of lanes output by the perception model, which is used to form the historical semantic observation sequence. The above historical trajectory and historical semantic sequence constitute the sliding observation window of the Hidden Markov Model and provide continuous observation basis for the probability calculation of the road candidate link.

[0041] Specifically, the historical data records historical operational data including global coordinates, road network matching coordinates, and perceived semantic information.

[0042] S4. Perform scene recognition and select candidate links based on the vehicle's current positioning stage; the vehicle's positioning stage includes the initial positioning stage and the non-initial positioning stage; when the vehicle is in the initial positioning stage, candidate links are determined by matching global coordinate positions and global road-level road network data, and the target link is determined by a geometric projection distance and direction weighting method; when the vehicle is in the non-initial positioning stage, candidate links with similar directions are searched based on the vehicle's historical data to determine whether the vehicle is in a complex scene; preferably, if it is the initial positioning, the single point (GPS position, direction) is directly matched with the candidate links (roads within 50m of GPS), and the coarse positioning road link result is determined by geometric projection distance and direction weighting. The specific determination method is as follows: the minimum value is obtained after linear normalization weighting to obtain the best matching road link.

[0043] X represents the normalized projected distance, and Y represents the normalized angular error. and The weights are set as segmented parameters based on the vehicle's speed.

[0044] When the vehicle is in a non-complex scene, the global coordinate position is used to locate and determine the target link using a Hidden Markov Model (HMM). When the vehicle is in a complex scene, the target link is located and determined using an HMM based on the road semantics and global coordinate position within the vehicle's field of vision, output by a perception model or end-to-end model. Specifically, in complex scenes, this application uses a specially designed HMM algorithm that integrates the current location information, the road environment classification and high-level semantic information output by the model, and combines the geometric, topological and attribute information of the route and road network to perform probability allocation calculation and probability update. Finally, based on the probabilities of the route and road, a specific road is selected and the longitudinal position projection on the selected road is calculated. The road-level localization probability update uses an HMM, which is a temporal probabilistic model that describes a random sequence of unobservable states generated by a hidden Markov chain, and then uses each state to generate an observation variable to infer the state.

[0045] Specifically, in step S4, road semantics includes one or more of the following: road type, whether it is elevated or ground level, number of lanes, and ramp attributes. The process of the Hidden Markov Model (HMM) determining the target link includes the following steps: S41, Observation Probability Calculation: The observation probabilities of the HMM include distance observation probability, heading observation probability, and road semantic observation probability when the vehicle is in a complex scenario; S42, Transition Probability Calculation: The transition probabilities of the HMM are determined based on road topology. When there is a topological connection between two candidate links, the state is allowed to transition from the previous candidate link to the next link; when there is no topological connection between two candidate links, their transition probability is set to zero, and state transition is prohibited; S43, Probability Calculation and Screening: A variable sliding window is used on the time axis, and forward probability calculation is performed only on observation sequences within a preset historical length. The travel distance corresponding to the historical length is not less than a preset threshold. The forward cumulative probabilities of candidate links within the window are normalized and sorted, and the maximum normalized probability is compared with a preset confidence threshold. When the maximum probability is greater than the confidence threshold, the corresponding candidate link is output as the target link.

[0046] Furthermore, in step S41, the formula for calculating the distance observation probability is:

[0047] Where dis is the geometric projection of the vehicle's global coordinate position onto the candidate link; The probability of distance observation is given by the formula; the probability of heading observation is calculated as follows: ;in, For the probability of heading observation, The heading is the angle between the vehicle's direction of travel and due north.

[0048] Furthermore, in step S41, the formula for calculating the semantic observation probability is:

[0049] in, Similarity of road types; Similarity of road attributes; To ensure consistency in the number of lanes; furthermore, road types may include: main roads, auxiliary roads, on-ramp / off-ramp, merging / exiting lanes, elevated roads, ground-level roads, roundabouts, urban expressways, highways, etc. The observation probability output by the Hidden Markov Model is calculated as follows: ;in, This represents the observation probability of a candidate link output by the Hidden Markov Model; when X=1, the vehicle is in a complex scene. 0.5; When X=0, the vehicle is in a non-complex scenario. 0.

[0050] Furthermore, in step S42, candidate links are numbered according to sequence and an adjacency matrix based on road topology is constructed. The transition probability of two candidate links is calculated using the adjacency matrix.

[0051] Preferably, the size of the variable sliding window is adjusted according to the vehicle speed to ensure that the historical driving distance covered by the window is greater than or equal to 300 meters; when the vehicle speed increases, the time length of the time axis is shortened, and when the vehicle speed decreases, the time length of the time axis is extended.

[0052] S5. Project the global coordinate position onto the target link, and output the target link's link ID, projected coordinates, and offset. Specifically, in this step, the global coordinate point is projected onto the candidate link output in S4 to obtain the projected coordinates on the link and the offset of the link's starting point. The final road candidate link ID, projection point x, y, and offset are output to the downstream module.

[0053] Unlike Example 1, Example 2 also designs a road-level localization system for complex scenarios suitable for autonomous driving systems, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute any of the road-level localization methods for complex scenarios suitable for autonomous driving systems described above.

[0054] In the aforementioned system, the system obtains the global coordinate position based on GNSS and IMU data obtained by autonomous driving, as well as vehicle speed, through ESKF filtering; it obtains a rough navigation position based on navigation route information transmitted from the cockpit domain and the current location; based on the global coordinate position of autonomous driving, it obtains global road-level network information (geometric, topological, road attribute, etc.) within a certain distance range (e.g., 1 km) around the vehicle; based on the autonomous driving model (BEV perception model (basic network + newly designed detection head) or end-to-end AI model), it obtains the classification of the road environment within the field of view and high-level semantic understanding; by recording the historical coordinate position of autonomous driving (global coordinate position and road network matching position), it obtains the original historical trajectory and the trajectory after road network matching; based on a specially designed HMM algorithm, it performs probability allocation calculation and probability update based on the current location information, the road environment classification and high-level semantic information output by the model, and combines the geometric, topological and attribute information of the route and road network; finally, it performs specific road selection and longitudinal position projection calculation on the selected road based on the probabilities of the route and road.

[0055] It is noteworthy that those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0056] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0057] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0058] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

[0060] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A road-level localization method for complex scenarios suitable for autonomous driving systems, characterized in that, The process includes the following steps: S1. Obtaining navigation route information and vehicle current location information transmitted from the cockpit domain to obtain vehicle navigation route location information; S2. Obtaining the vehicle's global coordinate position by performing ESKF filtering on the acquired GNSS information, IMU information, and vehicle speed information, and obtaining global road-level network data within a preset range around the vehicle based on the global coordinate position; S3. Recording the vehicle's historical data; S4. Performing scene recognition and selecting candidate links based on the vehicle's current positioning stage; wherein, the vehicle's current positioning stage includes the initial positioning stage and non-initial positioning stage; when the vehicle is in the initial positioning stage, matching the global coordinate position and the global road-level network data to determine the location... Candidate links are identified, and a target link is determined using a weighted method of geometric projection distance and direction. When the vehicle is not in the initial localization stage, candidate links with similar directions are searched based on the vehicle's historical data to determine whether the vehicle is in a complex scene. When the vehicle is in a non-complex scene, the global coordinate position is located using a Hidden Markov Model to determine the target link. When the vehicle is in a complex scene, the target link is located using a Hidden Markov Model based on the road semantics within the vehicle's field of view output by a perception model or an end-to-end model, and the global coordinate position. S5. The global coordinate position is projected onto the target link, and the link ID, projection coordinates, and offset of the target link are output.

2. The road-level localization method for complex scenarios applicable to autonomous driving systems according to claim 1, characterized in that, In step S4, the road semantics include one or more of the following: road type, whether it is elevated or ground level, number of lanes, and ramp attributes.

3. The road-level localization method for complex scenarios applicable to autonomous driving systems according to claim 2, characterized in that, The process of determining the target link using the Hidden Markov Model includes the following steps: S41, Observation probability calculation: The observation probability of the Hidden Markov Model includes distance observation probability, heading observation probability, and road semantic observation probability when the vehicle is in a complex scene; S42, Transition probability calculation: The transition probability of the Hidden Markov Model is determined based on road topology. When there is a topological connection between two candidate links, the state is allowed to transition from the previous candidate link to the next candidate link; when there is no topological connection between two candidate links, their transition probability is set to zero, and state transition is prohibited; S43, Probability calculation and filtering: A variable sliding window is used on the time axis to perform forward probability calculation only on the observation sequence within a preset historical length, where the driving distance corresponding to the historical length is not less than a preset threshold. The forward cumulative probabilities of the candidate links within the window are normalized and sorted, and the maximum normalized probability is compared with a preset confidence threshold. When the maximum probability is greater than the confidence threshold, the corresponding candidate link is output as the target link.

4. The road-level localization method for complex scenarios applicable to autonomous driving systems according to claim 3, characterized in that, In step S41, the formula for calculating the distance observation probability is: Where, dis is the geometric projection of the vehicle's global coordinate position onto the candidate link; The distance observation probability is given by the formula for calculating the heading observation probability. in, For the probability of heading observation, The heading is the angle between the vehicle's direction of travel and due north.

5. A road-level localization method for complex scenarios suitable for autonomous driving systems according to claim 4, characterized in that, In step S41, the formula for calculating the semantic observation probability is: in, Similarity of road types; Similarity of road attributes; To ensure consistency in the number of lanes, the observation probability output by the Hidden Markov Model is calculated as follows: ;in, The observation probability of the candidate link output by the Hidden Markov Model; when X=1, the vehicle is in a complex scene. 0.5; When X=0, the vehicle is in a non-complex scenario. 0。 6. A road-level localization method for complex scenarios applicable to autonomous driving systems according to claim 3, characterized in that, In step S42, the candidate links are numbered sequentially and an adjacency matrix based on road topology is constructed. The transition probability of two candidate links is calculated using the adjacency matrix.

7. A road-level localization method for complex scenarios applicable to autonomous driving systems according to claim 3, characterized in that, The size of the variable sliding window is adjusted according to the vehicle speed to ensure that the historical driving distance covered by the window is greater than or equal to 300 meters; when the vehicle speed increases, the time length of the time axis is shortened, and when the vehicle speed decreases, the time length of the time axis is extended.

8. A road-level localization method for complex scenarios applicable to autonomous driving systems according to claim 1, characterized in that, In step S4, the method for determining the target link using a geometric projection distance and orientation weighted method is to obtain the best-matching target link by finding the minimum value after linear normalization weighting; the calculation formula for the linear normalization weighting is: Where X is the normalized geometric projection distance, and Y is the normalized angular error. and The weights are set as segmented parameters based on vehicle speed.

9. A road-level localization method for complex scenarios applicable to autonomous driving systems according to claim 1, characterized in that, The historical data records historical operational data including global coordinates, road network matching coordinates, and perception semantic information.

10. A road-level positioning system for complex scenarios suitable for autonomous driving systems, characterized in that, include: At least one processor; The system includes a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform a road-level localization method for complex scenarios applicable to an autonomous driving system according to any one of claims 1 to 9.

Images