Vehicle long-distance track restoration method, system and equipment

Abstract

The invention relates to the technical field of intelligent traffic, and discloses a vehicle long-distance track restoration method, system and device. The method comprises the following steps: constructing a model training set based on vehicle tracking data, and training a path decision model based on the model training set; obtaining a license plate recognition result of the target vehicle in a preset time period, and determining all intersection positions through which the target vehicle passes based on the license plate recognition result to obtain a target intersection sequence; and based on the target intersection sequence and the trained path decision model, restoring to obtain a complete track of the target vehicle in a preset time period. Through the scheme of the invention, the problems of track fragmentation and incompleteness caused by only depending on a vehicle tracking algorithm are avoided, and the restoration effect of the long-distance continuous track of the vehicle is effectively improved, so that the active supervision requirement of the target vehicle in an intelligent network connection scene is met.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation technology, and in particular to a method, system and device for restoring a long-distance trajectory of a vehicle. Background Art

[0002] The full sample data of vehicle driving trajectories can support the acquisition and analysis of any traffic flow parameters on the road, and meet the active supervision needs of urban traffic management. At present, it is possible to use sensing devices such as laser radars and cameras installed on the roadside of intelligent connected roads to obtain the trajectories of all traffic participants. Furthermore, the existing technology uses a vehicle trajectory tracking algorithm to track the full sample data of vehicle driving trajectories, but this method is prone to problems such as tracking loss and vehicle switching, and the effect is poor when obtaining long-distance vehicle trajectories, resulting in incomplete vehicle trajectories, which cannot meet the active supervision needs of urban traffic management.

[0003] Therefore, it is urgent to propose a vehicle long-distance trajectory restoration method to overcome the above defects. Summary of the invention

[0004] In view of this, the present invention proposes a method, system and device for restoring a long-distance vehicle trajectory.

[0005] Based on the above purpose, one aspect of an embodiment of the present invention provides a method for restoring a long-distance vehicle trajectory, which specifically includes the following steps: Building a model training set based on vehicle tracking data, and training a path decision model based on the model training set; Obtaining a license plate recognition result of a target vehicle within a preset time period, and based on the license plate recognition result, determining positions of all intersections passed by the target vehicle to obtain a target intersection sequence; Based on the target intersection sequence and the trained path decision model, the complete trajectory of the target vehicle within the preset time period is restored.

[0006] In some embodiments, the step of restoring the complete trajectory of the target vehicle within the preset time period based on the target intersection sequence and the trained path decision model includes: Based on the target intersection sequence, two intersection positions at adjacent moments are taken as a target position group; Based on each of the target position groups, generating a corresponding target candidate path set; Determine the target path corresponding to each target position group based on each target candidate path set and the trained path decision model; Based on all the target paths, the complete trajectory of the target vehicle within the preset time period is restored.

[0007] In some implementations, the step of determining the target path corresponding to each target location group using each of the target candidate path sets and the trained path decision model includes: Based on each target candidate path in the target candidate path set, obtaining a path decision index of the target candidate path to construct a corresponding first index vector; Using each of the first indicator vectors as an input of the trained path decision model to obtain a model prediction result corresponding to each of the target candidate paths; Based on all model prediction results, the target path is determined.

[0008] In some implementations, the step of training the path decision model based on the model training set includes: Extract training set data from the model training set, and obtain a first position group corresponding to the training set data; Based on the first position group, generating a first candidate path set corresponding thereto; Based on the first candidate path set, obtaining a second indicator vector corresponding to each first candidate path, and obtaining a corresponding loss function of the path decision model according to each second indicator vector; The path decision model is trained based on each of the loss functions.

[0009] In some embodiments, the step of obtaining a license plate recognition result of a target vehicle within a preset time period, and determining the positions of all intersections passed by the target vehicle based on the license plate recognition result to obtain a target intersection sequence includes: Determining the license plate recognition result based on the image data of the target vehicle within a preset time period; Converting the two-dimensional pixel coordinates corresponding to the license plate recognition result into three-dimensional world coordinates; Based on a map matching algorithm and the three-dimensional world coordinates, determining the locations of all the intersections through which the target vehicle passes; All the intersection positions are sorted in chronological order to obtain the target intersection sequence.

[0010] In some embodiments, the step of constructing a model training set based on vehicle tracking data includes: Based on the vehicle tracking data, determining a first intersection sequence; Extracting a first intersection position from the first intersection sequence according to a preset rule to generate a first sequence; Based on the first sequence and the first intersection position corresponding thereto, obtaining training set data; The model training set is determined based on all the training set data obtained through multiple extractions.

[0011] In some embodiments, the step of determining a first intersection sequence based on the vehicle tracking data includes: Collecting the vehicle tracking data, and screening all trajectory sets that continuously pass through at least a preset number of intersections; Based on all the track sets obtained by screening, determining the intersection position corresponding to each of the intersections; All the corresponding intersection positions are sorted in chronological order to obtain the first intersection sequence.

[0012] In some embodiments, the vehicle tracking data is obtained by: Based on the roadside sensing device, identifying any vehicle on the road as a first vehicle; Based on a preset tracking algorithm, the vehicle tracking data corresponding to the first vehicle is obtained.

[0013] Another aspect of the embodiment of the present invention further provides a vehicle long-distance trajectory restoration system, comprising: A model training unit configured to construct a model training set based on the vehicle tracking data, and train a path decision model based on the model training set; A license plate recognition unit configured to obtain a license plate recognition result of a target vehicle within a preset time period, and based on the license plate recognition result, determine the positions of all intersections passed by the target vehicle to obtain a target intersection sequence; The trajectory restoration unit is configured to restore the complete trajectory of the target vehicle within the preset time period based on the target intersection sequence and the trained path decision model.

[0014] According to another aspect of an embodiment of the present invention, a computer device is provided, comprising: at least one processor; and a memory, wherein the memory stores a computer program executable on the processor, and the computer program implements the steps of the above method when executed by the processor.

[0015] The present invention has at least the following beneficial technical effects: the vehicle long-distance trajectory restoration method of the present invention, on the one hand, uses a vehicle tracking algorithm to obtain vehicle tracking data, and extracts path decision indicators therefrom to determine the vehicle's driving preference information as the basis for training a path decision model; on the other hand, it makes full use of the various license plate recognition results identified by the roadside sensing device in combination with the roadside decision model to complete the path between adjacent license plate recognition results, and finally realizes the long-distance trajectory restoration of the target vehicle, avoiding the problem of trajectory fragmentation and incompleteness caused by relying solely on the vehicle tracking algorithm, and can realize the restoration of missing trajectories, effectively improving the restoration effect of the vehicle's long-distance continuous trajectory, thereby meeting the active supervision needs of target vehicles in intelligent network scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other embodiments can be obtained based on these drawings without paying creative work.

[0017] Figure 1 A block diagram of an embodiment of a vehicle long-distance trajectory restoration method provided by the present invention; Figure 2 A schematic diagram of an embodiment of a model training set construction process provided by the present invention; Figure 3 A schematic diagram of an embodiment of the execution process of the vehicle long-distance trajectory restoration method provided by the present invention; Figure 4 A schematic diagram of an embodiment of a vehicle long-distance trajectory restoration system provided by the present invention; Figure 5 A schematic structural diagram of an embodiment of a computer device provided by the present invention. DETAILED DESCRIPTION

[0018] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the embodiments of the present invention are further described in detail below in combination with specific embodiments and with reference to the accompanying drawings.

[0019] It should be noted that all expressions using "first" and "second" in the embodiments of the present invention are for distinguishing two non-identical entities with the same name or non-identical parameters. It can be seen that "first" and "second" are only for the convenience of expression and should not be understood as limitations on the embodiments of the present invention. The subsequent embodiments will not explain this one by one.

[0020] Although it is possible to use roadside sensing devices such as lidar and cameras installed on the roadside of intelligent connected roads to obtain the trajectory of the vehicle, the trajectory obtained is fragmented. Furthermore, in the existing technology, the complete trajectory of the vehicle is predicted by a vehicle trajectory tracking algorithm. However, in practical applications, due to various factors, such as instability, occlusion, overlap of target detection, and limitations of the tracking algorithm itself, problems such as ID (identification) switching and ID loss may occur during long-distance trajectory tracking. ID switching refers to the sudden change of a vehicle's ID during the tracking process, which is usually caused by the tracking algorithm mistakenly confusing one vehicle with another. ID loss refers to the tracking algorithm losing track of a vehicle at a certain moment and can no longer assign a valid ID to it. This results in poor results in obtaining long-distance vehicle trajectories and does not completely solve the problem of trajectory fragmentation.

[0021] In order to solve the above defects, the first aspect of the embodiment of the present invention provides a method for restoring a long-distance vehicle trajectory. Figure 1 As shown, it includes the following steps: Step S100, constructing a model training set based on vehicle tracking data, and training a path decision model based on the model training set; Step S200, obtaining the license plate recognition result of the target vehicle within a preset time period, and based on the license plate recognition result, determining the positions of all intersections passed by the target vehicle to obtain a target intersection sequence; Step S300, based on the target intersection sequence and the trained path decision model, restore the complete trajectory of the target vehicle within a preset time period.

[0022] In some embodiments, the vehicle tracking data is obtained based on the roadside sensing device and the vehicle tracking algorithm, and the trajectory is fragmented. The vehicle tracking data is processed based on screening and construction to obtain a model training set. The model training set is further used to construct candidate paths, and the candidate paths and the real paths in the model training set are used together for the training of the path decision model. The path decision model is a classification model used to determine whether a candidate path is the real path of the vehicle. The target vehicle is a vehicle that needs to restore a continuous trajectory. The license plate recognition result is obtained by identifying the license plate of the target vehicle through a camera installed at the intersection, and the license plate recognition results of the target vehicle appearing at different intersections at different times within a preset time period are obtained through cameras at multiple intersections. The license plate recognition results are spatially matched to realize the conversion of the two-dimensional pixel coordinates in the license plate recognition results to the three-dimensional coordinates in the world coordinate system, and the matching intersection position is found using the three-dimensional coordinates to accurately locate the intersection through which the target vehicle passes. And all the matched intersection positions are sorted in the chronological order of the license plate recognition results to determine the target intersection sequence. The target intersection sequence is further used to construct multiple target candidate path sets, and the trained path decision model is used to select the optimal path in each target candidate path set, and all the selected optimal paths are spliced ​​at once to finally restore the complete trajectory corresponding to the target vehicle. Among them, the preset time period is the time range for the target vehicle to restore the trajectory, which can be set according to actual business needs and is not specifically limited here.

[0023] The vehicle long-distance trajectory restoration method of the present invention can accurately restore the complete trajectory of the target vehicle from the starting point to the end point by using the vehicle tracking data and the license plate recognition results for the training and application of the path decision model respectively, avoiding the problems of ID switching and ID loss caused by relying solely on the vehicle tracking algorithm, and effectively improving the restoration effect of the vehicle's long-distance continuous trajectory, thereby meeting the active supervision needs of target vehicles in intelligent network scenarios.

[0024] In some embodiments, based on the target intersection sequence, the trained path decision model and the target road network node set, the step of restoring the complete trajectory of the target vehicle within a preset time period includes: based on the target intersection sequence, taking the two intersection positions at adjacent moments as a target position group; based on each target position group, generating its corresponding target candidate path set; based on each target candidate path set and the trained path decision model, determining the target path corresponding to each target position group; based on all target paths, restoring the complete trajectory of the target vehicle within the preset time period.

[0025] In some embodiments, the intersection position sequence (i.e., the target intersection sequence) where the license plate of the target vehicle is continuously captured is completed in chronological order. That is, two intersection positions at adjacent moments in the target intersection sequence are obtained in chronological order as a target position group, for example, the intersection position at the first moment and the intersection position at the second moment are a target position group, and the intersection position at the second moment and the intersection position at the third moment are also a target position group. The KSP (K-Shortest Paths, K shortest paths) algorithm is used to generate multiple candidate paths between the two intersection positions corresponding to the target position group to form a target candidate path set. The selection probability of each target candidate path in the target candidate path set is obtained by the trained path decision model, and the target candidate path with the highest probability is selected as the target path, and the target path is used as the optimal trajectory between the two intersection positions corresponding to the target position group for restoration. And continue to obtain the next target position group in the target intersection sequence in chronological order, and repeat the above restoration process until the optimal trajectory between all adjacent intersection positions in the target intersection sequence is found, thereby restoring the complete long-distance trajectory corresponding to the target vehicle.

[0026] In some embodiments, multiple candidate paths between two intersection positions corresponding to the target position group are generated based on the KSP algorithm, that is, a target candidate path set is obtained. According to the time sequence, the two intersection positions corresponding to the target position group are respectively used as the starting intersection position and the ending intersection position. The step of generating the target candidate path set by the KSP algorithm includes: (1) Use Dijkstra's algorithm to find the shortest path between the starting intersection and the ending intersection. ; (2) Using the idea of ​​the recursive deviation path algorithm, find the previous Specifically, in solving the path When All intersection positions except the ending intersection position are considered as deviated nodes. The shortest path from each deviated node to the ending intersection position is calculated and then appended to the path from the starting intersection position to On the path that deviates from the node, the first Shortest Path ,forward The shortest path is the set of target candidate paths corresponding to the starting intersection position and the ending intersection position.

[0027] The vehicle long-distance trajectory restoration method of the present invention dynamically generates a target candidate path set based on the two intersection positions corresponding to the target position group, ensuring that the path selection is consistent with the actual road network structure, and by gradually processing the adjacent intersection positions in the target intersection sequence, the long-distance trajectory restoration problem is decomposed into multiple small-range path selection problems, thereby improving the restoration accuracy, effectively splicing the continuous trajectory of the target vehicle, and avoiding the incomplete trajectory problem in the traditional method.

[0028] In some embodiments, based on the trained path decision model, each target candidate path set and the trained path decision model, the step of determining the target path corresponding to each target location group includes: based on each target candidate path in the target candidate path set, obtaining the path decision index of the target candidate path to construct a corresponding first index vector; using each first index vector as an input of the trained path decision model to obtain the model prediction result corresponding to each target candidate path; and determining the target path based on all model prediction results.

[0029] In some embodiments, the path decision index includes the selected path length, driving time, number of straight trips, number of left turns, number of right turns, number of U-turns, path selection preference, and number of intersections. After all path decision indexes are standardized, they are used to construct a multidimensional index vector including all path decision indexes, i.e., a first index vector.

[0030] In some implementations, the standardization process of each path decision indicator is described below.

[0031] Select path length The corresponding formula for the standardized processing is as follows: ; in, Represented as the i-th target candidate path, is the target candidate path The path length, It is the median of the path lengths of all target candidate paths in the target candidate path set.

[0032] Driving time The corresponding formula for the standardized processing is as follows: ; in, is the target candidate path The driving time, It is the average speed of vehicles on the road network obtained by statistics.

[0033] Number of left turns The corresponding formula for the standardized processing is as follows: ; in, is the target candidate path The number of left turns, is the maximum number of left turns among all target candidate paths in the target candidate path set, It is the minimum number of left turns among all target candidate paths in the target candidate path set.

[0034] Number of right turns The corresponding formula for the standardized processing is as follows: ; in, is the target candidate path The number of right turns, is the maximum number of right turns among all target candidate paths in the target candidate path set, It is the minimum number of right turns among all target candidate paths in the target candidate path set.

[0035] Number of straight lines The corresponding formula for the standardized processing is as follows: ; in, is the target candidate path The number of straight lines, is the maximum number of straight-through times among all target candidate paths in the target candidate path set, It is the minimum number of straight lines among all target candidate paths in the target candidate path set.

[0036] Number of U-turns The corresponding formula for the standardized processing is as follows: ; in, is the target candidate path The number of U-turns, is the maximum number of U-turns among all target candidate paths in the target candidate path set, It is the minimum number of U-turns among all target candidate paths in the target candidate path set.

[0037] The path selection process of vehicles between intersections can be regarded as a Markov process. The state of the vehicle at the intersection is the state S of the Markov process. The process of vehicles moving between intersections can be regarded as the process of state transfer. , Indicates that the vehicle is at an intersection status, Indicates the road network node number corresponding to the intersection position. The probability of state transition between intersection positions can be represented by the state transition matrix P. In the process of path selection, a state transition can be regarded as a single turning process of the vehicle at the intersection position. Indicates that the vehicle reaches the intersection position from intersection position 1 through one turn. For positions at adjacent intersections, the state transition probability is the probability of the vehicle turning, and vehicle turning includes turning left, turning right, going straight, and turning around. For positions at non-adjacent intersections, the vehicle cannot complete the turn in one turn, and its state transition probability is 0. When the state of the intersection position is not transferred, it means that the vehicle has not moved, and the state transition probability is also 0. The situation of no state transfer is not considered in the present invention. By counting the probabilities of turning left, turning right, going straight, and turning around for vehicles at each intersection in the historical data, the state transition matrix P can be obtained: .

[0038] The target intersection sequence is For example, It represents the road network node number corresponding to the nth intersection position in the target intersection sequence, and its path selection preference PR can be expressed as: .

[0039] For target candidate paths , and its corresponding path selection preference The corresponding formula for the standardized processing is as follows: ; in, is the target candidate path , k is the total number of target candidate paths in the target candidate path set.

[0040] Number of intersections The corresponding formula for the standardized processing is as follows: ; in, is the target candidate path The number of intersections included, is the maximum number of intersections among all target candidate paths in the target candidate path set, It is the minimum number of intersections among all target candidate paths in the target candidate path set.

[0041] After the above standardization process, a multidimensional first index vector including multiple path decision indicators is obtained: , Considered as target candidate path of dimensional indicator vector, t corresponds to the number of path decision indicators. The path decision model can be an LR (Logistics Regression) path decision model. By inputting the first indicator vector into the trained path decision model , the output is the target candidate path The prediction formula is as follows: ; in, The target candidate path predicted by the model The selection probability ranges from [0,1], , is the model parameter.

[0042] Based on the above processing, the selection probability of each target candidate path in the target candidate path set is obtained, that is, the model prediction result corresponding to the target candidate path. The target candidate path with the largest probability is selected as the target path, that is, the optimal trajectory corresponding to the two adjacent intersection positions in the target intersection sequence is obtained.

[0043] The vehicle long-distance trajectory restoration method of the present invention predicts candidate paths through a trained path decision model, can perform a comprehensive evaluation based on multi-dimensional path decision indicators, and can quickly process multiple candidate paths and give prediction results using the trained path decision model. It is suitable for real-time path decision-making in large-scale traffic scenarios to quickly and accurately restore the long-distance complete trajectory of the target vehicle, providing important support for the active supervision and optimization decision-making of intelligent transportation systems.

[0044] In some embodiments, the step of training a path decision model based on a model training set includes: extracting training set data from the model training set to obtain a first position group corresponding to the training set data; based on the first position group, generating a corresponding first candidate path set; based on the first candidate path set, obtaining a second indicator vector corresponding to each first candidate path, and obtaining a corresponding loss function of the path decision model according to each second indicator vector; and training the path decision model based on each loss function.

[0045] In some implementations, a training set data is randomly selected from the model training set, and the first intersection position recorded in the training set data is obtained as the first position group. Multiple candidate paths of the first position group are generated based on the KSP algorithm, that is, a first candidate path set is obtained.

[0046] In some embodiments, the second indicator vector corresponding to the first candidate path is obtained by obtaining the path decision index of the first candidate path. The process of obtaining the path decision index of the first candidate path is the same as the process of obtaining the path decision index of the target candidate path, and the process of constructing the second indicator vector corresponding to the first candidate path is the same as the process of constructing the first indicator vector corresponding to the target candidate path. Selecting cross entropy as the loss function of the path decision model The loss function The formula is as follows: ; in, The first candidate path The prediction results, The first candidate path The real choice situation, ,like 1 means the first candidate path is selected , otherwise it means not selected.

[0047] The model training set is used to train the model parameters of the path decision model. The smaller the loss result obtained by the loss function during the training process, the closer the prediction result of the path decision model is to the actual choice of the path. , by using the gradient descent algorithm to continuously adjust the model parameters and , to minimize the loss function . Continuously extract training set data from the model training set and repeat the calculation of the above loss function. When the first candidate path The prediction results The actual selection situation When the training is close enough or reaches the preset number of iterations (such as 3000 times), the preset training termination condition is met, and the model training is terminated to obtain the trained path decision model.

[0048] The vehicle long-distance trajectory restoration method of the present invention uses real path data to construct a model training set to ensure that the path decision model is close to the actual traffic conditions. The generation method of the first candidate path set can cover more possible path conditions, improve the model's ability to cope with complex road network environments, and help the path decision model to master the global rules of path decision by learning different path decision characteristics in the model training set, avoid being limited to a single scenario or specific conditions, improve the reliability, accuracy and adaptability of the path decision model, and lay an important foundation for realizing vehicle trajectory restoration in a complex road network environment.

[0049] In some embodiments, the steps of obtaining a license plate recognition result of a target vehicle within a preset time period, and determining the positions of all intersections passed by the target vehicle based on the license plate recognition result to obtain a target intersection sequence include: determining the license plate recognition result based on image data of the target vehicle within the preset time period; converting the two-dimensional pixel coordinates corresponding to the license plate recognition result into three-dimensional world coordinates; determining the positions of all intersections passed by the target vehicle based on a map matching algorithm and the three-dimensional world coordinates; and sorting all intersection positions in chronological order to obtain a target intersection sequence.

[0050] In some embodiments, the roadside sensing device includes a roadside camera, which is mostly distributed at the intersection, and the license plate recognition result comes from the image data of the roadside camera. Due to the limited coverage of the camera, only a small area near the intersection can be recognized, so it is impossible to directly obtain a continuous trajectory sequence through the license plate recognition result. The license plate recognition result is spatially matched, and the spatial matching mainly converts the two-dimensional pixel coordinates corresponding to the license plate position in the license plate recognition result into three-dimensional coordinates of the world coordinate system. And the three-dimensional coordinates of the converted license plate position are used through a map matching algorithm to find the intersection position where the target vehicle passes. In the road network model, the intersection usually corresponds to the road network node, and each intersection has a unique road network node number to identify its position in the road network. The road network node numbers corresponding to all the intersection positions found are sorted in the chronological order of the license plate recognition results, so as to obtain the intersection position sequence where the license plate of the target vehicle is continuously captured, that is, the target intersection sequence.

[0051] In some implementations, the process of spatial matching of license plate recognition results is described: (1) Convert the two-dimensional pixel coordinates to camera coordinates. Assume that the two-dimensional pixel coordinates of the license plate position in the license plate recognition result are expressed as , the camera coordinates are expressed as , the conversion relationship is as follows, where is the focal length of the camera, is the optical center coordinate of the camera, is the height of the camera from the ground: ; ; ; (2) Convert the camera coordinates into the three-dimensional coordinates of the world coordinate system. Assume that the three-dimensional coordinates of the world coordinate system are expressed as , the conversion relationship is as follows, where R is the rotation matrix of the camera and T is the translation vector. For the camera coordinates: .

[0052] The vehicle long-distance trajectory restoration method of the present invention uniquely identifies the vehicle through the license plate information, ensures that subsequent trajectory restoration can be processed for the same vehicle, avoids the ID switching and ID loss problems caused by relying solely on the vehicle tracking algorithm, and comprehensively utilizes the license plate recognition result and the space matching algorithm to realize the conversion from pixel coordinates to world coordinates to obtain the target intersection sequence, providing effective input data for the path decision model, which is conducive to the optimization of the target path prediction result.

[0053] In some embodiments, the step of constructing a model training set based on tracking data results includes: determining a first intersection sequence based on vehicle tracking data; extracting a first intersection position from the first intersection sequence according to a preset rule to generate a first sequence; obtaining training set data based on the first sequence and its corresponding first intersection position; and determining a model training set based on all training set data obtained through multiple extractions.

[0054] In some implementations, the first intersection position is extracted from the first intersection sequence according to a preset rule to generate a first sequence, and then a training set data is constructed according to the first sequence. The preset rule is as follows: (1) Starting with the sampling number of 0, the intersection positions are continuously sampled in the first intersection sequence according to the sampling number; (2) Taking the extracted road network nodes as the first sequence, and constructing a training set of data based on the first sequence and the first intersection positions before and after the first sequence; (3) Add 1 to the number of extractions, and return to the step (1) to continuously extract intersection positions in the first intersection sequence according to the number of extractions, until the number of extractions is greater than the preset number.

[0055] The preset number here can be set according to actual business needs, and can be set to any value that does not exceed the total number of intersection positions in the first intersection sequence. It should be noted that the above step (1) can be drawn multiple times, and the number of draws can be set according to actual business needs. In step (1), continuous drawing means that when the number of draws is greater than 1, the multiple intersection positions drawn must be adjacent intersection positions in the first intersection sequence.

[0056] In some embodiments, 0 to a preset number of intersection positions are continuously extracted from random positions in the first intersection sequence, and the extracted intersection position sequence (i.e., the first sequence) and the intersection positions before and after the extracted positions (i.e., the first intersection position) are retained. This process is repeated to construct a model training set R, in which a single training set data content in the model data set includes the first sequence and its corresponding first intersection position.

[0057] In one example, the model data set can be expressed as R= ∪ ∪…∪ , where h is the preset number, It represents the data set corresponding to the training set data of 0 road network nodes extracted from the first road network node set, and so on. Figure 2 For example, if the preset number h is 5, the first intersection sequence is ,in Indicates the road network node number corresponding to the nth intersection position in the first intersection sequence, and explains the process of building the model training set: Figure 2 Scene 1 in the middle shows that when the number of extractions is 0, Extract 0 intersection positions from r as the first sequence, and keep the extracted first sequence and its first intersection position in turn to form , Represents a training set data, where Indicates the first intersection location, Indicates the intersection position before the extraction position, Indicates the intersection position after the extraction position, () indicates that the first sequence is empty, and so on; Figure 2 Scene 2 in the middle shows that when the number of extractions is 1, In the process of extracting one intersection position from r as the first sequence, the extracted first sequence and its first intersection position are retained in turn to form ; Figure 2 Scene 3 shows that when the number of draws is 2, In the process of extracting two intersection positions from r as the first sequence, the extracted first sequence and its first intersection position are retained in turn to form ; Increase the number of extractions in sequence, and continue to extract 3 to h intersections from r according to the above process to obtain to ,like Figure 2 As shown in Scenes 4 to 5. The model training set R obtained based on the first intersection sequence r is obtained by taking the union of all the data sets obtained above. According to the above process, the corresponding model training sets are continuously obtained for the first intersection sequences obtained by other vehicles until all vehicles generate corresponding model training sets, which are then merged to form the final model training set.

[0058] The vehicle long-distance trajectory restoration method of the present invention extracts different numbers of intersection positions to generate training set data, which helps the model adapt to various path decision scenarios. Through multi-level extraction, the model can take into account both short-path prediction and long-path restoration. The structure of the training set data provides favorable support for quickly obtaining the first and last intersection positions and the actual selection of candidate paths during the subsequent path decision model training process.

[0059] In some embodiments, the step of determining a first intersection sequence based on vehicle tracking data includes: collecting vehicle tracking data and screening all trajectory sets that continuously pass through at least a preset number of intersections; based on all the trajectory sets obtained by screening, determining the intersection position corresponding to each intersection therein; and sorting all the corresponding intersection positions in chronological order to obtain a first intersection sequence.

[0060] In some embodiments, the step of collecting vehicle tracking data includes: based on a roadside sensing device, identifying any vehicle on the road as a first vehicle; based on a preset tracking algorithm, obtaining vehicle tracking data corresponding to the first vehicle.

[0061] In some embodiments, the roadside sensing device includes a laser radar and a camera, etc., and the first vehicle is any motor vehicle on the road. First, the first vehicle is identified based on the roadside sensing device such as a laser radar and a camera and provides two-dimensional image information for identifying the appearance features and category of the first vehicle. This process can be completed using algorithms such as BEV (Bird's Eye View) and Transformer (a deep learning model). Then, a vehicle tracking algorithm is used to track each frame of the identified two-dimensional image information to obtain the trajectory of the first vehicle, that is, the vehicle tracking data. The trajectory in the vehicle tracking data generated at this time is fragmented.

[0062] In some embodiments, the fragmented trajectory data in the vehicle tracking data is filtered to filter out the trajectory set of the first vehicle that continuously passes through a preset number of intersections or more. The preset number here is set according to actual business needs and can be set to two. Each intersection position corresponds to a road network node number, which can be set by setting represents the road network node number, where m represents the total number of actual road network nodes. The first intersection sequence can be expressed as , where r is the first intersection sequence, Indicates the road network node number corresponding to the nth intersection position arranged in chronological order in the first intersection sequence.

[0063] The vehicle long-distance trajectory restoration method of the present invention obtains relatively comprehensive vehicle tracking data based on the existing vehicle tracking algorithm, filters out short-distance and discontinuous trajectory data of the vehicle tracking data through a set screening mechanism, improves data quality, reduces redundant path information, and obtains a time-ordered road network node set through arrangement and construction, providing a structured and high-quality data foundation for the subsequent construction of a model training set.

[0064] In some embodiments, Figure 3 Taking the example, the execution process of the vehicle long-distance trajectory restoration method of the present invention is explained.

[0065] Step 1: Obtain vehicle tracking data results. Use roadside sensing equipment (such as cameras and lidar) to identify motor vehicles in single-frame data, use vehicle tracking algorithms (such as SORT, Deepsort, etc.) to track the identified vehicle positions, obtain fragmented data of vehicle trajectories and filter them to obtain vehicle tracking data results; Step 2: Construct a model training set based on fragmented tracking data results. Perform data screening on the vehicle tracking data results to obtain a first intersection sequence, extract intersection positions from any position of the first intersection sequence according to preset rules to construct training set data, and obtain a model training set based on all the training set data obtained through multiple extractions; Step 3: Generate a candidate path set based on the KSP algorithm. Randomly extract a training set data from the model training set, and generate multiple candidate paths between the first intersection positions based on the KSP algorithm and the first intersection position recorded in the training set data to form the first candidate path set; Step 4: Calculate the candidate path decision index. Calculate the path decision index for each candidate path in the first candidate path set, and generate an index vector corresponding to each candidate path; Step 5: training the path decision model. Take the indicator vectors in step 4 as the input of the path decision model, determine the loss function according to the output results, and use the gradient descent algorithm to optimize the model parameters of the path decision model according to the loss function to train the path decision model; Step 6: Obtain the trained path decision model. Repeat steps 3 to 5 until the preset training termination condition is met, such as reaching the preset number of iterations, to obtain the trained path decision model; Step 7: Obtain license plate recognition results. The license plate recognition results of target vehicles passing through each intersection within a preset time period are collected through roadside cameras; Step 8: Perform spatial coordinate conversion based on pixel points. Convert the pixel coordinates (i.e. license plate position coordinates) in all license plate recognition results into three-dimensional coordinates in the world coordinate system; Step 9: Obtain the intersection sequence where the camera that continuously captures the license plate recognition results is located. The three-dimensional coordinates of all license plate recognition results processed in step 7 are used through a map matching algorithm to find the intersection position that matches each three-dimensional coordinate, thereby obtaining the intersection position sequence where the license plate is continuously captured, that is, the target intersection sequence; Step 10: Use KSP to calculate candidate paths for the intersection position sequence. Obtain two adjacent intersection positions in the target intersection sequence in chronological order, and use the KSP algorithm to generate multiple candidate paths between the two adjacent intersection positions to form a target candidate path set; Step 11: Calculate the candidate path decision index. Calculate the path decision index for each candidate path in the target candidate path set, and generate an index vector corresponding to each candidate path; Step 12: Obtain the optimal candidate trajectory of the target vehicle between two intersections. Use each indicator vector in step 10 as the input of the path decision model trained in step 6, and select the candidate path with the highest probability from all output results as the target path, that is, the optimal trajectory between two adjacent intersections; Step 13: Complete the complete trajectory of the target vehicle within the preset time period. Repeat steps 10 to 12 until the optimal trajectory between all adjacent intersections in the target intersection sequence is found to restore the complete long-distance trajectory of the target vehicle within the preset time period; Step 14: Get the long-distance trajectory results of all vehicles. Determine a new target vehicle and repeat steps 7 to 14 continuously to obtain the complete long-distance trajectory results of all vehicles within the preset time period.

[0066] The vehicle long-distance trajectory restoration method of the present invention, on the one hand, uses a vehicle tracking algorithm to obtain vehicle tracking data, and extracts path decision indicators therefrom to determine the vehicle's driving preference information as the basis for training a path decision model; on the other hand, it fully utilizes the various license plate recognition results recognized by the roadside sensing device in combination with the roadside decision model to complete the path between adjacent license plate recognition results, and finally realizes the long-distance trajectory restoration of the target vehicle, avoiding the problem of trajectory fragmentation and incompleteness caused by relying solely on the vehicle tracking algorithm, and can realize the restoration of missing trajectories, effectively improving the restoration effect of the vehicle's long-distance continuous trajectory, thereby meeting the active supervision needs of target vehicles in intelligent network scenarios.

[0067] Based on the same inventive concept, according to another aspect of the present invention, Figure 4 As shown, an embodiment of the present invention further provides a vehicle long-distance trajectory restoration system, comprising: A model training unit 110 is configured to construct a model training set based on the vehicle tracking data, and train a path decision model based on the model training set; The license plate recognition unit 120 is configured to obtain the license plate recognition result of the target vehicle within a preset time period, and based on the license plate recognition result, determine the positions of all intersections passed by the target vehicle to obtain a target intersection sequence; The trajectory restoration unit 130 is configured to restore the complete trajectory of the target vehicle within a preset time period based on the target intersection sequence and the trained path decision model.

[0068] The vehicle long-distance trajectory restoration system of the present invention, on the one hand, obtains vehicle tracking data using a vehicle tracking algorithm, and extracts path decision indicators therefrom to determine the vehicle's driving preference information as the basis for training a path decision model; on the other hand, it fully utilizes the license plate recognition results of the roadside sensing device in combination with the roadside decision model to complete the path between adjacent license plate recognition results, and ultimately achieves long-distance trajectory restoration of the target vehicle, avoiding the problem of trajectory fragmentation and incompleteness caused by relying solely on the vehicle tracking algorithm, and can achieve the restoration of missing trajectories, effectively improving the restoration effect of the vehicle's long-distance continuous trajectory, thereby meeting the needs of active supervision of target vehicles in intelligent network scenarios.

[0069] Based on the same inventive concept, according to another aspect of the present invention, Figure 5 As shown, an embodiment of the present invention further provides a computer device 30, which includes a processor 310 and a memory 320. The memory 320 stores a computer program 321 that can be run on the processor. When the processor 310 executes the program, the steps of the above method are performed.

[0070] Finally, it should be noted that a person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the storage medium of the program can be a disk, an optical disk, a read-only storage memory (ROM) or a random access memory (RAM), etc. The above-mentioned computer program embodiments can achieve the same or similar effects as the corresponding above-mentioned any method embodiments.

[0071] It will also be appreciated by those skilled in the art that various exemplary logic blocks, modules, circuits and algorithm steps described in conjunction with the disclosure herein can be implemented as electronic hardware, computer software or a combination of the two. In order to clearly illustrate this interchangeability of hardware and software, a general description has been given to the functions of various schematic components, blocks, modules, circuits and steps. Whether this function is implemented as software or hardware depends on specific applications and the design constraints imposed on the entire system. Those skilled in the art can implement the function in various ways for each specific application, but this implementation decision should not be interpreted as causing a departure from the disclosed scope of the embodiments of the present invention.

[0072] The above are exemplary embodiments disclosed in the present invention, but it should be noted that various changes and modifications may be made without departing from the scope of the embodiments disclosed in the present invention as defined in the claims. The functions, steps and / or actions of the method claims according to the disclosed embodiments described herein do not need to be performed in any particular order. The serial numbers of the embodiments disclosed in the above embodiments of the present invention are for description only and do not represent the advantages and disadvantages of the embodiments. In addition, although the elements disclosed in the embodiments of the present invention may be described or required in individual form, they may also be understood as multiple unless explicitly limited to the singular.

[0073] It should be understood that, as used herein, the singular forms "a", "an" are intended to include the plural forms as well, unless the context clearly supports an exception. It should also be understood that, as used herein, "and / or" refers to any and all possible combinations including one or more of the associated listed items.

[0074] A person skilled in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the disclosure of the embodiments of the present invention (including the claims) is limited to these examples; under the concept of the embodiments of the present invention, the technical features in the above embodiments or different embodiments can also be combined, and there are many other changes in different aspects of the above embodiments of the present invention, which are not provided in detail for the sake of simplicity. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present invention should be included in the protection scope of the embodiments of the present invention.

Claims

1. A method for restoring a long-distance vehicle trajectory, characterized in that: include: Building a model training set based on vehicle tracking data, and training a path decision model based on the model training set; Obtaining a license plate recognition result of a target vehicle within a preset time period, and based on the license plate recognition result, determining positions of all intersections passed by the target vehicle to obtain a target intersection sequence; Based on the target intersection sequence and the trained path decision model, the complete trajectory of the target vehicle within the preset time period is restored.

2. The vehicle long-distance trajectory restoration method according to claim 1, characterized in that: The step of restoring the complete trajectory of the target vehicle within the preset time period based on the target intersection sequence and the trained path decision model includes: Based on the target intersection sequence, two intersection positions at adjacent moments are taken as a target position group; Based on each of the target position groups, generating a corresponding target candidate path set; Determine the target path corresponding to each target position group based on each target candidate path set and the trained path decision model; Based on all the target paths, the complete trajectory of the target vehicle within the preset time period is restored.

3. The vehicle long-distance trajectory restoration method according to claim 2, characterized in that: The step of determining the target path corresponding to each target position group based on each target candidate path set and the trained path decision model comprises: Based on each target candidate path in the target candidate path set, obtaining a path decision index of the target candidate path to construct a corresponding first index vector; Using each of the first indicator vectors as an input of the trained path decision model to obtain a model prediction result corresponding to each of the target candidate paths; Based on all model prediction results, the target path is determined.

4. The vehicle long-distance trajectory restoration method according to claim 1, characterized in that: The step of training the path decision model based on the model training set includes: Extract training set data from the model training set, and obtain a first position group corresponding to the training set data; Based on the first position group, generating a first candidate path set corresponding thereto; Based on the first candidate path set, obtaining a second indicator vector corresponding to each first candidate path, and obtaining a corresponding loss function of the path decision model according to each second indicator vector; The path decision model is trained based on each of the loss functions.

5. The vehicle long-distance trajectory restoration method according to claim 1, characterized in that: The step of obtaining a license plate recognition result of a target vehicle within a preset time period, and determining the positions of all intersections passed by the target vehicle based on the license plate recognition result to obtain a target intersection sequence includes: Determining the license plate recognition result based on the image data of the target vehicle within a preset time period; Converting the two-dimensional pixel coordinates corresponding to the license plate recognition result into three-dimensional world coordinates; Based on a map matching algorithm and the three-dimensional world coordinates, determining the locations of all the intersections through which the target vehicle passes; All the intersection positions are sorted in chronological order to obtain the target intersection sequence.

6. The vehicle long distance trajectory restoration method according to claim 1, characterized in that: The step of constructing a model training set based on vehicle tracking data includes: Based on the vehicle tracking data, determining a first intersection sequence; Extracting a first intersection position from the first intersection sequence according to a preset rule to generate a first sequence; Based on the first sequence and the first intersection position corresponding thereto, obtaining training set data; The model training set is determined based on all the training set data obtained through multiple extractions.

7. The vehicle long distance trajectory restoration method according to claim 6, characterized in that: The step of determining a first intersection sequence based on the vehicle tracking data comprises: Collecting the vehicle tracking data, and screening all trajectory sets that continuously pass through at least a preset number of intersections; Based on all the track sets obtained by screening, determining the intersection position corresponding to each of the intersections; All the corresponding intersection positions are sorted in chronological order to obtain the first intersection sequence.

8. The vehicle long-distance trajectory restoration method according to claim 7, characterized in that: The vehicle tracking data is obtained by the following steps: Based on the roadside sensing device, identifying any vehicle on the road as a first vehicle; Based on a preset tracking algorithm, the vehicle tracking data corresponding to the first vehicle is obtained.

9. A vehicle long-distance trajectory restoration system, characterized in that: include: A model training unit configured to construct a model training set based on the vehicle tracking data, and train a path decision model based on the model training set; A license plate recognition unit configured to obtain a license plate recognition result of a target vehicle within a preset time period, and based on the license plate recognition result, determine the positions of all intersections passed by the target vehicle to obtain a target intersection sequence; The trajectory restoration unit is configured to restore the complete trajectory of the target vehicle within the preset time period based on the target intersection sequence and the trained path decision model.

10. A computer device comprising: at least one processor; as well as A memory storing a computer program executable on the processor, wherein the processor executes the steps of the method according to any one of claims 1 to 8 when executing the program.

Images