Unmanned vehicle high-precision perception blind area compensation method and system based on green sight constraint

Abstract

The application discloses an unmanned vehicle high-precision perception blind area compensation method and system based on green sight line constraint, and belongs to the unmanned vehicle field, wherein the method comprises the following steps: acquiring self-vehicle perception data, map data and positioning data of the unmanned vehicle in real time; constructing a virtual green sight line area based on the map data and the positioning data; performing blind area detection on the self-vehicle perception data according to the virtual green sight line area, and acquiring an unmanned vehicle blind area detection result; and performing dynamic compensation on the self-vehicle perception data according to the unmanned vehicle blind area detection result, and outputting an unmanned vehicle perception dynamic map. The application solves the technical problems that the existing technology has a high false detection rate of the unmanned vehicle blind area detection, the compensation result has motion jump, and traffic safety accidents are easily caused.

Description

Technical Field

[0001] This invention relates to the field of autonomous vehicles, specifically to a method and system for high-precision perception blind spot compensation for autonomous vehicles based on green line-of-sight constraints. Background Technology

[0002] Currently, the environmental perception of autonomous vehicles mainly relies on multi-sensor fusion arrays such as onboard LiDAR, millimeter-wave radar, and high-definition cameras. However, due to the physical detection boundaries of sensors, occlusion of the vehicle body structure, and occlusion of static and dynamic obstacles in the road scene, there are a large number of perception blind spots in high-frequency and complex urban road scenarios such as turning at intersections, parallel driving of large vehicles, and meeting oncoming traffic in narrow roads. Traditional blind spot compensation technologies mostly rely on simple trajectory extrapolation based on historical data of the vehicle's sensors, or only rely on single data for fragmented supplementation. This results in blurred blind spot boundaries, delayed updates of temporary blind spots caused by dynamic occlusion, and inaccurate blind spot risk classification. It is very easy to cause collision accidents due to missed blind spot detection, which seriously restricts the safe operation and large-scale deployment of autonomous vehicles. Summary of the Invention

[0003] This application provides a high-precision perception blind spot compensation method and system for unmanned vehicles based on green line-of-sight constraints, aiming to solve the technical problems of high false detection rate and motion jump in compensation results in the existing technology of unmanned vehicles, which can easily lead to traffic safety accidents.

[0004] In view of the above problems, this application provides a method and system for high-precision perception blind spot compensation for unmanned vehicles based on green line-of-sight constraints.

[0005] The first aspect disclosed in this application provides a high-precision perception blind spot compensation method for unmanned vehicles based on green line-of-sight constraints, the method comprising:

[0006] The system acquires real-time vehicle perception data, map data, and positioning data; constructs a virtual green line of sight area based on the map data and positioning data; performs blind spot detection on the vehicle perception data according to the virtual green line of sight area to obtain the blind spot detection results; and dynamically compensates the vehicle perception data according to the blind spot detection results to output an animated image of the vehicle perception.

[0007] Another aspect of this application discloses a high-precision perception blind spot compensation system for autonomous vehicles based on green line-of-sight constraints, the system comprising:

[0008] The system includes an acquisition module for acquiring real-time vehicle perception data, map data, and positioning data; a construction module for constructing a virtual green line of sight area based on the map data and positioning data; a detection module for performing blind spot detection on the vehicle perception data based on the virtual green line of sight area and obtaining the blind spot detection results; and a compensation module for dynamically compensating the vehicle perception data based on the blind spot detection results and outputting an animated perception graph of the autonomous vehicle.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0010] By using multi-source fusion positioning data from autonomous vehicles as the observation origin, a virtual green line of sight area combining static permanent occlusion and dynamic temporary occlusion is constructed. The boundaries of perception blind spots are accurately delineated and traffic conflict risks are classified. Then, virtual compensation trajectories are generated based on differentiated strategies, and a compensation transition smoothing evaluation and optimization mechanism is introduced. This effectively reduces the false detection rate of perception blind spots in complex traffic scenarios, enhances the autonomous vehicle's ability to predict potential risks in blind spots and its response stability, and improves the driving safety of autonomous vehicles.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0012] Figure 1 A flowchart illustrating the high-precision perception blind spot compensation method for unmanned vehicles based on green line-of-sight constraints is provided for embodiments of this application.

[0013] Figure 2 A schematic diagram of the structure of an unmanned vehicle high-precision perception blind spot compensation system based on green line of sight constraints is provided for the embodiments of this application.

[0014] Explanation of reference numerals in the attached diagram: Acquisition module 11, Construction module 12, Detection module 13, Compensation module 14. Detailed Implementation

[0015] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0016] The overall concept of the technical solution provided in this application is as follows:

[0017] This application provides a method and system for high-precision perception blind spot compensation for unmanned vehicles based on green line-of-sight constraints. It acquires the vehicle's perception data, map data, and positioning data in real time; constructs a virtual green line-of-sight region based on the map data and positioning data; performs blind spot detection on the vehicle's perception data according to the virtual green line-of-sight region to obtain blind spot detection results; and dynamically compensates the vehicle's perception data based on the blind spot detection results, outputting an animated perception graph of the unmanned vehicle. Ultimately, this enhances the unmanned vehicle's ability to predict potential risks in blind spots and improves its response stability, thereby enhancing the vehicle's driving safety.

[0018] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0019] Example 1, as Figure 1 As shown in the embodiment of this application, a method for high-precision perception blind spot compensation for unmanned vehicles based on green line-of-sight constraints is provided. The method includes:

[0020] S100: Acquires real-time self-perception data, map data, and positioning data of the autonomous vehicle.

[0021] Specifically, the first step is to collect vehicle perception data, which is the raw and pre-processed data obtained by the autonomous vehicle through various environmental perception sensors on its own, reflecting the real-time status of dynamic obstacles and static environment around the vehicle. The collection relies on a multi-sensor fusion array deployed around the vehicle. Among them, the lidar outputs point cloud data containing three-dimensional coordinates and reflection intensity information to accurately depict the geometric contours of obstacles, the millimeter-wave radar provides radial velocity and distance information of targets that are not obscured by rain, fog, or dust, the high-definition camera captures texture, color, and semantic information to identify traffic signs, lane lines, and pedestrian types, and the ultrasonic radar is responsible for detecting blind spots at close range around the vehicle. All raw data is transmitted to the domain controller through the onboard time-sensitive network. After ground point cloud removal, camera distortion correction, outlier filtering, and spatial registration between sensors are completed, standardized vehicle perception data frames are generated.

[0022] Secondly, map data is acquired, specifically centimeter-level precision pre-construction environment data for autonomous driving, collected and constructed in advance by professional surveying vehicles and supported in real-time dynamic updates. This includes static environmental information such as road geometry and topology, lane line types and widths, traffic light and sign positions, roadside guardrails, and building outlines, as well as parameters affecting vehicle driving such as road slope, curvature, and superelevation. The acquisition adopts a three-level architecture of local offline storage + cloud dynamic updates + V2X roadside supplementation. The pre-collected offline high-precision map covering frequently driven areas is stored in the vehicle's high-speed solid-state drive. At the same time, the 5G-V2X vehicle-road cooperative communication module obtains real-time updated data of dynamic traffic elements such as temporary construction, road closures, and accident scenes from the cloud map server, and automatically aligns the timestamps with the positioning data to ensure strict consistency between the map coordinate system and the vehicle coordinate system.

[0023] Finally, positioning data is obtained, which is the parameter data used to determine the real-time three-dimensional position, attitude and motion state of the unmanned vehicle in the global high-precision map coordinate system. It uses multi-source fusion positioning technology to fuse the absolute position information provided by the global satellite navigation system, the vehicle angular velocity and acceleration data measured by the inertial measurement unit, the driving distance and steering angle data output by the wheel speed odometer, and the position correction results obtained by matching feature points of the high-precision map, and finally outputs the positioning result.

[0024] S200: Based on the map data and the positioning data, construct a virtual green line of sight area.

[0025] Specifically, the initial green line-of-sight region is obtained by first calibrating the high-precision map using the observation origin as a reference. The observation origin refers to the three-dimensional coordinates of the optical center of the vehicle's main LiDAR, determined based on upstream multi-source fusion positioning data, combined with the vehicle's real-time heading angle, pitch angle, and roll angle attitude data to form a spatial observation reference point. Line-of-sight calibration refers to emitting dense virtual rays from the observation origin to the edge vertices of all static obstacles in the high-precision map, including curbs, guardrails, buildings, trees, and traffic cones. The critical angle and distance at which each ray is blocked by static obstacles are calculated. By fitting the angle-distance envelope in polar coordinates, the maximum visible range that the vehicle's sensors can theoretically cover when there are no dynamic obstacles blocking the view is defined, which is the initial green line-of-sight region. This region is essentially a theoretical visible boundary based on static prior information, which can eliminate permanent blind spots caused by fixed obstacles in advance.

[0026] Subsequently, real-time line-of-sight analysis is performed on the observation origin to obtain the real-time line-of-sight area. This process first retrieves location-related traffic data based on the current positioning data of the unmanned vehicle, namely dynamic traffic information strongly correlated with the vehicle's current position, driving direction, and time window. This includes data on obstacles in the blind spot of the intersection perceived by the V2X roadside unit, perception results shared by surrounding vehicles through vehicle-to-vehicle communication, and information on temporary construction barriers and accident scenes pushed from the cloud. Then, these multi-source heterogeneous data are enhanced by performing nanosecond-level time synchronization, centimeter-level spatial registration, outlier filtering, and confidence weighting to generate spatiotemporally unified associated traffic enhancement data. Based on the location and size information of dynamic obstacles in the enhancement data, the above virtual ray occlusion calculation process is repeated to obtain the temporary occlusion range caused by dynamic obstacles, i.e., the real-time line-of-sight area.

[0027] Finally, the initial green line of sight area and the real-time line of sight area are fused by a dynamic correction algorithm, and the intersection of the two is taken as the final virtual green line of sight area. This area is the actual visible range of the autonomous vehicle that combines static permanent occlusion and dynamic temporary occlusion. The space not covered by this area is the perception blind spot that needs to be compensated.

[0028] S300: Perform blind spot detection on the vehicle perception data based on the virtual green line of sight area to obtain the blind spot detection result of the unmanned vehicle.

[0029] Specifically, the first step is to identify obstacles based on the preprocessed vehicle perception data. This obstacle feature identification refers to the process of using a multi-sensor fusion algorithm to perform Euclidean clustering and boundary extraction on the 3D point cloud of the LiDAR, semantic segmentation and target detection on the high-definition camera images, and velocity filtering and target tracking on the millimeter-wave radar data. Finally, the fusion is used to obtain the 3D spatial position, geometric size, movement speed, movement direction, and category attributes of each obstacle, i.e., pedestrian, vehicle, non-motorized vehicle, static obstacle, etc., as well as confidence information. All obstacles identified under continuous timestamps are arranged in chronological order to form an obstacle perception sequence. This sequence fully records the spatiotemporal evolution characteristics of dynamic and static obstacles in the vehicle's surrounding environment.

[0030] Subsequently, each obstacle in the obstacle perception sequence is transformed from its own coordinate system to a global high-precision map coordinate system consistent with the virtual green line of sight area. Then, the center point, four corner points, and key sampling points on the boundary contour of each obstacle are projected onto the polar coordinate plane of the virtual green line of sight area. The angle, distance, occlusion status, whether it is located within the green line of sight area, and its relative positional relationship with the observation origin of each projection point are extracted to obtain multiple point projection features.

[0031] Next, based on the projection features of these points, traffic conflict risk association is performed. Traffic conflict risk association refers to first identifying obstacle areas that are completely or partially located outside the virtual green line of sight area by comparing the occlusion status of the projection points, which are initially determined as perception blind spots. Then, combined with the real-time driving speed, steering angle and planned driving trajectory of the autonomous vehicle, the spatiotemporal overlap probability between the blind spot and the future driving path of the vehicle is calculated. The blind spot is divided into risk-free blind spots, including areas behind the vehicle that do not affect driving, and high-risk blind spots, including areas in front of the vehicle that are occluded at intersections. At the same time, the risk level is classified according to the occlusion type of the blind spot, namely static permanent occlusion / dynamic temporary occlusion, range size and distance from the vehicle. Finally, the blind spot detection results of the autonomous vehicle containing the spatial coordinates of the blind spot, range boundary, occlusion type, risk level and corresponding obstacle information are generated. At the same time, different levels of blind spot warning signals are generated according to the level of high-risk blind spots.

[0032] S400: Based on the blind spot detection results of the autonomous vehicle, dynamically compensate the autonomous vehicle perception data and output an animated image of the autonomous vehicle perception.

[0033] Specifically, an initial perception map is first constructed based on the preprocessed vehicle perception data. This initial perception map refers to a static environmental snapshot in a unified bird's-eye view coordinate system generated with the autonomous vehicle as the center and by fusing the perception results of multiple sensors and prior information from high-precision maps. The construction process involves projecting the three-dimensional position, geometric dimensions, motion state, and category information of all detected obstacles obtained from obstacle feature identification onto the bird's-eye view plane, and overlaying static environmental elements such as lane lines, traffic signs, and roadside guardrails from the high-precision map. Each element is labeled with a corresponding timestamp and confidence level, forming a basic perception layer that only contains information detected by the vehicle's sensors at the current moment.

[0034] Secondly, virtual compensation trajectories are generated based on the blind spot detection results of autonomous vehicles. Virtual compensation trajectories refer to trajectory sequences generated based on historical perception data, obstacle motion models, and traffic rule constraints for high-risk blind spot areas, simulating the spatiotemporal evolution of potential obstacles in the blind spot. The generation logic is as follows: First, high-risk blind spots with risk levels higher than a preset threshold are selected, and the spatial range, occlusion type, static permanent occlusion / dynamic temporary occlusion, and historical observation data of the blind spots are extracted. For blind spots formed by dynamic occlusion, based on the movement speed and direction of the occluder, combined with the historical frequency and movement characteristics of obstacles in the area, the possible movement paths and state changes of potential obstacles are predicted using a long short-term memory network. For blind spots formed by static occlusion, the perception data shared by the V2X roadside unit and surrounding vehicles is preferentially integrated to extract real obstacle information. When there is no external data, a conservative virtual obstacle trajectory is generated based on traffic flow statistics. All virtual compensation trajectories are assigned differentiated confidence weights according to the blind spot risk level. The trajectory of a high-risk blind spot has a higher confidence level and a stronger impact on the decision-making system.

[0035] Next, the initial perception map is dynamically compensated based on the virtual compensation trajectory to generate an autonomous vehicle perception animation. Here, the autonomous vehicle perception animation refers to a visual sequence that can dynamically display the complete environment around the vehicle and the movement status of potential blind spot obstacles, which is formed by stitching together multiple consecutive frames of perception maps that have undergone blind spot compensation in chronological order. The generation process is to decompose the virtual compensation trajectory into discrete compensation frames at a perception frame rate of 10Hz or higher. Each compensation frame corresponds to the position and status of the virtual obstacle at a certain moment. Then, the virtual obstacle in each compensation frame is superimposed on the corresponding blind spot position of the initial perception map at the corresponding moment to form a single frame compensation perception map. Finally, the consecutive single frame compensation perception maps are stitched together in chronological order to generate a complete dynamic perception sequence containing real obstacles and virtual compensation obstacles.

[0036] Finally, the generated autonomous vehicle perception animation is evaluated for smoothness through compensation transition. The compensation transition smoothness coefficient is calculated, which is a quantitative index obtained by weighting the position difference, velocity difference, and acceleration difference of the same virtual compensation obstacle in two adjacent frames. The smaller the coefficient, the more natural and smooth the motion transition of the virtual obstacle. If the calculated compensation transition smoothness coefficient is less than the preset compensation transition smoothness threshold, the final autonomous vehicle perception animation is directly output. If it is greater than or equal to the threshold, the virtual compensation trajectory is smoothed and optimized using a cubic Bézier curve interpolation algorithm, and the position and velocity parameters of the virtual obstacle in adjacent frames are adjusted until the smoothness requirements are met.

[0037] Furthermore, in the method provided in the application embodiment, constructing a virtual green line-of-sight region based on the map data and the positioning data includes: using the positioning data as the observation origin, performing line-of-sight calibration on the map data to obtain an initial green line-of-sight region; performing real-time line-of-sight analysis on the observation origin to obtain a real-time line-of-sight region; and correcting the initial green line-of-sight region according to the real-time line-of-sight region to generate the virtual green line-of-sight region.

[0038] Specifically, the initial green line-of-sight region is first obtained by calibrating the map data using the positioning data as the observation origin. The observation origin is the three-dimensional global coordinates of the optical center of the vehicle's main LiDAR, which is accurately calculated based on upstream multi-source fusion positioning data. At the same time, the vehicle's real-time heading angle, pitch angle, and roll angle attitude data are superimposed to form a spatial observation reference point that is completely consistent with the actual observation perspective of the sensor. The line-of-sight calibration refers to emitting dense virtual rays from the observation origin to all static obstacles in the high-precision map, including the edge vertices and contour key points of curbs, guardrails, buildings, trees, traffic cones, and central dividers, with an angular resolution of 0.1 degrees. The coordinates of the intersection point of each ray with the first static obstacle are calculated to obtain the maximum visible distance at that angle. By fitting the envelope of angle-maximum visible distance in polar coordinates, the maximum visible range that the autonomous vehicle sensor can theoretically cover when there are no dynamic obstacles obstructing the view is defined, which is the initial green line-of-sight region. This region is essentially a theoretical visible boundary based on static prior information, which can eliminate permanent blind spots caused by fixed obstacles in advance.

[0039] Subsequently, real-time line-of-sight analysis is performed on the observation origin to obtain the real-time line-of-sight area. This process first retrieves location-related traffic data based on the current positioning data of the autonomous vehicle, namely, multi-source dynamic traffic information within a 500-meter radius centered on the current positioning coordinates of the autonomous vehicle and within a time window of ±1 second. This includes data on blind spot obstacles at intersections perceived by V2X roadside units, perception results shared by surrounding vehicles through vehicle-to-vehicle communication, temporary construction barriers pushed by cloud map servers, accident scene and road control information, etc. Next, these heterogeneous multi-source data are enhanced by performing nanosecond-level time synchronization, centimeter-level spatial registration, outlier filtering, and confidence-weighted fusion to eliminate spatiotemporal biases and noise from different data sources, generating spatiotemporally unified and quantifiable associated traffic enhancement data. Then, based on the three-dimensional position and geometric size information of dynamic obstacles in the associated traffic enhancement data, the above virtual ray occlusion calculation process is repeated to obtain the temporary occlusion range caused by dynamic obstacles, i.e., the real-time line-of-sight area. This area reflects the actual occlusion of the autonomous vehicle's line of sight by dynamic traffic participants at the current moment.

[0040] Finally, the initial green line of sight area is corrected based on the real-time line of sight area to generate a virtual green line of sight area. The initial green line of sight area and the real-time line of sight area are merged by a dynamic weighted correction algorithm, and the intersection of the two is taken as the final virtual green line of sight area. This is the actual visible range of the autonomous vehicle that combines static permanent occlusion and dynamic temporary occlusion. The space not covered by this area is the perception blind spot that needs to be compensated.

[0041] Furthermore, in the method provided in the application embodiment, real-time line-of-sight analysis is performed on the observation origin to obtain a real-time line-of-sight region, including: retrieving associated traffic data based on the positioning data to obtain positioning-associated traffic data; performing enhancement processing on the positioning-associated traffic data to obtain associated traffic enhancement data; and performing line-of-sight detection on the observation origin based on the associated traffic enhancement data to generate the real-time line-of-sight region.

[0042] Specifically, the system first retrieves associated traffic data based on the current location data of the autonomous vehicle to obtain location-related traffic data. This location-related traffic data refers to all multi-source dynamic traffic information strongly related to vehicle driving safety within a 500-meter radius centered on the observation origin and within a time window of ±1 second. The retrieval logic is as follows: the autonomous driving domain controller sends a data request with spatiotemporal tags to the 5G-V2X communication module based on the real-time location, driving direction, and speed information output from the upstream. At the same time, it pulls data of the corresponding range from the cloud map server and the local traffic information cache in parallel. This includes data on blind spot obstacles at intersections perceived by the V2X roadside unit, real-time perception results shared by surrounding vehicles through vehicle-to-vehicle communication, temporary construction barriers pushed from the cloud, traffic accident scenes, temporary road control information, and real-time phase and countdown data of traffic lights released by the traffic management platform.

[0043] Next, the acquired heterogeneous location-related traffic data is enhanced. Enhancement processing refers to standardized preprocessing and weighted fusion operations to address issues such as spatiotemporal deviations, noise interference, inconsistent confidence levels, and inconsistent data formats that exist in different data sources. Specifically, this includes: synchronizing all data to a unified nanosecond-level time base through the PTP precise time protocol to eliminate clock drift errors from different devices; mapping local coordinate system data from different data sources such as roadside units and surrounding vehicles to a unified global high-precision map coordinate system through a coordinate transformation matrix; removing outliers and duplicate detection data through statistical filtering and density clustering algorithms; and assigning differentiated confidence weights based on the perception accuracy and reliability of different data sources, including the confidence level of roadside unit perception data, the confidence level of shared data from surrounding vehicles, and the confidence level of cloud-pushed construction information. Finally, spatiotemporally unified, quantifiable, and standardized format-based enhanced traffic data is generated.

[0044] Finally, line-of-sight detection is performed on the observation origin based on the associated traffic augmentation data to generate a real-time line-of-sight region. Here, line-of-sight detection refers to the process of calculating the occlusion range of dynamic obstacles on the autonomous vehicle's sensor line of sight starting from the observation origin. Specifically, it is implemented as follows: extract the three-dimensional spatial position, geometric dimensions, contour key points, and motion state information of all dynamic obstacles from the associated traffic augmentation data; emit dense virtual rays in a horizontal 360-degree direction with an angular resolution of 0.1 degrees that is completely consistent with the initial green line-of-sight region calibration; calculate the coordinates of the intersection point of each ray with the first dynamic obstacle; obtain the maximum visible distance limited by the dynamic obstacle at this angle; and define the autonomous vehicle's visible range at the current moment, considering only the occlusion of dynamic obstacles, i.e., the real-time line-of-sight region, by fitting the envelope of angle-dynamic maximum visible distance in polar coordinates. This region will be updated in real time with a frequency of more than 10Hz as the dynamic obstacle moves, ensuring synchronization with the actual traffic scene.

[0045] Furthermore, in the method provided in the application embodiment, blind spot detection is performed on the vehicle perception data based on the virtual green line of sight area to obtain the blind spot detection result of the unmanned vehicle, including: identifying obstacle features based on the vehicle perception data to obtain an obstacle perception sequence; projecting the obstacle perception sequence onto the virtual green line of sight area to obtain multiple point projection features; and associating traffic conflict risks based on the multiple point projection features to generate the blind spot detection result of the unmanned vehicle.

[0046] Specifically, obstacle feature recognition is first performed based on preprocessed vehicle perception data. Obstacle feature recognition refers to the process of complementary processing and feature fusion of raw data from different sensors through multi-sensor fusion perception algorithms. This includes Euclidean clustering, boundary extraction, and point cloud registration of LiDAR 3D point clouds to obtain the precise geometric shape and 3D spatial coordinates of obstacles; convolutional neural network semantic segmentation and target detection of high-definition camera images to identify the category attributes of obstacles, including pedestrians, motor vehicles, non-motor vehicles, static obstacles, etc., as well as texture features; and constant false alarm rate filtering and target tracking of millimeter-wave radar data to obtain the radial velocity and motion trend of obstacles. Finally, the data is fused to obtain structured information such as the unique ID, 3D spatial position, geometric dimensions, motion speed, motion direction, category attributes, and perception confidence of each obstacle. All obstacles identified under continuous timestamps are associated and arranged with their IDs in chronological order to form an obstacle perception sequence. This sequence fully records the spatiotemporal evolution characteristics of dynamic and static obstacles in the vehicle's surrounding environment and can reflect the motion continuity and state change patterns of obstacles.

[0047] Subsequently, each obstacle in the obstacle perception sequence undergoes coordinate system transformation, mapping from the vehicle's own coordinate system centered on the autonomous vehicle to a global high-precision map coordinate system consistent with the virtual green line of sight area. Then, the center point, four corner points, and key sampling points evenly distributed on the boundary contour of each obstacle, typically one sampling point per meter, are projected onto the polar coordinate plane of the virtual green line of sight area. Information such as the polar angle, polar radius, occlusion status, whether it is located within the angle-distance envelope of the virtual green line of sight area, its relative position to the observation origin, and its corresponding obstacle ID are extracted for each projection point. This yields multiple point projection features. By using multi-point projection instead of a single center point projection, partially occluded obstacles can be accurately identified, avoiding the misjudgment that the entire obstacle is visible simply because the center point falls within the visible area.

[0048] Next, based on the projection features of these points, traffic conflict risk correlation is performed. This correlation refers to the spatiotemporal matching of the spatial location of the blind spot with the future driving trajectory of the autonomous vehicle, quantifying the impact of the blind spot on vehicle driving safety. Specifically: First, by comparing the occlusion status of all points of each obstacle, obstacles completely outside the virtual green line of sight area, partially outside the area, and visible obstacles completely within the area are identified, initially defining the spatial range and boundaries of the blind spot; then, combining the real-time driving speed, steering angle, and planned driving trajectory output by the autonomous driving decision system, the spatiotemporal overlap probability of the blind spot and the vehicle's driving path within the next 3 seconds is calculated, thus defining the blind spot... Blind spots are categorized into risk-free blind spots (areas behind the vehicle that do not affect driving), low-risk blind spots (areas to the side of the vehicle at a distance without conflict), medium-risk blind spots (areas to the side of the vehicle at a close distance where lane changes may occur), and high-risk blind spots (pedestrian crossings or oncoming lanes at intersections in front of the vehicle that are obstructed). The risk level is dynamically adjusted based on the type of obstruction (static permanent obstruction / dynamic temporary obstruction), the size of the area, and its distance from the vehicle. The system ultimately generates blind spot detection results for the autonomous vehicle, including a unique blind spot ID, spatial coordinates, area boundaries, obstruction type, risk level, information on the corresponding obstructing object, and predictions of potential obstacles. Simultaneously, different levels of blind spot warning signals are generated based on the risk level of high-risk blind spots and transmitted to the vehicle's decision-making and planning system and the human-machine interface.

[0049] Furthermore, in the method provided in the application embodiment, the autonomous vehicle perception data is dynamically compensated based on the blind spot detection result of the autonomous vehicle, and an autonomous vehicle perception animation is output, including: constructing an initial perception map based on the autonomous vehicle perception data; generating a virtual compensation trajectory based on the blind spot detection result of the autonomous vehicle; and dynamically compensating the initial perception map based on the virtual compensation trajectory to generate the autonomous vehicle perception animation.

[0050] Specifically, an initial perception map is first constructed based on the preprocessed vehicle perception data. This initial perception map refers to a single-frame static environmental snapshot in a unified bird's-eye view coordinate system generated with the autonomous vehicle as the center and by fusing the perception results of multiple sensors and prior information from high-precision maps. The construction process involves projecting the three-dimensional position, geometric dimensions, motion state, and category information of all detected obstacles obtained from obstacle feature identification onto the bird's-eye view plane through a coordinate transformation matrix. Static environmental elements such as lane lines, traffic signs, roadside guardrails, and road boundaries from the high-precision map are then overlaid. Each element is labeled with a corresponding timestamp and perception confidence level, forming a basic perception layer that only contains information detected by the vehicle's sensors at the current moment. This layer serves as the reference canvas for blind spot compensation.

[0051] Secondly, virtual compensation trajectories are generated based on the blind spot detection results of the autonomous vehicle. These virtual compensation trajectories are continuous trajectory sequences generated for high-risk blind spot areas, based on historical perception data, obstacle motion models, and traffic rule constraints. They simulate the spatiotemporal evolution of potential obstacles in the blind spot. The generation employs a differentiated strategy: first, high-risk blind spots with risk levels exceeding a preset threshold are selected; then, the spatial range, occlusion type, static permanent occlusion / dynamic temporary occlusion, and historical observation data are extracted. For blind spots formed by dynamic occlusion, including oncoming lanes obscured by large trucks ahead and pedestrian crossings obscured by vehicles in adjacent lanes, the real-time movement speed of the obstructing objects is considered. By combining the degree and direction with the frequency of historical obstacle occurrences and typical motion characteristics in the area, a long short-term memory network is used to predict the possible motion paths, speed changes, and acceleration changes of potential obstacles. For blind spots formed by static occlusions, including non-motorized lanes obscured by intersection buildings and turning areas obscured by green belts, real obstacle information is extracted by prioritizing the fusion of real-time perception data shared by V2X roadside units and surrounding vehicles. When there is no external data, a conservative virtual obstacle trajectory is generated based on traffic flow statistics. All virtual compensation trajectories are assigned differentiated confidence weights according to the blind spot risk level. Trajectories in high-risk blind spots have higher confidence and a stronger impact on the decision-making system.

[0052] Finally, the initial perception map is dynamically compensated based on the virtual compensation trajectory to generate an autonomous vehicle perception animation. Here, the autonomous vehicle perception animation refers to a visual sequence composed of multiple consecutive frames of perception maps after blind spot compensation, stitched together in chronological order, which can dynamically display the complete environment around the vehicle and the motion state of potential blind spot obstacles. The generation process is to decompose the virtual compensation trajectory into discrete compensation frames at a perception frame rate of 10Hz or higher. Each compensation frame corresponds to the position and state of the virtual obstacle at a certain moment. Then, the virtual obstacle in each compensation frame is superimposed on the corresponding blind spot position of the initial perception map at the corresponding moment to form a single frame compensation perception map. At the same time, the compensation transition smoothness evaluation is performed on the generated consecutive single frame compensation perception maps. The position difference, velocity difference, and acceleration difference of the same virtual compensation obstacle in two adjacent frames are calculated to obtain the compensation transition smoothness coefficient. If the coefficient is greater than a preset threshold, the cubic Bézier curve interpolation algorithm is used to smooth and optimize the virtual compensation trajectory. Finally, a complete dynamic perception sequence containing real obstacles and virtual compensation obstacles is generated.

[0053] Furthermore, in the method provided in the application embodiment, generating the autonomous vehicle perception animation includes: performing a compensation transition smoothing evaluation on the autonomous vehicle perception animation to obtain a compensation transition smoothing coefficient; determining whether the compensation transition smoothing coefficient is less than a compensation transition smoothing threshold; and if the compensation transition smoothing coefficient is less than the compensation transition smoothing threshold, performing compensation transition smoothing optimization on the autonomous vehicle perception animation.

[0054] Specifically, the first step is to evaluate the smoothness of the transition in the initially stitched animated image of the autonomous vehicle. This evaluation involves frame-by-frame quantitative assessment of the abrupt changes in the motion state of the same virtual obstacle in a continuous time series. The core of this evaluation is calculating the smoothness coefficient, a weighted quantitative index that integrates position, velocity, and acceleration. The coefficient ranges from 0 to 1; a larger value indicates a more abrupt transition, while a smaller value indicates a smoother transition. Specifically, a cross-frame ID association algorithm is used to uniquely identify and track all virtual obstacles, ensuring that the same virtual obstacle remains smooth in consecutive frames. The motion state of obstacles can be accurately matched. Then, the three-dimensional position, velocity and acceleration data of the same virtual obstacle with the same ID are extracted from two adjacent frames, namely time t and time t-1. The position difference ΔP, velocity difference ΔV and acceleration difference ΔA are calculated respectively. Differential weights are then assigned according to the degree of influence of each dimension on the autonomous driving decision system, including position difference weight, velocity difference weight and acceleration difference weight. Finally, the compensation transition smoothness coefficient of the virtual obstacle in the current frame is obtained through normalization. For scenarios with multiple virtual compensation obstacles, the maximum value of the smoothness coefficients of all obstacles is taken as the final compensation transition smoothness coefficient of the entire frame perception animation.

[0055] Next, it is determined whether the final compensation transition smoothing coefficient is less than the preset compensation transition smoothing threshold. This threshold is an empirical value calibrated through tens of thousands of hours of real-vehicle testing, typically set between 0.15 and 0.25. It represents the maximum range of motion changes that the autonomous driving decision-making system can stably handle without causing control jitter. If the calculated compensation transition smoothing coefficient is less than this threshold, it indicates that the motion transition of the virtual compensation obstacle is natural and smooth, and the final autonomous vehicle perception animation can be directly output without additional optimization. If the compensation transition smoothing coefficient is greater than or equal to this threshold, it is determined that there is a motion jump problem, and the autonomous vehicle perception animation needs to be optimized for compensation transition smoothing. Compensation transition smoothing optimization refers to the process of correcting the motion trajectory of the virtual compensation obstacle through a high-order interpolation algorithm to eliminate temporal jumps. Specifically... The optimization is achieved using a cubic Bézier curve interpolation algorithm. Specifically, the time window to be optimized is first determined, typically a continuous sequence of 5 to 7 frames, consisting of 2 to 3 frames before and after the transition. Then, a smooth cubic Bézier curve is generated, using the position, velocity, and acceleration of the virtual obstacle in the last frame before the transition as the starting control point and the position, velocity, and acceleration of the virtual obstacle in the first frame after the transition as the ending control point. This curve is then discretized at a perception frame rate of 10Hz to obtain the corrected position, velocity, and acceleration parameters of the virtual obstacle in each intermediate frame. Finally, the corrected parameters are used to replace the original transition frame parameters, and the sequence is reassembled to generate the optimized autonomous vehicle perception animation sequence. For virtual compensation obstacles in high-risk blind spots, the system automatically lowers the compensation transition smoothing threshold by 0.05, implementing a more stringent smoothing standard.

[0056] Furthermore, in the method provided in the application embodiment, obtaining the blind spot detection result of the unmanned vehicle includes: generating a blind spot warning signal based on the blind spot detection result of the unmanned vehicle.

[0057] Specifically, the blind spot detection results of the unmanned vehicles output from the upstream are first reconfirmed for blind spot risk level. Based on the spatial location, occlusion type, range size, distance from the vehicle, and spatiotemporal overlap probability with the planned driving trajectory of the blind spot, the results are verified against a blind spot risk level mapping table that has been calibrated through tens of thousands of hours of real vehicle testing. The blind spot risk level mapping table refers to a rule base that maps the quantitative risk indicators of blind spots to standardized warning levels. Blind spots are usually divided into four levels: no-risk blind spot, no conflict area at a long distance behind and to the side of the vehicle, low-risk blind spot, no lane change intention at a medium distance to the side of the vehicle, medium-risk blind spot, possible lane change area at a close distance to the side of the vehicle, and high-risk blind spot, the driving path area in front of the vehicle and at intersections that is obscured, corresponding to warning levels 0 to 3 respectively.

[0058] Next, based on the verified warning level, differentiated machine-executable warning commands and multimodal human-machine warning signals are generated. The machine-executable warning commands are structured signals output in a standardized CAN bus data format that can be directly parsed by the autonomous driving decision-making and planning system. They include a unique warning ID, risk level, blind spot 3D coordinate range, obstruction ID and type, and suggested response actions, including maximum response time requirements for deceleration, maintaining the current lane, prohibiting lane changes, and emergency braking. The end-to-end transmission delay of high-risk warning commands is strictly controlled within 50ms to ensure that the decision-making system can trigger safe operations in a timely manner. The multimodal human-machine warning signals are multi-sensory collaborative prompts designed for human drivers, including visual prompts of blind spot positions on the dashboard (green for low risk, yellow for medium risk, and red for high risk), directional voice broadcast prompts, and seat zone vibration prompts, including seat back vibration and steering wheel vibration prompts corresponding to the blind spot direction when the risk is high, i.e., triggered when there is a blind spot on the side during lane changes.

[0059] Subsequently, a warning suppression mechanism is activated. This mechanism is a filtering logic designed to avoid frequent warnings interfering with the driving experience and the stability of the decision-making system. Specifically, it includes: when the same blind spot appears repeatedly within a 500ms time window and the risk level does not change, the current warning status is maintained and the warning is not triggered again; when the risk level of the blind spot decreases, the warning is automatically downgraded or gradually deactivated; for risk-free blind spots, no warning signal is generated; when the vehicle is stationary or the speed is below 5km / h, unnecessary low-to-medium risk warnings are automatically suppressed; finally, the generated warning signals are synchronously output to the autonomous driving domain controller, the in-vehicle human-machine interface, and the vehicle remote monitoring platform, and the occurrence time, precise location, risk level, triggering reason, and system response of all warning events are completely recorded in the in-vehicle black box and cloud server.

[0060] In summary, the high-precision perception blind spot compensation method for unmanned vehicles based on green line-of-sight constraints provided in this application has the following technical effects:

[0061] By using multi-source fusion positioning data from autonomous vehicles as the observation origin, a virtual green line of sight area combining static permanent occlusion and dynamic temporary occlusion is constructed. The boundaries of perception blind spots are accurately delineated and traffic conflict risks are classified. Then, virtual compensation trajectories are generated based on differentiated strategies, and a compensation transition smoothing evaluation and optimization mechanism is introduced. This effectively reduces the false detection rate of perception blind spots in complex traffic scenarios, enhances the autonomous vehicle's ability to predict potential risks in blind spots and its response stability, and improves the driving safety of autonomous vehicles.

[0062] Example 2 is based on the same inventive concept as the high-precision perception blind spot compensation method for unmanned vehicles based on green line-of-sight constraints in the previous examples, such as... Figure 2As shown in the figure, this application provides a high-precision perception blind spot compensation system for unmanned vehicles based on green line-of-sight constraints. The system includes:

[0063] The acquisition module 11 is used to acquire the autonomous vehicle's perception data, map data, and positioning data in real time; the construction module 12 is used to construct a virtual green line of sight area based on the map data and the positioning data; the detection module 13 is used to perform blind spot detection on the autonomous vehicle's perception data according to the virtual green line of sight area and obtain the autonomous vehicle's blind spot detection results; the compensation module 14 is used to dynamically compensate the autonomous vehicle's perception data according to the autonomous vehicle's blind spot detection results and output an autonomous vehicle perception animation.

[0064] Furthermore, the construction module 12 is also used to perform the following steps: using the positioning data as the observation origin, performing line-of-sight calibration on the map data to obtain an initial green line-of-sight area; performing real-time line-of-sight analysis on the observation origin to obtain a real-time line-of-sight area; and correcting the initial green line-of-sight area based on the real-time line-of-sight area to generate the virtual green line-of-sight area.

[0065] Furthermore, the construction module 12 is also used to perform the following steps: retrieve associated traffic data based on the positioning data to obtain positioning-related traffic data; perform enhancement processing on the positioning-related traffic data to obtain associated traffic enhancement data; and perform line-of-sight detection on the observation origin based on the associated traffic enhancement data to generate the real-time line-of-sight region.

[0066] Furthermore, the detection module 13 is also used to perform the following steps: identify obstacle features based on the autonomous vehicle perception data to obtain an obstacle perception sequence; project the obstacle perception sequence onto the virtual green line of sight area to obtain multiple point projection features; and associate traffic conflict risks based on the multiple point projection features to generate the autonomous vehicle blind spot detection result.

[0067] Furthermore, the compensation module 14 is also used to perform the following steps: constructing an initial perception map based on the autonomous vehicle perception data; generating a virtual compensation trajectory based on the blind spot detection results of the autonomous vehicle; and dynamically compensating the initial perception map based on the virtual compensation trajectory to generate the autonomous vehicle perception animation.

[0068] Furthermore, the compensation module 14 is also used to perform the following steps: evaluate the compensation transition smoothing of the autonomous vehicle perception animation and obtain the compensation transition smoothing coefficient; determine whether the compensation transition smoothing coefficient is less than the compensation transition smoothing threshold; if the compensation transition smoothing coefficient is less than the compensation transition smoothing threshold, optimize the compensation transition smoothing of the autonomous vehicle perception animation.

[0069] Furthermore, the detection module 13 is also used to perform the following steps: generating a blind spot warning signal based on the blind spot detection results of the unmanned vehicle.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A high-precision perception blind spot compensation method for unmanned vehicles based on green line-of-sight constraints, characterized in that, The method includes: Real-time acquisition of autonomous vehicle perception data, map data, and location data; Based on the map data and the location data, a virtual green line of sight area is constructed; Based on the virtual green line of sight area, blind spot detection is performed on the vehicle's perception data to obtain the blind spot detection result of the unmanned vehicle; Based on the blind spot detection results of the autonomous vehicle, the vehicle perception data is dynamically compensated, and an animated image of the autonomous vehicle perception is output.

2. The method for compensating for blind spots in high-precision perception of unmanned vehicles based on green line-of-sight constraints as described in claim 1, characterized in that, Based on the map data and the location data, a virtual green line of sight area is constructed, including: Using the positioning data as the observation origin, the map data is calibrated to obtain the initial green line of sight area; Real-time line-of-sight analysis is performed on the observation origin to obtain the real-time line-of-sight area; The initial green line of sight area is corrected based on the real-time line of sight area to generate the virtual green line of sight area.

3. The method for compensating blind spots in high-precision perception of unmanned vehicles based on green line-of-sight constraints as described in claim 2, characterized in that, Real-time line-of-sight analysis is performed on the observation origin to obtain the real-time line-of-sight area, including: Based on the location data, retrieve associated traffic data to obtain location-related traffic data; The location-related traffic data is enhanced to obtain enhanced traffic data. Based on the associated traffic enhancement data, line-of-sight detection is performed on the observation origin to generate the real-time line-of-sight region.

4. The method for compensating for blind spots in high-precision perception of unmanned vehicles based on green line-of-sight constraints as described in claim 1, characterized in that, Based on the virtual green line-of-sight area, blind spot detection is performed on the vehicle's perception data to obtain the blind spot detection results for the autonomous vehicle, including: Obstacle feature recognition is performed based on the vehicle's perception data to obtain an obstacle perception sequence; The obstacle perception sequence is projected onto the virtual green line of sight area to obtain projection features at multiple points. Based on the projection features of the multiple points, traffic conflict risks are associated to generate the blind spot detection results of the unmanned vehicle.

5. The method for compensating for blind spots in high-precision perception of unmanned vehicles based on green line-of-sight constraints as described in claim 1, characterized in that, Based on the blind spot detection results of the autonomous vehicle, the autonomous vehicle's perception data is dynamically compensated, and an animated perception graph of the autonomous vehicle is output, including: Based on the vehicle's perception data, an initial perception map is constructed; Based on the blind spot detection results of the unmanned vehicle, a virtual compensation trajectory is generated; The initial perception map is dynamically compensated based on the virtual compensation trajectory to generate the autonomous vehicle perception animation.

6. The method for high-precision perception blind spot compensation of unmanned vehicles based on green line-of-sight constraints as described in claim 5, characterized in that, Generating the animated perception map of the autonomous vehicle includes: The compensation transition smoothness evaluation is performed on the perception animation of the unmanned vehicle to obtain the compensation transition smoothness coefficient; Determine whether the compensation transition smoothing coefficient is less than the compensation transition smoothing threshold; If the compensation transition smoothing coefficient is less than the compensation transition smoothing threshold, the autonomous vehicle perception animation is optimized by compensation transition smoothing.

7. The method for compensating for blind spots in high-precision perception of unmanned vehicles based on green line-of-sight constraints as described in claim 1, characterized in that, Obtain blind spot detection results for autonomous vehicles, including: Based on the blind spot detection results of the unmanned vehicle, a blind spot warning signal is generated.

8. A high-precision perception blind spot compensation system for unmanned vehicles based on green line-of-sight constraints, characterized in that, The system is used to implement the high-precision perception blind spot compensation method for unmanned vehicles based on green line-of-sight constraints as described in any one of claims 1 to 7, the system comprising: The acquisition module is used to acquire the autonomous vehicle's perception data, map data, and positioning data in real time. The construction module is used to construct a virtual green line of sight area based on the map data and the positioning data; The detection module is used to perform blind spot detection on the vehicle's perception data based on the virtual green line of sight area, and obtain the blind spot detection result of the unmanned vehicle; The compensation module is used to dynamically compensate the autonomous vehicle's perception data based on the blind spot detection results of the autonomous vehicle, and output an animated image of the autonomous vehicle's perception.

Images