Lift column interception system oriented to abnormal rush vehicle motion trajectory estimation

Abstract

This invention relates to the field of road safety management technology and discloses a rising bollard interception system for estimating the trajectory of abnormally rushing vehicles. The system includes roadside sensing equipment, control equipment, and rising bollards arranged in a discrete cylindrical array. The control equipment transforms the spatial arrangement of the rising bollards into a dynamic visible area; based on the acquired vehicle visible contour data, it extracts the gap exposure time-series slip, characterizing the continuous temporal shift of the time centroid; it uses this slip for particle swarm initialization and as an optimization parameter to participate in multi-objective adaptive optimization evolution, reconstructing the optimal occlusion trajectory; it incorporates the vehicle's lateral projection width into the calculation to obtain the earliest trigger time and sends a lifting command through an irreversible triggering mechanism; finally, it uses the actual lifting response to update the interception sensing model. This invention overcomes trajectory breakage caused by array occlusion and optical disturbances, achieving objective deduction of concealed steering and deterministic linkage interception.

Description

Technical Field

[0001] This invention relates to the field of road safety management technology, and more specifically, to a bollard interception system for estimating the trajectory of abnormally rushing vehicles. Background Technology

[0002] Anti-collision bollards are typically deployed in the form of discrete cylindrical arrays, capable of rapidly rising from a level position on the road surface. In practical applications, bollards inevitably incorporate highly reflective components such as stainless steel sleeves, reflective strips, and illuminated rings, and are often accompanied by safety detection accessories. When an intruding vehicle approaches and attempts to evade interception, the visible vehicle outline acquired by a fixed-view lateral sensing device is fragmented into discontinuous visible segments due to the rapid rise of the discrete array. Existing conventional tracking schemes typically rely on continuous, smooth observation data. Faced with time-varying occlusion scenarios modulated by discrete arrangement, rapid rise, and strong reflection disturbances, they are highly susceptible to misinterpreting changes in the vehicle outline as target loss or normal noise. When the driver of the intruding vehicle attempts a stealthy turn into a seemingly unclosed pseudo-passage in the final moments, traditional methods cannot accurately deduce this yaw transition from fragmented visible slices, leading to delays, inconsistencies, or incorrect target blocking areas when issuing interception commands. Summary of the Invention

[0003] This invention provides a rising bollard interception system for estimating the trajectory of abnormally rushing vehicles, which solves the technical problems mentioned in the background art.

[0004] This invention provides a rising bollard interception system for estimating the trajectory of abnormally rushing vehicles, comprising a roadside sensing device, a control device, and rising bollards arranged in a discrete cylindrical array. The control device is communicatively connected to the roadside sensing device and the rising bollards, and is used to perform the following: Acquire interception scene data and convert the spatial arrangement parameters and lifting status of the rising column into multiple dynamic visual ranges; Based on the vehicle visual contour data acquired by the roadside sensing device, the gap exposure time-series slip amount is extracted. The gap exposure time-series slip amount represents the continuous temporal shift of the time centroid of the vehicle visual contour data when the vehicle passes through each of the dynamic visual intervals. Particle swarm initialization is performed by revealing the timing slip amount using the aforementioned gap; The particle swarm evolution and occlusion trajectory reconstruction are performed under the constraints of the spatial arrangement parameters, wherein the gap exposure time slip is used as an optimization parameter to participate in the multi-objective adaptive optimization of the particle swarm evolution. The high-risk intrusion gap and arrival time are calculated based on the optimal occlusion trajectory obtained from the reconstruction. The lateral projection width of the vehicle is included in the calculation to obtain the earliest trigger time of each of the rising bollards. Based on the earliest triggering time, deterministic group column linkage execution and command locking are performed, and an irreversible lifting command is sent to the lifting column through an irreversible triggering mechanism; After the interception event ends, the extracted gap reveals the timing slip and the actual takeoff response to update the interception perception model.

[0005] The beneficial effects of this invention are as follows: Addressing the challenges of time-varying occlusion and perception chain disruption caused by the discrete configuration of the rising column array and high-reflectivity components, this invention extracts the temporal slip of the gap exposure and directly uses the center-of-gravity shift of the visible data as the vehicle passes through the dynamic visible range as an optimization parameter for particle swarm evolution. This objectively reconstructs the final turning trajectory of an abnormally intruding vehicle searching for a hidden gap without requiring conventional continuous tracking. Simultaneously, by incorporating the vehicle's lateral projection width into the intrusion risk assessment and employing a monotonic, pre-locked, irreversible triggering mechanism, it eliminates repeated oscillations at the control end caused by trajectory prediction jitter, ensuring the timeliness, determinism, and reliability of interception actions in complex intrusion scenarios. Attached Figure Description

[0006] Figure 1 This is a block diagram of the rising bollard interception system for estimating the trajectory of abnormally rushing vehicles according to the present invention. Detailed Implementation

[0007] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0008] like Figure 1 As shown, a bollard interception system for estimating the trajectory of abnormally rushing vehicles includes roadside sensing devices, control devices, and bollards arranged in a discrete cylindrical array. The control devices are communicatively connected to the roadside sensing devices and the bollards, and are used to perform the following: Acquire interception scene data and convert the spatial arrangement parameters and lifting status of the rising column into multiple dynamic visual ranges; Based on the vehicle visual contour data acquired by the roadside sensing device, the gap exposure time-series slip amount is extracted. The gap exposure time-series slip amount represents the continuous temporal shift of the time centroid of the vehicle visual contour data when the vehicle passes through each of the dynamic visual intervals. Particle swarm initialization is performed by revealing the timing slip amount using the aforementioned gap; The particle swarm evolution and occlusion trajectory reconstruction are performed under the constraints of the spatial arrangement parameters, wherein the gap exposure time slip is used as an optimization parameter to participate in the multi-objective adaptive optimization of the particle swarm evolution. The high-risk intrusion gap and arrival time are calculated based on the optimal occlusion trajectory obtained from the reconstruction. The lateral projection width of the vehicle is included in the calculation to obtain the earliest trigger time of each of the rising bollards. Based on the earliest triggering time, deterministic group column linkage execution and command locking are performed, and an irreversible lifting command is sent to the lifting column through an irreversible triggering mechanism; After the interception event ends, the extracted gap reveals the timing slip and the actual takeoff response to update the interception perception model.

[0009] Preferably, the interception scene data is acquired, and the spatial arrangement parameters and lifting status of the rising columns are converted into multiple dynamic visual ranges, including: The adjacent lifting columns are divided into multiple impact gaps. The normalized center position and normalized diameter of each lifting column are obtained. The normalized left boundary, normalized right boundary, and normalized width of the impact gap are calculated using the following formulas:

[0010]

[0011]

[0012] Based on the normalized lifting progress of adjacent lifting columns, the average lifting progress of the impact gap is calculated using the following formula:

[0013] Based on the normalized position of the vehicle's lateral trajectory at the normalized event time, the normalized left boundary of the impact gap, and the normalized width of the impact gap, the internal phase position of the vehicle gap is calculated using the following formula:

[0014] Combining the phase position within the vehicle gap with the average lifting progress of the impact gap, the dynamic visible range intensity is constructed using the following formula to form the dynamic visible range:

[0015] in, Indicates the first The normalized center position of each rising column. Indicates the first The normalized diameter of each rising column, Indicates the first The normalized left boundary of the rushing gap, Indicates the first The normalized right boundary of the rushing gap. Indicates the first The normalized width of the impact gap. Indicates the first The normalized center position of each adjacent rising column, Indicates the first The normalized diameter of each adjacent rising column, Indicates the normalized event time. Indicates the first The average lifting progress during each impact gap. Indicates the first Normalized lifting progress of each rising bollard Indicates the first Normalized lifting progress of adjacent lifting columns Indicates the phase position within the vehicle gap. This represents the normalized position of the vehicle's lateral trajectory at the normalized event time. Indicates the intensity of the dynamic visible region.

[0016] Interception scenario data is a dataset used to describe road geometry, bollard deployment, real-time bollard status, vehicle observation results, and a unified time reference during an abnormal intrusion event. It can be collected jointly using industrial cameras, millimeter-wave radar, lidar, bollard controller status readback interfaces, displacement sensors, and clock synchronization modules.

[0017] Spatial layout parameters are a set of deployment parameters used to describe the lateral coverage of the bollard array, the center position of each bollard, the diameter of the bollard, the number of bollards, the spacing between adjacent bollard centers, and the edge allowance. Under the condition that the effective lateral blocking width of the road is normalized to 1, the normalized diameter of a single bollard is preferably 0.04 to 0.10, the normalized spacing between adjacent bollard centers is preferably 0.10 to 0.25, and the edge allowance is preferably 0.02 to 0.05. This creates a calculable clearance while preventing excessively wide gaps that could allow small vehicles to pass through.

[0018] The lifting status is a quantity used to indicate whether each lifting column is currently descending, rising, or fully raised. It can be acquired through hydraulic cylinder stroke sensors, motor encoders, Hall position sensors, or controller feedback.

[0019] The clearance gap is the lateral gap between two adjacent bollards that may allow vehicles to pass or pass through.

[0020] The normalized center position of the i-th lifting column is the dimensionless position value of the center of the i-th lifting column in a unified horizontal coordinate system.

[0021] The normalized diameter of the i-th lifting column is the dimensionless diameter value of the outer diameter of the i-th lifting column under a uniform lateral scale.

[0022] The normalized left boundary of the g-th impact gap is the normalized coordinate of the effective opening boundary on the left side of the g-th impact gap.

[0023] The normalized right boundary of the g-th impact gap is the normalized coordinate of the effective opening boundary on the right side of the g-th impact gap.

[0024] The normalized width of the g-th impact gap is the effective width of the g-th impact gap in the normalized lateral coordinate system. The larger the value, the stronger the apparent traffic guidance of the gap for the vehicle.

[0025] The normalized center position of the (i+1)th adjacent lifting column is the dimensionless position value of the center of the next lifting column adjacent to the i-th lifting column in a unified horizontal coordinate system.

[0026] The normalized diameter of the (i+1)th adjacent lifting column is the dimensionless diameter value of the outer diameter of the next lifting column adjacent to the i-th lifting column under a uniform horizontal scale.

[0027] Normalized event time is a time variable that maps the start and end times of an intercepted event to the interval between 0 and 1.

[0028] The average lifting progress of the g-th impact gap is the average of the lifting progress of the adjacent lifting columns on both sides of the g-th impact gap, which is used to characterize the overall degree to which the gap is blocked at the current moment.

[0029] The normalized lifting progress of the i-th bollard is the normalized ratio of the current actual lifting height of the i-th bollard to the rated maximum lifting height. The value is generally between 0 and 1. The larger the value, the closer the bollard is to being fully raised.

[0030] The normalized lifting progress of the (i+1)th adjacent lifting column is the normalized ratio of the current lifting degree of the next lifting column adjacent to the i-th lifting column.

[0031] The normalized position of the vehicle's lateral trajectory at normalized event time is the estimated lateral position of the vehicle at the current event time.

[0032] The phase position inside the vehicle gap is a dimensionless phase quantity representing the relative position of the vehicle from left to right within the g-th gap. A value closer to 0 indicates that the vehicle is closer to the left side of the gap, and a value closer to 1 indicates that the vehicle is closer to the right side of the gap.

[0033] The dynamic visibility interval intensity is a visual exposure intensity obtained by combining the phase position inside the vehicle gap and the average lifting progress of the gap. The larger the value, the more likely the vehicle is to be observed by the roadside sensing device when passing through the gap.

[0034] The dynamic visible interval is a set of observable time intervals formed by the change in the intensity of the dynamic visible interval over time.

[0035] In specific implementation, regarding the specific composition, coordinate system definition, time synchronization method, and update cycle of the interception scene data, the interception scene data should at least include the coordinates of the left boundary of the array, the coordinates of the right boundary of the array, the number of each rising bollard, the center coordinates of each rising bollard, the outer diameter of each rising bollard, the real-time lifting height of each rising bollard, the longitudinal and lateral positions of the vehicle in each frame, the timestamp of the roadside sensing device, the timestamp of the controller, and the event start and end markers; the lateral coordinates increase to the right from the left side of the vehicle's direction of travel, and the longitudinal coordinates are positive from the direction in which the vehicle approaches the rising bollard; the industrial camera, LiDAR, and rising bollard controller should all use the same millisecond-level clock or be aligned to the same time base via a network time synchronization module; the update cycle is preferably 20 to 40 milliseconds, with the controller refreshing once every 25 milliseconds, and each refresh simultaneously writing the vehicle observation and the lifting status of each bollard in the current frame, thereby ensuring that the subsequent normalized event time and the vehicle's visible contour data can correspond one-to-one.

[0036] In specific implementation, regarding the specific content, normalization benchmark, and calibration method of spatial layout parameters, the spatial layout parameters should at least include the total number of lifting bollards, the lateral coordinates of the center of each lifting bollard, the actual outer diameter of each lifting bollard, the left boundary of the array, the right boundary of the array, the width of the impassable area at the edge, and the maximum rated lifting height. The normalized lateral coordinates should use the effective blocking width of the array as the denominator, and the normalized lifting progress should use the rated maximum lifting height as the denominator. During calibration, after construction is completed, the center position and outer diameter of each lifting bollard should be measured using a total station, measuring tape, or laser rangefinder, and then the control equipment should save it as the array configuration.

[0037] In specific implementation, regarding the numbering rules for the rush gaps and the handling method for the boundary gaps, first sort the normalized center positions of each bollard from smallest to largest, then define the gap between the i-th bollard and the (i+1)-th adjacent bollard as the g-th rush gap, where g and i maintain a sequential correspondence; for scenarios where there are curbs, guardrails, or walls at both ends of the road and vehicles cannot pass, no boundary rush gaps are set; for scenarios where there are passable shoulder areas at both ends, virtual boundary posts can be established at the edges before dividing the boundary rush gaps.

[0038] In specific implementation, for the abnormal handling of the normalized left boundary, normalized right boundary, and normalized width of the impact gap when the gap is zero or negative, when the calculated normalized width of the impact gap is less than or equal to 0, the impact gap is judged as a closed gap and is not included in the dynamic visible interval generation and subsequent risk assessment; when the normalized width of the impact gap is greater than 0 but less than the minimum traversable threshold, it can be retained for occlusion calculation, but a low passage penalty is added when screening high-risk impact gaps; the minimum traversable threshold can be preferably set to 0.02 to 0.05. When the outer edges of the two columns have already contacted and the calculated normalized width of the impact gap is negative, the dynamic visible interval intensity of the gap should be directly set to 0.

[0039] In practice, for the definition of the start and end of the normalized event time, the start time of the event is preferably the moment when the vehicle first enters the monitoring area in front of the array and is stably identified, and the end time of the event is preferably the earliest moment when the vehicle has passed through the array section, completely left the monitoring area, been successfully stopped, or had no effective observation for more than 1 second. The normalized event time is obtained by subtracting the start time of the event from the actual time and then dividing by the total duration of the event, and the result is cropped to 0 to 1.

[0040] In specific implementation, regarding the acquisition method of the normalized position of the vehicle's lateral trajectory under occlusion under normalized event time, when direct observation is possible, the lateral coordinates of the vehicle center output by industrial cameras or LiDAR are given priority; when the vehicle is partially occluded by rising pillars, short-term extrapolation is performed using the two most recent effective observation points, combined with millimeter-wave radar lateral velocity or visual optical flow correction; if continuous occlusion exceeds the set upper limit, the last reliable lateral velocity is retained and the extrapolation result is clipped at the boundary.

[0041] In specific implementation, regarding the threshold, clipping, and continuity determination rules for the dynamic visible interval intensity, the phase position inside the vehicle gap is first limited to the range of 0 to 1 within the gap. When the vehicle is outside the gap, the dynamic visible interval intensity is directly set to 0. Then, the dynamic visible interval intensity at each moment is compared with the threshold. When the intensity is greater than or equal to 0.15 for two consecutive frames, it is recorded as entering the dynamic visible interval. When the intensity is less than 0.10 for two consecutive frames, it is recorded as leaving the dynamic visible interval. If the break duration between two adjacent segments does not exceed two frames, they are merged into the same dynamic visible interval.

[0042] Preferably, based on the vehicle visual contour data acquired by the roadside sensing device, a gap exposure time-series slip is extracted. This gap exposure time-series slip characterizes the continuous temporal shift of the temporal centroid of the vehicle visual contour data as the vehicle passes through each of the dynamic visual intervals, including: The currently acquired vehicle visual contour data is used as the original vehicle visual contour data for extreme value mapping, and the normalized vehicle visual contour data is obtained through the following formula:

[0043] Extract multiple time segments of the visible intervals segmented by the dynamic visible intervals, and calculate the local normalized time within the visible segment using the following formula, based on the normalized event time, the left boundary of the visible interval time segment, and the right boundary of the visible interval time segment:

[0044] Combining the local normalized time within the visible segment with the normalized vehicle visible contour data, the time centroid of the visible contour data within each visible time segment is calculated using the following formula:

[0045] Based on the duration proportion of each visible time segment, the normalized weight of the visible segment duration is calculated using the following formula:

[0046] Based on the difference in temporal centroids of the visible contour data within adjacent visible segments, and combined with the normalized weights of the visible segment durations, a segment-by-segment cumulative offset calculation is performed. The temporal slip of the gap exposure is then extracted using the following formula:

[0047] in, This represents normalized vehicle visible outline data. This represents the original visible vehicle outline data. This represents the minimum visible vehicle outline data. This represents the maximum visible vehicle outline data. This represents the normalized event time. This represents the local normalized time within a visible segment. Indicates the left boundary of the visible time segment. Indicates the right boundary of the visible time segment. Indicates the first The temporal centroid of the visible outline data within each visible time segment corresponds to a visible time segment. Indicates the first A visible time segment Indicates the normalized weight of the duration of the visible segment. This indicates the total number of time segments within the visible interval. Indicates the first The left boundary of a visible time segment Indicates the first The right boundary of a visible time segment Indicates the amount of slippage when the gap is exposed. Indicates the first The temporal centroid of the visible contour data within the visible segment corresponding to each adjacent visible time segment of the visible interval.

[0048] Vehicle visible contour data refers to the visible portion of a vehicle's shape that roadside sensing equipment can still observe even when the vehicle is obstructed by bollards or suffers from reflection interference. This data can be acquired through side-mounted industrial cameras, binocular cameras, infrared cameras, or LiDAR point cloud projection results.

[0049] The raw vehicle visual contour data is an unnormalized observation of the vehicle visual contour. It can be obtained by directly measuring the area of ​​the visible region of the vehicle, the contour length, the bounding box width, or the lateral span of the point cloud in each frame of the image.

[0050] Normalized vehicle visual profile data is a dimensionless observation obtained by performing extreme value mapping on the original vehicle visual profile data. It is used to eliminate dimensional differences caused by different vehicle scales and different sensor gains.

[0051] The minimum vehicle visible profile data is the minimum value of the vehicle visible profile data within the current event's valid observation window, used to determine the lower limit of the extreme value mapping.

[0052] The maximum vehicle visible profile data is the maximum value of the vehicle visible profile data within the current event's valid observation window, used to determine the upper limit of the extreme value mapping.

[0053] A visible interval time segment is a continuous observable segment obtained by dividing a dynamic visible interval on the time axis. It is used to break down intermittently visible vehicle outline data into multiple local analysis units.

[0054] The left boundary of the visible interval time segment is the normalized time value at the start of the g-th visible interval time segment, which is used to define the starting point of the local time coordinate.

[0055] The right boundary of the visible interval time segment is the normalized time value at the end of the g-th visible interval time segment, which is used to define the endpoint of the local time coordinate.

[0056] The local normalized time within a visible segment is the amount of local time after remapping the time within the g-th visible interval time segment to the interval between 0 and 1. It is used to eliminate the impact of inconsistent durations of different visible segments.

[0057] The g-th visible time segment is the set of time samples corresponding to the g-th visible time segment, used to limit the range of samples participating in the statistics when calculating the time centroid.

[0058] The temporal centroid of the visible contour data within the visible segment corresponding to the g-th visible interval time segment is the centroid position of the vehicle's visible contour data relative to the local time axis within the g-th visible interval time segment, used to characterize whether the vehicle appears earlier or later in this segment.

[0059] The normalized weight of the visible segment duration is a weight value determined according to the proportion of the duration of each visible interval time segment, which is used to increase the contribution of longer-duration segments in the calculation of temporal slip.

[0060] The total number of visible interval time segments is the total number of visible interval time segments obtained by dynamic visual interval segmentation in the current event, used to limit the overall scale of the segment-by-segment cumulative calculation.

[0061] The left boundary of the u-th visible time segment is the start time of the visible time segment identified by the sequence number u, which is used to construct the cumulative denominator of the normalized weights.

[0062] The right boundary of the u-th visible time segment is the end time of the visible time segment identified by the sequence number u, which is used to calculate the duration of the u-th visible time segment.

[0063] The gap exposure time slip is a normalized time-series characteristic that represents the continuous shift trend of the exposed center of gravity when a vehicle passes through multiple dynamic visible sections. The closer the value is to 0, the more concentrated the exposed center of gravity is in the forward segment, and the closer it is to 1, the more concentrated the exposed center of gravity is in the backward segment.

[0064] In specific implementation, the vehicle visual contour data can be any one or a combination of the following, depending on the specific data format, sampling frequency, and preprocessing method: visible area, minimum bounding rectangle width, number of contour points, edge length, or horizontal span of point cloud projection. The sampling frequency is preferably consistent with the frame rate of the main sensor, preferably 25 to 50 frames per second. During preprocessing, background subtraction, reflection highlight suppression, moving target segmentation, and abnormal burr removal are performed first, and then the visible contour data consistent with the vehicle body is extracted.

[0065] In practice, the time window for extreme value mapping, invalid sample removal rules, and processing when the maximum vehicle visible contour data equals amin are addressed. The minimum and maximum vehicle visible contour data used for extreme value mapping should be statistically analyzed on all valid observation samples of the current event, rather than independently within a single segment. Invalid samples include those with completely missing frames, reflection saturation, contour area increases by more than 3 times compared to the previous frame, or detection confidence is below the threshold. When the maximum vehicle visible contour data equals the minimum vehicle visible contour data, it indicates insufficient change in visible contours within the current event. In this case, the normalized vehicle visible contour data should be uniformly set to 0.5 or the normalization result of the previous event template should be used.

[0066] In practice, the threshold for dividing the dynamic visible interval into visible interval time segments, the shortest segment duration, and the rules for merging adjacent segments are defined. When the intensity of the dynamic visible interval is continuously greater than or equal to 0.15, a visible interval time segment is started to be recorded, and when the intensity of the dynamic visible interval is continuously lower than 0.10, the segment ends. The shortest segment duration is preferably no less than 2 frames, and segments with less than 2 frames are considered noise and discarded. If the break duration between two adjacent segments does not exceed 2 frames and the temporal center trends at both ends are consistent, they are merged into a longer visible interval time segment.

[0067] In specific implementation, regarding the discrete implementation method of the time centroid of the visible contour data within the visible segment and the handling when the denominator is zero, for the g-th visible interval time segment, the product of the local normalized time within the visible segment and the normalized vehicle visible contour data at each discrete moment within the segment is summed, and then divided by the sum of the normalized vehicle visible contour data within the segment; if the denominator is 0, it means that although the segment is cut out by the dynamic visible interval, there is no effective visible contour amount. At this time, the time centroid of the visible contour data within the visible segment corresponding to the g-th visible interval time segment is set to 0.5 or set to the time centroid of the previous effective segment.

[0068] In specific implementation, regarding the summation range of the denominator of the normalized weight for the duration of visible segments and the processing of single-segment scenes, the denominator of the normalized weight for the duration of visible segments should sum the durations of all valid visible time segments, rather than summing only some segments; when the total number of visible time segments is equal to 1, the normalized weight for the duration of the visible segment of the unique segment is directly set to 1, and the gap exposure time sequence slip is directly set to 0.5 as the unbiased initial value.

[0069] In practice, for the boundary clipping, denoising and smoothing rules of the time shift amount revealed by the gap, the time centroid difference between adjacent segments is accumulated segment by segment according to the normalized weight of the visible segment duration, and the final result is linearly mapped to the range of 0 to 1, and clipped to 0 to 1 again before output; if the absolute value of the time centroid difference between adjacent segments exceeds 0.8 and the duration of the corresponding segment is less than half of the average duration of all segments, the difference is regarded as occasional noise and processed with a weight of 0.5.

[0070] Preferably, particle swarm initialization is performed by revealing the timing slip using the gap, including: Based on the normalized longitudinal and lateral positions of the first historical effective observation point, and the normalized longitudinal and lateral positions of the second historical effective observation point, the geometric trajectory extrapolation coefficient and the geometric prediction normalized lateral endpoint are calculated using the following formulas:

[0071]

[0072] Based on the normalized sorting value of the impact gap and the time-series slip of the gap exposure, the time-series slip similarity bias weight is calculated using the following formula:

[0073]

[0074] Combining the temporal slip similarity bias weights with the normalized lateral position of the center of the impact gap, the normalized lateral endpoint of the slip bias is calculated using the following formula:

[0075] The temporal slip intensity is calculated based on the gap exposure time slip amount. Combining the geometrically predicted normalized lateral endpoint and the slip offset normalized lateral endpoint, the basic fusion normalized lateral endpoint is obtained by weighting using the following formula:

[0076]

[0077] Based on the total number of impact gaps and the total number of visible time segments, the initial total number of particles is calculated using the following formula:

[0078] Based on the normalized longitudinal position and normalized lateral position of the first historical effective observation point, the initial particle trajectory normalization starting point is set. Based on the basic fusion normalized lateral ending point, the initial particle trajectory normalization ending point is set. Within the range of longitudinal coordinate envelope difference and lateral coordinate envelope difference, combined with the basic curve parameter vector and low-difference random sequence value, the initial particle normalization parameter vector is generated by the following formula, and the initial particle normalization parameter vector is used as the particle trajectory control point parameter vector:

[0079]

[0080]

[0081]

[0082]

[0083] in, This indicates the normalized vertical position of the first historical valid observation point. Indicates the normalized lateral position of the first historical valid observation point. This indicates the normalized vertical position of the second historical valid observation point. This indicates the normalized lateral position of the second historical valid observation point. Represents the extrapolation coefficients of the geometric trajectory. Indicates the geometrically predicted normalized lateral endpoint. Indicates the number of the gap during the rush. This represents the normalized sorted value of the impact gap. This indicates the total number of impact gaps. This indicates the amount of time-series slippage when the gap is exposed. Indicates the first Temporally glide similarity bias weights Indicates the first The normalized lateral position of the center of the impact gap in each impact gap. Indicates the normalized lateral endpoint of the slip bias. Indicates the time-series slip intensity. Indicates the lateral endpoint of basic fusion normalization. This represents the total number of visible time segments within the visible time segment. This represents the initial total number of particles. This represents the normalized starting point of the initial particle trajectory. This represents the normalized endpoint of the initial particle trajectory. This represents the difference in the vertical coordinate envelope. This represents the difference in lateral coordinate envelope. Represents the basic curve parameter vector. Indicates the value of a low-dispersion random sequence. Indicates the first The initial particle normalization parameter vector of each particle.

[0084] The normalized longitudinal position of the first historical valid observation point is the normalized position value of the most recent valid observation point in a unified longitudinal coordinate system, which is used to determine the initial longitudinal state of the initial particle trajectory.

[0085] The normalized lateral position of the first historical valid observation point is the normalized position value of the most recent valid observation point in a unified horizontal coordinate system, which is used to determine the initial lateral state of the initial particle trajectory.

[0086] The normalized longitudinal position of the second-to-last historical valid observation point is the normalized position value of the second-to-last valid observation point in a unified longitudinal coordinate system, used to provide historical longitudinal change information required for trajectory extrapolation.

[0087] The normalized lateral position of the second-to-last historical valid observation point is the normalized position value of the second-to-last valid observation point in a unified horizontal coordinate system, used to provide historical lateral change information required for trajectory extrapolation.

[0088] The geometric trajectory extrapolation coefficient is an extrapolation scale calculated based on the longitudinal distance between two adjacent historical valid observation points. It is used to control the degree of amplification of the geometric prediction endpoint relative to the historical observation amount.

[0089] The geometrically predicted normalized lateral endpoint is a predicted value of the vehicle's lateral endpoint obtained by extrapolating the geometric trend of historical observations. It reflects the natural steering continuity direction of the vehicle without considering the amount of slip during the clearance exposure time.

[0090] The normalized sorting value of the impact gap is a dimensionless sorting value obtained by mapping the impact gap from left to right or from near to far according to a predetermined order. It is used to convert the discrete impact gap number into a continuous position quantity that can be compared with the time-series slip of the gap exposure.

[0091] The total number of rushing gaps is the total number of available rushing gaps in the current rising and falling column array, used to complete the rushing gap sorting normalization and the initial particle total number configuration.

[0092] The g-th temporal slip similarity bias weight is the weight value of the similarity between the g-th impact gap sorting position and the gap exposure temporal slip amount. The larger the value, the more likely the vehicle is to be biased towards that impact gap.

[0093] The normalized lateral position of the center of the g-th impact gap is the normalized position value of the center point of the g-th impact gap in a unified lateral coordinate system, which is used to construct the normalized lateral endpoint of the slip offset.

[0094] The slip-off bias normalized lateral endpoint is an estimate of the lateral endpoint obtained by weighting the temporal slip-off similarity bias of each passage gap, reflecting the tendency of the vehicle to deflect toward the hidden passage gap.

[0095] The temporal slip intensity is a fusion intensity coefficient calculated based on the gap exposure temporal slip amount, used to adjust the fusion ratio between the geometric prediction normalized lateral endpoint and the slip offset normalized lateral endpoint.

[0096] The basic fusion normalized lateral endpoint is a basic estimate of the vehicle's lateral endpoint obtained by integrating geometric prediction information and temporal slip information, and is used to determine the endpoint envelope center of the initial particle swarm.

[0097] The initial total number of particles is the number of particles that need to be generated during the initialization phase of the particle swarm, which determines the balance between the initial search coverage and the solution cost.

[0098] The normalized starting point of the initial particle trajectory is the initial control point of the initial particle trajectory in the normalized coordinate system, which is directly determined by the most recent valid historical observation point.

[0099] The normalized endpoint of the initial particle trajectory is the termination control point of the initial particle trajectory in the normalized coordinate system, which is determined by the basic fusion normalized lateral endpoint when the vehicle reaches the cross section of the rising column array.

[0100] The longitudinal coordinate envelope difference is the adjustable span of the initial particle trajectory from the starting point to the ending point in the longitudinal direction, used to limit the perturbation amplitude of the particle control point in the longitudinal direction.

[0101] The lateral coordinate envelope difference is the magnitude of the lateral difference between the slip bias normalized lateral endpoint and the geometric prediction normalized lateral endpoint, used to limit the perturbation magnitude of the particle control point in the lateral direction.

[0102] The base curve parameter vector is a set of reference control parameters used to generate the initial particle normalized parameter vector. Preferably, it is a 5-dimensional vector corresponding to a smooth linear cubic curve, with the first dimension being 0.33, the second dimension being 0, the third dimension being 0.67, the fourth dimension being 0, and the fifth dimension being 0. This ensures that the initial curve remains smooth when there is no significant lateral maneuver, while reserving symmetrical adjustment space for subsequent lateral disturbances.

[0103] The low-difference random sequence value u1 is a low-difference sampled value used to generate the first dimension of the perturbation of the initial particle normalization parameter vector. It takes a value between 0 and 1 and is used to improve the uniformity of the initial particle coverage in the search space.

[0104] The low-difference random sequence value u2 is a low-difference sampled value used to generate the second dimension of the perturbation in the initial particle normalization parameter vector. It takes a value between 0 and 1 and is used to stabilize the lateral offset of the first intermediate control point.

[0105] The low-difference random sequence value u3 is a low-difference sampled value used to generate the third dimension of the perturbation in the initial particle normalization parameter vector. It takes a value between 0 and 1 and is used to stabilize the longitudinal offset of the second intermediate control point.

[0106] The low-difference random sequence value u4 is a low-difference sampled value used to generate the fourth dimension of the perturbation in the initial particle normalization parameter vector. It takes a value between 0 and 1 and is used to stabilize the lateral offset of the second intermediate control point.

[0107] The low-difference random sequence value u5 is a low-difference sampled value used to generate the fifth dimension perturbation of the initial particle normalization parameter vector. It takes a value between 0 and 1 and is used to supplement the search degrees of freedom for terminal lateral changes.

[0108] The initial particle normalization parameter vector of the k-th particle is the set of dimensionless control parameters corresponding to the k-th particle at the initialization time, which is used to carry the specific shape information of the initial candidate trajectory.

[0109] The particle trajectory control point parameter vector is a set of control parameters used to uniquely determine the shape of a single particle's trajectory.

[0110] In practice, regarding the selection criteria and validity determination of the first and second historical valid observation points, all observation samples are traced back from the current moment, and the two most recent observation points that meet the following conditions in reverse time order are selected as the first and second historical valid observation points: the detection confidence level is not lower than 0.6, the observation points are not judged as reflection anomalies, the observation points do not exceed the road boundary, and the time interval between adjacent observation points does not exceed 0.2 seconds. If only one valid observation point is found, that point is used as the starting point, and the geometric trajectory extrapolation coefficient is set to 0.

[0111] In specific implementation, for the stabilization processing of the geometric trajectory extrapolation coefficient when the denominator approaches zero, during backward movement, or during jump observation, when the absolute value of the difference between the normalized longitudinal position of the first historical effective observation point and the normalized longitudinal position of the second historical effective observation point is less than 0.02, it indicates that the longitudinal change is too small. At this time, the geometric trajectory extrapolation coefficient is limited to a safe range of 0 to 3 and 0 is preferred. When backward movement or historical point jump exceeds the maximum lateral maneuverability of a normal vehicle, the average of the most recent 3 frames is used to replace the difference of a single frame.

[0112] In specific implementation, for the negative value processing, normalization processing, and multi-gap parallel processing of temporal slip similarity bias weights, the original similarity value is first obtained by subtracting the absolute difference between the normalized sorting value of the gap and the temporal slip amount of the gap exposure from 1, and then the negative values ​​are directly clipped to 0; then all the g-th temporal slip similarity bias weights are summed and normalized so that the sum of all weights equals 1; when multiple gaps have the same maximum weight, the gap that is closer to the lateral endpoint of the geometric prediction normalization is given priority to increase the weight.

[0113] In specific implementation, regarding the clipping rules for the temporal slip intensity and the base fusion normalized lateral endpoint, the temporal slip intensity, after being mapped from the gap exposure temporal slip, should be clipped to the range of 0 to 1; the base fusion normalized lateral endpoint, after being weighted by the geometric prediction normalized lateral endpoint and the slip bias normalized lateral endpoint, should also be clipped to the range of 0 to 1; when the gap exposure temporal slip is close to 0.5, the temporal slip intensity approaches 0, indicating that more reliance should be placed on geometric prediction; when the gap exposure temporal slip is close to 0 or 1, the temporal slip intensity approaches 1, indicating that more reliance should be placed on the slip bias result.

[0114] In specific implementation, regarding the applicable scope of the formula for the initial total number of particles and the minimum particle count guarantee, the initial total number of particles is preferably taken as the sum of the total number of rushing gaps and the total number of time segments in the visible interval plus 2, so as to ensure that the number of particles increases with the increase of scene complexity; at the same time, a minimum particle count guarantee is set, preferably no less than 12, and the maximum number of particles is preferably no more than 80, in order to balance search coverage and real-time performance.

[0115] In practice, regarding how to map the basic curve parameter vector to the particle trajectory control point parameter vector, after fixing the initial particle trajectory normalized starting point and the initial particle trajectory normalized ending point, the first and third dimensions of the basic curve parameter vector represent the position ratio of the first and second intermediate control points within the longitudinal envelope, respectively; the second and fourth dimensions represent the lateral offset of the corresponding intermediate control points, respectively; and the fifth dimension represents the endpoint tangential offset or the terminal lateral fine-tuning coefficient. Based on these five parameters, four control points can be recovered, thereby forming subsequent candidate occlusion trajectories.

[0116] In specific implementation, regarding the generation method, dimensional correspondence, random seed, and experiment reproducibility rules of low-difference random sequences, the Sobol sequence or Holton sequence is preferred for generating low-difference random sequences. Each particle uses 5-dimensional samples to correspond to the 5 perturbation dimensions of the basic curve parameter vector. To ensure debugging and experiment reproducibility, the control device records the sequence start index and random seed at each event initialization, and preferably skips the first 16 samples before starting to take values.

[0117] In specific implementation, regarding the setting of upper and lower bounds for the longitudinal coordinate envelope difference and the lateral coordinate envelope difference, as well as the control point over-boundary clipping rules, the longitudinal coordinate envelope difference should at least ensure a minimum forward advance, preferably not less than 0.05; the lateral coordinate envelope difference should be constrained by the effective road width and array boundary, preferably not exceeding 0.35; each control point recovered from the basic curve parameter vector and low-difference random sequence value should be clipped to the range of 0 to 1 for both longitudinal and lateral coordinates after mapping.

[0118] Preferably, the particle swarm evolution and occlusion trajectory reconstruction constrained by the spatial arrangement parameters are performed, wherein the gap exposure time-series slip is used as an optimization parameter in the multi-objective adaptive optimization of the particle swarm evolution, including: Based on the particle trajectory control point parameter vector, candidate occlusion trajectories are generated using the following formula:

[0119] Extract the phase positions within the gaps in the candidate trajectories, and construct the predicted visual contour data using the following formula:

[0120] The observation geometric fit residual of the candidate occlusion trajectory, the exposure temporal centroid fit residual extracted by combining the temporal centroid of the visible contour data inside the visible segment, the dynamic smoothness residual, and the high-risk gap penetration fit residual are calculated using the following formulas:

[0121]

[0122]

[0123]

[0124] The adaptive weights of the residual channels are obtained based on the variance of the residual channels, and the comprehensive optimization target residual is fused using the following formula:

[0125]

[0126] Combining individual evolutionary cognitive weights and global evolutionary collaborative weights, the particle's current normalized parameter vector is updated using the comprehensive optimization objective residual through the following formula to reconstruct the optimal occlusion trajectory:

[0127]

[0128] in, Indicates the first Candidate occlusion trajectories of individual particles Represents the baseline function of the trajectory curve. This represents the particle trajectory control point parameter vector. This represents the predicted visual contour data. Indicates the phase position within the gap of the candidate trajectory. This indicates the average lifting progress during the corresponding impact gap. Represents the residuals of the geometric fit of the observations. Indicates the availability weight of the observation point. The normalized longitudinal position of the candidate trajectory represents the candidate occlusion trajectory. This represents the normalized lateral position of the candidate trajectory for the candidate occlusion trajectory. Indicates the normalized longitudinal position of the effective observation points. Indicates the normalized lateral position of the effective observation point. This indicates the residuals showing the consistency of the time-series centroids. This represents the normalized weight of the duration of the visible segment. This indicates the temporal centroid of the visible contour data within the visible segment. Indicates the first The temporal centroid within the visible fragment of the candidate trajectory of each particle. Represents dynamic smoothness residuals. This represents the longitudinal second-order difference change. This represents the lateral second-order difference change. This indicates the residual of high-risk gap penetration fit. This indicates that the candidate trajectory has reached the normalization time. Indicates the first The normalized lateral position of the center of the impact gap in each impact gap. Indicates the first The normalized width of the impact gap. This represents the normalized lateral projection width of the candidate trajectory. This indicates the adaptive weights of the residual channels. This represents the variance of the residual channel differences. This represents the residual of the overall optimization objective. This represents the normalized individual residual. Indicates the rate of particle renewal and evolution. Represents evolutionary inertia weights. This indicates the current evolution speed of the particle. Represents the cognitive weight of an evolved individual. Represents the first evolutionary random number. This represents the optimal parameter vector for an individual particle. This represents the current normalized parameter vector of the particle. Represents the global collaborative weights in the evolution. Represents the second evolutionary random number. This represents the global optimal parameter vector of the particle swarm. This represents the particle update normalization parameter vector.

[0129] The candidate occlusion trajectory of the kth particle is the candidate motion trajectory of the vehicle during the occlusion phase generated by the control parameters of the kth particle.

[0130] The trajectory curve reference function is a basis function used to map the particle trajectory control point parameter vector to a continuous candidate occlusion trajectory. A cubic Bessel basis function is preferred, as it allows for the expression of smooth steering and concealed yaw in the final stage using a finite number of control points, while also offering low computational complexity and easy boundary constraints.

[0131] The predicted visual contour data is the theoretical visible exposure amount derived from the candidate occlusion trajectory of the k-th particle.

[0132] The phase position inside the candidate trajectory gap is the relative lateral position of the candidate occlusion trajectory in the corresponding impact gap.

[0133] The observation geometry fit residual is the residual amount of the degree of inconsistency between the candidate occlusion trajectory of the k-th particle and the effective observation point in terms of geometric position. The smaller the value, the closer the candidate trajectory is to the actual observation geometry.

[0134] The observation point availability weight is the weight of the contribution of the effective observation points at the current observation time to the geometric residual.

[0135] The normalized vertical position of the candidate occlusion trajectory is the normalized vertical coordinate value of the candidate occlusion trajectory of the k-th particle at a certain moment.

[0136] The normalized lateral position of the candidate occlusion trajectory is the normalized lateral coordinate value of the candidate occlusion trajectory of the k-th particle at a certain moment.

[0137] The normalized longitudinal position of the effective observation point is the normalized coordinate value of the effective observation point in a unified longitudinal coordinate system obtained from the actual observation data.

[0138] The normalized lateral position of the effective observation point is the normalized coordinate value of the effective observation point in a unified horizontal coordinate system obtained from the actual observation data.

[0139] The revealed temporal centroid alignment residual is the difference between the candidate visible segment temporal centroid of the k-th particle and the actual visible segment temporal centroid.

[0140] The temporal centroid of the candidate trajectory visual segment of the k-th particle is the temporal centroid of the predicted visual contour data of the k-th particle within the g-th visual segment.

[0141] Dynamic smoothness residual is a residual quantity used to measure whether there is excessive jitter or abrupt change in the discrete time sequence of the candidate occlusion trajectory of the k-th particle. The smaller the value, the smoother the candidate trajectory.

[0142] The longitudinal second-order difference change is the second-order change of the discrete sequence of the longitudinal position of the candidate occlusion trajectory, which is used to evaluate whether there is an unreasonable abrupt change in acceleration in the longitudinal motion.

[0143] The lateral second-order difference change is the second-order change of the discrete sequence of the lateral position of the candidate occlusion trajectory, which is used to evaluate whether there is an unreasonable sharp swing in the lateral steering.

[0144] The high-risk gap penetration fit residual is a residual used to measure the degree of geometric alignment between the k-th particle candidate trajectory and the high-risk gap. The smaller the value, the more the candidate trajectory matches the behavior pattern of the vehicle attempting to rush into the dangerous gap.

[0145] The normalized time when the candidate trajectory reaches the normalized time is the time when the candidate trajectory of the k-th particle reaches the cross section of the rising and falling column array.

[0146] The normalized lateral projection width of the candidate trajectory is the estimated lateral occupancy width of the vehicle corresponding to the arrival time of the k-th particle candidate trajectory.

[0147] The adaptive weight of the residual channel is a weight value that is adaptively allocated based on the stability and discriminative power of different residual channels. It is used to fuse multiple residual channels to form a comprehensive optimization target residual.

[0148] The residual channel difference variance is a statistical variance measure used to characterize the ability of a residual channel to distinguish between different particles. The larger the value, the better the residual channel can distinguish between superior and inferior particles.

[0149] The comprehensive optimization objective residual is the overall objective function value of the k-th particle obtained by fusing multiple residual channels, and it is the direct basis for particle swarm update and global optimal selection.

[0150] The normalized individual residual is the normalized residual value of the k-th particle in the r-th type residual channel, which is used to eliminate the weight imbalance caused by the inconsistency of the dimensions of different residual channels.

[0151] The particle update evolution velocity is the velocity vector of the k-th particle in the next round of evolution, used to control the update magnitude and direction of the particle's current normalized parameter vector.

[0152] Evolutionary inertia weight is a weighting parameter that controls the degree to which a particle's velocity from the previous round is inherited by its current velocity. The preferred value is between 0.4 and 0.9, with an initial value preferably of 0.7. A higher value is beneficial for expanding the search range in the early stages, while a lower value is beneficial for stable convergence in the later stages.

[0153] The current evolution velocity of a particle is the velocity vector of the k-th particle in the current round of evolution, which is used together with the individual experience term and the group cooperation term to generate the velocity of the next round.

[0154] The evolutionary individual cognitive weight is a cognitive coefficient that controls the degree to which a particle reverts to its historical best position. The optimal value is between 1.5 and 2.5, with an initial value preferably set at 2.0, thus balancing individual exploration ability and convergence stability.

[0155] The first evolutionary random number is a random modulation coefficient that acts on individual cognitive terms, with a value between 0 and 1, used to prevent all particles from updating along the exact same trajectory.

[0156] The optimal parameter vector for an individual particle is the best historical parameter vector obtained so far for the k-th particle, used to preserve the successful search experience of a single particle.

[0157] The current normalized parameter vector of a particle is the parameter state of the k-th particle in the current iteration round, which determines the specific shape of the current candidate occlusion trajectory.

[0158] The evolutionary global cooperative weight is a cooperative coefficient that controls the degree to which particles move toward the global optimal position of the particle swarm. The preferred value is 1.5 to 2.5, with an initial value preferably of 2.0, so as to improve convergence efficiency by utilizing the optimal information of the swarm, while avoiding premature convergence caused by excessive swarm pull.

[0159] The second evolutionary random number is a random modulation coefficient that acts on the global cooperative term. Its value is between 0 and 1 and is used to maintain the random diversity of the particle swarm update process.

[0160] The global optimal parameter vector of the particle swarm is the optimal parameter vector that achieves the minimum comprehensive optimization objective residual among all particles up to the current iteration round. It is used to guide the entire particle swarm to converge toward the vicinity of the current optimal solution.

[0161] The particle update normalized parameter vector is the parameter state of the k-th particle after the next iteration. After boundary clipping, it is used to generate new candidate occlusion trajectories.

[0162] In specific implementation, regarding the particle trajectory control point parameter vector and the control point reconstruction method and curve type of the candidate occlusion trajectory, the initial particle trajectory normalized starting point is taken as the first control point, and the initial particle trajectory normalized ending point is taken as the fourth control point. Then, the second and third control points are restored based on the current normalized parameter vector of the particle. Finally, continuous candidate occlusion trajectories are generated through the trajectory curve reference function. The trajectory curve reference function preferably adopts the cubic Bézier curve basis function because this curve can ensure that the endpoints are fixed and can express the final steering of the vehicle through a finite number of control points.

[0163] In specific implementation, the process for obtaining the phase position inside the candidate trajectory gap and the predicted visual contour data involves sampling each candidate particle across all discrete normalized event times to obtain the normalized longitudinal position and normalized lateral position of the candidate trajectory at each moment. Then, based on the gap into which the normalized lateral position of the candidate trajectory falls, the corresponding normalized left boundary and normalized width of the gap are found to obtain the phase position inside the candidate trajectory gap. If the normalized lateral position of the candidate trajectory does not fall within any gap, the predicted visual contour data at that moment is directly set to 0.

[0164] In specific implementation, regarding the sampling time matching, observation point availability weight source, and missing sample processing for the observation geometric fit residual, the discrete sampling time of the candidate occluded trajectory should be aligned with the actual observation time, preferably using the sensing device timestamp directly; the observation point availability weight can be jointly determined by the detection confidence, the integrity of the visible contour, and whether the sample is contaminated by strong reflection, preferably between 0 and 1; when a certain observation time is missing, the corresponding observation point availability weight is directly set to 0 and does not participate in the calculation of the observation geometric fit residual.

[0165] In specific implementation, regarding the calculation method of the temporal centroid within the candidate trajectory visual segment of the k-th particle in the revealed temporal centroid matching residual, for each g-th visual interval time segment, within a time window that is exactly the same as the actual segment, the predicted visual contour data of the k-th particle is extracted, and the sum of the product of the locally normalized time and the predicted visual contour data is divided by the sum of the predicted visual contour data; if the sum of the predicted visual contour data in this segment is 0, then the temporal centroid within the candidate trajectory visual segment of the k-th particle is set to 0.5.

[0166] In specific implementation, for the discrete sampling interval, second-order difference boundary point processing, and mean calculation window of the dynamic smoothness residual, the candidate occlusion trajectory should be sampled on a unified discrete normalized event time grid, preferably at least 50 discrete time points for each event; the longitudinal second-order difference change and the lateral second-order difference change are only calculated at the internal sampling points, and the first and last two boundary points do not participate in the second-order difference; the dynamic smoothness residual adopts the average value of the absolute value of the second-order difference of all valid internal points.

[0167] In specific implementation, for the stabilization of the high-risk gap penetration matching residual under the scenarios of multi-gap competition and excessively small denominators, for each candidate particle, at the time when the candidate trajectory reaches the normalization point, the geometric alignment degree between it and the center of all impact gaps is calculated, and the maximum matching value is taken to form the high-risk gap penetration matching residual; when the sum of the normalized width of the impact gap and the normalized lateral projection width of the candidate trajectory is too small, a stabilization term of not less than 0.001 should be added to the denominator; if multiple impact gaps obtain approximately the same maximum matching value, the impact gap that is consistent with the current lateral movement direction of the vehicle is selected first.

[0168] In practice, regarding the statistical caliber of residual channel difference variance, residual channel adaptive weights, and normalized individual residuals, in each iteration, the observed geometric fit residual, the revealed temporal centroid fit residual, the dynamic smoothness residual, and the high-risk gap penetration fit residual are first normalized from minimum to maximum value on all particles to obtain normalized individual residuals; then, the variance of each type of normalized individual residual is calculated on all particles to obtain the residual channel difference variance; subsequently, the residual channel difference variances of each type are normalized by summation to obtain the residual channel adaptive weights; if all variances are very small, then uniform weights are used instead.

[0169] In specific implementation, regarding the individual particle optimal parameter vector, the global optimal parameter vector of the particle swarm, the stopping criterion, and the iteration termination condition, when the comprehensive optimization target residual of the k-th particle in the current round is less than its historical minimum value, the individual particle optimal parameter vector is updated with the particle's current normalized parameter vector; the global optimal parameter vector of the particle swarm is taken as the one with the smallest comprehensive optimization target residual among all individual particle optimal parameter vectors; the stopping criterion is preferably reaching the maximum number of iterations of 30 to 80, or the global optimal improvement amount is less than 0.001 for 5 consecutive rounds.

[0170] In practice, regarding the dynamic adjustment strategy for evolutionary inertia weight, individual evolutionary cognition weight, and global evolutionary coordination weight, the evolutionary inertia weight is preferably linearly decreased from 0.9 to 0.4 to achieve expanded search in the early stage and stable convergence in the later stage; the individual evolutionary cognition weight is preferably slowly decreased from 2.2 to 1.6, and the global evolutionary coordination weight is preferably increased from 1.6 to 2.2, so as to emphasize individual exploration first and then gradually emphasize group convergence.

[0171] Preferably, the high-risk intrusion gap and arrival time are calculated based on the optimal occlusion trajectory obtained from the reconstruction, and the lateral projection width of the vehicle is included in the calculation to obtain the earliest trigger time of each of the rising bollards, including: Based on the normalized longitudinal position of the optimal occlusion trajectory, the time when the optimal occlusion trajectory reaches normalization is extracted using the following formula:

[0172] Combining the normalized vehicle physical width, normalized vehicle physical length, and the normalized cosine and sine values ​​of the vehicle's arrival heading angle, the normalized vehicle lateral projection width is calculated using the following formula:

[0173] Based on the normalized lateral position of the arrival time of the optimal occlusion trajectory, the normalized lateral position of the center of the intrusion gap, the normalized width of the intrusion gap, and the normalized lateral projection width of the vehicle, the intrusion risk value of the intrusion gap is calculated using the following formula to screen out the highest-risk intrusion gap number:

[0174]

[0175] Combining the intrusion risk value of the aforementioned gap with the normalized center position of the rising bollard, the original influence weight of the rising bollard is calculated using the following formula, and then normalized to obtain the normalized influence weight of the rising bollard:

[0176]

[0177] Based on the normalized arrival time of the optimal occlusion trajectory, the normalized influence weight of the rising column, and the normalized lifting load of the rising column, the earliest triggering normalized time of each rising column is calculated using the following formula:

[0178] in, This indicates that the optimal occlusion trajectory reaches the normalized time. This represents the normalized longitudinal position of the optimal occlusion trajectory. This represents the normalized lateral projection width of the vehicle. This represents the normalized physical width of the vehicle. This represents the normalized cosine value of the vehicle's heading angle. This represents the normalized physical length of the vehicle. This represents the normalized sine value of the vehicle's heading angle upon arrival. Indicates the first The intrusion risk value of each impact gap. This represents the normalized lateral position at the arrival time of the optimal occlusion trajectory. Indicates the first The normalized lateral position of the center of each impact gap. Indicates the first The normalized width of the impact gap. This indicates the highest-risk gap number. Indicates the first The original influence weight of each rising and falling column. Indicates the relationship with the first The set of adjacent impact gaps of the rising columns. Indicates the first The normalized center position of each rising column Indicates the first The normalized impact weight of each rising and falling column. Indicates the first The earliest normalization time is triggered by the rising column. Indicates the first The lifting load of each rising bollard is normalized.

[0179] The normalized vertical position of the optimal occlusion trajectory is the vertical coordinate value of the optimal occlusion trajectory at any normalization time after particle swarm optimization.

[0180] The time when the optimal occlusion trajectory reaches the normalization point is the time when the longitudinal position of the optimal occlusion trajectory reaches the array cross section.

[0181] Normalized vehicle physical width is a dimensionless width value of the vehicle's physical width under a uniform lateral scale, used to estimate the lateral occupancy of the vehicle at the time of arrival.

[0182] Normalized vehicle physical length is a dimensionless length value of the vehicle's physical length under a uniform longitudinal or comprehensive scale, used to compensate for the lateral projection width when the vehicle has a yaw angle.

[0183] The normalized cosine value of the vehicle's heading angle is the directional characteristic value of the vehicle's heading angle cosine after uniform normalization when the vehicle arrives at the array section. It is used to reflect the relationship between the vehicle's orientation and lateral projection at the time of arrival.

[0184] The normalized sine value of the vehicle's heading angle is the directional characteristic value of the vehicle's heading angle sine value after being uniformly normalized when the vehicle arrives at the array section.

[0185] The normalized lateral projection width of a vehicle is the equivalent width occupied by the vehicle in the lateral direction when it arrives at the array section. The larger the value, the more difficult it is for the vehicle to pass through narrow gaps.

[0186] The normalized lateral position at the arrival time of the optimal occlusion trajectory is the lateral coordinate value corresponding to when the vehicle arrives at the array section along the optimal occlusion trajectory. It is used to determine which gap the vehicle is most likely to rush into.

[0187] The risk value of the g-th intrusion gap is a measure of the risk that a vehicle will intrude into the g-th intrusion gap when it arrives at the array section. The larger the value, the higher the degree of alignment between the vehicle and the center of the intrusion gap and the greater the probability of passage.

[0188] The highest-risk intrusion gap number is the identifier of the intrusion gap with the highest intrusion risk value among all intrusion gaps, and is used as the key target for group-column coordinated interception.

[0189] The original impact weight of the i-th rising bollard is the original effect strength of the i-th rising bollard on blocking the highest risk of rushing through the gap and the trajectory of nearby vehicles, and is used to assign different early trigger priorities among multiple rising bollards.

[0190] The set of impact gaps adjacent to the i-th rising column is the set of impact gaps that are geometrically directly adjacent to or directly related to the i-th rising column in terms of control.

[0191] The normalized influence weight of the i-th rising and falling column is the weight value after uniformly normalizing the original influence weights of each rising and falling column, which is used to eliminate the scale difference of the original influence weights of different rising and falling columns.

[0192] The normalized lifting burden of the i-th rising column is the normalized amount of the time and travel cost required for the i-th rising column to reach the effective blocking state from its current state. The larger the value, the earlier the rising column needs to be triggered.

[0193] The earliest trigger normalized time for the i-th rising bollard is the earliest trigger time calculated for the i-th rising bollard to ensure effective blocking before the vehicle arrives.

[0194] In specific implementation, regarding the search accuracy of the normalized arrival time of the optimal occlusion trajectory, the handling of parallel optimal points, and the discrete-time interpolation rules, the normalized longitudinal position of the optimal occlusion trajectory is sampled on a unified discrete-time grid, preferably with no less than 100 discrete points. If two adjacent sampling points cross the array cross section one after the other, linear interpolation is used to obtain a more accurate normalized arrival time of the optimal occlusion trajectory. If multiple times are at the same distance from the array cross section, the earliest time is taken as the normalized arrival time of the optimal occlusion trajectory.

[0195] In specific implementation, for the acquisition of normalized vehicle physical width, normalized vehicle physical length, and normalized components of vehicle arrival heading angle, the vehicle physical width and vehicle physical length can be obtained from vehicle type recognition results, visual 3D bounding boxes, LiDAR length and width estimation, or known vehicle templates. Then, they are divided by the lateral normalized reference width and the longitudinal normalized reference length, respectively, to obtain the normalized vehicle physical width and normalized vehicle physical length. The vehicle arrival heading angle can be calculated from the tangent direction of the final segment of the optimal occlusion trajectory, visual orientation estimation, or multi-frame position difference, and then the cosine and sine are mapped to normalized components.

[0196] In practice, the normalized cosine and sine values ​​in the formula for the normalized vehicle lateral projection width, as well as the normalized cosine and sine values ​​of the vehicle's heading angle, are derived from the original cosine and sine values ​​after mapping from -1 to 1 to 0 to 1. Therefore, when calculating the normalized vehicle lateral projection width, it is necessary to first restore the signed directional component by subtracting 1 from twice the normalized value, and then take the absolute value to represent the projection contribution of width and length in the lateral direction. When the vehicle's heading angle is completely aligned with the road's longitudinal direction, the width term contributes the most while the length term contributes less. When the vehicle veers significantly, the contribution of the length term to the lateral projection width increases.

[0197] In practice, regarding the threshold for the intrusion risk value of the rushing gap, the handling of the highest-risk rushing gaps, and the risk screening rules, after calculating the intrusion risk value of all rushing gaps, firstly, retain the valid risk items that are greater than 0; if there is a unique maximum value, it directly corresponds to the highest-risk rushing gap number; if the difference between the intrusion risk values ​​of two or more rushing gaps does not exceed 0.02, then the rushing gap that is closer to the normalized lateral position at the arrival time of the optimal shielding trajectory is selected first; if it is still impossible to distinguish, then the one with the lower average lifting progress of the adjacent lifting bollards is selected.

[0198] In specific implementation, for the modeling rules of the set of rushing gaps adjacent to the i-th rising column, the rising column located inside the array is adjacent to both the rushing gap on its left and the rushing gap on its right, and the boundary rising column located at the leftmost or rightmost end is adjacent to the rushing gap on only one side; if virtual boundary columns are used to construct edge rushing gaps, the boundary rising column can also be associated with an edge rushing gap at the same time.

[0199] In practical implementation, regarding the components, calibration methods, and range limitations of the normalized lifting load of the bollard, the normalized lifting load of the bollard should at least consider the current remaining lifting stroke ratio, the rated lifting speed of the actuator, the controller confirmation delay, the ambient temperature correction, and the equipment health status correction. During calibration, the time required for each bollard to rise from different initial heights to the effective blocking height should be recorded through bench tests, and then divided by the maximum allowable response time of the event to obtain the normalized value. This parameter should be cropped to the range of 0 to 1.

[0200] In specific implementation, for the correction rules for the earliest trigger normalized time of the rising bollard in negative, advanced, and multi-bollard synchronous scenarios, the result obtained by subtracting the weighted terms of the normalized lifting burden and normalized influence of the rising bollard from the normalized time of arrival of the optimal occlusion trajectory should be cropped to the range of 0 to 1. If the calculation result is less than 0, the earliest trigger normalized time of the i-th rising bollard is set to 0, indicating that it should be triggered immediately at the earliest time of the event. If the earliest trigger normalized times of multiple rising bollards fall within the same control cycle, these rising bollards are allowed to trigger synchronously.

[0201] Preferably, deterministic group column linkage execution and command locking are performed based on the earliest trigger time, and an irreversible lifting command is sent to the lifting column through an irreversible triggering mechanism, including: Obtain the monotonic locking trigger normalized time of the previous cycle's rising and falling columns, and compare it with the earliest trigger normalized time of the current cycle's rising and falling columns, which is updated from the earliest trigger normalized time of each of the aforementioned rising and falling columns. Take the minimum value of the two values ​​using the following formula as the monotonic locking trigger normalized time of the current cycle's rising and falling columns:

[0202] Determine whether the normalized event time of the current cycle reaches the monotonic locking trigger normalization time of the current cycle's lifting column to generate a lifting trigger determination indicator. Compare the lifting command status of the previous cycle's lifting column with the lifting trigger determination indicator, and take the maximum value of the two using the following formula as the lifting command status of the current cycle's lifting column to achieve irreversible command locking:

[0203] Obtain the normalized lifting progress of the lifting column in the previous cycle, and input the difference between the normalized event time of the current cycle and the normalized trigger time of the monotonic lock of the lifting column in the current cycle into the normalized lifting response curve of the lifting column to obtain the dynamic lifting mapping progress. Update the normalized lifting progress of the lifting column in the current cycle by taking the maximum value of the normalized lifting progress of the lifting column in the previous cycle and the dynamic lifting mapping progress using the following formula, so as to drive the lifting column to perform irreversible lifting:

[0204] in, Indicates the first The current cycle of the rising column triggers the normalization time of monotonic locking. Indicates the first The normalization time of the monotonic lock trigger of the previous cycle of each rising column. This indicates the number of times the normalization time of the earliest triggering of the rising column is updated. The earliest time that the current cycle of each rising column triggers normalization. Indicates the first The current cycle lifting command status of each bollard. Indicates the first The previous cycle's lifting command status for each bollard. This indicates the lifting trigger determination indicator. This represents the normalized event time for the current period. Indicates the first The current cycle of the normalized lifting progress of each bollard. Indicates the first The normalized lifting progress of the bollard in the previous cycle. Indicates the first Normalized lifting response curve of each rising bollard.

[0205] The normalized trigger time of the monotonic locking of the i-th lifting column in the current cycle is the trigger time of the i-th lifting column after being locked at the minimum value in the current control cycle.

[0206] The normalized time of monotonic lock trigger of the i-th rising column in the previous cycle is the lock trigger time that the i-th rising column has already formed in the previous control cycle.

[0207] The earliest trigger normalized time of the i-th rising column in the current cycle is the time value after the current control cycle by discretizing and mapping the continuously calculated trigger time.

[0208] The current cycle lifting command status of the i-th lifting column is a state variable indicating whether the i-th lifting column has entered the lifting command lock state in the current control cycle, and it is usually 0 or 1.

[0209] The lifting command status of the i-th lifting column in the previous cycle is the lifting command status recorded by the i-th lifting column in the previous control cycle.

[0210] The lifting trigger determination indicator is a determination quantity that converts the current cycle normalized event time into a discrete Boolean result to determine whether the lock trigger time has been reached.

[0211] The normalized event time for the current cycle is the event time sample value corresponding to the nth update cycle of the controller, which is used to drive command status updates and lift progress mapping.

[0212] The normalized lifting progress of the i-th lifting column in the current cycle is the actual lifting degree of the i-th lifting column in the current control cycle. The larger the value, the closer the column is to being fully raised.

[0213] The normalized lifting progress of the i-th lifting column in the previous cycle is the lifting degree recorded by the i-th lifting column in the previous control cycle.

[0214] The normalized lifting response curve of the i-th lifting bollard is a monotonic response curve that maps the time difference after triggering to the lifting progress of the bollard. Preferably, it is a piecewise smooth monotonically increasing curve, with the starting hysteresis zone preferably ranging from 0 to 0.10, the rapid rise zone preferably ranging from 0.10 to 0.80, and the top buffer zone preferably ranging from 0.80 to 1.00. This matches the three mechanical stages typically present in hydraulic or electromechanical lifting bollards: starting hysteresis, accelerated lifting, and top buffer. The complete lifting time, converted to actual time, is preferably between 1.5 and 5 seconds.

[0215] The dynamic lifting mapping progress is an instantaneous lifting progress value obtained by inputting the difference between the current cycle normalized event time and the locked trigger time into the normalized lifting response curve of the lifting column. It is used to convert discrete trigger commands into a continuous column lifting process.

[0216] In specific implementation, the control cycle is preferably between 10 and 20 milliseconds, considering the control cycle length, the discretization method of the normalized event time, and the update order of the current cycle. The normalized event time of the current cycle is obtained by subtracting the event start time from the actual time of the current control cycle and then dividing by the total duration of the event. Within each control cycle, the vehicle observation and bollard status are updated first, then the earliest trigger normalized time of the i-th bollard in the current cycle is calculated, and then the monotonic lock trigger normalized time of the i-th bollard, the lifting command status, and the normalized lifting progress of the bollard in the current cycle are updated.

[0217] In specific implementation, for the discrete rule that maps the earliest trigger normalized time of the current cycle rising and falling column from continuous time to periodic time, the earliest trigger normalized time of the i-th rising and falling column is mapped to the discrete control cycle according to the principle of not being later than the theoretical trigger time; when the theoretical trigger time is located between two cycle boundaries, the earlier cycle boundary is taken as the earliest trigger normalized time of the current cycle rising and falling column of the i-th rising and falling column to ensure that there is no delay.

[0218] In specific implementation, regarding the implementation form, anti-jitter conditions, and communication confirmation mechanism of the lifting trigger determination indicator, the lifting trigger determination indicator is preferably represented as a discrete Boolean value of 0 and 1 within the controller. When the normalized event time of the current cycle reaches or exceeds the normalized time of the monotonic locking trigger of the lifting column in the current cycle, the lifting trigger determination indicator is immediately set to 1. To avoid communication glitches causing misjudgment at the control end, the controller can be required to receive confirmation feedback from the lifting column driver after sending the lifting command. If no confirmation is received in the current cycle, the command will be resent in the next cycle without canceling the already formed locking state.

[0219] In practice, regarding the calibration method, monotonicity constraint, and multi-device difference compensation for the normalized lifting response curve of the bollard, multiple lifting tests are conducted on each bollard after installation. The actual height at each moment after triggering is recorded, and then divided by the rated maximum lifting height to obtain the normalized lifting progress curve. The same bollard is tested at least 5 times, the average curve is taken, and it is fitted with monotonic splines to ensure that the normalized lifting response curve of the bollard is always monotonic and does not decrease. Different bollards are allowed to use their own independent normalized lifting response curves to compensate for hydraulic differences, load differences, and temperature differences.

[0220] In practice, the dynamic lifting mapping progress output by the normalized lifting response curve of the lifting column is clipped to the range of 0 to 1, based on the limiting rules, power failure retention logic, and fault tolerance handling for the dynamic lifting mapping progress. If a power supply jitter or feedback loss occurs in a certain control cycle, the normalized lifting progress of the i-th lifting column in the current cycle will at least maintain the value of the previous cycle without decreasing. If the height sensor fails, but the actuator has received the lifting command, the controller will continue to conservatively advance according to the normalized lifting response curve of the lifting column and record the fault alarm.

[0221] In practical implementation, regarding the execution interface, retransmission strategy, and safety interlocking conditions for irreversible lifting commands, when the control equipment sends an irreversible lifting command to the bollard, it can be sent via industrial Ethernet, control bus, or hard-wired digital output interface. Once the lifting command status of the i-th bollard in the current cycle is locked to 1, the algorithm layer is not allowed to cancel the command except in manual maintenance mode and independent hardware emergency stop circuit. If no execution confirmation is received in the current cycle, it should be retransmitted for at least 3 consecutive cycles. At the same time, it should be hardware-level interlocked with anti-pinch injury, access control interlock, and vehicle passage permission signal to ensure that it will not be triggered erroneously in normal passage mode.

[0222] Preferably, after the interception event ends, the interception perception model is updated by using the extracted gap-revealed timing slip and the actual takeoff response, including: The mean of the normalized vehicle visible contour data and the mean of the available indicators of the observed samples are obtained, and the completeness of the current event assessment is calculated using the following formula:

[0223] Combining the completeness assessments of the historical model and the current event, the update weights for the current event model and the retention weights for the historical model are calculated using the following formulas:

[0224]

[0225] The current event scene feature vector is constructed by combining the gap exposure time-series slip, the current event normalized lift response data, the current event normalized occlusion template data, and the current event normalized reflection template data using the following formula:

[0226] By using the updated weights of the current event model and the maintained weights of the historical model, the feature vectors of the historical scene and the current event scene are weighted and fused, and the updated scene feature vector is calculated using the following formula:

[0227] By using the updated weights of the current event model and the maintained weights of the historical model, weighted fusion is performed on the historical occlusion mesh model and the current event occlusion mesh model, and the historical reflection mesh model and the current event reflection mesh model, respectively. The updated occlusion mesh model and the updated reflection mesh model are calculated using the following formulas:

[0228]

[0229] in, Indicates the completeness of the current event assessment. This represents the mean of the normalized vehicle visible outline data. The mean of the available indicators for the observed sample is represented. This indicates that the current event model updates the weights. Indicates the completeness of the historical model assessment. This indicates that the historical model retains its weights. This represents the feature vector of the current event scenario. This indicates the amount of time-series slippage when the gap is exposed. This represents the normalized lift response data for the current event. This represents the normalized occlusion template data for the current event. This represents the normalized reflection template data for the current event. This indicates updating the scene feature vector. Represents the feature vector of a historical scene. This indicates an update to the occlusion mesh model. Represents a historical occlusion mesh model. This indicates the current event occlusion mesh model. This indicates an update to the reflection mesh model. Represents a historical reflection mesh model. This represents the current event reflection mesh model.

[0230] The mean of normalized vehicle visibility data is an overall visibility index obtained by averaging the normalized vehicle visibility data throughout the entire process of the current event. It is used to reflect the overall observation clarity of the current event.

[0231] The observation sample availability indicator is a discrete availability flag used to indicate whether the observation sample at a certain moment can be used for subsequent statistics and model updates. It is usually set to 0 or 1 and is used to mask missing samples, abnormal reflection samples, and severely occluded samples.

[0232] The mean of available indicators for the observed samples is a sample availability index obtained by averaging the available indicators of the observed samples over the entire event timeline. It is used to measure the coverage of available observations in the current event.

[0233] The current event assessment completeness is an event reliability score obtained by combining the mean of normalized vehicle visible contour data and the mean of available indications of the observed samples. The higher the value, the more suitable the current event is for model updates.

[0234] The completeness of historical model assessment is a cumulative score of the reliability of existing historical models.

[0235] The current event model update weight is the contribution ratio of the current event calculated based on the completeness of the current event assessment and the completeness of the historical model assessment. The larger the value, the stronger the impact of the current event on the model update.

[0236] The historical model retention weight is the proportion of the historical model that is retained in this fusion update.

[0237] The current event normalized lifting response data is a set of normalized response data obtained by organizing the actual lifting process of each lifting column in this interception event.

[0238] The current event normalized occlusion template data is normalized template data constructed based on the spatiotemporal distribution of vehicles being occluded by rising bollards in this event, and is used to describe the occlusion pattern of this event.

[0239] The normalized reflection template data for the current event is normalized template data constructed based on the perceived disturbances caused by reflective sleeves, light rings, or other reflection sources in this event, and is used to describe the reflection interference pattern of this event.

[0240] The current event scene feature vector is a scene representation vector obtained by splicing the gap exposure timing slip amount, the current event normalized lift response data, the current event normalized occlusion template data, and the current event normalized reflection template data in a unified order. It is used to carry the comprehensive scene information of this event.

[0241] Historical scene feature vectors are scene representation vectors that the system has saved before the model is updated.

[0242] Updating the scene feature vector is a new scene representation vector obtained by fusing the historical scene feature vector and the current event scene feature vector according to weights.

[0243] The historical occlusion mesh model is an occlusion probability space mesh model that the system had stored before this update.

[0244] The current event occlusion mesh model is an occlusion mesh model constructed based on the observation results of the current event, used to characterize the actual occlusion distribution of the current event.

[0245] The updated occlusion mesh model is a new occlusion mesh model obtained by weighted fusion of the historical occlusion mesh model and the current event occlusion mesh model.

[0246] The historical reflection mesh model is a spatial mesh model of reflection perturbation that the system had stored before this update.

[0247] The current event reflection grid model is a reflection grid model constructed based on the observation results of the reflection disturbance of this event.

[0248] The updated reflection mesh model is a new reflection mesh model obtained by weighted fusion of the historical reflection mesh model and the current event reflection mesh model.

[0249] In practice, the criteria for determining the end of an interception event, the evaluation window, and the timing for triggering updates are as follows: the event is considered to be over when the vehicle has passed through the array section, completely left the monitoring area, been effectively stopped, or there has been no effective observation for more than 1 second; the evaluation window lasts from the start of the event until the lifting status of all relevant bollards stabilizes; the model feedback update should be triggered after the event ends and the data is written to disk to avoid modifying online parameters during the event.

[0250] In practice, based on the rules for generating the available indicator of the observed sample, the criteria for identifying abnormal samples, and the statistical range of the mean, for each sampling time, the availability of the observed sample at that time is determined according to the detection confidence level, whether severe overexposure reflection occurs, whether the contour area changes abruptly, whether the target leaves the road area, and whether the time continuity is reasonable. If the conditions are met, the available indicator of the observed sample is set to 1; otherwise, it is set to 0. The mean of the available indicator of the observed sample should be statistically analyzed within the entire evaluation window.

[0251] In specific implementation, regarding the initialization method, recursive update method, and decay mechanism of historical model evaluation completeness, the historical model evaluation completeness is initialized to 0.5 as a neutral prior when the system is first deployed. After each event update, the historical model evaluation completeness of the previous round and the current event evaluation completeness can be recalculated in an exponential sliding manner, with a preferred retention coefficient of 0.8 to 0.95. When the system has not experienced any effective events for a long period of time or the sensor configuration has changed, the historical model evaluation completeness can be appropriately reduced.

[0252] In specific implementation, regarding the construction methods of the current event normalized lifting response data, the current event normalized occlusion template data, and the current event normalized reflection template data, the current event normalized lifting response data is composed of the actual lifting progress sequence of each lifting column within the evaluation window, resampled according to a uniform time length and then stitched together; the current event normalized occlusion template data can be constructed from the occlusion ratio or visibility inverse value of the vehicle at different times and different lateral positions; the current event normalized reflection template data can be constructed from the statistical values ​​of high bright spot density, saturated pixel ratio, or reflection anomaly score on the same spatiotemporal grid.

[0253] In practice, regarding the dimensional alignment, missing value imputation, and standardization rules for the current event scene feature vector and the historical scene feature vector, the scene feature vector adopts a fixed dimension and a fixed order. It is preferable to store the gap exposure time-series slip first, followed by the current event normalized rise response data, the current event normalized occlusion template data, and the current event normalized reflection template data. If some components are missing, they are filled with 0, nearest neighbor interpolation, or historical mean, but the filling rules must be fixed within the same system. All components should be clipped to the range of 0 to 1 before being written.

[0254] In practice, for the historical occlusion mesh model, the current event occlusion mesh model, the historical reflection mesh model, and the current event reflection mesh model, all mesh models should use a unified road coordinate system, a unified horizontal and vertical resolution, and a unified time layering method; the spatial mesh resolution is preferably 0.05 meters to 0.20 meters, or a fixed mesh such as 64 by 32 is used; the current event mesh must be geometrically registered and resampled with the historical mesh before fusion.

[0255] In specific implementation, regarding the storage replacement, version rollback, and subsequent calling rules for updating scene feature vectors, updating occlusion mesh models, and updating reflection mesh models, after each model feedback update is completed, the system should simultaneously save the updated scene feature vectors, updated occlusion mesh models, and updated reflection mesh models, and record the event number, update time, and current event evaluation completeness; the main model is only allowed to be overwritten when the current event evaluation completeness is not lower than a preset threshold; at the same time, the most recent 3 to 10 historical models are retained to support version rollback.

[0256] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A bollard interception system for estimating the trajectory of abnormally rushing vehicles, comprising roadside sensing equipment, control equipment, and bollards arranged in a discrete cylindrical array, wherein the control equipment is communicatively connected to the roadside sensing equipment and the bollards, characterized in that, The control device is used to perform: Acquire interception scene data and convert the spatial arrangement parameters and lifting status of the rising column into multiple dynamic visual ranges; Based on the vehicle visual contour data acquired by the roadside sensing device, the gap exposure time-series slip amount is extracted. The gap exposure time-series slip amount represents the continuous temporal shift of the time centroid of the vehicle visual contour data when the vehicle passes through each of the dynamic visual intervals. Particle swarm initialization is performed by revealing the timing slip amount using the aforementioned gap; The particle swarm evolution and occlusion trajectory reconstruction are performed under the constraints of the spatial arrangement parameters, wherein the gap exposure time slip is used as an optimization parameter to participate in the multi-objective adaptive optimization of the particle swarm evolution. The high-risk intrusion gap and arrival time are calculated based on the optimal occlusion trajectory obtained from the reconstruction. The lateral projection width of the vehicle is included in the calculation to obtain the earliest trigger time of each of the rising bollards. Based on the earliest triggering time, deterministic group column linkage execution and command locking are performed, and an irreversible lifting command is sent to the lifting column through an irreversible triggering mechanism; After the interception event ends, the extracted gap reveals the timing slip and the actual takeoff response to update the interception perception model.

2. The bollard interception system for estimating the trajectory of abnormally rushing vehicles according to claim 1, characterized in that, Acquire interception scene data, and convert the spatial arrangement parameters and lifting status of the rising columns into multiple dynamic visual ranges, including: The adjacent lifting columns are divided into multiple impact gaps. The normalized center position and normalized diameter of each lifting column are obtained. The normalized left boundary, normalized right boundary and normalized width of the impact gap are calculated respectively. Based on the normalized lifting progress of the adjacent lifting columns, the average lifting progress of the impact gap is calculated. Based on the normalized position of the vehicle's lateral trajectory under normalized event time, the normalized left boundary of the impact gap, and the normalized width of the impact gap, the phase position inside the vehicle gap is calculated. By combining the phase position inside the vehicle gap with the average lifting progress of the impact gap, a dynamic visible range intensity is constructed to form the dynamic visible range.

3. The rising bollard interception system for estimating the trajectory of abnormally rushing vehicles according to claim 2, characterized in that, Based on the vehicle visual contour data acquired by the roadside sensing device, the gap exposure time-series slip is extracted. This gap exposure time-series slip characterizes the continuous temporal shift of the vehicle visual contour data's temporal centroid as the vehicle passes through each of the dynamic visual intervals, including: The currently acquired vehicle visual contour data is used as the original vehicle visual contour data for extreme value mapping to obtain normalized vehicle visual contour data. Extract multiple time segments of the visible intervals divided by the dynamic visible intervals, and calculate the local normalized time inside the visible segment based on the normalized event time, the left boundary of the visible interval time segment, and the right boundary of the visible interval time segment. By combining the local normalized time within the visible segment with the normalized vehicle visible contour data, the time centroid of the visible contour data within the visible segment corresponding to each visible interval time segment is calculated. Based on the duration proportion of each visible time segment, the normalized weight of the visible segment duration is calculated. Based on the difference in the temporal centroid of the visible contour data within adjacent visible segments, and combined with the normalized weight of the duration of the visible segments, a segment-by-segment cumulative offset calculation is performed to extract the temporal slip amount of the gap exposure.

4. The rising bollard interception system for estimating the trajectory of abnormally rushing vehicles according to claim 3, characterized in that, Particle swarm initialization is performed by revealing the timing slip using the gap, including: Based on the normalized longitudinal position and normalized lateral position of the first historical effective observation point, the normalized longitudinal position and normalized lateral position of the second historical effective observation point are calculated to obtain the geometric trajectory extrapolation coefficient and the geometric prediction normalized lateral endpoint. Based on the normalized sorting value of the impact gap and the time slip of the gap exposure, the time slip similarity bias weight is calculated. By combining the temporal slip similarity bias weight with the normalized lateral position of the center of the impact gap, the normalized lateral endpoint of the slip bias is calculated. The temporal slip intensity is calculated based on the gap exposure temporal slip amount, and the basic fusion normalized lateral endpoint is obtained by weighting the geometric prediction normalized lateral endpoint and the slip offset normalized lateral endpoint. The initial total number of particles is calculated based on the total number of the rushing gaps and the total number of the visible time segments of the visible time segment; The initial particle trajectory normalization starting point is set based on the normalized longitudinal position and the normalized lateral position of the first historical effective observation point. The initial particle trajectory normalization ending point is set based on the basic fusion normalized lateral ending point. Within the range of the difference between the longitudinal coordinate envelope and the difference between the lateral coordinate envelope, the initial particle normalization parameter vector is generated by combining the basic curve parameter vector and the low-difference random sequence value. The initial particle normalization parameter vector is used as the particle trajectory control point parameter vector.

5. The rising bollard interception system for estimating the trajectory of abnormally rushing vehicles according to claim 4, characterized in that, Performing particle swarm evolution and occlusion trajectory reconstruction constrained by the spatial arrangement parameters, wherein the gap exposure temporal slip is used as an optimization parameter in the multi-objective adaptive optimization of the particle swarm evolution, including: Candidate occlusion trajectories are generated by combining the particle trajectory control point parameter vector, and the phase positions inside the gaps between the candidate trajectories are extracted to construct predicted visual contour data; Calculate the observation geometric fit residual of the candidate occlusion trajectory, the exposure temporal centroid fit residual extracted by combining the temporal centroid of the visible contour data inside the visible segment, the dynamic smoothness residual, and the high-risk gap penetration fit residual; The residual channel adaptive weights are obtained based on the residual channel difference variance to fuse the comprehensive optimization target residual; By combining the individual evolutionary cognitive weights and the global evolutionary collaborative weights, the particle's current normalized parameter vector is updated through the comprehensive optimization target residual to reconstruct the optimal occlusion trajectory.

6. The rising bollard interception system for estimating the trajectory of abnormally rushing vehicles according to claim 5, characterized in that, Based on the optimal occlusion trajectory obtained from the reconstruction, the high-risk intrusion gap and arrival time are calculated. The lateral projection width of the vehicle is included in the calculation to obtain the earliest trigger time of each of the aforementioned bollards, including: Based on the normalized longitudinal position of the optimal occlusion trajectory, the time when the optimal occlusion trajectory reaches the normalization is extracted. The normalized vehicle lateral projection width is calculated by combining the normalized vehicle physical width, normalized vehicle physical length, normalized cosine value of vehicle arrival heading angle and normalized sine value of vehicle arrival heading angle. Based on the normalized lateral position of the arrival time of the optimal shielding trajectory, the normalized lateral position of the center of the intrusion gap, the normalized width of the intrusion gap, and the normalized lateral projection width of the vehicle, the intrusion gap intrusion risk value is calculated to screen out the highest risk intrusion gap number. By combining the intrusion risk value of the impact gap with the normalized center position of the rising column, the original influence weight of the rising column is calculated, and then normalized to obtain the normalized influence weight of the rising column. Based on the normalized arrival time of the optimal occlusion trajectory, the normalized influence weight of the rising column, and the normalized lifting load of the rising column, the earliest triggering normalized time of each rising column is calculated.

7. The rising bollard interception system for estimating the trajectory of abnormally rushing vehicles according to claim 6, characterized in that, Based on the earliest trigger time, deterministic group column linkage execution and command locking are performed, and an irreversible lifting command is sent to the lifting column through an irreversible triggering mechanism, including: Obtain the monotonic locking trigger normalization time of the previous cycle's rising column, compare it with the earliest trigger normalization time of the current cycle's rising column, which is updated from the earliest trigger normalization time of each of the rising columns, and take the minimum value of the two as the monotonic locking trigger normalization time of the current cycle's rising column. Determine whether the normalized event time of the current cycle reaches the normalized time of the monotonic locking trigger of the current cycle lifting column, so as to generate a lifting trigger determination indicator. The lifting command status of the previous cycle is compared with the lifting trigger determination indicator, and the maximum value of the two is taken as the lifting command status of the current cycle, so as to achieve irreversible locking of the command. Obtain the normalized lifting progress of the lifting column in the previous cycle, and input the difference between the normalized event time of the current cycle and the normalized trigger time of the monotonic lock of the lifting column in the current cycle into the normalized lifting response curve of the lifting column to obtain the dynamic lifting mapping progress. Take the maximum value of the normalized lifting progress of the lifting column in the previous cycle and the dynamic lifting mapping progress to update the normalized lifting progress of the lifting column in the current cycle, so as to drive the lifting column to perform irreversible lifting.

8. The rising bollard interception system for estimating the trajectory of abnormally rushing vehicles according to claim 7, characterized in that, After the interception event concludes, the extracted gap-revealed timing slip and the actual takeoff response are used to update the interception perception model, including: The mean of the normalized vehicle visible contour data and the mean of the available indicators of the observation samples are obtained to calculate the completeness of the current event assessment. By combining the completeness of the historical model assessment with the completeness of the current event assessment, the update weight of the current event model and the retention weight of the historical model are calculated respectively. The gap exposure timing slip, the current event normalized lift response data, the current event normalized occlusion template data, and the current event normalized reflection template data are combined to construct the current event scene feature vector; By using the updated weights of the current event model and the maintained weights of the historical model, the feature vectors of the historical scene and the current event scene are weighted and fused to obtain the updated scene feature vector. By using the updated weights of the current event model and the weights of the historical model, the historical occlusion mesh model and the current event occlusion mesh model, and the historical reflection mesh model and the current event reflection mesh model are respectively weighted and fused to obtain the updated occlusion mesh model and the updated reflection mesh model.

Images