Trajectory division model training method and device, trajectory division method and device

Abstract

An embodiment of the present application provides a model training method and device for trajectory segmentation, the method comprising: obtaining a sample trajectory, wherein the sample trajectory includes a plurality of trajectory points. When traversing to the first trajectory point among the plurality of trajectory points, determining the predicted excitation corresponding to the first trajectory point according to the prediction model, the predicted excitation comprising a first predicted excitation for determining the first trajectory point as a segmentation point, and a second predicted excitation for determining the first trajectory point as a non-segmentation point. Execute a first operation according to the first predicted excitation and the second predicted excitation. Determine the real excitation corresponding to the execution of the first operation according to the sub-trajectory parameters corresponding to the segmentation point before the execution of the first operation and the sub-trajectory parameters corresponding to the segmentation point after the execution of the first operation. Update the model parameters of the prediction model according to the real excitation and the predicted excitation. The method provided by the present application can realize trajectory segmentation simply and effectively.

Description

Technical Field

[0001] The embodiments of the present application relate to computer technology, and in particular to a model training method and device for trajectory segmentation, and a trajectory segmentation method and device. Background Art

[0002] The mining and analysis of trajectories has important research significance, because long trajectories contain more features, so trajectory segmentation for long trajectories is the basis of many trajectory studies.

[0003] In the existing technology, trajectory segmentation is usually performed by manually formulating segmentation rules based on the actual segmentation requirements, and then designing a corresponding segmentation algorithm based on the formulated segmentation rules. However, the implementation process of formulating segmentation rules requires manual planning of complex segmentation algorithms and the corresponding design of complex parameters, which leads to a high complexity in trajectory segmentation implementation. Summary of the Invention

[0004] The embodiments of the present application provide a model training method and device for trajectory segmentation, and a trajectory segmentation method and device to overcome the problem of high complexity in implementing trajectory segmentation.

[0005] In a first aspect, an embodiment of the present application provides a model training method for trajectory segmentation, comprising:

[0006] Acquire a sample trajectory, wherein the sample trajectory includes a plurality of trajectory points;

[0007] When traversing to a first trajectory point among the plurality of trajectory points, determining, according to a prediction model, a predicted stimulus corresponding to the first trajectory point, the predicted stimulus including a first predicted stimulus for determining the first trajectory point as a division point, and a second predicted stimulus for determining the first trajectory point as a non-division point, wherein adjacent division points determined among the plurality of trajectory points are used to constitute a sub-trajectory corresponding to the sample trajectory;

[0008] performing a first operation according to the first predicted stimulus and the second predicted stimulus, wherein the first operation is determining the first trajectory point as a division point, or the first operation is determining the first trajectory point as a non-division point;

[0009] determining a real stimulus corresponding to the execution of the first operation based on sub-trajectory parameters corresponding to the division point before the execution of the first operation and sub-trajectory parameters corresponding to the division point after the execution of the first operation, wherein the sub-trajectory parameters are used to indicate the length of the sub-trajectory and the difference between the sub-trajectory and the sample trajectory;

[0010] The model parameters of the prediction model are updated according to the real stimulus and the predicted stimulus.

[0011] In a second aspect, an embodiment of the present application provides a trajectory division method, including:

[0012] Acquire a target trajectory to be divided, wherein the target trajectory includes a plurality of trajectory points;

[0013] When traversing to a first trajectory point among the multiple trajectory points, determining, according to a prediction model, a predicted stimulus corresponding to the first trajectory point, the predicted stimulus including a first predicted stimulus that determines the first trajectory point as a dividing point and a second predicted stimulus that determines the first trajectory point as a non-dividing point, the prediction model being trained according to the model training method described in the first aspect above;

[0014] performing a first operation according to the predicted stimulus, wherein the first operation is determining the first trajectory point as a division point, or the first operation is determining the first trajectory point as a non-division point;

[0015] The target trajectory is divided according to the trajectory points determined as division points to obtain a plurality of sub-trajectories.

[0016] In a third aspect, an embodiment of the present application provides a model training device for trajectory segmentation, comprising:

[0017] An acquisition module, configured to acquire a sample trajectory, wherein the sample trajectory includes a plurality of trajectory points;

[0018] a determination module, configured to, when traversing to a first trajectory point among the plurality of trajectory points, determine, based on a prediction model, a predicted stimulus corresponding to the first trajectory point, the predicted stimulus including a first predicted stimulus for determining the first trajectory point as a partition point and a second predicted stimulus for determining the first trajectory point as a non-partition point, wherein adjacent partition points determined among the plurality of trajectory points are used to constitute a sub-trajectory corresponding to the sample trajectory;

[0019] a processing module, configured to perform a first operation according to the first prediction stimulus and the second prediction stimulus, wherein the first operation is determining the first trajectory point as a division point, or the first operation is determining the first trajectory point as a non-division point;

[0020] The determining module is configured to determine a real stimulus corresponding to the execution of the first operation based on sub-trajectory parameters corresponding to the division point before the execution of the first operation and sub-trajectory parameters corresponding to the division point after the execution of the first operation, wherein the sub-trajectory parameters are used to indicate the length of the sub-trajectory and the difference between the sub-trajectory and the sample trajectory;

[0021] An updating module is used to update the model parameters of the prediction model according to the real stimulus and the predicted stimulus.

[0022] In a fourth aspect, an embodiment of the present application provides a trajectory division device, including:

[0023] An acquisition module, configured to acquire a target trajectory to be divided, wherein the target trajectory includes a plurality of trajectory points;

[0024] a determination module, configured to, when traversing to a first trajectory point among the plurality of trajectory points, determine, based on a prediction model, a predicted stimulus corresponding to the first trajectory point, the predicted stimulus including a first predicted stimulus for determining the first trajectory point as a dividing point and a second predicted stimulus for determining the first trajectory point as a non-dividing point, the prediction model being trained according to the model training method described in the first aspect above;

[0025] a processing module, configured to perform a first operation according to the predicted stimulus, wherein the first operation is determining the first trajectory point as a division point, or the first operation is determining the first trajectory point as a non-division point;

[0026] The processing module is further configured to divide the target trajectory according to the trajectory points determined as dividing points to obtain a plurality of sub-trajectories.

[0027] In a fifth aspect, an embodiment of the present application provides an electronic device, including:

[0028] Memory, used to store programs;

[0029] A processor is used to execute the program stored in the memory. When the program is executed, the processor is used to execute the methods described in the first and second aspects above.

[0030] In a sixth aspect, an embodiment of the present application provides a computer-readable storage medium comprising instructions, which, when executed on a computer, enables the computer to execute the methods described in the first and second aspects above.

[0031] In a seventh aspect, an embodiment of the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the methods described in the first and second aspects above.

[0032] The embodiments of the present application provide a model training method and device for trajectory segmentation. The method traverses each trajectory point in a sample trajectory, and then determines whether each trajectory point is a predicted stimulus corresponding to a segmentation point based on the prediction model dependency. The method then selects to execute a corresponding action based on the predicted stimulus, and then determines the real stimulus based on the sub-trajectory parameters before and after the execution of the corresponding action. The real stimulus can effectively measure the effect of executing the corresponding action. Therefore, the parameters of the prediction model are then updated based on the predicted stimulus and the real stimulus, thereby ensuring the accuracy of the predicted stimulus corresponding to the two execution actions output by the trained prediction model, and thus can simply and effectively implement trajectory segmentation based on the prediction model.

[0033] In addition, an embodiment of the present application provides a trajectory division method and device, which traverses each trajectory point in the target trajectory, and then predicts a first prediction stimulus that determines the trajectory point as a connection point based on a trained prediction model, and a second prediction stimulus that determines the trajectory point as a non-connection point, and then determines whether to determine the traversed trajectory point as a connection point based on the first prediction stimulus and the second prediction stimulus. After traversing each trajectory point, multiple trajectory points can be obtained, and then the sub-trajectory division results are determined based on the trajectory points, so that the trajectory division can be realized simply and effectively, while ensuring that the simplicity and accuracy of the divided sub-trajectories are relatively good. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0035] Figure 1 A schematic diagram of the trajectory provided in the embodiment of the present application;

[0036] Figure 2 A schematic diagram of the distance between trajectories provided in an embodiment of the present application;

[0037] Figure 3 Flowchart of the model training method for trajectory segmentation provided in an embodiment of the present application;

[0038] Figure 4 The process of the model training method for trajectory division provided in the embodiment of this application Figure 2 ;

[0039] Figure 5 A schematic diagram of an implementation of determining a division point provided in an embodiment of the present application;

[0040] Figure 6 A schematic diagram of the implementation of the prediction model training provided in the embodiment of the present application;

[0041] Figure 7 A flow chart of the trajectory division method provided in an embodiment of the present application;

[0042] Figure 8 A schematic diagram of implementing trajectory division provided in an embodiment of the present application;

[0043] Figure 9 A schematic diagram of the structure of a model training device for trajectory segmentation provided in an embodiment of the present application;

[0044] Figure 10 A schematic diagram of the structure of a trajectory division device provided in an embodiment of the present application;

[0045] Figure 11 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0046] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0047] In order to better understand the technical solution of this application, the relevant technologies involved in this application are further introduced in detail below.

[0048] With the rapid development of 4G (fourth-generation mobile communication technology), 5G (fifth-generation mobile communication technology), satellite positioning, and mobile internet technologies, coupled with the widespread adoption of mobile devices, the amount of spatiotemporal data generated by various mobile objects, including pedestrians, vehicles, and aircraft, has exploded. The mining and analysis of trajectory data has become a crucial research topic. This type of data can be applied to a variety of scenarios, including location-based services, traffic forecasting, and motion behavior research.

[0049] Typically, recorded complete trajectories are relatively long, and long trajectories contain many features. Therefore, segmenting long trajectories is the basis of many trajectory applications. For example, segmenting long trajectories can be used for trajectory simplification and sub-trajectory clustering.

[0050] Taking sub-trajectory clustering as an example, for trajectories that are long and have complex motion, clustering based on the entire trajectory as the basic unit will cause the following problems: (1) local features of complex trajectories may be ignored; (2) it is difficult to discover common sub-trajectory patterns.

[0051] For example, if two complete trajectories have low similarity, but some of their subtrajectories are similar, clustering directly based on the long trajectories will result in poor clustering results. Therefore, we can study the clustering of subtrajectories. In the implementation of subtrajectory clustering, we can first divide the trajectory into subtrajectories according to certain division rules, and then cluster the subtrajectories based on their similarity.

[0052] The following describes the implementation of trajectory division in related technologies. Existing trajectory division is usually done by manually formulating certain rules. For example, the trajectory can be divided according to changes in direction, speed, angle, etc. Complex parameters usually need to be formulated in this process, and the time complexity is relatively high.

[0053] Among them, the existing sub-trajectory division method is mainly divided into the following two steps:

[0054] The first step is to propose a classification criterion based on the purpose of the classification. Usually, the classification criterion will take into account factors such as the accuracy of the sub-trajectory, the simplicity of the sub-trajectory, and the characteristics of the original trajectory represented by the sub-trajectory, such as the speed and direction of the sub-trajectory.

[0055] The second step is to propose a corresponding partitioning algorithm based on the proposed partitioning criteria. Usually, the time complexity of obtaining the optimal partitioning solution based on the partitioning algorithm is O(2 n ), this time complexity is not realistic in practical applications, so an approximate solution or local optimal solution is usually obtained based on the partitioning algorithm.

[0056] The sub-trajectory partitioning method described above has two main issues: First, the algorithm implementation involves parameter selection, such as threshold parameters and coefficient parameters, which vary with different partitioning trajectory datasets. Second, different partitioning criteria may require different algorithms, and different partitioning criteria will also cause parameter changes.

[0057] Therefore, the implementation method of trajectory division in the existing technology requires setting different division algorithms for different division requirements, and also requires manual operations such as complex algorithm design and parameter setting, which leads to a high complexity in the implementation of trajectory division.

[0058] In response to the problems in the existing technology, this application proposes the following technical concept: a prediction model is trained through the reinforcement learning method, and then each trajectory point in the original trajectory is traversed. The prediction model can output a prediction incentive for performing sub-trajectory division or not performing sub-trajectory division at each trajectory point. Then, according to the predicted incentive, the corresponding action can be selected to achieve the division of the sub-trajectory. By taking advantage of the data-driven characteristics of reinforcement learning, the division can be learned according to different data features and dynamic changes to ensure the effectiveness of model training. After the model training is completed, the incentive for executing the corresponding action can be simply and effectively based on the model prediction. Then, the corresponding sub-trajectory can be determined according to the division point determined by the execution of the action, so that the trajectory division can be simply and effectively achieved.

[0059] Before introducing the specific methods provided in this application, the relevant concepts involved in this application are first explained.

[0060] 1. Trajectory. Trajectory refers to a series of time-ordered trajectory points generated by a moving object during its movement, which can be expressed as T = <p1,p2,…p n >, where the trajectory point p i It can be expressed as (lat, lon, t), where (lat, lon) represents the spatial position of the trajectory point, corresponding to longitude and latitude respectively, and t represents the sampling time of the trajectory point. The length of the trajectory can be expressed as the number of trajectory points. The length of the trajectory T above is n.

[0061] For example, you can combine Figure 1 Understand the trajectory. Figure 1 A schematic diagram of the trajectory provided in an embodiment of the present application.

[0062] like Figure 1 As shown, it is assumed that there is currently a trajectory, which includes trajectory point p1, trajectory point p2, trajectory point p3, trajectory point p4 and trajectory point p5. Figure 1 The trajectory described in includes 5 trajectory points, so it can be understood that the length of this trajectory is 5.

[0063] 2. Sub-trajectory. A sub-trajectory can refer to a sub-trajectory composed of a subset of continuous trajectory points in a trajectory, that is, sub_T(i,j)= <p i ,p i+1 ,…p j >, where 1≤i≤j≤n, where the sub-trajectory sub_T(i,j) is the trajectory point p i To trajectory point p j The trajectory is composed of a subset of continuous trajectory points.

[0064] For example, still combined Figure 1 Understand, assume Figure 1The subset composed of the three trajectory points p1, p2, p3, and p4 can be understood as a sub-trajectory.

[0065] Alternatively, a sub-track may be formed by determining division points among multiple track points and then connecting adjacent division points.

[0066] For example, still combined Figure 1 To understand, assume that Figure 1 In the trajectory shown, the trajectory point p1 and the trajectory point p4 are determined as the division points, and then the line segment obtained by connecting the trajectory point p1 and the trajectory point p4 can form a sub-trajectory.

[0067] 3. Distance. Distance is usually used to measure the similarity between two trajectories. The larger the distance, the lower the similarity, and the smaller the distance, the higher the similarity. In this embodiment, for example, the similarity between two trajectories can be measured by parallel distance, perpendicular distance, and angular distance.

[0068] The following combination Figure 2 These three distances are explained separately. Figure 2 A schematic diagram of the distance between trajectories provided in an embodiment of the present application.

[0069] like Figure 2 As shown in FIG, assuming that there is currently a track 1 consisting of track points s1 and e1, and a track 2 consisting of track points s2 and e2. Based on this, for example, a perpendicular line can be drawn from the two endpoints of one track to the other track, for example, a perpendicular line is usually drawn from the short side to the long side.

[0070] Reference Figure 2 , the perpendicular distance from the track point s1 of track 1 to the track 2 can be expressed as L ⊥1 , and the perpendicular distance from the trajectory point e1 of trajectory 1 to the trajectory 2 can be expressed as L ⊥2 .

[0071] Among them, the vertical distance between track 1 and track 2 can be expressed as:

[0072] And, refer to Figure 2 , the distance between the perpendicular point of the track point s1 of track 1 to track 2 and the adjacent track point s2 in track 2 is a parallel distance, which can be expressed as L ||1 , and the distance between the perpendicular point of the track point e1 of track 1 to track 2 and the adjacent track point e2 in track 2 is another parallel distance, which can be expressed as L ||2 .

[0073] Among them, the parallel distance between track 1 and track 2 can be expressed as: d ||=min(L ||1 ,L ||2 ).

[0074] and reference Figure 1 , the angle between track 1 and track 2 is θ, and the length of track 1 is L j .

[0075] Among them, the angular distance between track 1 and track 2 can be expressed as: d θ =||L j ||×sin(θ).

[0076] 4. Minimum Description Length (MDL).

[0077] In this application, it is assumed that the original trajectory can be expressed as T= <p1,p2,…p n >, and assuming that the partition point determined among the multiple trajectory points of the original trajectory can be expressed as These m+1 partitioning points can constitute m sub-trajectories.

[0078] In a possible implementation, in the process of solving the MDL, a sub-trajectory can be understood as a trajectory formed by connecting two adjacent partition points.

[0079] The minimum description length can be expressed as MDL=L(H)+L(D|H).

[0080] Among them, L(H) can indicate the length of the sub-trajectory and can be expressed as in, It is a trajectory point and trajectory points The Euclidean distance between .

[0081] For example, based on Figure 1 To understand the trajectory shown, assume that Figure 1 Two partition points are determined, namely p1 and p4, then L(H)=log2(len(p1,p4).

[0082] And, L(D|H) can indicate the distance between the sub-trajectory and the original trajectory, which can be expressed as

[0083]

[0084] in, Can be understood as trajectory points and trajectory points The trajectory composed of p k p k+1 It can be understood as the trajectory point p kand trajectory point p k+1 The trajectory is composed of, where the value range of k is c j to c j+1 -1, which can be understood as the original trajectory from the trajectory point To track point and multiple trajectory points in between.

[0085] For example, based on Figure 1 To understand the trajectory shown, assume that Figure 1 Two dividing points are determined, namely p1 and p4, then

[0086] L(D|H)=log2(d ⊥ (p1p4,p1p2)+d ⊥ (p1p4,p2p3)+d ⊥ (p1p4,p3p4))+log2(d θ (p1p4,p1p2)+d θ (p1p4,p2p3)+d θ (p1p4,p3p4)).

[0087] Building on the above introduction, the following section further details the role of MDL. Trajectory segmentation refers to the process of dividing a long original trajectory (or original trajectory) into several shorter sub-trajectories using a specific partitioning scheme. Typically, there are requirements for both accuracy and simplicity in the segmentation. Accuracy refers to the degree of difference between the sub-trajectories and the original trajectory, while simplicity refers to the number of sub-trajectories. These two requirements are somewhat contradictory.

[0088] It is understandable that if the trajectory is divided at every point, the difference between the obtained sub-trajectory and the original trajectory is minimal, and the accuracy is the highest, but the number of sub-trajectories is too large and not concise enough; if the trajectory is divided only at the starting point and the end point, only one sub-trajectory can be obtained, that is, the sub-trajectory obtained by directly connecting the starting point and the end point. In this case, the sub-trajectory is the simplest, but the difference between the sub-trajectory and the original trajectory is too large and not accurate enough.

[0089] The MDL provided in this application can effectively balance the accuracy and simplicity of trajectory division. It can be understood that the L(H) in the MDL described above can indicate the length of the sub-trajectory. If the number of sub-trajectories obtained by dividing the original trajectory is relatively large, then the L(H) obtained by summing up the lengths of the sub-trajectories will be relatively large. Conversely, if the number of sub-trajectories obtained by dividing the original trajectory is relatively small, then the L(H) obtained by summing up the lengths of the sub-trajectories will be relatively small. This can be understood in conjunction with a specific example. For example, in the above Figure 2In the trajectory shown, the following two situations are assumed:

[0090] Case 1: The trajectory point p1 and the trajectory point p3 are determined as the division points, and then the trajectory point p1 and the trajectory point p3 are connected to determine the sub-trajectory 1.

[0091] Case 2: The trajectory point p1, the trajectory point p2 and the trajectory point p3 are all determined as the division points, and then the trajectory point p1 and the trajectory point p2 are connected to determine the sub-trajectory 2, and the trajectory point p2 and the trajectory point p3 are connected to determine the sub-trajectory 3.

[0092] In both cases above, the length of sub-trajectory 1 is smaller than the sum of the lengths of sub-trajectory 2 and sub-trajectory 3. Therefore, the L(H) corresponding to sub-trajectory 1 is smaller than the L(H) corresponding to sub-trajectory 1 and sub-trajectory 2. In other words, the fewer sub-trajectories there are, the smaller L(H), indicating a more concise sub-trajectory. Therefore, L(H) can effectively reflect the simplicity of trajectory partitioning.

[0093] Furthermore, the L(D|H) in the MDL described above can indicate the difference between the sub-trajectory and the original trajectory. If the difference between the sub-trajectory and the original trajectory is relatively small, then the similarity between the corresponding sub-trajectory and the original trajectory is relatively large. Therefore, L(D|H) can be understood as reflecting the accuracy of trajectory division.

[0094] Based on the above, it can be understood that the simplicity and accuracy of trajectory division are two contradictory indicators. When the trajectory division is the simplest, the accuracy is the worst (that is, the difference between the sub-trajectory and the original trajectory is the largest). In other words, when L(H) is the smallest, L(D|H) is the largest.

[0095] Furthermore, when the trajectory division is most accurate (that is, the difference between the sub-trajectory and the original trajectory is the smallest), the simplicity is the worst. In other words, when L(H) is the largest, L(D|H) is the smallest.

[0096] Therefore, for MDL, whether L(H) or L(D|H) is minimized, it will never reach its minimum value. When MDL reaches its minimum value, it must strike a balance between simplicity and accuracy, minimizing both L(H) and L(D|H). Therefore, if the MDL corresponding to the sub-trajectory obtained by partitioning is minimized, it can ensure that the sub-trajectory has a high similarity with the original trajectory and a relatively small number of sub-trajectories.

[0097] In order to obtain the trajectory partitioning scheme that minimizes the MDL, it is necessary to list all possible partitioning situations. However, the time complexity of this operation is O(2 n ), which is unacceptable in practical applications.

[0098] Therefore, in one implementation, a greedy approximation algorithm can be used. For example, each trajectory point is sequentially traversed. When the MDL obtained by partitioning at a trajectory point is less than the MDL obtained by not partitioning at that trajectory point, it is considered that partitioning at that trajectory point is a better choice. However, the simplicity of the sub-trajectory partitioning must also be considered. Therefore, the traversal continues until the MDL obtained by partitioning at a certain trajectory point is greater than the MDL obtained by not partitioning at that trajectory point. At this point, it is determined that not partitioning at that trajectory point is a better choice, and the trajectory point before this trajectory point can be determined as the sub-trajectory partitioning location.

[0099] However, this implementation method only yields approximate solutions and requires manual rule-making. For example, when the MDL obtained by partitioning a trajectory point is less than the MDL obtained by not partitioning at that point, the traversal proceeds to the next point; when the MDL obtained by partitioning a trajectory point is greater than the MDL obtained by not partitioning at that point, the previous point is determined as the sub-trajectory partition location. These rules are still manually formulated.

[0100] This implementation method requires artificial formulation of division rules and is based on the sub-trajectory division results obtained based on approximate solutions. Therefore, it cannot guarantee the universality of the sub-trajectory division method or the sub-trajectory division effect.

[0101] The following is an introduction to the model training method for trajectory segmentation provided in this application in conjunction with specific embodiments. It should be noted that the execution subjects of the various methods provided in the embodiments of this application can be, for example, servers, processors, chips, and other devices with data processing functions. This embodiment does not limit the specific execution subjects, which can be selected and set according to actual needs. Any device with data processing functions can be used as the execution subject of the embodiments of this application.

[0102] First combine Figure 3 To explain, Figure 3 Flowchart of the model training method for trajectory segmentation provided in an embodiment of the present application.

[0103] like Figure 3 As shown, the method includes:

[0104] S301: Acquire a sample trajectory, where the sample trajectory includes multiple trajectory points.

[0105] In this embodiment, a sample trajectory is a trajectory used for model training and may include multiple trajectory points. During actual model training, multiple sample trajectories may be obtained. The number of sample trajectories, the number of trajectory points included in each sample trajectory, and the shape of the sample trajectories can all be selected and determined based on actual needs. This embodiment does not limit the specific implementation of the sample trajectories.

[0106] S302. When traversing to the first trajectory point among multiple trajectory points, determine the prediction excitation corresponding to the first trajectory point according to the prediction model, the prediction excitation including a first prediction excitation for determining the first trajectory point as a division point and a second prediction excitation for determining the first trajectory point as a non-division point, wherein the adjacent division points determined among the multiple trajectory points are used to constitute a sub-trajectory corresponding to the sample trajectory.

[0107] The processing method for each sample trajectory is similar, so this embodiment uses the processing of any sample trajectory as an example for explanation. During the model training process, each trajectory point in the sample trajectory can be traversed in sequence, where the traversal order can be, for example, based on the time sequence of each trajectory point.

[0108] For any sample trajectory, when traversing to the first trajectory point among multiple trajectory points, this embodiment can determine the predicted incentive corresponding to the first trajectory point according to the prediction model. Among them, the predicted incentive is the incentive output by the prediction model, wherein the incentive (reward), or reward, can be understood as feedback information after performing a certain operation, and the feedback information here is used to indicate the pros and cons of performing the operation. For example, when performing a certain operation is an operation encouraged by the system, the incentive is greater. What specific operation the system encourages to perform depends on the current reinforcement learning purpose. For example, in this embodiment, the purpose of reinforcement learning is to obtain the sub-trajectory division result with the minimum MLD. If the MLD can be effectively reduced after performing a certain operation, then this operation will receive more incentives accordingly.

[0109] The first trajectory point may be any one of the multiple trajectory points. Since the processing methods for each trajectory point are similar, the following description will only take the first trajectory point as an example.

[0110] The prediction model in this embodiment is a model for outputting prediction excitations corresponding to trajectory points, wherein the prediction excitations may include a first prediction excitation that determines the first trajectory point as a division point, and may also include a second prediction excitation that determines the first trajectory point as a non-division point.

[0111] In this embodiment, adjacent division points determined from multiple trajectory points are used to form sub-trajectories corresponding to the sample trajectory. For example, division points can be determined from multiple trajectory points, and then adjacent division points can be sequentially connected to obtain multiple sub-trajectories. Alternatively, adjacent division points can be used as endpoints, and the trajectories corresponding to the two endpoints in the original trajectory can be determined as sub-trajectories. Sub-trajectories in this embodiment can also be called division trajectories, target trajectories, and so on.

[0112] Therefore, if the first track point is determined as a division point, the first track point can be connected with the adjacent division points to form a sub-track. If the first track point is determined as a non-division point, when determining the sub-track, the first track point will be skipped and will not be connected with the division point to form a sub-track.

[0113] S303: Perform a first operation according to the first predicted stimulus and the second predicted stimulus, where the first operation is to determine the first trajectory point as a division point, or the first operation is to determine the first trajectory point as a non-division point.

[0114] After determining the first predicted stimulus and the second predicted stimulus, for example, a first operation may be performed according to the first predicted stimulus and the second predicted stimulus.

[0115] In one possible implementation, for example, an action with a larger stimulus may be selected for execution. Thus, for example, when the first predicted stimulus is greater than the second predicted stimulus, the first operation performed may be determining the first trajectory point as a split point. When the first predicted stimulus is not greater than the second predicted stimulus, the first operation performed may be determining the first trajectory point as a non-split point.

[0116] S304. Determine the real stimulus corresponding to the first operation based on the sub-trajectory parameters corresponding to the division point before the first operation and the sub-trajectory parameters corresponding to the division point after the second operation. The sub-trajectory parameters are used to indicate the length of the sub-trajectory and the difference between the sub-trajectory and the sample trajectory.

[0117] It is understandable that after the first operation is performed, the division points determined among the multiple track points may change. For example, if the first track point is determined as the division point, then one more division point is determined.

[0118] In this embodiment, corresponding sub-trajectory parameters can be determined for the sub-trajectory formed by the dividing points, where the sub-trajectory parameters can indicate the length of the sub-trajectory and the difference between the sub-trajectory and the sample trajectory. In one possible implementation, the sub-trajectory parameters can be, for example, the MDL described above. Alternatively, in actual implementation, the sub-trajectory parameters can be obtained by adding corresponding parameters to the MDL determination method described above. Alternatively, the sub-trajectory parameters can be selected and set according to actual needs, as long as the sub-trajectory parameters can reflect the length of the sub-trajectory and the difference between the sub-trajectory and the sample trajectory. And the sub-trajectory parameters in this embodiment can also be called trajectory feature parameters, for example.

[0119] When the division point changes, the corresponding sub-trajectory parameters also change. Therefore, in this embodiment, the sub-trajectory parameters corresponding to the division point before the first operation is performed can be determined, and the sub-trajectory parameters corresponding to the division point after the first operation is performed can also be determined. Based on the above description, it can be determined that in actual implementation, the smaller the sub-trajectory parameter, the better the corresponding sub-trajectory division effect.

[0120] Therefore, for example, the real stimulus corresponding to the execution of the first operation can be determined based on the sub-trajectory parameters corresponding to the division points before the execution of the first operation and the sub-trajectory parameters corresponding to the division points after the execution of the first operation.

[0121] In a possible implementation, in this embodiment, for example, the difference between the sub-trajectory parameter corresponding to the division point before the first operation is performed and the sub-trajectory parameter corresponding to the division point after the first operation is performed can be determined as the real excitation.

[0122] During model training, this embodiment sequentially traverses each trajectory point, and then performs the first operation for each trajectory point to determine whether it is a partition point or a non-partition point. It is understandable that for each trajectory point, the sub-trajectory parameters corresponding to the partition point after the first operation is performed are actually the sub-trajectory parameters corresponding to the partition point before the first operation is performed for the next trajectory point.

[0123] Therefore, the cumulative real stimulus after the real stimulus corresponding to each trajectory point is accumulated is actually the difference between the sub-trajectory parameter corresponding to the first trajectory point before the first operation (which is 0) and the sub-trajectory parameter corresponding to the last trajectory point after the first operation. It can be understood that the purpose of performing the corresponding operation for each trajectory point is to maximize the cumulative real stimulus. When the cumulative real stimulus is maximized, the sub-trajectory parameter corresponding to the last trajectory point after the first operation is minimized.

[0124] If the sub-trajectory parameter corresponding to the last trajectory point after executing the first operation is the smallest, it means that the sub-trajectory corresponding to the division point determined when traversing to the last trajectory point has the smallest sub-trajectory length and the difference between the sub-trajectory and the sample trajectory. In other words, the simplicity and accuracy of the sub-trajectory are optimal. Therefore, the real excitation in this embodiment can effectively reflect the execution effect of executing the first operation in the sub-trajectory division process.

[0125] S305: Update the model parameters of the prediction model according to the real incentive and the predicted incentive.

[0126] Among them, the real incentive is the accurate incentive result after executing the first operation, and the predicted incentive is the incentive result predicted for executing the first operation. Therefore, the model parameters of the prediction model can be updated according to the real incentive and the predicted incentive, thereby effectively improving the prediction accuracy of the prediction model.

[0127] The model training method for trajectory segmentation provided in an embodiment of the present application includes: obtaining a sample trajectory, wherein the sample trajectory includes multiple trajectory points. When traversing to the first trajectory point among the multiple trajectory points, determining a predicted stimulus corresponding to the first trajectory point according to a prediction model, the predicted stimulus including a first predicted stimulus for determining the first trajectory point as a segmentation point and a second predicted stimulus for determining the first trajectory point as a non-segmentation point, wherein adjacent segmentation points determined among the multiple trajectory points are used to constitute a sub-trajectory corresponding to the sample trajectory. Based on the first predicted stimulus and the second predicted stimulus, performing a first operation, wherein the first operation is to determine the first trajectory point as a segmentation point, or the first operation is to determine the first trajectory point as a non-segmentation point. Based on the sub-trajectory parameters corresponding to the segmentation point before performing the first operation and the sub-trajectory parameters corresponding to the segmentation point after performing the first operation, determining the true stimulus corresponding to performing the first operation, wherein the sub-trajectory parameters are used to indicate the length of the sub-trajectory and the difference between the sub-trajectory and the sample trajectory. Based on the true stimulus and the predicted stimulus, updating the model parameters of the prediction model. By traversing each trajectory point in the sample trajectory, each trajectory point is determined to be the predicted stimulus corresponding to the segmentation point based on the prediction model dependency, and then the corresponding action is selected according to the predicted stimulus. The real stimulus is determined based on the sub-trajectory parameters before and after the corresponding action is executed. The real stimulus can effectively measure the effect of executing the corresponding action. Therefore, the parameters of the prediction model are updated based on the predicted stimulus and the real stimulus, thereby ensuring the accuracy of the predicted stimulus corresponding to the two execution actions output by the trained prediction model, and thus trajectory segmentation can be realized simply and effectively based on the prediction model.

[0128] Based on the above introduction, Figures 4 to 6The model training method for trajectory division provided in this application is further introduced in detail. Figure 4 The process of the model training method for trajectory division provided in the embodiment of this application Figure 2 , Figure 5 This is a schematic diagram of an implementation of determining a division point provided in an embodiment of the present application. Figure 6 Schematic diagram of the implementation of the prediction model training provided in the embodiment of the present application.

[0129] like Figure 4 As shown, the method includes:

[0130] S401: Acquire a sample trajectory, where the sample trajectory includes multiple trajectory points.

[0131] The implementation of S401 is similar to the implementation of S301 described above, and will not be described in detail here.

[0132] The following combination Figure 5 Take a specific sample trajectory as an example to illustrate. Figure 5 As shown, assuming that there is currently Figure 5 The sample trajectory shown includes five trajectory points, namely trajectory point p1, trajectory point p2, trajectory point p3, trajectory point p4 and trajectory point p5.

[0133] S402: When traversing to the first track point among the multiple track points, obtain a target track point that has been determined as a division point among the multiple track points.

[0134] In this embodiment, each trajectory point in the sample trajectory is traversed in sequence, and a first operation is performed on each traversed trajectory point. The first operation can be to determine the trajectory point as a division point or to determine the trajectory point as a non-division point.

[0135] It is understandable that when the first of the multiple trajectory points is traversed, some partitioning points may have already been determined before the first trajectory point. Therefore, when the first of the multiple trajectory points is traversed, the target trajectory point that has been determined as a partitioning point can be obtained from the multiple trajectory points. It is understandable that because each trajectory point is traversed sequentially, the target trajectory points that have been determined as partitioning points are all trajectory points before the first trajectory point.

[0136] For example, you can combine Figure 5To understand this, suppose we are currently traversing to point p5, which is actually the first point in this example. Therefore, points p1 to p4 have already been traversed. Assuming that points p1 and p3 were identified as split points during the previous traversal, then when we traverse to point p5, the target points will include points p1 and p3. It should be noted that the first point in the trajectory is always identified as a split point.

[0137] S403 : Determine at least one first sub-trajectory according to the target trajectory point, and determine sub-trajectory parameters corresponding to the target trajectory point according to the trajectory length of the at least one first sub-trajectory and the distance between the at least one first sub-trajectory and the sample trajectory.

[0138] After determining the target trajectory point, in this embodiment, it is necessary to determine the sub-trajectory parameters corresponding to the target trajectory point. For example, at least one first sub-trajectory can be determined based on the target trajectory point. The first sub-trajectory is a sub-trajectory formed by connecting adjacent target trajectory points.

[0139] For example, Figure 5 In the example, the target trajectory points include p1 and p3, then the trajectory point p1 and the trajectory point p3 can be connected to form Figure 5 The trajectory 501 is shown in FIG.

[0140] For another example, if the target trajectory points include trajectory point p1, trajectory point p2, and trajectory point p3, then trajectory point p1 and trajectory point p2 can be connected to form a first sub-trajectory, and trajectory point p2 and trajectory point p3 can be connected to form another first sub-trajectory.

[0141] After determining the first sub-trajectory according to the target trajectory point, in this embodiment, the sub-trajectory parameters corresponding to the target trajectory point may be determined according to the trajectory length of the first sub-trajectory and the distance between at least one first sub-trajectory and the sample trajectory.

[0142] For example, the sub-trajectory parameter in this embodiment can be the MDL introduced above. Based on the above introduction, it can be determined that MDL is the sum of L(H) and L(D|H), where L(H) corresponds to the trajectory length of the currently introduced sub-trajectory, and L(D|H) corresponds to the distance between the currently introduced sub-trajectory and the sample trajectory.

[0143] For example, Figure 5 In the example, the trajectory length of the first sub-trajectory is L(H)=log2(len(p1,p3), and the distance between the first sub-trajectory and the sample trajectory is: L(D|H)=log2(d ⊥ (p1p3,p1p2)+d ⊥ (p1p3,p2p3))+log2(dθ (p1p3,p1p2)+d θ (p1p3,p2p3)).

[0144] Then, the sum of the trajectory length L(H) of the first sub-trajectory and the distance L(D|H) between the first sub-trajectory and the sample trajectory can be determined as the MDL corresponding to the target trajectory point, that is, the sub-trajectory parameter corresponding to the target trajectory point in this embodiment. For example, the sub-trajectory parameter corresponding to the current target trajectory point can be recorded as total_mdl.

[0145] S404: Determine a second sub-trajectory based on the first trajectory point and the last target trajectory point among the multiple target trajectory points, and determine sub-trajectory parameters corresponding to determining the first trajectory point as a division point based on the trajectory length of the second sub-trajectory and the distance between the second sub-trajectory and the sample trajectory.

[0146] After determining the target trajectory point, in this embodiment, it is also necessary to determine the sub-trajectory parameters corresponding to when the first trajectory point is determined as the division point. For example, the second sub-trajectory can be determined based on the first trajectory point and the last target trajectory point among multiple target trajectory points, where the second sub-trajectory is the sub-trajectory formed by connecting the first trajectory point and the last target trajectory point.

[0147] For example, Figure 5 In the example, the target track points include p1 and p3, the last target track point is p3, and the first track point in the current example is p5, then the track points p3 and p5 can be connected to form Figure 5 The trajectory 502 is shown in FIG.

[0148] After determining the second sub-trajectory based on the first trajectory point and the last target trajectory point, in this embodiment, the sub-trajectory parameters corresponding to determining the first trajectory point as the division point can be determined based on the trajectory length of the second sub-trajectory and the distance between the second sub-trajectory and the sample trajectory.

[0149] Similarly, for example, the sub-trajectory parameters in this embodiment can be the MDL introduced above, for example, Figure 5 In the example, the trajectory length of the second sub-trajectory is L(H)=log2(len(p3,p5), and the distance between the first sub-trajectory and the sample trajectory is L(D|H)=log2(d ⊥ (p3p5,p3p4)+d ⊥ (p3p5,p4p5))+log2(d θ ((p3p5,p3p4))+d θ (p3p5,p4p5)).

[0150] Then, the sum of the trajectory length L(H) of the second sub-trajectory and the distance L(D|H) between the second sub-trajectory and the sample trajectory can be determined as the MDL corresponding to the second sub-trajectory, that is, the sub-trajectory parameter corresponding to the first trajectory point being determined as the division point in this embodiment. For example, the sub-trajectory parameter corresponding to the first trajectory point being determined as the division point can be recorded as mdl_par.

[0151] S405. Determine at least one third sub-trajectory based on the first trajectory point, the last target trajectory point among the multiple target trajectory points, and the trajectory points between the first trajectory point and the last target trajectory point, and determine sub-trajectory parameters corresponding to determining the first trajectory point as a non-division point based on the trajectory length of the at least one third sub-trajectory and the distance between the at least one third sub-trajectory and the sample trajectory.

[0152] After determining the target trajectory point, in this embodiment, it is also necessary to determine the sub-trajectory parameters corresponding to when the first trajectory point is determined as a non-division point. For example, a third sub-trajectory can be determined based on the first trajectory point, the last target trajectory point among multiple target trajectory points, and the trajectory points between the first trajectory point and the last target trajectory point. The third sub-trajectory is a sub-trajectory formed by connecting the first trajectory point, the last target trajectory point among multiple target trajectory points, and the trajectory points between the first trajectory point and the last target trajectory point.

[0153] For example, Figure 5 In the example, the target trajectory points include p1 and p3, the last target trajectory point is p3, and the first trajectory point in the current example is p5. Then, the first trajectory point p3, the last target trajectory point p5, and the trajectory point p4 between the trajectory point p3 and the trajectory point p5 can be connected in sequence to form two third sub-trajectories, namely sub-trajectory p3p4 and sub-trajectory p4p5.

[0154] After determining the third sub-trajectory, in this embodiment, the sub-trajectory parameters corresponding to determining the first trajectory point as a non-division point may be determined based on the trajectory length of the third sub-trajectory and the distance between the third sub-trajectory and the sample trajectory.

[0155] Similarly, for example, the sub-trajectory parameters in this embodiment can be the MDL introduced above, for example, Figure 5 In the example, the trajectory length of the third sub-trajectory L(H)=log2(len(p3,p4)+len(p4,p 45 ), and it can be understood that, in this case, the third sub-trajectory and the corresponding partial sample trajectory are completely overlapped, so when the first trajectory point is determined as a non-division point, L(D|H)=0.

[0156] Then, the sum of the trajectory length L(H) of the third sub-trajectory and the distance L(D|H) between the third sub-trajectory and the sample trajectory can be determined as the MDL corresponding to the second sub-trajectory, that is, the sub-trajectory parameter corresponding to determining the first trajectory point as a non-division point in this embodiment. For example, the sub-trajectory parameter corresponding to determining the first trajectory point as a non-division point can be recorded as mdl_nopar.

[0157] In actual implementation, the determination of sub-trajectory parameters is not limited to the aforementioned implementation method. For example, corresponding parameters can be added to the aforementioned MDL determination method, or the formula can be transformed by identity, and the corresponding parameters can be understood as the sub-trajectory parameters in this embodiment. Alternatively, any parameters derived from the length of the sub-trajectory and the distance between the sub-trajectory and the sample trajectory can be understood as the sub-trajectory parameters in this embodiment.

[0158] In this embodiment, the sub-trajectory parameters corresponding to the target trajectory point determined above, the sub-trajectory parameters corresponding to determining the first trajectory point as a division point, and the sub-trajectory parameters corresponding to determining the first trajectory point as a non-division point can be determined as state information.

[0159] It is understandable that when each trajectory point is traversed, corresponding status information can be determined for the traversed trajectory point according to the implementation method introduced above.

[0160] It should also be noted that the implementation described above uses the MDL standard, which considers both accuracy and simplicity, to measure the effectiveness of trajectory segmentation. In actual implementation, the selection of other metrics for measuring segmentation effectiveness can be expanded to include different metrics based on different application scenarios and actual needs. Furthermore, other metrics can be defined by adding dimensions such as time and text.

[0161] S406: Input the state information into the prediction model so that the prediction model outputs a prediction incentive.

[0162] After determining the status information, refer to Figure 6 In this embodiment, the state information can be input into the prediction model, and then the prediction model can output the predicted reward.

[0163] Reference Figure 6 , the prediction reward includes a first prediction reward for determining the first trajectory point as a division point, and a second prediction reward for determining the first trajectory point as a non-division point.

[0164] The prediction model in this embodiment can be implemented based on DQN (deep q-learning network), for example, or the specific method of implementing the prediction model can be selected and set according to actual needs, and this embodiment does not limit this.

[0165] S407 . Execute a first operation according to the first predicted stimulus and the second predicted stimulus, where the first operation is to determine the first trajectory point as a division point, or the first operation is to determine the first trajectory point as a non-division point.

[0166] In this embodiment, it can be determined what specific first operation to be performed based on the first predicted reward and the second predicted reward.

[0167] In one possible implementation, for example, an operation may be selected to be performed when the predicted stimulus is greater. For example, when the first predicted stimulus is greater than the second predicted stimulus, the first operation of determining the first trajectory point as a split point is performed. Furthermore, when the first predicted stimulus is less than or equal to the second predicted stimulus, the first operation of determining the first trajectory point as a non-split point is performed.

[0168] S408 : Determine the difference between the sub-trajectory parameter corresponding to the division point before the first operation is performed and the sub-trajectory parameter corresponding to the division point after the first operation is performed as the real excitation.

[0169] After the first operation is completed, for example, the division point before the first operation and the division point after the first operation can be obtained.

[0170] For example, Figure 5 In the example, if the first operation is to determine the first trajectory point p5 as the division point, then the division points before the first operation include p1 and p3, and the division points after the first operation include p1, p3 and p5.

[0171] Then, following the aforementioned method for determining sub-trajectory parameters, we can determine the sub-trajectory parameters corresponding to the split point before the first operation is performed, and the sub-trajectory parameters corresponding to the split point after the first operation is performed. The difference between these two sub-trajectory parameters is then used as the actual reward after the first operation is performed.

[0172] It is understandable that when traversing each trajectory point, each trajectory point corresponds to a state. This state is how the current division point is determined. For example, the state can be represented as s, and the state corresponding to the i-th trajectory point is represented as s i .

[0173] For example Figure 5In the example, when traversing to the trajectory point p5, it can be determined that the trajectory point p1 and the trajectory point p3 have been determined as the division point. Then it can be determined that the state s5 corresponding to the trajectory point p5 is that the division point includes p1 and p3.

[0174] Then, after performing the first operation on the trajectory point p5, for example, determining the trajectory point p5 as the division point, when traversing to the trajectory point p6, it can be determined that the state s6 corresponding to the trajectory point p6 is the division point including p1, p3 and p5.

[0175] Furthermore, assuming that s.total_mdl represents the sub-trajectory parameter corresponding to the partition point before the first operation is performed, and s`.total_mdl represents the sub-trajectory parameter corresponding to the partition point after the first operation is performed, then the actual reward after the first operation can be expressed as reward = s.total_mdl - s`.total_mdl.

[0176] And it can be further understood that, in the example introduced above, the sub-trajectory parameter s5`.total_mdl corresponding to the division point after the first operation is performed on p5 is actually the sub-trajectory parameter s6.total_mdl corresponding to the division point before the first operation is performed on p6.

[0177] Then the actual excitation determined after the first operation is performed on each trajectory point is accumulated and can be expressed as

[0178] R = s0.total mdl -s1.total mdl +s1·total mdl -s2·total mdl +…+s n-1 ·total mdl -s n ·total mdl =s0·total mdl -s n ·total mdl .

[0179] It is understandable that the purpose of model training, that is, reinforcement learning, is to maximize the cumulative reward, where s0.total mdl =0, then the cumulative reward is the largest, which means s n ·total mdl Therefore, by setting up such a reward mechanism to train the model, it can be ensured that the sub-trajectory parameters corresponding to the division points determined after traversing each trajectory point are minimized, and thus it can be ensured that a trajectory division scheme that minimizes the MDL can be obtained.

[0180] S409: Update the model parameters of the prediction model according to the real incentive and the predicted incentive.

[0181] In order to make the predicted stimulus output by the prediction model as close as possible to the real stimulus, in this embodiment, the model parameters of the prediction model may be updated according to the real stimulus and the predicted stimulus.

[0182] In a possible implementation, if the first operation is to determine the first trajectory point as a dividing point, then referring to Figure 6 , a first loss function value can be determined according to the real stimulus and the first predicted stimulus, and then the model parameters of the prediction model are updated according to the first loss function value.

[0183] If the first operation is to determine the first trajectory point as a non-division point, then refer to Figure 6 , the second loss function value can be determined according to the real stimulus and the first predicted stimulus, and then the model parameters of the prediction model are updated according to the second loss function value.

[0184] The model training method for trajectory segmentation provided in the embodiment of the present application obtains the target trajectory point that has been determined as a segmentation point among the previously traversed trajectory points when traversing each trajectory point, then determines the sub-trajectory parameters determined by the target trajectory point, and the sub-trajectory parameters corresponding to the traversed trajectory point being determined as a segmentation point, and the sub-trajectory parameters corresponding to the traversed trajectory point being determined as a non-segmentation point. These three sub-trajectory parameters are then determined as the state information corresponding to the currently traversed trajectory point, and the state information is then input into the prediction model so that the prediction model has sufficient information as a reference for processing, thereby effectively improving the effectiveness of the output first prediction stimulus and the second prediction stimulus. Furthermore, the first operation to be executed is then determined based on the first prediction stimulus and the second prediction stimulus, so that the operation with the larger prediction stimulus can be effectively selected for execution, thereby ensuring that the connection point determined after executing the first operation is more optimal. In this embodiment, the actual stimulus is determined based on the sub-trajectory parameters corresponding to the connection point before the first operation and the sub-trajectory parameters corresponding to the connection point after the first operation. This ensures that the sub-trajectory parameters corresponding to the connection point determined after the first operation corresponding to the last trajectory point are minimized, based on the principle of maximizing the accumulated actual stimulus. This ensures that the sub-trajectory division points are determined in a globally optimal manner. The prediction model parameters are then updated based on the actual stimulus and the predicted stimulus, effectively ensuring the accuracy and effectiveness of the predicted stimulus output by the prediction model.

[0185] The above embodiment introduces the model training process of trajectory segmentation. After the model training is completed, the trajectory segmentation can be implemented based on the trained prediction model. Therefore, based on the above introduction, the following Figure 7 and Figure 8 The trajectory division method provided in this application is further introduced in detail.

[0186] Figure 7 This is a flow chart of the trajectory division method provided in an embodiment of the present application. Figure 8 Schematic diagram of the implementation of trajectory division provided in an embodiment of the present application.

[0187] like Figure 7 As shown, the method includes:

[0188] S701: Obtain a target trajectory to be divided, where the target trajectory includes multiple trajectory points.

[0189] In this embodiment, a target trajectory to be segmented can be obtained, and multiple trajectory points can be included in the target estimation. This embodiment does not limit the specific implementation of the target trajectory, and any trajectory that needs to be segmented can be used as the target estimation in this embodiment.

[0190] S702. When traversing to the first trajectory point among multiple trajectory points, determine the predicted stimulus corresponding to the first trajectory point according to the prediction model, the predicted stimulus including a first predicted stimulus that determines the first trajectory point as a dividing point, and a second predicted stimulus that determines the first trajectory point as a non-dividing point.

[0191] During the trajectory segmentation process, each trajectory point in the target trajectory to be segmented is traversed in sequence. The processing method for each trajectory point is similar. The following uses any one of the multiple trajectory points as an example to illustrate. For example, traversing to the first trajectory point among the multiple trajectory points, where the first trajectory point can be any trajectory point.

[0192] In this embodiment, the predicted stimulus corresponding to the first trajectory point can be determined based on the prediction model, where the predicted stimulus includes a first predicted stimulus for determining the first trajectory point as a partition point, and a second predicted stimulus for determining the first trajectory point as a partition point. This implementation is similar to that in the aforementioned model training process and will not be further described here. Furthermore, in the trajectory partitioning process, all implementations other than model parameter updates are similar to those described in the aforementioned model training process and will not be further described here.

[0193] The prediction model in this embodiment is trained using the model training method described above, so the prediction stimulus output by the prediction model is guaranteed to be correct. That is, if the first prediction stimulus is relatively large, then the operation of determining the first trajectory point as a split point is guaranteed to be optimal. Furthermore, if the second prediction stimulus is relatively large, then the operation of determining the first trajectory point as a non-split point is guaranteed to be optimal.

[0194] S703: Execute a first operation according to the predicted stimulus, where the first operation is to determine the first trajectory point as a division point, or the first operation is to determine the first trajectory point as a non-division point.

[0195] Therefore, after the predicted stimulus is determined, the first operation can be performed according to the predicted stimulus.

[0196] In one possible implementation, if the first predicted stimulus is greater than the second predicted stimulus, then a first operation of determining the first trajectory point as a split point may be performed. Alternatively, if the first predicted stimulus is less than or equal to the second predicted stimulus, then a second operation of determining the first trajectory point as a non-split point may be performed.

[0197] S704 : Divide the target trajectory according to the trajectory points determined as division points to obtain multiple sub-trajectories.

[0198] After traversing each trajectory point in the target trajectory, multiple division points can be obtained.

[0199] In a possible implementation, for example, each trajectory point determined as a division point may be connected in sequence, wherein every two adjacent division points may form a sub-trajectory, thereby obtaining a plurality of sub-trajectories.

[0200] For example, reference Figure 8 understand, Figure 8 , which shows the target trajectory to be divided, including trajectory points p1 to p6.

[0201] And after the traversal is completed for each trajectory point, it is assumed that the trajectory point p1, the trajectory point p3, the trajectory point p5 and the trajectory point p6 are determined as the division points.

[0202] So, for example, we can be sure Figure 8 The first sub-trajectory is formed by connecting the trajectory point p1 and the trajectory point p3, the second sub-trajectory is formed by connecting the trajectory point p3 and the trajectory point p5, and the third sub-trajectory is formed by connecting the trajectory point p5 and the trajectory point p6.

[0203] Or in another possible implementation, for example, each trajectory point determined as a division point can be divided into a sub-trajectory by dividing the partial trajectory in the original target trajectory, thereby obtaining multiple sub-trajectories.

[0204] For example, reference Figure 8 understand, Figure 8 , which shows the target trajectory to be divided, including trajectory points p1 to p6.

[0205] And after the traversal is completed for each trajectory point, it is assumed that the trajectory point p1, the trajectory point p3, the trajectory point p5 and the trajectory point p6 are determined as the division points.

[0206] So, for example, we can be sure Figure 8 The partial trajectory between trajectory point p1 and trajectory point p3 in the target trajectory illustrated in FIG is the first sub-trajectory, the partial trajectory between trajectory point p3 and trajectory point p5 in the target trajectory is the second sub-trajectory, and the partial trajectory between trajectory point p5 and trajectory point p6 in the target trajectory is the third sub-trajectory.

[0207] The trajectory division method provided in the embodiment of the present application traverses each trajectory point in the target trajectory, and then predicts a first prediction stimulus that determines the trajectory point as a connection point based on a trained prediction model, and a second prediction stimulus that determines the trajectory point as a non-connection point. Then, based on the first prediction stimulus and the second prediction stimulus, it is determined whether to determine the traversed trajectory point as a connection point. After traversing each trajectory point, multiple trajectory points can be obtained, and then the sub-trajectory division results are determined based on the trajectory points. In this way, trajectory division can be achieved simply and effectively, while ensuring that the simplicity and accuracy of the divided sub-trajectories are relatively good.

[0208] It is understandable that the model training process introduced above can be understood as reinforcement learning. Therefore, reinforcement learning is used in this application to solve the problem of dividing the trajectory into sub-trajectories. Because reinforcement learning is data-driven, it can adapt to the different dynamic changes of the problem according to the intrinsic characteristics of the data. In addition, the above-mentioned traversal of each trajectory point in sequence, performing the corresponding operation on each trajectory point, and determining the state after the operation as the initial state corresponding to the next trajectory point can actually be understood as modeling the environment as a Markov decision process (MDP), where the core problem of MDP is to find an optimal strategy for the agent, which is used to guide the agent in a specific state to choose what action to maximize the cumulative reward.

[0209] Corresponding to the implementation method introduced above, that is, the prediction model can output accurate prediction incentives, and then the device can choose to execute the operation that maximizes the cumulative reward to ensure that after traversing each trajectory point, it can select the operation that maximizes the reward according to the current connection point state, thereby ensuring.

[0210] In summary, the trajectory division method provided by this application regards trajectory division as a sequential decision-making process. The device can be considered as the executor of the action of determining the trajectory point as the division point. It then scans each trajectory point in sequence and makes a decision on whether to divide it at each trajectory point. Through learning, it can obtain the optimal trajectory division strategy. Compared with other sub-trajectory division methods, this method does not require the manual formulation of complex corresponding algorithms and the selection of corresponding parameters. Instead, it utilizes the data-driven characteristics of reinforcement learning. It can learn the division strategy based on the characteristics and dynamic changes of different data, thereby achieving a simple and effective trajectory division while ensuring the effect of trajectory division.

[0211] Figure 9 This is a schematic diagram of the structure of the model training device for trajectory division provided in the embodiment of the present application. Figure 9 As shown, the device 90 includes: an acquisition module 901 , a determination module 902 , a processing module 903 and an update module 904 .

[0212] An acquisition module 901 is configured to acquire a sample trajectory, wherein the sample trajectory includes a plurality of trajectory points;

[0213] a determination module 902 configured to, when traversing to a first trajectory point among the plurality of trajectory points, determine, based on a prediction model, a predicted stimulus corresponding to the first trajectory point, the predicted stimulus including a first predicted stimulus for determining the first trajectory point as a partition point and a second predicted stimulus for determining the first trajectory point as a non-partition point, wherein adjacent partition points determined among the plurality of trajectory points are used to constitute a sub-trajectory corresponding to the sample trajectory;

[0214] A processing module 903 is configured to perform a first operation according to the first prediction stimulus and the second prediction stimulus, where the first operation is to determine the first trajectory point as a division point, or the first operation is to determine the first trajectory point as a non-division point;

[0215] The determining module 902 is configured to determine a real stimulus corresponding to the execution of the first operation based on sub-trajectory parameters corresponding to the division point before the execution of the first operation and sub-trajectory parameters corresponding to the division point after the execution of the first operation, wherein the sub-trajectory parameters are used to indicate the length of the sub-trajectory and the difference between the sub-trajectory and the sample trajectory;

[0216] The updating module 904 is configured to update the model parameters of the prediction model according to the real stimulus and the predicted stimulus.

[0217] In one possible design, the determining module 902 is specifically configured to:

[0218] Acquire a target trajectory point determined as a dividing point from the plurality of trajectory points;

[0219] determining state information based on the target trajectory point and the first trajectory point, the state information including a sub-trajectory parameter corresponding to the target trajectory point, a sub-trajectory parameter corresponding to determining the first trajectory point as a dividing point, and a sub-trajectory parameter corresponding to determining the first trajectory point as a non-dividing point;

[0220] The state information is input into the prediction model so that the prediction model outputs the predicted stimulus.

[0221] In one possible design, the determining module 902 is specifically configured to:

[0222] determining at least one first sub-trajectory according to the target trajectory point, and determining a sub-trajectory parameter corresponding to the target trajectory point according to a trajectory length of the at least one first sub-trajectory and a distance between the at least one first sub-trajectory and the sample trajectory;

[0223] determining a second sub-trajectory based on the first trajectory point and a last target trajectory point among the plurality of target trajectory points, and determining a sub-trajectory parameter corresponding to determining the first trajectory point as a dividing point based on a trajectory length of the second sub-trajectory and a distance between the second sub-trajectory and the sample trajectory;

[0224] At least one third sub-trajectory is determined based on the first trajectory point, the last target trajectory point among multiple target trajectory points, and the trajectory points between the first trajectory point and the last target trajectory point. Based on the trajectory length of the at least one third sub-trajectory and the distance between the at least one third sub-trajectory and the sample trajectory, the sub-trajectory parameters corresponding to determining the first trajectory point as a non-division point are determined.

[0225] In one possible design, the processing module 903 is specifically configured to:

[0226] If the first predicted stimulus is greater than the second predicted stimulus, performing a first operation of determining the first trajectory point as a division point; or

[0227] If the first predicted stimulus is less than or equal to the second predicted stimulus, a first operation of determining the first trajectory point as a non-division point is performed.

[0228] In one possible design, the determining module 902 is specifically configured to:

[0229] A difference between a sub-trajectory parameter corresponding to a division point before performing the first operation and a sub-trajectory parameter corresponding to a division point after performing the first operation is determined as the real excitation.

[0230] In one possible design, the updating module 904 is specifically configured to:

[0231] If the first operation is to determine the first trajectory point as a division point, determining a first loss function value according to the real stimulus and the first predicted stimulus, and updating the model parameters of the prediction model according to the first loss function value; or

[0232] If the first operation is to determine the first trajectory point as a non-division point, a second loss function value is determined according to the real stimulus and the first predicted stimulus, and the model parameters of the prediction model are updated according to the second loss function value.

[0233] The device provided in this embodiment can be used to execute the technical solution of the above method embodiment. Its implementation principle and technical effects are similar and will not be described in detail in this embodiment.

[0234] Figure 10 This is a schematic diagram of the structure of the track division device provided in the embodiment of the present application. Figure 10 As shown, the device 100 includes: an acquisition module 1001 , a determination module 1002 and a processing module 1003 .

[0235] An acquisition module 1001 is configured to acquire a target trajectory to be divided, wherein the target trajectory includes a plurality of trajectory points;

[0236] a determination module 1002 for determining, when traversing to a first trajectory point among the plurality of trajectory points, a predicted stimulus corresponding to the first trajectory point based on a prediction model, the predicted stimulus including a first predicted stimulus for determining the first trajectory point as a partition point and a second predicted stimulus for determining the first trajectory point as a non-partition point, the prediction model being trained according to the model training method described in the above embodiment;

[0237] A processing module 1003 is configured to perform a first operation according to the predicted stimulus, where the first operation is to determine the first trajectory point as a division point, or the first operation is to determine the first trajectory point as a non-division point;

[0238] The processing module is further configured to divide the target trajectory according to the trajectory points determined as dividing points to obtain a plurality of sub-trajectories.

[0239] In one possible design, the determining module 1002 is specifically configured to:

[0240] Acquire a target trajectory point determined as a dividing point from the plurality of trajectory points;

[0241] determining state information based on the target trajectory point and the first trajectory point, the state information including a sub-trajectory parameter corresponding to the target trajectory point, a sub-trajectory parameter corresponding to determining the first trajectory point as a dividing point, and a sub-trajectory parameter corresponding to determining the first trajectory point as a non-dividing point;

[0242] The state information is input into the prediction model so that the prediction model outputs the predicted stimulus.

[0243] The device provided in this embodiment can be used to execute the technical solution of the above method embodiment. Its implementation principle and technical effects are similar and will not be described in detail in this embodiment.

[0244] Figure 11 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application is shown in FIG. Figure 11 As shown, the electronic device 110 of this embodiment includes: a processor 1101 and a memory 1102;

[0245] Memory 1102, for storing computer-executable instructions;

[0246] The processor 1101 is configured to execute computer-executable instructions stored in the memory to implement the model training method for trajectory segmentation and the various steps performed by the trajectory segmentation method in the above embodiment. For details, please refer to the relevant description in the above method embodiment.

[0247] Optionally, the memory 1102 may be independent or integrated with the processor 1101 .

[0248] When the memory 1102 is independently provided, the electronic device further includes a bus 1103 for connecting the memory 1102 and the processor 1101 .

[0249] An embodiment of the present application also provides a computer-readable storage medium, which stores computer-executable instructions. When a processor executes the computer-executable instructions, the model training method and trajectory division method of the trajectory division performed by the electronic device as described above are implemented.

[0250] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the modules is merely a logical function division. In actual implementation, there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, device or module, which can be electrical, mechanical or other forms.

[0251] The above-mentioned integrated module implemented in the form of a software functional module can be stored in a computer-readable storage medium. The above-mentioned software functional module is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to perform some steps of the method described in various embodiments of the present application.

[0252] It should be understood that the processor described above may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), or application-specific integrated circuits (ASICs). A general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the present invention may be directly executed by a hardware processor or by a combination of hardware and software modules within the processor.

[0253] The memory may include a high-speed RAM memory, and may also include non-volatile storage NVM, such as at least one disk memory, and may also be a USB flash drive, a mobile hard disk, a read-only memory, a magnetic disk or an optical disk.

[0254] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. Buses can be classified into address buses, data buses, and control buses. For ease of illustration, the buses in the drawings of this application are not limited to just one bus or just one type of bus.

[0255] The storage medium may be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium may be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0256] Those skilled in the art will appreciate that all or part of the steps in the above-described method embodiments can be implemented using hardware associated with program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0257] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A trajectory segmentation model training method, comprising: Obtain a sample trajectory, wherein the sample trajectory includes multiple trajectory points; the trajectory refers to a series of trajectory points sequenced in time generated by the moving object during the movement process, and the trajectory point sequence includes n trajectory points , trajectory point Expressed as , Represents the spatial position of the trajectory point, Indicates longitude, represents latitude, t represents the trajectory point The sampling time is, i is a positive integer less than or equal to n; the length of the trajectory can be represented by the number of trajectory points; the moving object is a pedestrian, a vehicle or an aircraft; When traversing to a first trajectory point among the plurality of trajectory points, a target trajectory point that has been determined as a division point is obtained from the plurality of trajectory points; state information is determined based on the target trajectory point and the first trajectory point, the state information including a sub-trajectory parameter corresponding to the target trajectory point, a sub-trajectory parameter corresponding to determining the first trajectory point as a division point, and a sub-trajectory parameter corresponding to determining the first trajectory point as a non-division point; the state information is input into a prediction model so that the prediction model outputs a prediction stimulus, the prediction stimulus including a first prediction stimulus for determining the first trajectory point as a division point and a second prediction stimulus for determining the first trajectory point as a non-division point, wherein adjacent division points determined from the plurality of trajectory points are used to constitute a sub-trajectory corresponding to the sample trajectory; If the first predicted stimulus is greater than the second predicted stimulus, a first operation of determining the first trajectory point as a division point is performed; or if the first predicted stimulus is less than or equal to the second predicted stimulus, a first operation of determining the first trajectory point as a non-division point is performed, the first operation being to determine the first trajectory point as a division point, or the first operation being to determine the first trajectory point as a non-division point; Determining a difference between a sub-trajectory parameter corresponding to a division point before performing the first operation and a sub-trajectory parameter corresponding to a division point after performing the first operation as a true stimulus, wherein the sub-trajectory parameter is used to indicate a length of the sub-trajectory and a difference between the sub-trajectory and the sample trajectory; The model parameters of the prediction model are updated according to the real stimulus and the predicted stimulus.

2. The method according to claim 1, characterized in that The determining of state information according to the target trajectory point and the first trajectory point includes: determining at least one first sub-trajectory according to the target trajectory point, and determining a sub-trajectory parameter corresponding to the target trajectory point according to a trajectory length of the at least one first sub-trajectory and a distance between the at least one first sub-trajectory and the sample trajectory; determining a second sub-trajectory based on the first trajectory point and a last target trajectory point among the plurality of target trajectory points, and determining a sub-trajectory parameter corresponding to determining the first trajectory point as a dividing point based on a trajectory length of the second sub-trajectory and a distance between the second sub-trajectory and the sample trajectory; At least one third sub-trajectory is determined based on the first trajectory point, the last target trajectory point among multiple target trajectory points, and the trajectory points between the first trajectory point and the last target trajectory point. Based on the trajectory length of the at least one third sub-trajectory and the distance between the at least one third sub-trajectory and the sample trajectory, the sub-trajectory parameters corresponding to determining the first trajectory point as a non-division point are determined.

3. The method according to claim 1, characterized in that The updating of the model parameters of the prediction model according to the real stimulus and the predicted stimulus includes: If the first operation is to determine the first trajectory point as a division point, determining a first loss function value according to the real stimulus and the first predicted stimulus, and updating the model parameters of the prediction model according to the first loss function value; or If the first operation is to determine the first trajectory point as a non-division point, a second loss function value is determined according to the real stimulus and the first predicted stimulus, and the model parameters of the prediction model are updated according to the second loss function value.

4. A trajectory division method, characterized in that: include: Get the target trajectory to be divided, the target trajectory includes multiple trajectory points; for any trajectory point, the trajectory point is represented as , Represents the spatial position of the trajectory point, Indicates longitude, represents the latitude, and t represents the sampling time of the trajectory point; When traversing to a first trajectory point among the multiple trajectory points, determining, according to a prediction model, a predicted stimulus corresponding to the first trajectory point, the predicted stimulus including a first predicted stimulus for determining the first trajectory point as a partition point and a second predicted stimulus for determining the first trajectory point as a non-partition point, the prediction model being trained according to the model training method for trajectory partitioning according to any one of claims 1 to 3; performing a first operation according to the predicted stimulus, wherein the first operation is determining the first trajectory point as a division point, or the first operation is determining the first trajectory point as a non-division point; The target trajectory is divided according to the trajectory points determined as division points to obtain a plurality of sub-trajectories.

5. A trajectory segmentation model training device, characterized in that: include: An acquisition module is used to acquire a sample trajectory, wherein the sample trajectory includes multiple trajectory points; the trajectory refers to a series of trajectory points sequenced in time generated by a moving object during its movement; the trajectory point sequence includes n trajectory points , trajectory point Expressed as , Represents the spatial position of the trajectory point, Indicates longitude, represents latitude, t represents the trajectory point The sampling time is, i is a positive integer less than or equal to n; the length of the trajectory can be represented by the number of trajectory points; the moving object is a pedestrian, a vehicle or an aircraft; a determination module configured to, when traversing to a first trajectory point among the plurality of trajectory points, obtain a target trajectory point determined as a partition point among the plurality of trajectory points; determine state information based on the target trajectory point and the first trajectory point, the state information including a sub-trajectory parameter corresponding to the target trajectory point, a sub-trajectory parameter corresponding to determining the first trajectory point as a partition point, and a sub-trajectory parameter corresponding to determining the first trajectory point as a non-partition point; input the state information into a prediction model so that the prediction model outputs a prediction stimulus, the prediction stimulus including a first prediction stimulus for determining the first trajectory point as a partition point and a second prediction stimulus for determining the first trajectory point as a non-partition point, wherein adjacent partition points determined among the plurality of trajectory points are used to constitute a sub-trajectory corresponding to the sample trajectory; a processing module, configured to, if the first predicted stimulus is greater than the second predicted stimulus, perform a first operation of determining the first trajectory point as a division point; or, if the first predicted stimulus is less than or equal to the second predicted stimulus, perform a first operation of determining the first trajectory point as a non-division point, wherein the first operation is determining the first trajectory point as a division point, or the first operation is determining the first trajectory point as a non-division point; The determining module is configured to determine a difference between a sub-trajectory parameter corresponding to a division point before performing the first operation and a sub-trajectory parameter corresponding to a division point after performing the first operation as a true stimulus, wherein the sub-trajectory parameter is used to indicate a length of the sub-trajectory and a difference between the sub-trajectory and the sample trajectory; An updating module is used to update the model parameters of the prediction model according to the real stimulus and the predicted stimulus.

6. A trajectory division device, characterized in that: include: An acquisition module, configured to acquire a target trajectory to be divided, wherein the target trajectory includes a plurality of trajectory points; a determination module, configured to, when traversing to a first trajectory point among the plurality of trajectory points, determine, based on a prediction model, a predicted stimulus corresponding to the first trajectory point, the predicted stimulus including a first predicted stimulus for determining the first trajectory point as a partition point and a second predicted stimulus for determining the first trajectory point as a non-partition point, the prediction model being trained using the trajectory partitioning model training method according to claims 1 to 3; a processing module, configured to perform a first operation according to the predicted stimulus, wherein the first operation is determining the first trajectory point as a division point, or the first operation is determining the first trajectory point as a non-division point; The processing module is further configured to divide the target trajectory according to the trajectory points determined as dividing points to obtain a plurality of sub-trajectories.

7. An electronic device, characterized in that: include: Memory, used to store programs; A processor is configured to execute the program stored in the memory; when the program is executed, the processor is configured to execute the method according to any one of claims 1 to 3 or claim 4.

8. A computer-readable storage medium, characterized in that The method comprises instructions which, when executed on a computer, cause the computer to execute the method according to any one of claims 1 to 3 or claim 4.

9. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 3 or claim 4 is implemented.

Images