Label splicing method after vehicle identification failure in semi-closed space

Abstract

The invention discloses a label splicing method after vehicle identification failure in a semi-closed space, and the method comprises the following steps: carrying out the joint calibration of vehicles through a camera and a radar, and endowing each vehicle entering the semi-closed space with an entry label; constructing a plane gridding base map for the semi-closed space, and constructing existence probability distribution of the target vehicle in the base map; if the detection of the target vehicle is lost, predicting and correcting the position of the target vehicle; and when the target vehicle is identified by the radar again, track connection is carried out on the current observation item and the historical base map item with the highest matching degree. The method has the beneficial effects that by introducing a base map track memory and re-recognition matching mechanism, the label information of the vehicle in a multi-radar and multi-camera cooperative scene can be continuously continued. Compared with the problem of re-numbering after shielding or broken identification in a traditional method, the method can effectively reduce the phenomena of track breakage and wrong numbering, and remarkably improves the continuity and reliability of vehicle global tracking.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent traffic monitoring, in particular to a method for label continuation after vehicle disrecognition in a semi-closed space. BACKGROUND

[0002] The semi-closed space includes a high-speed service area and an open parking lot, etc. Taking the high-speed service area as an example, it plays a key role in vehicle distribution, temporary parking, charging and energy supplement, and safety control, etc. as an important part of the modern highway traffic system. With the increase of traffic flow and the diversification of vehicle types, the traffic state inside the service area is becoming more and more complex, and problems such as high vehicle density, frequent entry and exit, and non-standard parking significantly increase the difficulty of the perception system and the data processing burden.

[0003] Laser radar (LiDAR) is widely used in three-dimensional perception and vehicle monitoring in service area and other scenes due to its high precision, anti-light, and all-weather working advantages. The existing technology mainly focuses on two directions for the application of laser radar. One direction is to focus on improving the accuracy of vehicle recognition and classification, such as enhancing point cloud target detection through deep learning or recognizing vehicle types based on visual features. The other direction is to focus on multi-frame point cloud target tracking, which realizes short-time continuous recognition through Kalman filtering or multi-target matching. However, these methods generally have the following shortcomings: (1) Lack of label continuation mechanism in disrecognition: once a vehicle is not recognized in several frames, its label is reset and cannot be associated with previous vehicle information. (2) Over-reliance on visual data: some solutions rely on multiple cameras in the field to continuously identify and compare, which requires large data transmission and high installation and maintenance costs. (3) Not considering the spatial constraint characteristics of the service area scene: such as the constraint relationship of the vehicle driving by the lane direction, buildings and green belts, etc. (4) Bottom map information does not form a long-term memory mechanism: most algorithms are based on short-time trajectory prediction, but lack of fusion and update of historical distribution probability, resulting in insufficient stability.

[0004] In summary, the existing vehicle tracking methods still have difficulty in maintaining label continuity after vehicle disrecognition in semi-closed spaces such as service areas, affecting the accuracy of vehicle global tracking and event correlation analysis. SUMMARY

[0005] One of the purposes of the present application is to provide a method for label continuation after vehicle disrecognition in a semi-closed space, which can solve at least one of the defects in the background art.

[0006] To achieve the above at least one purpose, the technical solution adopted by the present application is as follows: a method for label continuation after vehicle disrecognition in a semi-closed space, comprising the following steps: S100: Joint calibration of vehicles by cameras and radars at the entrance of semi-enclosed space, so that each vehicle entering the semi-enclosed space is given a target tag containing complete information about itself; S200: Building a planar gridded base map of the semi-enclosed space and establishing environmental constraints; based on the position of the target vehicle in the base map and the identification result of the target vehicle by radar, constructing and dynamically maintaining the existence probability distribution of the target vehicle in the base map to form a base map entry; S300: If the radar loses detection of the target vehicle during the continuous tracking stage, generate a position prediction of the target vehicle by Kalman filtering; after correcting the position prediction according to the driving direction of the vehicle and the environmental attributes, fuse it with the historical base map entry; S400: When the target vehicle is identified again by radar, continue the trajectory with the historical base map entry with the highest matching degree; if there are multiple historical base map entries that meet the matching degree requirement, prioritize based on historical trajectory, direction vector, and base map probability distribution.

[0007] Preferably, in step S100, the camera is responsible for providing the appearance information of the vehicle, including the license plate number, accurate vehicle model, color, and structural features; the radar is responsible for providing the spatial position information and rough vehicle model of the vehicle; the target boxes of the vehicle in the camera coordinate system and the radar coordinate system are unified in the coordinate system, and the IOU value is calculated; if the calculated IOU value is greater than the set threshold value, and the vehicle category meets the fuzzy matching condition, it is determined that the camera and radar identification results correspond to the same vehicle, and a target tag containing complete information about the vehicle is assigned to the vehicle.

[0008] Preferably, in step S200, the base map is divided into different spatial semantic partitions according to environmental functions, and different probability weights β are assigned to different spatial semantic partitions; the occupancy probability of the target vehicle at any grid position in the base map is calculated , the occupancy probability of all grids is summarized , and the existence probability distribution of the target vehicle in the base map is obtained; the calculation formula of the occupancy probability is: ; ; wherein, represents the occupancy probability of the target vehicle under the identification of the i-th radar, k represents the total number of radars that can identify the target vehicle in the semi-enclosed space, (x0, y0) represents the coordinates of the center of the target vehicle in the base map coordinate system, (x, y) represents the center coordinates of any grid in the base map coordinate system, σ p represents the diffusion parameter.

[0009] Preferably, in step S200, the bottom map retains the position, existence probability distribution and target label of each vehicle to form a bottom map entry; during the continuous tracking phase of the target vehicle in the semi-closed space, the update of the bottom map entry includes the following process: the target vehicle trajectory output by the radar in each frame is matched with all bottom map entries based on the matching score calculation of trajectory overlap, distance difference, probability distribution and appearance similarity; if the calculated matching score is greater than the set first matching threshold, it is determined that the target vehicle trajectory of the current frame is associated with the corresponding bottom map entry, and the position of the bottom map entry is updated through the matching of the target vehicle trajectory; at the same time, the existence probability distribution in the bottom map entry is updated through the confidence probability observed by the current frame radar and the predicted occupancy probability of the historical frame.

[0010] Preferably, the matching score S i is calculated according to the following formula: ; The update expression of the existence probability distribution of the bottom map entry is as follows: ; Wherein, represents the overlap degree of the target vehicle trajectory and the bottom map entry in the BEV view, c T represents the vehicle center point coordinate of the target vehicle trajectory, c i represents the vehicle center point coordinate corresponding to the i-th bottom map entry, σ d represents a normalization parameter, P i represents the historical probability corresponding to the i-th bottom map entry, type score and plate score respectively represent the vehicle type and license plate matching scores, w1 to w5 all represent corresponding weight coefficients, represents the confidence probability update value of the target vehicle in the bottom map at the current t time, represents the existence probability prediction of the bottom map entry corresponding to the current matched target vehicle at t time in the historical frame, represents the confidence probability observed by the current frame radar, and λ represents the fusion weight of the historical information and the new observation information.

[0011] Preferably, the step S300 comprises the following specific process: if the radar loses detection of the target vehicle in the continuous tracking stage, marking the corresponding map entry of the target vehicle; for each target vehicle whose map entry is marked, using Kalman filtering to perform multi-hypothesis position prediction to obtain the original prediction probability of the corresponding grid position of the target vehicle in the map; combining the environmental constraints of the grid position where the target vehicle is located and the weight of the vehicle motion direction to correct the original prediction probability to obtain the single-vehicle prediction probability; based on the vehicle motion direction, obtaining the candidate grid range of the target vehicle in the map, and then normalizing the single-vehicle prediction probability according to the obtained grid range to obtain the single-vehicle prediction probability distribution; fusing the obtained single-vehicle prediction probability distribution with the existence probability distribution of the historical map entry and writing back the historical map entry or updating the existence probability distribution of the historical map entry.

[0012] Preferably, a retention mechanism is set for the map entry marked by the radar losing detection of the target vehicle, and the retention mechanism comprises the following contents: if the marked map entry is from the camera at the entrance position and the license plate confidence is greater than a set confidence threshold, or the difference between the last recorded radar observation time of the marked map entry and the current time is not more than a set interval threshold, the marked map entry is retained; the marked map entry is deleted when it is determined that the target vehicle has left or the manual / system display is cancelled; for the marked map entry that is only temporarily detected by the radar but not determined by the camera at the entrance position, if the marked map entry has no further observation or matching within a set duration, the confidence probability of the marked map entry is exponentially decayed until it is finally deleted.

[0013] Preferably, the step S400 comprises the following process: when the target vehicle is recognized by the radar again, calculating the matching score of the current observation entry and the historical map entry, and if the calculated matching score is greater than a set second matching threshold, associating and matching the current observation entry with the corresponding historical map entry; if there are multiple historical map entries that satisfy the second matching threshold, selecting the historical map entry with the highest matching score for association and performing a probability suppression operation on the remaining historical map entries; if the matching scores of the historical map entries that satisfy the second matching threshold differ by less than a set error threshold, comparing the time difference between the last radar observation time of each historical map entry and the current frame, and selecting the historical map entry with the smallest time difference for association.

[0014] Preferably, when the trajectory continuation of the current observation entry and the historical map entry is performed, the continuity score C of the current trajectory corresponding to the current observation entry and the historical trajectory corresponding to the historical map entry is calculated; if the calculated continuity score C exceeds a set continuity threshold, the current trajectory and the historical trajectory are spliced and updated in the map; the calculation formula of the continuity score C is as follows: ; Where Δt, Δd, and Δθ represent the time interval, spatial interval, and difference in direction of motion between the current trajectory and the historical trajectory, respectively; , as well as These represent the attenuation coefficients for the time interval, spatial interval, and direction of motion difference between the current trajectory and historical trajectories, respectively; w t w d and w θ All of these represent weighting coefficients.

[0015] Preferably, when the target vehicle leaves the semi-enclosed space, if the target vehicle identified by the camera based on the exit position is inconsistent with the base map entry assigned by the radar, a swap verification is triggered. Specifically, the process includes the following steps: within a set short time window, other marked base map entries are searched to replace the base map entry assigned by the radar to the target vehicle, and a matching score is calculated; if there is a marked base map entry that improves the matching score, the marked base map entry is swapped with the base map entry assigned by the radar to the target vehicle, and the reason and evidence for the swap are recorded.

[0016] Compared with the prior art, the beneficial effects of this application are as follows: (1) This application introduces a base map trajectory memory and re-identification matching mechanism, enabling the continuous continuation of vehicle tag information in multi-radar and multi-camera collaborative scenarios. Even in the event of short-term radar miss detection, occlusion, or field of view switching, the system can still maintain stable association of vehicle tags based on historical probability distribution and Kalman prediction information. Compared with the problem of re-numbering after occlusion or discontinuation of recognition in traditional methods, this application can effectively reduce trajectory breakage and misnumbering, and significantly improve the continuity and reliability of vehicle full-domain tracking.

[0017] (2) In the vehicle re-identification stage, this application comprehensively considers multiple features such as spatial overlap, vehicle category, consistency score, and license plate matching, forming a multi-factor fusion matching scoring system. At the same time, an exit interchange verification logic is introduced, which can automatically identify and correct label assignment errors when license plate mismatch or target intersection occurs. This mechanism effectively prevents problems such as multi-target mismatch and cross-pairing in complex traffic scenarios, ensuring the label consistency and trajectory stability of the system under high-density traffic conditions.

[0018] (3) The base map mechanism of this application adopts a probability retention and exponential decay strategy to maintain a certain probability of existence for temporarily missing vehicles and not immediately clear entries due to short-term detection loss. When a vehicle re-enters the visible area or is identified by the exit camera, the system can quickly match and restore tags based on the residual probability distribution and historical trajectory information. This design ensures that the system can maintain continuous perception of intermittently observed targets during long-term monitoring, providing basic support for all-time scene coverage. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall working steps of this application. Detailed Implementation

[0020] The present application will now be further described in conjunction with specific embodiments. It should be noted that, in the description of this specification, the use of terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicates that the specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms should not be construed as necessarily referring to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0021] In the description of this application, it should be noted that the terms "center", "lateral", "longitudinal", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", etc., which indicate the orientation and positional relationship based on the orientation or positional relationship shown in the accompanying drawings, are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and should not be construed as limiting the specific protection scope of this application.

[0022] It should be noted that the terms "first," "second," etc., in the specification and claims of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0023] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "joining," and "fixing," etc., should be interpreted broadly. For example, they can refer to a connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0024] In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature being directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature being directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0025] The terms “comprising” and “having”, and any variations thereof, in the specification and claims of this application are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0026] One preferred embodiment of this application, such as Figure 1 As shown, a method for reconnecting a vehicle tag after identification failure in a semi-enclosed space includes the following steps: S100: At the entrance of the semi-enclosed space, vehicles are jointly calibrated using cameras and radar, so that each vehicle entering the semi-enclosed space is assigned an entry label containing its complete information.

[0027] Understandably, semi-enclosed spaces typically deploy checkpoint cameras and radar at entrances and exits, while multiple fixed LiDAR units are deployed within the semi-enclosed space for vehicle identification and tracking. The checkpoint cameras are responsible for identifying vehicle appearance information, including license plate number, precise vehicle model, color, and structural features. The radar detects the time of vehicle entry and exit from the semi-enclosed space, its spatial location, and a rough model, and performs joint calibration of the vehicle after aligning with the camera's spatial coordinate system. The radar within the semi-enclosed space performs full-area vehicle tracking, primarily for identifying the vehicle's spatial location and rough model.

[0028] S200: Construct a planar gridded base map for the semi-enclosed space and establish environmental constraints; based on the position of the target vehicle in the base map, construct and dynamically maintain the probability distribution of the target vehicle's presence in the base map through radar identification results to form base map entries.

[0029] It is understandable that the base map is based on the planar projection of the semi-enclosed space onto the ground, and the entire projection area is divided into several two-dimensional grids of a set specification. The specific size of the grid can be selected according to the actual needs of those skilled in the art, for example, the size of each grid is 1m×1m.

[0030] Once a vehicle is identified by the checkpoint camera and enters the semi-enclosed space, a set of grid probability distributions can be initialized for it on the base map. This is the probability that a vehicle may exist in each grid within a local area, and can also be called the existence probability distribution. In the grid probability distribution, the probability of the vehicle's center location is 1, and the neighboring grids are distributed based on the uncertainty of the vehicle's size and location. That is, except for the grid where the vehicle is currently located, the existence probability of the vehicle in other grids can be regarded as the probability that the vehicle may drive to that location in the future.

[0031] Based on the probability distribution of the target vehicle's presence in the base map, each grid can store several attributes to achieve base map trajectory memorization, providing an information basis for subsequent vehicle re-identification after discontinuity identification. The specific attributes stored in each grid include the probability that the current grid is occupied by a certain vehicle, vehicle information associated with that vehicle (license plate number, vehicle type, and trajectory ID, etc.), and the vehicle's historical trajectory, speed, and direction information.

[0032] S300: If the radar loses detection of the target vehicle during the continuous tracking phase, a position prediction of the target vehicle is generated through Kalman filtering; the position prediction is then corrected based on the vehicle's driving direction and environmental attributes and fused with historical base map entries.

[0033] As can be understood from the foregoing, the existence probability distribution can predict the driving status of a vehicle over a future period. Since the existence probability distribution radiates in multiple directions from the vehicle's center point, it can provide trajectory predictions for vehicles in different directions. However, a vehicle's actual driving trajectory can only be one. Therefore, based on the vehicle's historical driving data, a Kalman filter can be used to predict the target vehicle's position with relatively high accuracy. This prediction can then be corrected by combining environmental attributes and the vehicle's driving direction, making the predicted trajectory close to the actual trajectory the vehicle traveled during the radar detection and loss process. After predicting the trajectory of the target vehicle during the radar detection and loss process, this predicted trajectory can be fused with historical base map entries, allowing the target vehicle's full-domain trajectory within a semi-enclosed space to be continued.

[0034] S400: When the target vehicle is identified by radar again, the current observation entry is connected to the historical base map entry with the highest matching degree for trajectory continuation; if multiple historical base map entries meet the matching degree requirements, priority is determined based on historical trajectory, direction vector and base map probability distribution.

[0035] It is understandable that when radar loses identification of a target vehicle, meaning the radar has a blind spot in a localized area within a semi-enclosed space, then every vehicle passing through that area will be lost to recognition. If multiple vehicles are simultaneously lost and then re-identified in that area, multiple vehicles may need to have their trajectories continued. Therefore, it is necessary to match the current observation entries of different vehicles with different historical base map entries, and then associate each vehicle with the historical base map entry with the highest matching degree to achieve trajectory continuation.

[0036] However, since radar can only roughly identify the vehicle type, the target vehicle may have a high degree of matching with multiple historical map entries. Therefore, in order to reduce or avoid misjudgment, it is necessary to combine the spatiotemporal characteristics of the target vehicle to determine the priority of each historical map entry, and then connect the current observation entry with the historical map entry with the highest priority.

[0037] The technical solution of this application introduces a base map trajectory memory and re-identification matching mechanism, enabling continuous vehicle tag information in multi-radar, multi-camera collaborative scenarios. Even in cases of short-term radar misses, occlusion, or field-of-view switching, the system can still maintain stable vehicle tag association based on historical probability distribution and Kalman prediction information. Compared to traditional methods that require re-numbering after occlusion or discontinuation of identification, this application effectively reduces trajectory breaks and misnumbering, significantly improving the continuity and reliability of vehicle tracking across the entire domain.

[0038] In this embodiment, during step S100, since the camera can recognize color information, it can accurately distinguish the vehicle type and color. Therefore, the checkpoint camera is responsible for recognizing the vehicle's appearance information, including license plate number, precise vehicle type (sedan, SUV, van, truck, bus, etc.), color, and structural features. Since the radar can only provide point coordinate information, its accuracy in recognizing vehicles with similar physical structures is not high. Therefore, the checkpoint radar provides the vehicle's spatial location information and a rough vehicle type (large vehicle, small vehicle). The radar output is a target bounding box and vehicle type information based on a bird's-eye view (BEV). Based on the working methods of the camera and radar, the following will describe in detail the specific process of assigning entry tags to the target vehicle when it enters the entrance of the semi-enclosed space using the checkpoint camera and checkpoint radar.

[0039] Specifically, assigning tags to entries involves the following process: S110: By jointly calibrating the camera and radar, the extrinsic parameter matrix between their coordinate systems is obtained. .

[0040] It is important to know that for extrinsic parameters... The specific expression for obtaining the result is: .

[0041] Where, p l and p c These represent the vehicle's three-dimensional coordinates in the radar coordinate system and the camera coordinate system, respectively.

[0042] S120: Based on the obtained extrinsic parameter matrix, the target bounding box of the vehicle in the camera coordinate system and the radar coordinate system can be unified.

[0043] It should be noted that the unification of coordinate systems can be achieved by projecting the camera coordinate system onto the radar coordinate system or by projecting the radar coordinate system onto the camera coordinate system. Considering that the camera coordinate system is a pixel coordinate system and the radar coordinate system is a local coordinate system, it is more convenient to project the pixel coordinate system onto the local coordinate system. Therefore, in this embodiment, the camera detection frame is projected onto the local coordinate system.

[0044] S130: When the target vehicle enters the entrance of the semi-enclosed space, the crossover ratio (IOU) of the target frame detected by the camera and radar at the same time is calculated in the BEV plane. If the calculated IOU value is greater than the set threshold and the vehicle category meets the fuzzy matching condition, it is determined that the camera and radar recognition results correspond to the same vehicle. Then, the complete recognition information of the vehicle is written into the database and a unique vehicle tag ID is generated, i.e., tag entry.

[0045] It's important to know that the formula for calculating the IOU value is as follows: .

[0046] Among them, A overlap A represents the area of ​​the overlapping region between the camera detection box and the radar detection box in the BEV plane. union This represents the area of ​​the union region of the camera detection frame and the radar detection frame in the BEV plane.

[0047] It is also important to know that the specific value of the threshold used for IOU value determination can be selected according to the actual needs of those skilled in the art, for example, it can be 0.5. For fuzzy matching conditions for vehicle categories, it can be seen that the vehicle models identified by the camera and radar tend to be consistent. For example, if the radar identifies the target vehicle as a "large vehicle", while the camera identifies the target vehicle as a "hazardous goods vehicle / truck / bus", then the identification of vehicle categories by the camera and radar can be regarded as a match.

[0048] In this embodiment, during step S200, to enhance the accuracy of prediction and continuation, the base map can be spatially semantically partitioned based on environmental functions during the generation stage. The spatial semantic partitions include: drivable areas, parking areas, and non-drivable areas. Drivable areas mainly correspond to driving lanes, ring roads, etc., allowing vehicle movement; therefore, no constraints are required for drivable areas, meaning the initial value of the probability weight β for calculating the existence probability distribution in drivable areas is 1. Parking areas indicate that vehicles move slowly or remain stationary, reflecting a low probability of vehicle movement; therefore, certain environmental constraints are needed for parking areas, i.e., the probability weight β for calculating the existence probability distribution in parking areas can be set with a decay weight, meaning the probability weight β can decay over time within a given range. The specific decay range can be selected according to the actual needs of those skilled in the art, for example, a decay range of [0.3, 0.7]. Non-drivable areas mainly include buildings, green belts, etc., where vehicles are unlikely to appear; therefore, strict environmental constraints are needed for non-drivable areas, i.e., the probability weight β for calculating the existence probability distribution in non-drivable areas can be a fixed probability value of 0. This type of partitioning information is imported into the system through structured maps or on-site annotations to provide boundary constraints for subsequent Kalman prediction and probability updates.

[0049] In this embodiment, the calculation of the probability distribution of the target vehicle's existence in step S200 can be achieved by first calculating the probability of the target vehicle occupying any grid position in the base map based on the coordinates (x0, y0) of the target vehicle's center in the base map coordinate system. Then, by summing the probability of the target vehicle occupying all grids in the base map, the required probability distribution of existence can be obtained. There are various methods for calculating the probability of the target vehicle occupying grids in the base map. In this embodiment, a two-dimensional Gaussian kernel function is preferred for calculation. The specific calculation formula is as follows: .

[0050] Among them, P occ (x, y) represents the probability that the target vehicle occupies any grid center (x, y) in the base map coordinate system; σ p Denotes the diffusion parameter, σ p The value of σ depends on the size of the vehicle and the detection error range. When the vehicle is identified as being in a parked state, the diffusion parameter σ can be automatically adjusted based on the vehicle's stationary time. p The value of is chosen to prevent excessive drift in multi-frame prediction of the radar.

[0051] Based on the above formula for calculating the occupancy probability, when the target vehicle is driving continuously in the drivable area, its occupancy probability distribution is as follows: the occupancy probability of the central grid actually occupied by the target vehicle is 1, and the occupancy probability of the surrounding grids decreases with distance; for grids in the non-drivable area, the occupancy probability is always 0 based on the value of the probability weight β.

[0052] It is important to note that when a vehicle is driving in a semi-enclosed space, it may be detected by multiple radars. Generally, the occupancy probability is calculated using the identification result of the main radar closest to the target vehicle. However, to increase the system's robustness to partial obstruction or single radar failure, this embodiment updates the target vehicle's occupancy probability using the identification results of the remaining radars after calculating the occupancy probability based on the main radar. Let the total number of radars capable of identifying the target vehicle in the semi-enclosed space be k. Then, the occupancy probability of the target vehicle on any grid in the base map under the identification of k radars is... The calculation formula is: .

[0053] in, Let i represent the probability of a target vehicle being occupied by the i-th radar, where i∈{1,2,...,k}.

[0054] In this embodiment, during step S200, the base map retains the position, existence probability distribution, and entry labels of each vehicle to form base map entries. During the continuous tracking phase of the target vehicle within a semi-enclosed space, to ensure the vehicle label information continues even with discontinuous identification, a dual verification mechanism combining the base map and radar output can be used to update the base map entries in real time, achieving continuous updating and error correction of vehicle information. The updating of base map entries includes the following process: S210: The radar outputs the trajectory of the currently detected target vehicle in each frame, and the base map retains the base map entry information for each vehicle; in order to correlate the radar output with the base map information, the target vehicle trajectory output by the radar in each frame is matched with all base map entries based on trajectory overlap, spacing difference, probability distribution, and appearance similarity score S. i The calculation.

[0055] Specifically, the matching score S i The calculation formula is as follows: .

[0056] in, c represents the degree of overlap between the target vehicle trajectory and the base map entry from the BEV perspective. T c represents the coordinates of the vehicle's center point on the target vehicle's trajectory. iThis represents the coordinates of the vehicle's center point corresponding to the i-th base map entry. σ represents the squared Euclidean distance between the coordinates of the vehicle center point of the target vehicle trajectory and the coordinates of the vehicle center point corresponding to the i-th base map entry. d P represents the normalization parameter. i This represents the historical probability corresponding to the i-th base map entry, that is, the probability that this base map entry was actually passed by a vehicle in history. For example, if 30 vehicles passed near this base map entry, and 15 of them actually passed by the base map entry, then the historical probability of this base map entry is 15 / 30 = 0.5; score and plate score The values ​​represent the vehicle model and license plate matching scores, respectively, ranging from 0 to 1. These scores represent the matching based on shape and text, respectively. The specific matching methods are well-known to those skilled in the art and will not be described in detail here. w1 to w5 represent the corresponding weight coefficients, and the specific values ​​can be selected according to the actual needs of those skilled in the art.

[0057] S220: Set a first matching threshold and compare the calculated matching score with the first matching threshold; if the calculated matching score is greater than the set first matching threshold, determine that the target vehicle trajectory in the current frame is associated with the corresponding base map entry, and then update the position of the base map entry through the matching of the target vehicle trajectory; at the same time, update the existence probability distribution in the base map entry through the confidence probability of radar observation in the current frame and the predicted occupancy probability of historical frames.

[0058] It is understandable that the specific value of the first matching threshold can be selected according to the actual needs of those skilled in the art; for example, the value of the first matching threshold can be 0.8 to 0.9. As mentioned above, the base map entry needs to store the vehicle's position, existence probability distribution, and entry label. The entry label displays the vehicle's appearance information, which remains constant. Therefore, the update of the association between the target vehicle trajectory and the base map entry in the current frame mainly involves updating the vehicle's position and existence probability distribution. The update expression for the existence probability distribution of the base map entry is as follows: .

[0059] in, This represents the updated confidence probability of the target vehicle in the base map at time t. This represents the predicted probability of the existence of the base map entry corresponding to the currently matched target vehicle at time t in the historical frame. The confidence probability of radar observation in the current frame can be represented by the occupancy probability calculated from the vehicle position output by the radar in the current frame and the overlap with the BEV plane. The result is obtained by multiplication; λ represents the fusion weight of historical information and new observation information.

[0060] It's important to understand that in the above calculation formula, using the max function to take the maximum of the two probabilities ensures that the occupancy probability of the current target vehicle's base map entry will not unexpectedly decrease due to the fusion calculation. Because... It directly reflects the confidence probability of the current radar observation. Therefore, even if the historical prediction occupancy probability is low, it can ensure that the probability of the matching target vehicle entry in the base map remains at least the current observation value, thus stably reflecting the presence status of the target vehicle detected by the radar.

[0061] In this embodiment, when executing step S300, if the radar loses detection of the target vehicle during the continuous tracking phase, the corresponding base map entry for the target vehicle can be marked. To ensure that the marked base map entry can achieve trajectory continuity when the target vehicle is re-identified by the radar, a retention mechanism needs to be set to prevent the marked base map entry from being deleted in a short period of time. The specific implementation process of the retention mechanism is as follows: S310: Determine whether the marked base map entry meets the following conditions: (1) The base map entry is from the camera at the entrance location and the license plate confidence level is greater than the set confidence threshold.

[0062] (2) If the difference between the most recent recorded radar observation time and the current time of the marked base map entry does not exceed the set interval threshold, the marked base map entry shall be retained.

[0063] It is understandable that the specific values ​​of the confidence threshold and the time threshold can be determined by those skilled in the art based on their actual needs; for example, the confidence threshold can be 0.85 and the time threshold can be 300s.

[0064] S320: If a marked base map entry can satisfy any of the conditions given in step S310, the base map entry can be retained in the marked state until the target vehicle is determined to have left the site or the manual / system display indicates cancellation, at which point the marked base map entry will be deleted.

[0065] S330: For marked base map entries that do not meet all the conditions given in step S310, i.e., those that are only briefly detected by radar but not determined by the camera at the entrance position, if the marked base map entries are not further observed or matched within a set duration, the confidence probability of the marked base map entries is exponentially decayed until they are finally deleted.

[0066] It is understood that the specific value of the duration required in step S330 can be selected according to the actual needs of those skilled in the art; for example, the duration can be 60~120s. It should be noted that for vehicles that have not been detected for a long time in steps S320 and S330, their occupancy probability decays exponentially, but a certain threshold is retained to support rapid matching and recovery when the vehicle reappears; the corresponding base map entry is only deleted from the base map when it is clearly determined that the vehicle has left or the occupancy probability is below the lower limit of the threshold.

[0067] Compared to traditional methods, the base map mechanism in this application employs a probability retention and exponential decay strategy, maintaining a certain probability of presence for temporarily missing vehicles and preventing immediate deletion of entries due to short-term detection omissions. When a vehicle re-enters the visible area or is identified by the exit camera, the system can quickly match and restore tags based on the residual probability distribution and historical trajectory information. This design ensures that the system maintains continuous perception of intermittently observed targets even during long-term monitoring, providing fundamental support for all-day scene coverage.

[0068] In this embodiment, during step S300, to ensure rapid trajectory reconnection when the target vehicle is re-identified by the radar, it is necessary to predict the position of the marked base map entries during the target vehicle detection loss process, ensuring that the target vehicle corresponds to the base map entries in both spatial and temporal dimensions; the specific process is as follows: S340: For each target vehicle marked in the base map entry, use Kalman filtering to perform multi-hypothesis position prediction, and obtain the original predicted probability of the target vehicle at the corresponding grid i position in the base map. .

[0069] It's important to understand that Kalman filtering-based target vehicle position prediction primarily utilizes historical information from labeled base map entries. This position prediction can be viewed as an optimal extrapolation trajectory based on historical information. However, this trajectory carries Gaussian uncertainty, meaning the grid positions corresponding to this optimal extrapolation trajectory are represented as probabilities, which are the original prediction probabilities. .

[0070] S350: By combining the environmental constraints of the grid location of the target vehicle and the weight of the vehicle's direction of motion, the original prediction probability is corrected to obtain the single-vehicle prediction probability.

[0071] It is important to note that the location prediction in step S340 does not consider the environmental information during the loss of the target vehicle or its motion state before the loss. This introduces a degree of uncertainty into the location prediction obtained in step S340. Therefore, the original prediction probability can be adjusted by considering environmental constraints and the weights of the vehicle's motion direction. Corrections are made. Environmental constraints can be represented using a site semantic mask M, which takes values ​​of 0 and 1. 0 indicates that vehicles are not allowed at grid i (non-drivable area), and 1 indicates that vehicles are allowed at grid i (drivable area and parking area). Vehicle motion direction weights are also considered. This can represent the probability of a vehicle at grid i traveling in different directions. For example, if grid i is in a straight lane, the weight of the vehicle in the straight direction can tend to be 1, and the weight in other directions can tend to be 0. If grid i is at an intersection, the weights for the vehicle in the straight and turning directions can be set according to the actual needs of the environment. Based on the above description, the single-vehicle prediction probability of the target vehicle at grid i is... The expression is: .

[0072] S360: Based on the vehicle's direction of motion, obtain the candidate grid range of the target vehicle in the base map, and then normalize the single vehicle prediction probability according to the obtained grid range to obtain the single vehicle prediction probability distribution.

[0073] It's important to understand that the single-vehicle prediction probability distribution obtained in step S360 is the distribution of occupancy probabilities across different grids on a single trajectory, while the existence probability distribution mentioned earlier is the occupancy probability distribution over a multi-directional radiation range centered on the vehicle. The single-vehicle prediction probability distributions for different lost target vehicles are independent and do not undergo cross-vehicle normalization; the same grid can simultaneously accommodate the prediction probabilities of multiple vehicles. The expression for the normalized single-vehicle prediction probability distribution is as follows: .

[0074] in, Let represent the single-vehicle prediction probability of any grid j in the predicted position of the target vehicle, and m represent the total number of grids corresponding to the predicted position of the target vehicle. In simpler terms, the position prediction of the target vehicle based on Kalman filtering is represented as a trajectory segment consisting of m consecutive grids; grid i is one of the grids in this trajectory segment, and grid j is any grid in this trajectory segment.

[0075] S370: Fuse the obtained single-vehicle prediction probability distribution with the existence probability distribution of the historical base map entry and write it back to the historical base map entry or update the existence probability distribution of the historical base map entry.

[0076] It is understandable that the existence probability distribution can be viewed as the probability of a target vehicle forming a trajectory in any direction, centered on the target vehicle. That is, the existence probability distribution may include the single-vehicle prediction probability distribution obtained in step S360, or it may not. If it includes it, the single-vehicle prediction probability distribution can be directly written back into the existence probability distribution; if it does not include it, the existence probability distribution in the historical base map entries can be updated. For both cases, a priority retention strategy can be adopted to ensure that grids with historically high occupancy probabilities are not dragged down by predictions in a short period, while allowing predictions to gradually build confidence on grids that were not historically dominant. The priority retention strategy can be specifically expressed by the following expression: .

[0077] in, This represents the fusion probability of a historical base map entry at time t+1; β represents the historical probability of a historical basemap entry at time t in the current frame; hist The lower limit for historical retention is indicated, and the specific value can be set according to the actual needs of those skilled in the art, for example, the value range is [0.9, 1]; μ represents the weight for the fusion of old and new information, and the value can be 0.8~0.9.

[0078] Compared to traditional methods, this application proposes a probabilistic fusion mechanism that prioritizes historical data retention during the base map update process. This mechanism sets a lower limit coefficient β for historical data retention. hist By combining the old and new information fusion factors (the old and new information fusion weight μ), the high-confidence historical grid is protected while the new prediction results are gradually absorbed. This fusion method avoids the interference of short-term prediction fluctuations on long-term high-confidence trajectories, enabling the system to maintain both response sensitivity and decision stability in dynamic environments, and enhancing the reliability and consistency of the prediction update process.

[0079] In this embodiment, the execution of step S400 specifically includes the following process: S410: When the target vehicle is identified by radar again, all possible matching historical map entries can be searched on the base map.

[0080] It is important to know that the matching criteria include the historical probability P of the base map entries. i The overlap between the current observation trajectory and the base map entries in the BEV plane Model matching score score And license plate matching score plate scoreSince the historical map entries matched with the target vehicle are all predicted, the historical map entries corresponding to the target vehicle may differ significantly from the actual driving direction of the target vehicle. Therefore, the distance penalty term between the historical map entries and the actual position of the target vehicle is removed during the matching process.

[0081] S420: Based on the matching criteria in step S420, calculate the matching score S between the current observation entry and the historical base map entry. i The specific calculation formula is as follows: .

[0082] S430: If the calculated matching score is greater than the set second matching threshold, the current observation entry is associated with the corresponding historical base map entry. The specific value of the second matching threshold can be set according to the actual needs of those skilled in the art, for example, a value of 0.8 to 0.9.

[0083] It is important to note that since radar can only identify the rough outline of a target vehicle, there are multiple historical map entries with a high degree of matching with the current observation entry. This means that several historical map entries may have a matching score greater than the second matching threshold. In this case, the historical map entry with the highest matching score can be selected for association, and probability suppression operations should be performed on the remaining historical map entries to prevent duplicate label assignment or cross-over.

[0084] It is also important to note that if the difference in matching scores between historical base map entries that meet the second matching threshold is less than the set error threshold, the specific value of the error threshold can be set according to the actual needs of those skilled in the art, for example, a value of 0.05; for example, if there are three historical base map entries with matching scores of 0.91, 0.92, and 0.93 with the current observation entry, all of which meet the second matching threshold requirement and the error threshold requirement, then further decision-making can be made through time continuity constraints, that is, comparing the time difference between the most recent radar observation time of each historical base map entry and the current frame, and selecting the historical base map entry with the smallest time difference for association. That is, in the case of multiple historical base map entries, priority is given to selecting historical base map entries with shorter time intervals and closer locations for final matching, thereby improving the stability of matching.

[0085] In this embodiment, after matching the current observation entry with the historical base map entry, when performing trajectory continuation between the current observation entry and the historical base map entry, the continuity score C of the current trajectory corresponding to the current observation entry and the historical trajectory corresponding to the historical base map entry can be calculated to determine whether they belong to the same vehicle's movement process. That is, if the calculated continuity score C exceeds a set continuity threshold, it indicates that they belong to the same vehicle's movement process, and the current trajectory and historical trajectory can be stitched together and updated in the base map. The formula for calculating the continuity score C is as follows: .

[0086] Where Δt, Δd, and Δθ represent the time interval, spatial interval, and difference in direction of motion between the current trajectory and the historical trajectory, respectively; , as well as These represent the attenuation coefficients for the time interval, spatial interval, and direction of motion difference between the current trajectory and historical trajectories, respectively; w t w d and w θ All of these represent weighting coefficients.

[0087] It is important to know that the specific value of the continuity threshold used for determining the continuity score C can be selected according to the actual needs of those skilled in the art; for example, it can be 0.8 to 0.9. Through the above-mentioned continuation mechanism, when a vehicle is detected again, its historical label can be restored immediately, achieving seamless continuation of the vehicle in multi-radar collaborative scenarios and avoiding mismatches and chain breaks caused by differences in multi-source detection or label conflicts.

[0088] In this embodiment, the above-mentioned continuation of the current observation trajectory and historical base map entries is performed under radar detection scenarios. This may result in a discrepancy between the target vehicle identified by the camera at the exit position and the base map entry assigned by the radar when the target vehicle leaves the semi-enclosed space. This could cause the base map entry corresponding to a vehicle still in the semi-enclosed space to be matched to a vehicle at the exit position, leading to inaccurate vehicle trajectories within the semi-enclosed space. Therefore, in the technical solution of this application, a swap verification mechanism can be set at the exit position of the semi-enclosed space, specifically including the following process: When the target vehicle identified by the camera at the exit location is inconsistent with the base map entry assigned by the radar, a swap check is triggered. At this time, other marked base map entries can be searched within a set short time window to replace the base map entry assigned to the target vehicle by the radar, and a matching score is calculated. If there is a marked base map entry that improves the matching score, the marked base map entry is swapped with the base map entry assigned to the target vehicle by the radar, and the reason and evidence for the swap are recorded.

[0089] Compared to traditional methods, the technical solution of this application introduces exit interchange verification logic, which can automatically identify and correct label assignment errors when license plate mismatch or target overlap occurs. This mechanism effectively prevents problems such as multi-target mismatch and cross-pairing in complex traffic scenarios, ensuring label consistency and trajectory stability of the system under high-density traffic conditions.

[0090] The basic principles, main features, and advantages of this application have been described above. Those skilled in the art should understand that this application is not limited to the above embodiments. The embodiments and descriptions in the specification are merely the principles of this application. Various changes and modifications can be made to this application without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claims. The scope of protection claimed by this application is defined by the appended claims and their equivalents.

Claims

1. A method for reconnecting tags after vehicle identification failure in a semi-enclosed space, characterized in that, Includes the following steps: S100: At the entrance of the semi-enclosed space, vehicles are jointly calibrated using cameras and radar, so that each vehicle entering the semi-enclosed space is assigned an entry label containing its complete information. S200: Construct a planar gridded base map for a semi-enclosed space and establish environmental constraints; based on the position of the target vehicle in the base map, construct and dynamically maintain the probability distribution of the target vehicle's presence in the base map to form base map entries through radar identification results of the target vehicle; S300: If the radar loses detection of the target vehicle during the continuous tracking phase, the position prediction of the target vehicle is generated through Kalman filtering; The location prediction is corrected based on the vehicle's direction of travel and environmental attributes, and then merged with historical base map entries. S400: When the target vehicle is identified by radar again, the current observation entry is connected to the historical base map entry with the highest matching degree for trajectory continuation; if multiple historical base map entries meet the matching degree requirements, priority is determined based on historical trajectory, direction vector and base map probability distribution.

2. The method for reconnecting tags after vehicle identification failure in a semi-enclosed space as described in claim 1, characterized in that, In step S100, the camera is responsible for providing the vehicle's appearance information, including license plate number, precise vehicle model, color, and structural features; The radar is responsible for providing the vehicle's spatial location information and a rough model number; Unify the target bounding boxes of the vehicle in the camera coordinate system and the radar coordinate system, and calculate the IOU value; If the calculated IOU value is greater than the set threshold and the vehicle category meets the fuzzy matching conditions, it is determined that the camera and radar recognition results correspond to the same vehicle, and the vehicle is assigned an entry label containing its complete information.

3. The method for reconnecting tags after vehicle identification failure in a semi-enclosed space as described in claim 1, characterized in that, In step S200, the base map is divided into spatial semantic partitions according to environmental functions, and different spatial semantic partitions are assigned different probability weights β; Calculate the probability of the target vehicle occupying any grid position in the base map. Summarize the occupancy probability of all grids This yields the probability distribution of the target vehicle's presence in the base map. Occupancy probability The calculation formula is: ； ； in, Let represent the occupancy probability of the target vehicle under the identification of the i-th radar, k represent the total number of radars capable of identifying the target vehicle within the semi-enclosed space, (x0, y0) represent the coordinates of the center of the target vehicle in the base map coordinate system, and (x, y) represent the center coordinates of any grid in the base map coordinate system. σ p This represents the diffusion parameter.

4. The method for reconnecting tags after vehicle identification failure in a semi-enclosed space as described in claim 1, characterized in that, In step S200, the base map retains the location, existence probability distribution, and entry label of each vehicle to form base map entries; During the phase of continuous tracking of the target vehicle within a semi-enclosed space, the updating of base map entries includes the following process: The target vehicle trajectory output by each radar frame is matched with all base map entries based on trajectory overlap, spacing difference, probability distribution, and appearance similarity score. If the calculated matching score is greater than the set first matching threshold, the target vehicle trajectory of the current frame is determined to be associated with the corresponding base map entry, and then the position of the base map entry is updated by matching the target vehicle trajectory. Simultaneously, the existence probability distribution in the base map entries is updated using the confidence probability of radar observations in the current frame and the predicted occupancy probability of historical frames.

5. The method for reconnecting tags after vehicle identification failure in a semi-enclosed space as described in claim 4, characterized in that, Matching score S i The calculation formula is as follows: ； The update expression for the probability distribution of the existence of base map entries is as follows: ； in, c represents the degree of overlap between the target vehicle trajectory and the base map entry from the BEV perspective. T c represents the coordinates of the vehicle's center point on the target vehicle's trajectory. i σ represents the coordinates of the vehicle center point corresponding to the i-th base map entry. d P represents the normalization parameter. i The type represents the historical probability corresponding to the i-th base map entry. score and plate score These represent the vehicle model and license plate matching scores, respectively, with w1 to w5 representing the corresponding weighting coefficients. This represents the updated confidence probability of the target vehicle in the base map at time t. This represents the predicted probability of the existence of the base map entry corresponding to the currently matched target vehicle at time t in the historical frame. λ represents the confidence probability of radar observations in the current frame, and λ represents the fusion weight of historical information and new observation information.

6. The method for reconnecting tags after vehicle identification failure in a semi-enclosed space as described in any one of claims 1-5, characterized in that, Step S300 includes the following specific processes: If the radar loses detection of the target vehicle during the continuous tracking phase, the corresponding base map entry for that target vehicle will be marked. For each target vehicle marked in the base map entry, Kalman filtering is used to perform multiple hypothesis position prediction to obtain the original predicted probability of the target vehicle at the corresponding grid position in the base map. By combining the environmental constraints of the grid location of the target vehicle and the weight of the vehicle's direction of motion, the original prediction probability is corrected to obtain the single-vehicle prediction probability. Based on the vehicle's direction of motion, the candidate grid range of the target vehicle is obtained from the base map. Then, the prediction probability of a single vehicle is normalized according to the obtained grid range to obtain the prediction probability distribution of a single vehicle. The obtained bicycle prediction probability distribution is merged with the existence probability distribution of the historical base map entry, and then written back to the historical base map entry or the existence probability distribution of the historical base map entry is updated.

7. The method for reconnecting tags after vehicle identification failure in a semi-enclosed space as described in claim 6, characterized in that, A mechanism is in place to retain base map entries that are lost during radar detection of target vehicles. This mechanism includes the following: If the marked base map entry originates from a camera at the entrance location and the license plate confidence level is greater than the set confidence threshold, or if the difference between the most recent recorded radar observation time and the current time does not exceed the set interval threshold, the marked base map entry will be retained; when it is determined that the target vehicle has left the site or the manual / system display is canceled, the marked base map entry will be deleted. For marked base map entries that are only briefly detected by radar but not identified by cameras at the entrance location, if the marked base map entries are not further observed or matched within a set duration, the confidence probability of the marked base map entries is exponentially decayed until they are eventually deleted.

8. The method for reconnecting tags after vehicle identification failure in a semi-enclosed space as described in claim 4, characterized in that, Step S400 includes the following process: When the target vehicle is identified by the radar again, the matching score between the current observation entry and the historical base map entry is calculated. If the calculated matching score is greater than the set second matching threshold, the current observation entry is associated with the corresponding historical base map entry for matching. If multiple historical base map entries meet the second matching threshold, select the historical base map entry with the highest matching score for association, and perform probability suppression operation on the remaining historical base map entries; If the matching score difference of historical base map entries that meet the second matching threshold is less than the set error threshold, compare the time difference between the most recent radar observation time of each historical base map entry and the current frame, and select the historical base map entry with the smallest time difference for association.

9. The method for reconnecting tags after vehicle identification failure in a semi-enclosed space as described in claim 1, characterized in that, When connecting the current observation entry with the historical base map entry, calculate the continuity score C between the current trajectory corresponding to the current observation entry and the historical trajectory corresponding to the historical base map entry; if the calculated continuity score C exceeds the set continuity threshold, the current trajectory and the historical trajectory are spliced ​​together and updated in the base map. The formula for calculating the continuity score C is as follows: ； Where Δt, Δd, and Δθ represent the time interval, spatial interval, and difference in direction of motion between the current trajectory and the historical trajectory, respectively; , as well as These represent the attenuation coefficients for the time interval, spatial interval, and direction of motion difference between the current trajectory and historical trajectories, respectively; w t w d and w θ All of these represent weighting coefficients.

10. The method for reconnecting tags after vehicle identification failure in a semi-enclosed space as described in claim 4, characterized in that, When the target vehicle leaves the semi-enclosed space, the target vehicle identified by the camera based on the exit position is inconsistent with the base map entry assigned by the radar, triggering a cross-check, which includes the following process: Within a set short time window, other marked base map entries are retrieved from the base map to replace the base map entry assigned to the target vehicle by the radar, and a matching score is calculated. If there is a marked map entry that improves the matching score, swap the marked map entry with the map entry assigned to the target vehicle by the radar, and record the reason and evidence for the swap.

Images