Intelligent lifting and rotating signboard early warning method and system based on visual recognition

Abstract

The application discloses an intelligent lifting and rotating signboard early warning method and system based on visual identification, which comprises the following steps: collecting continuous road scene image data of a target road area and performing pretreatment; extracting road disturbance features and generating a road disturbance feature set; generating source disturbance positions and source propagation directions by inversion; constructing risk propagation correlation and generating a risk propagation prediction field; performing car sight line projection and road visual field obstruction and generating driving visual field blind area distribution results; constructing a digital twin road mirror image of the target road area and generating a signboard candidate action set; mapping the signboard candidate action set to the digital twin road mirror image for pre-performance; generating a target early warning action result; controlling the intelligent lifting and rotating signboard to perform actions and completing early directional early warning arrangement. The application adopts visual identification and digital twin pre-performance methods, realizes directional early warning arrangement of the signboard, and has the advantages of high precision, high efficiency and small interference.

Description

Technical Field

[0001] This invention relates to the field of road traffic safety early warning, and in particular to an intelligent lifting and rotating sign warning method and system based on visual recognition. Background Technology

[0002] Existing road traffic early warning technologies mainly rely on fixed signs, preset time periods, manual patrols to issue warning information, or trigger prompts based on single video detection results. Although some solutions can identify vehicle, lane, and road occupancy status and can also combine with electronic sign equipment to issue information, they are mostly limited to the identification and display control of local abnormal states, lacking collaborative analysis of the formation and propagation process of road disturbances and the visibility conditions of oncoming vehicles.

[0003] The shortcomings of existing technologies are that it is difficult to invert the location and direction of the source disturbance based on continuous road scene data, and it is impossible to predict the range and sequence of risk propagation in advance. At the same time, there is insufficient consideration for blind spots in the driver's field of vision, and the warning direction, height and timing of sign warnings are usually not targeted, which can easily lead to problems such as delayed warning release, mismatch of warning areas, and additional disturbances to normal traffic. Therefore, it is difficult to achieve early targeted warning deployment for dynamic risk propagation. Summary of the Invention

[0004] One objective of this invention is to propose an intelligent lifting and rotating sign warning method based on visual recognition. This invention uses visual recognition and digital twin pre-simulation methods to achieve directional warning deployment of signs, which has the advantages of accuracy, efficiency and low interference.

[0005] A visual recognition-based intelligent lifting and rotating sign warning method according to an embodiment of the present invention includes the following steps: Collect continuous road scene image data of the target road area, perform preprocessing, and generate a standardized road scene dataset; Based on a standardized road scene dataset, road disturbance features are extracted to generate a road disturbance feature set; Based on the road disturbance feature set, deceleration inheritance relationship, avoidance inheritance relationship, trajectory deviation inheritance relationship, traffic flow compression inheritance relationship and stagnation transmission relationship are extracted, and the source disturbance location and source propagation direction are generated by combining the standardized road scene dataset; Based on the road disturbance feature set, the source disturbance location and the source propagation direction, the risk propagation correlation relationship corresponding to each spatial location within the target road area is constructed to generate a risk propagation prediction field. Based on a standardized road scene dataset, oncoming vehicle line-of-sight projection and road visual field occlusion are performed to generate the distribution results of blind spots in the driver's visual field. Based on standardized road scene datasets, risk propagation prediction fields, and blind spot distribution results in the driver's field of vision, a digital twin road image of the target road area is constructed, and a set of candidate actions for signs is generated. The candidate action set of signs is mapped to the digital twin road mirror for pre-playing, and the risk reduction results, blind spot reduction results and traffic disturbance results corresponding to each candidate action of the sign are extracted to generate early warning action evaluation results; Based on the evaluation results of the early warning actions, generate the target early warning action results; Based on the target warning action results, the intelligent lifting and rotating signboard is controlled to perform lifting and rotating actions, and the advance directional warning deployment is completed.

[0006] Optionally, the generation of the standardized road scene dataset specifically includes: Acquire continuous road scene image data of the target road area, and perform time alignment, spatial calibration, image denoising and distortion correction on the continuous road scene image data; Based on the processed continuous road scene image data, vehicle target information, lane structure information, road boundary information, occlusion information and environmental state information are extracted at each time index through visual recognition; Target association is performed on vehicle target information, and anomaly removal is performed on the data entries corresponding to vehicle target information, lane structure information, road boundary information, occlusion information and environmental state information to form a standardized road scene dataset.

[0007] Optionally, the generation of the road disturbance feature set specifically includes: Read vehicle target information and lane structure information under each time index in the standardized road scene dataset, and organize the data entries of the same vehicle target under continuous time index in sequence according to the vehicle target correspondence; Based on the road position coordinate changes of the same vehicle target under adjacent time indices, the position change features corresponding to each vehicle target are extracted. Based on the magnitude and time interval of the road position coordinate changes of the same vehicle target under continuous time index, the speed change features corresponding to each vehicle target are extracted; By combining the motion direction change and lane connection direction change of the same vehicle target under continuous time index, the directional change features corresponding to each vehicle target are extracted. Based on the change in road position coordinate interval between adjacent vehicle targets under the same time index, extract the vehicle distance change features corresponding to each vehicle target; Based on the lane passage area switching status of each vehicle target under the continuous time index, extract the lane change change features corresponding to each vehicle target; Based on the road position maintenance state and motion amplitude maintenance state of each vehicle target under the continuous time index, extract the stagnation change features corresponding to each vehicle target; The characteristics of position change, speed change, direction change, distance change, lane change, and stationary change are organized according to vehicle target and time index to form a road disturbance feature set.

[0008] Optionally, the generation of the source disturbance location and the source propagation direction specifically includes: Read the position change features, speed change features, direction change features, distance change features, lane change features, and stationary change features corresponding to each vehicle target in the road disturbance feature set, and organize them according to the vehicle target and time index; Based on the speed change characteristics of preceding and subsequent vehicle targets under adjacent time indices, deceleration continuity is extracted. Combined with the directional change characteristics and lane change characteristics between adjacent vehicle targets, avoidance continuity is extracted. Based on the position change characteristics and directional change characteristics between adjacent vehicle targets, trajectory offset continuity is extracted. Based on the vehicle distance and speed change characteristics of multiple vehicle targets under continuous time index, the traffic flow compression succession relationship is extracted. Based on the stagnation change characteristics of preceding and subsequent vehicle targets under continuous time index, the stagnation transmission relationship is extracted. Read the road position coordinates of each vehicle target in the standardized road scene dataset under each time index, and map the deceleration relationship, avoidance relationship, trajectory offset relationship, traffic flow compression relationship and stagnation relationship to the corresponding road position; Reverse convergence localization is performed on each mapped connection along the continuous time index to determine the common starting road location, generate the source disturbance location, and generate the source propagation direction according to the expansion order of the road locations corresponding to each connection.

[0009] Optionally, the generation of the risk propagation prediction field specifically includes: Starting from the location of the source disturbance, and combining the direction of source propagation, the road locations distributed along the direction of source propagation within the target road area are sequentially arranged and divided into multiple spatial location units; Read the position change features, speed change features, direction change features, distance change features, lane change features and stationary change features corresponding to each vehicle target in the road disturbance feature set, and map each vehicle target to each spatial location unit according to the time index and road location to generate a spatial location disturbance feature distribution sequence. Based on the spatial location disturbance feature distribution sequence, the order of appearance of various road disturbance features between the previous spatial location unit and the next spatial location unit is compared item by item to extract the continuation results of position change features, speed change features, direction change features, distance change features, lane change features, and stagnation features. Continuous statistics are performed on the results of the continuity of various road disturbance features extracted between spatial location units. Adjacent spatial location unit combinations that satisfy the continuous continuity of two or more types of road disturbance features are selected, and these adjacent spatial location unit combinations are identified as risk propagation associations. The various risk propagation relationships are connected in chronological order according to their spatial location to generate a risk propagation path that extends outward from the source disturbance location along the source propagation direction; The risk propagation sequence of each spatial location unit in the risk propagation path is determined based on the connection order of each spatial location unit in the target road area, and the risk propagation coverage is determined based on the range of spatial location units reached by the risk propagation path. The risk propagation sequence and risk propagation coverage are then organized to generate a risk propagation prediction field.

[0010] Optionally, the generation of the blind spot distribution results in the driving field of vision specifically includes: Read vehicle target information, lane structure information, road boundary information, occlusion information and environmental status information under each time index in the standardized road scene dataset, filter vehicle targets that enter the warning impact range along the travel direction of the target road area, and generate a set of oncoming vehicle targets; Based on the road position and movement direction of each incoming vehicle target in the set of incoming vehicle targets, and combined with the lane passage area direction in the lane structure information, line of sight extension matching is performed on the road position corresponding to each incoming vehicle target in front of the target road area to generate a set of incoming vehicle line of sight projection paths. Based on road boundary information, the effective projection range in the set of projection paths of each oncoming vehicle is bounded, and the occlusion position on the projection path of each oncoming vehicle is checked in combination with the occlusion information, so as to generate a set of road vision occlusion results corresponding to each oncoming vehicle target. Based on the lighting conditions, weather conditions, visibility conditions, and road surface appearance conditions in the environmental status information, visible distance correction and occlusion boundary correction are performed on the road visual occlusion result set corresponding to each incoming vehicle target to generate the invisible road location set corresponding to each incoming vehicle target. The set of invisible road locations corresponding to each incoming vehicle target is summarized in order of road location to generate the driving blind spot distribution result corresponding to the target road area.

[0011] Optionally, the generation of the candidate action set for the signboard specifically includes: Read the standardized road scene dataset, risk propagation prediction field, and driving blind spot distribution results, and establish a unified correspondence based on time index and road location; Based on vehicle target information, lane structure information, road boundary information, and occlusion information, scene reconstruction is performed on the vehicle distribution location, lane passage area, road edge location, and occlusion location within the target road area. Combined with environmental state information, environmental state mapping is performed to generate a digital twin road image. The sequence of risk propagation, the coverage of risk propagation, and the distribution of blind spots in the driving field are mapped to the corresponding road locations in the digital twin road mirror, forming a digital twin road mirror that includes the risk propagation status and the driving blind spot status. Based on the coverage of risk propagation, the warning area of ​​the sign is determined in the digital twin road mirror; based on the order of risk propagation, the triggering order of the sign action is determined; based on the distribution of blind spots in the driver's field of vision, the direction of the warning action of the sign and the range of the sign's raising and lowering action are determined. Based on the warning area of ​​the sign, the triggering sequence of the sign action, the warning direction of the sign, and the lifting range of the sign, different lifting heights, different rotation directions, different action timings, different warning release states, and different warning pointing states are combined to generate a set of candidate sign actions.

[0012] Optionally, the generation of the early warning action evaluation result specifically includes: Read the digital twin road mirror and the candidate action set of the sign, and use the road position where the sign is located in the digital twin road mirror as the candidate action loading position, and call the candidate actions of each sign in sequence; In the digital twin road mirror, a baseline pre-play scene without loading candidate actions for signs and an action pre-play scene with loading candidate actions for signs are generated respectively. The sign raising and lowering actions, rotation actions, and warning release actions are pre-played, and the pre-play scene results corresponding to each candidate action for signs are generated. Based on the baseline pre-simulation scenario and the results of the pre-simulation scenario, the risk propagation sequence and risk propagation coverage corresponding to each spatial location unit in the risk propagation prediction field are called. The risk propagation coverage reduction and risk propagation sequence adjustment before and after loading each sign candidate action are compared to generate the risk reduction results corresponding to each sign candidate action. Based on the distribution of blind spots in the driving field of vision, the changes in the number and range of blind spot road locations corresponding to the warning area of ​​the sign in the baseline pre-simulation scenario and the pre-simulation scenario results are statistically analyzed to generate blind spot reduction results corresponding to each sign candidate action. Extract the position change features, speed change features, and lane change features corresponding to the vehicle target. Compare the position change feature disturbance, speed change feature disturbance, and lane change feature disturbance before and after loading the candidate actions of each sign to generate the traffic disturbance results corresponding to each candidate action of the sign. The risk reduction results, blind spot reduction results, and traffic disturbance results corresponding to each candidate action of the sign are correlated and organized to generate the early warning action assessment results.

[0013] Optionally, the generation of the target early warning action result specifically includes: Read the risk reduction results, blind spot reduction results, and traffic disturbance results corresponding to each sign candidate action in the early warning action assessment results, and establish a correspondence according to the candidate lifting height, candidate rotation direction, candidate action timing, candidate early warning release status, and candidate early warning pointing status; Based on the risk reduction results, blind spot reduction results, and traffic disturbance results corresponding to each sign's candidate action, the optimal candidate action is generated; The candidate elevation height, candidate rotation direction, candidate action timing, candidate warning release status, and candidate warning pointing status corresponding to the optimal candidate action are respectively determined as the target elevation height, target rotation direction, target action timing, target warning release status, and target warning pointing status, and then associated and organized to generate the target warning action result.

[0014] An intelligent lifting and rotating sign warning system based on visual recognition according to an embodiment of the present invention includes: The road scene acquisition module is used to acquire continuous road scene image data of the target road area, perform preprocessing, and generate a standardized road scene dataset. The road disturbance feature extraction module is used to extract road disturbance features based on a standardized road scene dataset and generate a road disturbance feature set. The source disturbance inversion module is used to extract deceleration inheritance relationships, avoidance inheritance relationships, trajectory deviation inheritance relationships, traffic flow compression inheritance relationships and stagnation transmission relationships based on the road disturbance feature set, and invert to generate the source disturbance location and source propagation direction by combining the standardized road scene dataset; The risk propagation prediction module is used to construct the risk propagation correlation between various spatial locations within the target road area based on the road disturbance feature set, the source disturbance location, and the source propagation direction, and generate a risk propagation prediction field. The driving vision blind spot analysis module is used to project oncoming vehicle lines of sight and road vision occlusion based on a standardized road scene dataset, and generate driving vision blind spot distribution results. The digital twin mirror and candidate action generation module is used to construct a digital twin road mirror of the target road area based on a standardized road scene dataset, risk propagation prediction field, and blind spot distribution results in the driver's field of vision, and to generate a set of candidate actions for signs. The pre-simulation evaluation module is used to map the set of candidate actions for signs to a digital twin road mirror for pre-simulation, extract the risk reduction results, blind spot reduction results and traffic disturbance results corresponding to each candidate action for signs, and generate early warning action evaluation results; The target action generation module is used to generate target early warning action results based on the early warning action evaluation results; The sign control module is used to control the intelligent lifting and rotating sign to perform lifting and rotating actions based on the target warning action results, and to complete the early directional warning deployment.

[0015] The beneficial effects of this invention are: This invention does not rely on a single abnormal event to trigger an early warning. Instead, it uses continuous road scene image data as a foundation. First, it extracts road disturbance features, then inverts the source disturbance location and propagation direction through deceleration, avoidance, trajectory deviation, traffic flow compression, and stagnation transmission relationships. Based on this, it constructs a risk propagation prediction field, transforming the early warning generation process from a reactive response to a proactive assessment and judgment of the risk propagation process. This technical approach enables the identification of road disturbance propagation trends before they spread to a wider area, improving the proactiveness, continuity, and targeting of road risk identification, thereby providing a more reliable data foundation for subsequent early warning actions.

[0016] This invention further integrates the distribution results of blind spots in the driver's field of vision with the risk propagation prediction field to construct a digital twin road mirror, and performs pre-analysis of candidate sign actions within the digital twin road mirror. This not only determines the warning area and action triggering sequence of the sign based on the risk propagation coverage and sequence, but also determines the warning direction and lifting / lowering range of the sign by combining the driver's blind spot distribution results. This ensures that the lifting, rotating, and warning release actions of the sign match the risk propagation state and oncoming vehicle visibility state in the real road scenario. This improves the directional accuracy, timing adaptability, and spatial matching of warning deployment, reducing the problem of insufficient warning effectiveness caused by improper warning location, orientation deviation, or action timing offset.

[0017] This invention also comprehensively evaluates the risk reduction results, blind spot reduction results, and traffic disturbance results corresponding to each candidate sign action to generate a target warning action result, and then controls the intelligent lifting and rotating sign to complete the advance directional warning deployment. This scheme not only focuses on risk reduction capabilities but also simultaneously considers the improvement effect on the driver's field of vision and the impact on normal traffic conditions, ensuring that the final determined target warning action balances safety warning effectiveness and road operation stability. Therefore, this invention can improve the control accuracy, warning effectiveness, and scene adaptability of intelligent lifting and rotating signs in complex road scenarios, enhancing the initiative and intelligence level of road traffic warning. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an intelligent lifting and rotating sign warning method based on visual recognition proposed in this invention; Figure 2 This is a schematic diagram illustrating the generation of a twin road mirror and a set of candidate actions for a visual recognition-based intelligent lifting and rotating sign warning method proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 and Figure 2 A visual recognition-based intelligent lifting and rotating sign warning method includes the following steps: Collect continuous road scene image data of the target road area, perform preprocessing, and generate a standardized road scene dataset; Based on a standardized road scene dataset, road disturbance features are extracted to generate a road disturbance feature set; Based on the road disturbance feature set, deceleration inheritance relationship, avoidance inheritance relationship, trajectory deviation inheritance relationship, traffic flow compression inheritance relationship and stagnation transmission relationship are extracted, and the source disturbance location and source propagation direction are generated by combining the standardized road scene dataset; Based on the road disturbance feature set, the source disturbance location and the source propagation direction, the risk propagation correlation relationship corresponding to each spatial location within the target road area is constructed to generate a risk propagation prediction field. Based on a standardized road scene dataset, oncoming vehicle line-of-sight projection and road visual field occlusion are performed to generate the distribution results of blind spots in the driver's visual field. Based on standardized road scene datasets, risk propagation prediction fields, and blind spot distribution results in the driver's field of vision, a digital twin road image of the target road area is constructed, and a set of candidate actions for signs is generated. The candidate action set of signs is mapped to the digital twin road mirror for pre-playing, and the risk reduction results, blind spot reduction results and traffic disturbance results corresponding to each candidate action of the sign are extracted to generate early warning action evaluation results; Based on the evaluation results of the early warning actions, generate the target early warning action results; Based on the target warning action results, the intelligent lifting and rotating signboard is controlled to perform lifting and rotating actions, and the advance directional warning deployment is completed.

[0021] In this embodiment, the generation of the standardized road scene dataset specifically includes: Acquire continuous road scene image data of the target road area, and perform time alignment, spatial calibration, image denoising and distortion correction on the continuous road scene image data; Based on the processed continuous road scene image data, vehicle target information, lane structure information, road boundary information, occlusion information and environmental state information are extracted at each time index through visual recognition; Target association is performed on vehicle target information, and anomaly removal is performed on the data entries corresponding to vehicle target information, lane structure information, road boundary information, occlusion information and environmental state information to form a standardized road scene dataset.

[0022] In this embodiment, the generation of the road disturbance feature set specifically includes: Read vehicle target information and lane structure information under each time index in the standardized road scene dataset, and organize the data entries of the same vehicle target under continuous time index in sequence according to the vehicle target correspondence; Based on the road position coordinate changes of the same vehicle target under adjacent time indices, the position change features corresponding to each vehicle target are extracted. Based on the magnitude and time interval of the road position coordinate changes of the same vehicle target under continuous time index, the speed change features corresponding to each vehicle target are extracted; By combining the motion direction change and lane connection direction change of the same vehicle target under continuous time index, the directional change features corresponding to each vehicle target are extracted. Based on the change in road position coordinate interval between adjacent vehicle targets under the same time index, extract the vehicle distance change features corresponding to each vehicle target; Based on the lane passage area switching status of each vehicle target under the continuous time index, extract the lane change change features corresponding to each vehicle target; Based on the road position maintenance state and motion amplitude maintenance state of each vehicle target under the continuous time index, extract the stagnation change features corresponding to each vehicle target; Position change features are the changes in road position coordinates of the same vehicle target under adjacent time indices; speed change features are the changes in the speed of movement of the same vehicle target under continuous time indices based on the magnitude and time interval of changes in road position coordinates; direction change features are the changes in the direction of movement of the same vehicle target under continuous time indices relative to the lane connection direction; distance change features are the changes in the interval between road position coordinates of adjacent vehicle targets under the same time index; lane change change features are the changes in the switching of lane passage areas of the same vehicle target under continuous time indices; and stagnation change features are the stagnation change features characterized by the road position maintenance state and the motion amplitude maintenance state of the same vehicle target under continuous time indices. The characteristics of position change, speed change, direction change, distance change, lane change, and stationary change are organized according to vehicle target and time index to form a road disturbance feature set.

[0023] In this embodiment, the generation of the source disturbance location and the source propagation direction specifically includes: Read the position change features, speed change features, direction change features, distance change features, lane change features, and stationary change features corresponding to each vehicle target in the road disturbance feature set, and organize them according to the vehicle target and time index; Based on the speed change characteristics of preceding and subsequent vehicle targets under adjacent time indices, deceleration continuity is extracted. Combined with the directional change characteristics and lane change characteristics between adjacent vehicle targets, avoidance continuity is extracted. Based on the position change characteristics and directional change characteristics between adjacent vehicle targets, trajectory offset continuity is extracted. The system extracts position change features, speed change features, direction change features, and lane change features of each vehicle target under the same road segment and adjacent time indices from the road disturbance feature set, and establishes a continuous correspondence between vehicle targets and time indices. Based on the sequential road position along the vehicle's direction of travel under the same time index, the vehicle target in front is identified as the preceding vehicle target, and the vehicle target behind is identified as the following vehicle target. The preceding and following vehicle targets that maintain an adjacent road position in the next time index are selected as the objects for determining the succession relationship. First, it checks whether the preceding vehicle target shows a decrease in speed change features from the current time index to the next time index. Then, it checks whether the following vehicle target shows a decrease in speed change features consistent with the preceding vehicle target in subsequent adjacent time indices. If the speed change features of the preceding vehicle target decrease first, and the speed change features of the following vehicle target follow suit along the continuous time index, then the change process between the preceding and following vehicle targets is determined as... Deceleration continuity relationship; When extracting avoidance continuity relationship, continue to check the directional change characteristics and lane change characteristics between adjacent vehicle targets. If the preceding vehicle target first changes its directional change characteristics or switches its lane change characteristics, and the subsequent vehicle target shows a responsive directional change characteristic deflection in response to the disturbance direction of the preceding vehicle target in the subsequent adjacent time index, or shows a lane change characteristic switch consistent with the preceding vehicle target, then the change process between the preceding vehicle target and the subsequent vehicle target is determined as an avoidance continuity relationship; When extracting trajectory offset continuity relationship, read the position change characteristics and directional change characteristics between adjacent vehicle targets. If the preceding vehicle target first shows a position change characteristic deviating from the original travel path direction, accompanied by a directional change characteristic deflection, and the subsequent vehicle target shows a continuous position change characteristic offset along the offset side consistent with the preceding vehicle target in the subsequent adjacent time index, accompanied by a directional change characteristic deflection, then the continuous offset process between the preceding vehicle target and the subsequent vehicle target is determined as a trajectory offset continuity relationship; Based on the vehicle distance and speed change characteristics of multiple vehicle targets under continuous time index, the traffic flow compression succession relationship is extracted. Based on the stagnation change characteristics of preceding and subsequent vehicle targets under continuous time index, the stagnation transmission relationship is extracted. The system extracts the distance change features, speed change features, and stagnation change features of each vehicle target under the same road segment and continuous time index from the road disturbance feature set, and establishes a continuous correspondence according to the vehicle target, time index, and road position. When extracting the traffic flow compression continuity relationship, multiple vehicle targets are first arranged sequentially along the vehicle travel direction under the same time index. Then, it checks whether the distance change features between adjacent vehicle targets continuously show a decrease in the difference between vehicle distances and an increase in the amount of distance contraction. At the same time, it checks whether the speed change features of the corresponding vehicle targets continuously show a downward trend. If the vehicle target in front first shows a contraction in distance change features and a decrease in speed change features, the vehicle target behind will subsequently show the same contraction in distance change features and speed change features under the continuous time index. If the change in the vehicle characteristics decreases and this change is continuously transmitted between multiple vehicle targets along the vehicle's direction of travel, then the continuous contraction process formed between these multiple vehicle targets is identified as a traffic flow compression succession relationship. When extracting the stagnation transmission relationship, according to the method for determining the preceding and subsequent vehicle targets, it is checked whether the preceding vehicle target first shows a stagnation change characteristic under the continuous time index, and then it is determined whether the subsequent vehicle target shows a stagnation change characteristic consistent with the preceding vehicle target under the subsequent continuous time index. If the preceding vehicle target first shows a stagnation change characteristic, and the subsequent vehicle target subsequently shows an increase in the continuous stagnation duration or an increase in the stagnation frequency, and continues along the continuous time index after the preceding vehicle target, then the stagnation continuation process between the preceding and subsequent vehicle targets is identified as a stagnation transmission relationship. Read the road position coordinates of each vehicle target in the standardized road scene dataset under each time index, and map the deceleration relationship, avoidance relationship, trajectory offset relationship, traffic flow compression relationship and stagnation relationship to the corresponding road position; Reverse convergence localization is performed on each mapped connection along the continuous time index to determine the common starting road location, generate the source disturbance location, and generate the source propagation direction according to the expansion order of the road locations corresponding to each connection. First, read the deceleration, avoidance, trajectory deviation, traffic flow compression, and stagnation transmission relationships mapped to the corresponding road locations, and arrange them in reverse order from back to front according to the time index. Using the latest time index corresponding to each relationship as the starting point for reverse tracing, trace back along the continuous time index to the road location corresponding to the relationship under the previous time index, recording the road location change trajectory of each relationship under the continuous time index. After completing the reverse tracing of each relationship, compare the road location change trajectories of each relationship within the same road segment, and filter those that fall into two categories. The earliest road location that all the above connection relationships point to is determined as the common starting road location. After determining the common starting road location, it is determined as the source disturbance location. When generating the source propagation direction, the source disturbance location is used as the starting point. The order of expansion of each connection relationship on subsequent road locations is checked along the continuous time index. The subsequent road locations are connected sequentially along the expansion order of the road locations corresponding to each connection relationship. The direction of road location arrangement extending outward from the source disturbance location is determined, and this road location arrangement direction is determined as the source propagation direction.

[0024] In this embodiment, the generation of the risk propagation prediction field specifically includes: Starting from the location of the source disturbance, and combining the direction of source propagation, the road locations distributed along the direction of source propagation within the target road area are sequentially arranged and divided into multiple spatial location units; A spatial location unit is a set of adjacent road locations formed by continuously dividing the target road area along the direction of propagation from the source. Read the position change features, speed change features, direction change features, distance change features, lane change features and stationary change features corresponding to each vehicle target in the road disturbance feature set, and map each vehicle target to each spatial location unit according to the time index and road location to generate a spatial location disturbance feature distribution sequence. Based on the spatial location disturbance feature distribution sequence, the order of appearance of various road disturbance features between the previous spatial location unit and the next spatial location unit is compared item by item to extract the continuation results of position change features, speed change features, direction change features, distance change features, lane change features, and stagnation features. The results of the continuation of position change features, speed change features, direction change features, distance change features, lane change features, and stagnation features are all judgment results generated on whether the corresponding road disturbance features are continuously transmitted between the previous spatial position unit and the next spatial position unit. Among them, continuous transmission means that the corresponding road disturbance feature that appears first in the previous spatial position unit reappears in the next spatial position unit along the subsequent time index, and the order of appearance is consistent with the direction of propagation from the source. Continuous statistics are performed on the results of the continuity of various road disturbance features extracted between spatial location units. Adjacent spatial location unit combinations that satisfy the continuous continuity of two or more types of road disturbance features are selected, and these adjacent spatial location unit combinations are identified as risk propagation associations. A single road disturbance feature may be caused by normal following, local avoidance or short-term congestion during road operation, and cannot stably represent risk propagation. However, when two or more types of road disturbance features continue together along a continuous time index between the preceding and following spatial location units, it can indicate that the disturbance has been transformed from a single behavioral change into a continuous spatial transmission process. Therefore, the combination of these adjacent spatial location units can be identified as a risk propagation correlation. Risk propagation correlation is the continuous transmission relationship of disturbances between the preceding and following spatial location units. It is used to characterize the road disturbance features that have appeared in the preceding spatial location unit and have been transmitted to the following spatial location unit along the source propagation direction. The various risk propagation relationships are connected in chronological order according to their spatial location to generate a risk propagation path that extends outward from the source disturbance location along the source propagation direction; The risk propagation sequence of each spatial location unit in the risk propagation path is determined based on the connection order of each spatial location unit in the target road area, and the risk propagation coverage is determined based on the range of spatial location units reached by the risk propagation path. The risk propagation sequence and risk propagation coverage are then organized to generate a risk propagation prediction field.

[0025] In this embodiment, the generation of the blind spot distribution result in the driving field of vision specifically includes: Read vehicle target information, lane structure information, road boundary information, occlusion information and environmental status information under each time index in the standardized road scene dataset, filter vehicle targets that enter the warning impact range along the travel direction of the target road area, and generate a set of oncoming vehicle targets; Based on the road position and movement direction of each incoming vehicle target in the set of incoming vehicle targets, and combined with the lane passage area direction in the lane structure information, line of sight extension matching is performed on the road position corresponding to each incoming vehicle target in front of the target road area to generate a set of incoming vehicle line of sight projection paths. Line-of-sight matching starts with the road position and movement direction of each incoming vehicle target, and continuously matches the road positions in the target road area along its forward travel direction to determine whether each road position is within the forward line-of-sight coverage of the incoming vehicle target. Based on road boundary information, the effective projection range in the set of projection paths of each oncoming vehicle is bounded, and the occlusion position on the projection path of each oncoming vehicle is checked in combination with the occlusion information, so as to generate a set of road vision occlusion results corresponding to each oncoming vehicle target. Based on the lighting conditions, weather conditions, visibility conditions, and road surface appearance conditions in the environmental status information, visible distance correction and occlusion boundary correction are performed on the road visual occlusion result set corresponding to each incoming vehicle target to generate the invisible road location set corresponding to each incoming vehicle target. The set of invisible road locations is a set of road locations that cannot be directly observed by each oncoming vehicle target after road visual field occlusion judgment and environmental state correction on the corresponding line of sight projection path. The set of invisible road locations corresponding to each incoming vehicle target is summarized in order of road location to generate the driving blind spot distribution result corresponding to the target road area.

[0026] In this embodiment, the generation of the candidate action set for the signboard specifically includes: Read the standardized road scene dataset, risk propagation prediction field, and driving blind spot distribution results, and establish a unified correspondence based on time index and road location; Based on vehicle target information, lane structure information, road boundary information, and occlusion information, scene reconstruction is performed on the vehicle distribution location, lane passage area, road edge location, and occlusion location within the target road area. Combined with environmental state information, environmental state mapping is performed to generate a digital twin road image. Digital twin road mirroring is a virtual road scene formed by synchronously reconstructing the real road scene based on the vehicle target information, lane structure information, road boundary information, occlusion information and environmental state information of the target road area under the corresponding time index. It is used to reflect the vehicle distribution state, road structure state, occlusion state and environmental state of the target road area in the virtual environment, and to carry the risk propagation prediction field, the distribution results of driver's blind spot, and the pre-rehearsal process of sign candidate actions. The sequence of risk propagation, the coverage of risk propagation, and the distribution of blind spots in the driving field are mapped to the corresponding road locations in the digital twin road mirror, forming a digital twin road mirror that includes the risk propagation status and the driving blind spot status. Based on the coverage of risk propagation, the warning area of ​​the sign is determined in the digital twin road mirror; based on the order of risk propagation, the triggering order of the sign action is determined; based on the distribution of blind spots in the driver's field of vision, the direction of the warning action of the sign and the range of the sign's raising and lowering action are determined. First, the risk propagation coverage area, risk propagation sequence, and blind spot distribution results in the digital twin road mirror are read, and the risk propagation coverage area is mapped to each road location in the digital twin road mirror. The relative positional relationship between each road location within the risk propagation coverage area and the road location of the sign is examined, and the road locations within the risk propagation coverage area that are continuously distributed in front of the sign along the source propagation direction are determined as the sign warning area. After determining the sign warning area, the arrival sequence of each road location within the risk propagation coverage area is examined along the risk propagation sequence, and the road location that first arrives at the sign warning area within the risk propagation coverage area is used as the corresponding warning area. The time index serves as the starting time for triggering the sign action, and the sign action triggering sequence is formed according to the risk propagation arrival order of each subsequent road location. The distribution of blind spot road locations corresponding to the sign warning area is read from the blind spot distribution results in the driver's field of vision. The directional relationship between the road location where the sign is located and the concentrated distribution area of ​​blind spot road locations is checked, and the direction from the road location where the sign is located to the concentrated distribution area of ​​blind spot road locations is determined as the warning direction of the sign. The road distance range and visible coverage height requirements corresponding to each blind spot road location within the warning area of ​​the sign are statistically analyzed, and the lifting range that can cover all blind spot road locations within the warning area of ​​the sign is determined as the lifting range of the sign. Based on the warning area of ​​the sign, the triggering sequence of the sign action, the warning direction of the sign, and the lifting range of the sign, a set of candidate sign actions is generated by combining different lifting heights, different rotation directions, different action timings, different warning release states, and different warning pointing states. Different warning issuance states refer to the different warning output states of the sign during the rehearsal or actual control process, including the warning not issued, the warning issued, and the warning continuous. Different warning directional states refer to the different warning action directional states of the sign during the rehearsal or actual control process, including the directional states towards the left-hand traffic direction, towards the middle-hand traffic direction, and towards the right-hand traffic direction.

[0027] In this embodiment, the generation of the early warning action evaluation result specifically includes: Read the digital twin road mirror and the candidate action set of the sign, and use the road position where the sign is located in the digital twin road mirror as the candidate action loading position, and call the candidate actions of each sign in sequence; In the digital twin road mirror, a baseline pre-play scene without loading candidate actions for signs and an action pre-play scene with loading candidate actions for signs are generated respectively. The sign raising and lowering actions, rotation actions, and warning release actions are pre-played, and the pre-play scene results corresponding to each candidate action for signs are generated. The generation of the pre-simulation scenario results specifically includes: first, reading the road location of the sign in the digital twin road mirror, and, while keeping the road location of the sign, vehicle distribution location, lane passage area, road edge location, occlusion location, and environmental state unchanged, determining the digital twin road mirror without loading any sign candidate actions as the baseline pre-simulation scenario; sequentially reading each sign candidate action in the sign candidate action set, and loading the candidate lifting height, candidate rotation direction, candidate action timing, candidate warning release status, and candidate warning pointing status of the current sign candidate action into the road location of the sign in the digital twin road mirror; under the time index corresponding to the candidate action timing, first controlling the sign to move from the initial height to the candidate lifting height in the digital twin road mirror, then controlling the sign to rotate from the initial orientation to the candidate rotation direction, and then... The candidate warning release status is executed to display the warning content, and the effect direction of the warning content is adjusted according to the candidate warning direction status to form the action pre-play scene corresponding to the current sign candidate action; after the loading of a single sign candidate action is completed, the candidate's lifting height, rotation direction, warning release status, and warning direction status are maintained along the continuous time index after the candidate action timing. The vehicle distribution position, occlusion position, risk propagation status, and driving blind spot status in the digital twin road mirror are updated synchronously to obtain the complete pre-play scene result corresponding to the sign candidate action under the continuous time index; after the current sign candidate action pre-play is completed, the digital twin road mirror is restored to the baseline pre-play scene without loading the sign candidate action, and then the next sign candidate action is loaded and the above process is repeated until the pre-play scene result corresponding to each sign candidate action is generated; Based on the baseline pre-simulation scenario and the results of the pre-simulation scenario, the risk propagation sequence and risk propagation coverage corresponding to each spatial location unit in the risk propagation prediction field are called. The risk propagation coverage reduction and risk propagation sequence adjustment before and after loading each sign candidate action are compared to generate the risk reduction results corresponding to each sign candidate action. The generation of risk reduction results specifically includes: first, reading the baseline pre-simulation scenario, the pre-simulation scenario results, and the risk propagation sequence and coverage of each spatial location unit in the risk propagation prediction field, and establishing a one-to-one correspondence according to the spatial location unit order and time index order; using the baseline pre-simulation scenario as a reference scenario for candidate actions without loading signs, counting the number of spatial location units within the risk propagation coverage area in the baseline pre-simulation scenario and the arrival sequence of risk propagation for each spatial location unit, forming the baseline risk propagation distribution result; performing the same processing on the pre-simulation scenario results, counting the number of spatial location units covered after loading sign candidate actions and the arrival sequence of risk propagation for each spatial location unit, forming the action risk propagation distribution result; and then... The risk propagation distribution results are compared item by item with the action risk propagation distribution results. First, the changes in the number of spatial location units within the risk propagation coverage area before and after the candidate action of loading the sign are compared to determine the reduction in the risk propagation coverage area. Then, the changes in the arrival order of risk propagation for each spatial location unit are compared to determine the adjustment of the risk propagation order. When the number of spatial location units covered after loading the candidate action of loading the sign is less than the number of spatial location units covered in the baseline pre-simulation scenario, and the arrival order of risk propagation for each spatial location unit is shifted relative to the baseline pre-simulation scenario, the reduction in the risk propagation coverage area and the adjustment of the risk propagation order corresponding to the candidate action of the sign are associated and organized to generate the risk reduction result corresponding to the candidate action of the sign. Based on the distribution of blind spots in the driving field of vision, the changes in the number and range of blind spot road locations corresponding to the warning area of ​​the sign in the baseline pre-simulation scenario and the pre-simulation scenario results are statistically analyzed to generate blind spot reduction results corresponding to each sign candidate action. The generation of blind spot reduction results specifically includes: first, reading the blind spot road locations corresponding to the warning area of ​​the sign in the blind spot distribution results of the driving field of vision; using the baseline pre-playing scenario without the candidate action of the sign as a control scenario, counting the number of blind spot road locations within the warning area of ​​the sign in the baseline pre-playing scenario, and determining the start and end positions of the blind spot road locations in the road position sequence to form the baseline blind spot distribution results; performing the same processing on the pre-playing scenario results after the candidate action of the sign is added, counting the number of blind spot road locations within the warning area of ​​the sign in the pre-playing scenario results, and determining the corresponding blind spots. The starting and ending positions of the road locations in the road location sequence form the blind spot distribution result. After extracting the baseline blind spot distribution result and the action blind spot distribution result, the two are compared item by item. First, the difference in the number of blind spot road locations before and after the candidate action of the sign is loaded is compared to determine the change in the number of blind spot road locations. Then, the changes in the starting and ending positions of the blind spot road locations in the road location sequence are compared to determine the change in the range of blind spot road locations. The changes in the number of blind spot road locations and the changes in the range of blind spot road locations corresponding to the candidate action of the sign are associated and organized to generate the blind spot reduction result corresponding to the candidate action of the sign. Extract the position change features, speed change features, and lane change features corresponding to the vehicle target. Compare the position change feature disturbance, speed change feature disturbance, and lane change feature disturbance before and after loading the candidate actions of each sign to generate the traffic disturbance results corresponding to each candidate action of the sign. The generation of traffic disturbance results specifically includes: first, using the baseline pre-play scene without candidate actions for signboards as a control scene, continuously tracking vehicle targets in the baseline pre-play scene according to time index and road position order, and extracting the position change features, speed change features, and lane change features corresponding to each vehicle target to form baseline traffic feature results; performing the same processing on the pre-play scene results after loading candidate actions for signboards, extracting the position change features, speed change features, and lane change features corresponding to each vehicle target according to the same vehicle target identifier, time index, and road position order as the baseline pre-play scene to form action traffic feature results; and finally, after completing the extraction of baseline traffic feature results and action traffic feature results... Then, a one-to-one correspondence is established between the two based on the vehicle target identifier and the time index. First, the differences in position change characteristics of each vehicle target before and after the loading of the sign candidate action are compared to determine the position change characteristic disturbance. Then, the differences in speed change characteristics of each vehicle target before and after the loading of the sign candidate action are compared to determine the speed change characteristic disturbance. Next, the differences in lane change characteristics of each vehicle target before and after the loading of the sign candidate action are compared to determine the lane change characteristic disturbance. The position change characteristic disturbance, speed change characteristic disturbance, and lane change characteristic disturbance are then associated and organized according to the vehicle target identifier, time index, and road position order to generate the traffic disturbance result corresponding to the sign candidate action. The risk reduction results, blind spot reduction results, and traffic disturbance results corresponding to each candidate action of the sign are correlated and organized to generate the early warning action assessment results.

[0028] In this embodiment, the generation of the target warning action result specifically includes: Read the risk reduction results, blind spot reduction results, and traffic disturbance results corresponding to each sign candidate action in the early warning action assessment results, and establish a correspondence according to the candidate lifting height, candidate rotation direction, candidate action timing, candidate early warning issuance status, and candidate early warning pointing status; Based on the risk reduction results, blind spot reduction results, and traffic disturbance results corresponding to each sign's candidate action, the optimal candidate action is generated; The generation of optimal candidate actions specifically includes: reading the risk reduction results, blind spot reduction results, and traffic disturbance results corresponding to each sign's candidate action in the early warning action evaluation results, and extracting the risk propagation coverage reduction value, risk propagation sequence adjustment value, blind spot road location quantity change value, blind spot road location range change value, location change characteristic disturbance value, speed change characteristic disturbance value, and lane change characteristic disturbance value corresponding to each sign's candidate action; and processing the risk propagation coverage reduction value, risk propagation sequence adjustment value, blind spot road location quantity change value, and blind spot road location range change value. Normalization is performed on the disturbance values ​​of position change, speed change, and lane change characteristics. Then, the normalized values ​​for the reduction in risk propagation coverage, the adjustment in the order of risk propagation, the change in the number of blind spot road locations, the change in the range of blind spot road locations, the disturbance values ​​of position change, speed change, and lane change characteristics are weighted and summed to generate a comprehensive evaluation value for each signpost candidate action. The comprehensive evaluation values ​​for each signpost candidate action are compared, and the signpost candidate action with the highest comprehensive evaluation value is selected as the optimal candidate action. The candidate elevation height, candidate rotation direction, candidate action timing, candidate warning release status, and candidate warning pointing status corresponding to the optimal candidate action are respectively determined as the target elevation height, target rotation direction, target action timing, target warning release status, and target warning pointing status, and then associated and organized to generate the target warning action result.

[0029] A vision-recognition-based intelligent lifting and rotating sign warning system includes: The road scene acquisition module is used to acquire continuous road scene image data of the target road area, perform preprocessing, and generate a standardized road scene dataset. The road disturbance feature extraction module is used to extract road disturbance features based on a standardized road scene dataset and generate a road disturbance feature set. The source disturbance inversion module is used to extract deceleration inheritance relationships, avoidance inheritance relationships, trajectory deviation inheritance relationships, traffic flow compression inheritance relationships and stagnation transmission relationships based on the road disturbance feature set, and invert to generate the source disturbance location and source propagation direction by combining the standardized road scene dataset; The risk propagation prediction module is used to construct the risk propagation correlation between various spatial locations within the target road area based on the road disturbance feature set, the source disturbance location, and the source propagation direction, and generate a risk propagation prediction field. The driving vision blind spot analysis module is used to project oncoming vehicle lines of sight and road vision occlusion based on a standardized road scene dataset, and generate driving vision blind spot distribution results. The digital twin mirror and candidate action generation module is used to construct a digital twin road mirror of the target road area based on a standardized road scene dataset, risk propagation prediction field, and blind spot distribution results in the driver's field of vision, and to generate a set of candidate actions for signs. The pre-simulation evaluation module is used to map the set of candidate actions for signs to a digital twin road mirror for pre-simulation, extract the risk reduction results, blind spot reduction results and traffic disturbance results corresponding to each candidate action for signs, and generate early warning action evaluation results; The target action generation module is used to generate target early warning action results based on the early warning action evaluation results; The sign control module is used to control the intelligent lifting and rotating sign to perform lifting and rotating actions based on the target warning action results, and to complete the early directional warning deployment.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a two-way road in a mountainous area where a sharp bend along a cliff meets a downhill slope. This section of road is obstructed by a continuous mountainside on one side and has a protective slope on the other. Trees, guardrails, and temporary construction barriers are also present ahead of the bend. Before entering the bend, oncoming vehicles cannot promptly observe whether vehicles behind them are slowing down, yielding, or partially stopping. This easily leads to situations where disturbances have already formed and are spreading along the road, while subsequent vehicles continue to enter the bend area according to their original driving patterns. Especially in rainy weather, low light in the early morning, and with strong road surface reflection, drivers' ability to detect risks ahead is further delayed. Traditional fixed signs can only provide static warnings and cannot be adjusted synchronously based on the location of the disturbance, its direction of propagation, and the blind spots of oncoming vehicles. Therefore, there are problems such as misaligned warning locations, inaccurate warning directions, and delayed warning timing, making it difficult to effectively intervene in road risks in advance.

[0031] In this scenario, image acquisition devices are deployed above and to the sides of the road to continuously collect continuous road scene image data of the target road area, and the acquisition results are input into the method flow of this invention. The system first performs time alignment, spatial calibration, image denoising, and distortion correction on the continuous road scene image data to form a standardized road scene dataset; then, it extracts vehicle target information, lane structure information, road boundary information, occlusion information, and environmental state information from it, and further extracts position change features, speed change features, direction change features, distance change features, lane change features, and stagnation change features to form a road disturbance feature set. When vehicles continuously decelerate near the exit of a curve, and then following vehicles decelerate, deviate from their tracks, and experience traffic flow compression, the system does not directly issue a simple alarm. Instead, it first performs reverse convergence positioning of the disturbance origin based on deceleration inheritance relationships, avoidance inheritance relationships, track deviation inheritance relationships, traffic flow compression inheritance relationships, and stagnation transmission relationships to determine the source disturbance location, and determines the source propagation direction by combining the expansion of each inheritance relationship along the road space. Based on this, the system continues to construct a risk propagation prediction field to determine which spatial units the disturbance has spread outward along, and which road locations will be affected first in the future.

[0032] After completing the risk propagation prediction, the system further combines the vehicle's direction of travel, road boundaries, mountain obstructions, fence obstructions, and the current environmental conditions to perform line-of-sight projection and road visual field occlusion analysis on oncoming vehicles entering the warning impact range, generating the driver's blind spot distribution results. Subsequently, the system maps the risk propagation prediction field and the driver's blind spot distribution results onto a digital twin road mirror, determining the warning area, action triggering sequence, warning direction, and lifting range of the sign within the virtual road scene, and generating multiple candidate sign actions. These candidate actions are not directly executed but are first rehearsed in the digital twin road mirror, analyzing the impact of different lifting heights, rotation directions, action timings, and warning orientations on reducing the risk propagation range, improving blind spot coverage, and controlling normal traffic disturbances. After comparison, the system selects the candidate action with the highest overall suitability as the target warning action result, then controls the intelligent lifting and rotating sign to complete the corresponding lifting and rotation deployment, oriented the sign towards the direction of oncoming vehicles that need to be alerted, and entering the effective visibility range at a more appropriate time.

[0033] To verify the performance of the present invention, it was compared with the traditional method. The comparison results are shown in Table 1.

[0034] Table 1. Comparison of the effectiveness of intelligent lifting and rotating sign warning methods with traditional fixed warning methods. Mountain winding road section Traditional fixed early warning methods 2.8 12.4 72.1 12.6 70.3 Mountain winding road section Method of the present invention 6.7 44.6 91.2 4.5 91.8 Temporary construction blocking road sections Traditional fixed early warning methods 2.6 10.1 69.8 13.4 68.5 Temporary construction blocking road sections Method of the present invention 6.3 41.7 89.6 5.2 89.4 Low light conditions at night Traditional fixed early warning methods 2.3 8.9 67.5 14.1 66.8 Low light conditions at night Method of the present invention 5.9 38.8 87.9 5.7 88.1 As shown in Table 1, the method of this invention outperforms traditional fixed warning methods in various road environments, and the improvement effect is consistent. In mountainous curved road sections, the risk identification lead time of the method of this invention is increased from 2.8 seconds to 6.7 seconds, and the effective warning reach rate is increased from 72.1% to 91.2%. This indicates that the invention can complete risk identification and warning action deployment in the early stages of road disturbance formation and expansion, enabling oncoming vehicles to receive effective warning information earlier. Simultaneously, the blind spot coverage improvement rate is increased from 12.4% to 44.6%, indicating that the invention can adjust the warning direction and lifting range of the sign based on the blind spot distribution results in the driver's field of vision, thereby significantly enhancing the visual warning effect in complex curved road sections.

[0035] In temporary construction-obstructed road sections, traditional fixed early warning methods are insufficiently adaptable to changes in obstructions due to the relatively fixed warning location and orientation. Therefore, the effective warning reach rate is only 69.8%, and the accuracy rate of warning action matching is only 68.5%. In contrast, the method of this invention improves these rates to 89.6% and 89.4%, respectively. This demonstrates that the invention can construct a digital twin road mirror based on a standardized road scene dataset, a risk propagation prediction field, and the distribution results of blind spots in the driver's field of vision. Furthermore, it pre-screens candidate actions for signs within the digital twin road mirror, resulting in a higher consistency between the final generated target warning action and the actual road risk state. Therefore, it can improve the targeting and adaptability of early warning deployment.

[0036] In low-light road sections at night, the method of this invention maintains a high level of effectiveness, with blind spot coverage improvement rate increasing from 8.9% to 38.8%, effective warning reach rate increasing from 67.5% to 87.9%, and traffic disturbance increase decreasing from 14.1% to 5.7%. This indicates that the invention not only improves warning effectiveness but also reduces additional impact on normal traffic flow while enhancing warning capabilities. The main reason for this is that the invention does not directly trigger a single warning action. Instead, it first rehearses different candidate actions for different signs and then evaluates the results of risk reduction, blind spot reduction, and traffic disturbance before determining the target warning action. Therefore, it achieves a better balance between risk warning capability and road operational stability.

[0037] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for early warning of intelligent lifting and rotating signs based on visual recognition, characterized in that, Includes the following steps: Collect continuous road scene image data of the target road area, perform preprocessing, and generate a standardized road scene dataset; Based on a standardized road scene dataset, road disturbance features are extracted to generate a road disturbance feature set; Based on the road disturbance feature set, deceleration inheritance relationship, avoidance inheritance relationship, trajectory deviation inheritance relationship, traffic flow compression inheritance relationship and stagnation transmission relationship are extracted, and the source disturbance location and source propagation direction are generated by combining the standardized road scene dataset; Based on the road disturbance feature set, the source disturbance location and the source propagation direction, the risk propagation correlation relationship corresponding to each spatial location within the target road area is constructed to generate a risk propagation prediction field. Based on a standardized road scene dataset, oncoming vehicle line-of-sight projection and road visual field occlusion are performed to generate the distribution results of blind spots in the driver's visual field. Based on standardized road scene datasets, risk propagation prediction fields, and blind spot distribution results in the driver's field of vision, a digital twin road image of the target road area is constructed, and a set of candidate actions for signs is generated. The candidate action set of signs is mapped to the digital twin road mirror for pre-playing, and the risk reduction results, blind spot reduction results and traffic disturbance results corresponding to each candidate action of the sign are extracted to generate early warning action evaluation results; Based on the evaluation results of the early warning actions, generate the target early warning action results; Based on the target warning action results, the intelligent lifting and rotating signboard is controlled to perform lifting and rotating actions, and the advance directional warning deployment is completed; The generation of the early warning action evaluation results specifically includes: Read the digital twin road mirror and the candidate action set of the sign, and use the road position where the sign is located in the digital twin road mirror as the candidate action loading position, and call the candidate actions of each sign in sequence; In the digital twin road mirror, a baseline pre-play scene without loading candidate actions for signs and an action pre-play scene with loading candidate actions for signs are generated respectively. The sign raising and lowering actions, rotation actions, and warning release actions are pre-played, and the pre-play scene results corresponding to each candidate action for signs are generated. Based on the baseline pre-simulation scenario and the results of the pre-simulation scenario, the risk propagation sequence and risk propagation coverage corresponding to each spatial location unit in the risk propagation prediction field are called. The risk propagation coverage reduction and risk propagation sequence adjustment before and after loading each sign candidate action are compared to generate the risk reduction results corresponding to each sign candidate action. Based on the distribution of blind spots in the driving field of vision, the changes in the number and range of blind spot road locations corresponding to the warning area of ​​the sign in the baseline pre-simulation scenario and the pre-simulation scenario results are statistically analyzed to generate blind spot reduction results corresponding to each sign candidate action. Extract the position change features, speed change features, and lane change features corresponding to the vehicle target. Compare the position change feature disturbance, speed change feature disturbance, and lane change feature disturbance before and after loading the candidate actions of each sign to generate the traffic disturbance results corresponding to each candidate action of the sign. The risk reduction results, blind spot reduction results, and traffic disturbance results corresponding to each candidate action for a sign are correlated and organized to generate an early warning action assessment result. The generation of the target early warning action result specifically includes: Read the risk reduction results, blind spot reduction results, and traffic disturbance results corresponding to each sign candidate action in the early warning action assessment results, and establish a correspondence according to the candidate lifting height, candidate rotation direction, candidate action timing, candidate early warning release status, and candidate early warning pointing status; Based on the risk reduction results, blind spot reduction results, and traffic disturbance results corresponding to each sign's candidate action, the optimal candidate action is generated; The candidate elevation height, candidate rotation direction, candidate action timing, candidate warning release status, and candidate warning pointing status corresponding to the optimal candidate action are respectively determined as the target elevation height, target rotation direction, target action timing, target warning release status, and target warning pointing status, and then associated and organized to generate the target warning action result.

2. The intelligent lifting and rotating sign warning method based on visual recognition according to claim 1, characterized in that, The generation of the standardized road scene dataset specifically includes: Acquire continuous road scene image data of the target road area, and perform time alignment, spatial calibration, image denoising and distortion correction on the continuous road scene image data; Based on the processed continuous road scene image data, vehicle target information, lane structure information, road boundary information, occlusion information and environmental state information are extracted at each time index through visual recognition; Target association is performed on vehicle target information, and anomaly removal is performed on the data entries corresponding to vehicle target information, lane structure information, road boundary information, occlusion information and environmental state information to form a standardized road scene dataset.

3. The intelligent lifting and rotating sign warning method based on visual recognition according to claim 1, characterized in that, The generation of the road disturbance feature set specifically includes: Read vehicle target information and lane structure information under each time index in the standardized road scene dataset, and organize the data entries of the same vehicle target under continuous time index in sequence according to the vehicle target correspondence; Based on the road position coordinate changes of the same vehicle target under adjacent time indices, the position change features corresponding to each vehicle target are extracted. Based on the magnitude and time interval of the road position coordinate changes of the same vehicle target under continuous time index, the speed change features corresponding to each vehicle target are extracted; By combining the motion direction change and lane connection direction change of the same vehicle target under continuous time index, the directional change features corresponding to each vehicle target are extracted. Based on the change in road position coordinate interval between adjacent vehicle targets under the same time index, extract the vehicle distance change features corresponding to each vehicle target; Based on the lane passage area switching status of each vehicle target under the continuous time index, extract the lane change change features corresponding to each vehicle target; Based on the road position maintenance state and motion amplitude maintenance state of each vehicle target under the continuous time index, extract the stagnation change features corresponding to each vehicle target; The characteristics of position change, speed change, direction change, distance change, lane change, and stationary change are organized according to vehicle target and time index to form a road disturbance feature set.

4. The intelligent lifting and rotating sign warning method based on visual recognition according to claim 1, characterized in that, The generation of the source disturbance location and the source propagation direction specifically includes: Read the position change features, speed change features, direction change features, distance change features, lane change features, and stationary change features corresponding to each vehicle target in the road disturbance feature set, and organize them according to the vehicle target and time index; Based on the speed change characteristics of preceding and subsequent vehicle targets under adjacent time indices, deceleration continuity is extracted. Combined with the directional change characteristics and lane change characteristics between adjacent vehicle targets, avoidance continuity is extracted. Based on the position change characteristics and directional change characteristics between adjacent vehicle targets, trajectory offset continuity is extracted. Based on the vehicle distance and speed change characteristics of multiple vehicle targets under continuous time index, the traffic flow compression succession relationship is extracted. Based on the stagnation change characteristics of preceding and subsequent vehicle targets under continuous time index, the stagnation transmission relationship is extracted. Read the road position coordinates of each vehicle target in the standardized road scene dataset under each time index, and map the deceleration relationship, avoidance relationship, trajectory offset relationship, traffic flow compression relationship and stagnation relationship to the corresponding road position; Reverse convergence localization is performed on each mapped connection along the continuous time index to determine the common starting road location, generate the source disturbance location, and generate the source propagation direction according to the expansion order of the road locations corresponding to each connection.

5. The intelligent lifting and rotating sign warning method based on visual recognition according to claim 1, characterized in that, The generation of the risk propagation prediction field specifically includes: Starting from the location of the source disturbance, and combining the direction of source propagation, the road locations distributed along the direction of source propagation within the target road area are sequentially arranged and divided into multiple spatial location units; Read the position change features, speed change features, direction change features, distance change features, lane change features and stationary change features corresponding to each vehicle target in the road disturbance feature set, and map each vehicle target to each spatial location unit according to the time index and road location to generate a spatial location disturbance feature distribution sequence. Based on the spatial location disturbance feature distribution sequence, the order of appearance of various road disturbance features between the previous spatial location unit and the next spatial location unit is compared item by item to extract the continuation results of position change features, speed change features, direction change features, distance change features, lane change features, and stagnation features. Continuous statistics are performed on the results of the continuity of various road disturbance features extracted between spatial location units. Adjacent spatial location unit combinations that satisfy the continuous continuity of two or more types of road disturbance features are selected, and these adjacent spatial location unit combinations are identified as risk propagation associations. The various risk propagation relationships are connected in chronological order according to their spatial location to generate a risk propagation path that extends outward from the source disturbance location along the source propagation direction; The risk propagation sequence of each spatial location unit in the risk propagation path is determined based on the connection order of each spatial location unit in the target road area, and the risk propagation coverage is determined based on the range of spatial location units reached by the risk propagation path. The risk propagation sequence and risk propagation coverage are then organized to generate a risk propagation prediction field.

6. The intelligent lifting and rotating sign warning method based on visual recognition according to claim 1, characterized in that, The generation of the driving blind spot distribution results specifically includes: Read vehicle target information, lane structure information, road boundary information, occlusion information and environmental status information under each time index in the standardized road scene dataset, filter vehicle targets that enter the warning impact range along the travel direction of the target road area, and generate a set of oncoming vehicle targets; Based on the road position and movement direction of each incoming vehicle target in the set of incoming vehicle targets, and combined with the lane passage area direction in the lane structure information, line of sight extension matching is performed on the road position corresponding to each incoming vehicle target in front of the target road area to generate a set of incoming vehicle line of sight projection paths. Based on road boundary information, the effective projection range in the set of projection paths of each oncoming vehicle is bounded, and the occlusion position on the projection path of each oncoming vehicle is checked in combination with the occlusion information, so as to generate a set of road vision occlusion results corresponding to each oncoming vehicle target. Based on the lighting conditions, weather conditions, visibility conditions, and road surface appearance conditions in the environmental status information, visible distance correction and occlusion boundary correction are performed on the road visual occlusion result set corresponding to each incoming vehicle target to generate the invisible road location set corresponding to each incoming vehicle target. The set of invisible road locations corresponding to each incoming vehicle target is summarized in order of road location to generate the driving blind spot distribution result corresponding to the target road area.

7. The intelligent lifting and rotating sign warning method based on visual recognition according to claim 1, characterized in that, The generation of the candidate action set for the signboard specifically includes: Read the standardized road scene dataset, risk propagation prediction field, and driving blind spot distribution results, and establish a unified correspondence based on time index and road location; Based on vehicle target information, lane structure information, road boundary information, and occlusion information, scene reconstruction is performed on the vehicle distribution location, lane passage area, road edge location, and occlusion location within the target road area. Combined with environmental state information, environmental state mapping is performed to generate a digital twin road image. The sequence of risk propagation, the coverage of risk propagation, and the distribution of blind spots in the driving field are mapped to the corresponding road locations in the digital twin road mirror, forming a digital twin road mirror that includes the risk propagation status and the driving blind spot status. Based on the coverage of risk propagation, the warning area of ​​the sign is determined in the digital twin road mirror; based on the order of risk propagation, the triggering order of the sign action is determined; based on the distribution of blind spots in the driver's field of vision, the direction of the warning action of the sign and the range of the sign's raising and lowering action are determined. Based on the warning area of ​​the sign, the triggering sequence of the sign action, the warning direction of the sign, and the lifting range of the sign, different lifting heights, different rotation directions, different action timings, different warning release states, and different warning pointing states are combined to generate a set of candidate sign actions.

8. A vision-based intelligent lifting and rotating sign warning system, comprising the vision-based intelligent lifting and rotating sign warning method according to any one of claims 1 to 7, characterized in that, include: The road scene acquisition module is used to acquire continuous road scene image data of the target road area, perform preprocessing, and generate a standardized road scene dataset. The road disturbance feature extraction module is used to extract road disturbance features based on a standardized road scene dataset and generate a road disturbance feature set. The source disturbance inversion module is used to extract deceleration inheritance relationships, avoidance inheritance relationships, trajectory deviation inheritance relationships, traffic flow compression inheritance relationships and stagnation transmission relationships based on the road disturbance feature set, and invert to generate the source disturbance location and source propagation direction by combining the standardized road scene dataset; The risk propagation prediction module is used to construct the risk propagation correlation between various spatial locations within the target road area based on the road disturbance feature set, the source disturbance location, and the source propagation direction, and generate a risk propagation prediction field. The driving vision blind spot analysis module is used to project oncoming vehicle lines of sight and road vision occlusion based on a standardized road scene dataset, and generate driving vision blind spot distribution results. The digital twin mirror and candidate action generation module is used to construct a digital twin road mirror of the target road area based on a standardized road scene dataset, risk propagation prediction field, and blind spot distribution results in the driver's field of vision, and to generate a set of candidate actions for signs. The pre-simulation evaluation module is used to map the set of candidate actions for signs to a digital twin road mirror for pre-simulation, extract the risk reduction results, blind spot reduction results and traffic disturbance results corresponding to each candidate action for signs, and generate early warning action evaluation results; The target action generation module is used to generate target early warning action results based on the early warning action evaluation results; The sign control module is used to control the intelligent lifting and rotating sign to perform lifting and rotating actions based on the target warning action results, and to complete the early directional warning deployment.

Images