Multi-possibility track prediction method and system
Through the multi-possible track prediction model based on GAN, the track prediction data set is extracted and constructed, and the problem of insufficient credibility of the activity trend and prediction results in the prior art that are difficult to predict the target's future possibilities is solved, and the credibility of the prediction results is achieved. Accurate multi-possibility prediction and adaptive prediction capabilities for tracks.
Patent Information
- Application Number
- CN202411442172.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-16
- Publication Date
- 2025-02-18
AI Technical Summary
The existing track prediction methods are difficult to effectively deal with the activity trends of multiple possibilities in the future of the target, and the credibility and reliability of the prediction results are insufficient, so they cannot adapt to different data situations to make long-term and short-term predictions.
Using a multi-possible track prediction model based on GAN, the track prediction data set is constructed by extracting the key routes and route connection relationships in the historical track data, and multiple rounds of confrontation training are carried out through custom loss functions to generate multiple lines. A future track with uncertain lengths and outputs its confidence and credibility for each track.
Accurate prediction of the target's possible tracks in the future is achieved, which improves the credibility and reliability of the prediction results, and can adapt to different data situations to make predictions in the long and short periods, providing better decision support.
Smart Images
Figure CN119479375A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of intelligent analysis, big data mining and track prediction, and in particular to a multi-possibility track prediction method and system. Background Art
[0002] Track prediction is a key technology in aviation and maritime traffic. It aims to effectively track and predict aircraft or ships. It is of great significance for intelligent traffic management, intelligent route planning, and track anomaly warning. Existing track prediction research mainly includes track prediction methods based on physical models, statistical models, and machine learning or deep learning models. Different track prediction methods have their own advantages and disadvantages and applicable scenarios. How to choose a suitable track prediction method and improve the accuracy and reliability of the track prediction model is still a hot topic and difficulty in current research.
[0003] The existing research on track prediction has the following shortcomings:
[0004] 1. Multiple possibilities of track prediction results: In practical applications, moving targets have multiple different possible activity trends in the future multi-scale space-time range. Existing track prediction models generally only give the best prediction results, ignoring the possibility of other activity trends of the target in the future, which may affect the analyst's next decision-making plan.
[0005] 2. The credibility of the track prediction results is unclear: In different data scenarios, the distribution of track data is different and the data quality varies greatly. The track prediction model always gives the current best prediction result, but there is a lack of accurate evaluation and analysis of the reliability or credibility of the prediction results.
[0006] 3. The output length of the track prediction result is fixed: Under different target types or different geographical conditions, the target's movement characteristics and change patterns are different, resulting in different predictability of the target's future track. For example, when the target activity has a long-term regularity, long-term prediction can be made, while when the target movement trend changes rapidly and complexly, only short-term prediction can be made. The existing track prediction model cannot make adaptive long-term and short-term predictions for different data situations.
[0007] On the other hand, in ground rail transit, there are more relatively mature pedestrian or vehicle trajectory prediction studies, which have certain reference significance for trajectory prediction research. When migrating and transforming related technologies, it is necessary to focus on the differences between trajectory and ground trajectory at the data feature level:
[0008] 1. Data volume and data quality are different: Compared with pedestrian and vehicle trajectory data, the data volume of track data is larger. In the track prediction scenario, the same target generates more track data points from the start to the end of the activity. At the same time, the data quality of track data is worse, with many problems such as missing data, noise, and uncertainty.
[0009] 2. Track data lacks accurate road network information: Rail transit has relatively clear road network information, and the movement of pedestrians or vehicles is subject to road traffic rules. In the field of track research, route planning is relatively complex, and it is difficult to obtain accurate route information and maritime or aviation traffic rules. At the same time, track data may deviate from known routes and may contain free navigation or data anomalies. Summary of the invention
[0010] In view of the above-mentioned problems, the present invention aims to provide a fast and robust multi-possibility track prediction method and system with accurate track prediction, good algorithm model robustness, the ability to predict multiple future tracks of the target, and the ability to provide confidence and credibility for the prediction results.
[0011] The present invention provides a method for predicting multiple possible flight paths, comprising the following steps:
[0012] S01, extracting key routes and connection relationships between routes: reading historical track data in the historical track database, performing data preprocessing on the historical track data, and extracting key routes and connection relationships between routes in the historical track data;
[0013] S02, constructing a track prediction data set: dividing the historical track data in the historical track database into a track observation sequence and a real predicted track sequence, and constructing a track prediction data set in combination with the key routes and the connection relationship between the routes;
[0014] S03, construct a multi-possible trajectory prediction model: construct a multi-possible trajectory prediction model based on GAN;
[0015] S04, training a multi-possible track prediction model: performing multiple rounds of adversarial training on the GAN-based multi-possible track prediction model using the track prediction dataset through a custom loss function;
[0016] S05, applying a multi-possibility trajectory prediction model: using the trained multi-possibility trajectory prediction model, the future trajectory of the target is predicted in real time, multiple future trajectories of indefinite length are output, and the confidence and credibility of each predicted trajectory are output respectively.
[0017] Furthermore, in step S01, extracting key routes and connection relationships between routes includes the following sub-steps:
[0018] S01.01, read historical track data from the historical track database; each track is represented by a track point sequence, and the track points in the track point sequence include longitude, latitude, altitude, speed, heading and time;
[0019] S01.02, cleaning track data: first remove invalid values and abnormal values in historical track data, and then use Kalman filter algorithm to smooth the track data;
[0020] S01.03, Track Compression: Through track compression, the historical track data is divided into different straight track segments as candidate key routes;
[0021] S01.04, extracting key routes: clustering the candidate key routes using a clustering algorithm, and outputting the clustering results as key routes; in the process of clustering the candidate key routes, equidistant interpolation is performed on the candidate key routes according to the set distance threshold based on the head and tail positions, and each route is represented as a coordinate sequence setCoor;
[0022] S01.05, extract key routes and connection relationships between routes: based on the key routes, traverse the candidate key routes in the historical track database again, perform key route matching on the candidate key routes, count and supplement the attribute characteristics PropR of the key routes and the historical connection status ConnR with other routes, so that each key route is finally represented as R = {setCoor, PropR, ConnR}.
[0023] Furthermore, in step S02, constructing a track prediction data set includes the following sub-steps:
[0024] S02.01, performing equidistant sampling on the historical track data after data cleaning in step S01.02 at a fixed time interval T;
[0025] S02.02, based on the equidistant sampling data, by generating and dividing the historical observation track sequence and the real prediction track sequence, and combining the key routes and the connection relationship between the routes, construct a track prediction data set;
[0026] Furthermore, an asymmetric historical observation track sequence and a true prediction track sequence are used:
[0027] S02.02.01, when constructing a historical observation track sequence, according to the different types of targets, select the activity time threshold and distance threshold of other targets that have an impact on the activity trend of the current target, as well as the distance threshold of the surrounding key routes, and obtain the surrounding target tracks and surrounding key routes of the current target based on the threshold screening; wherein the track information includes longitude, latitude, altitude, speed and heading, and the route information includes the attribute characteristics of the route and the connection relationship between routes;
[0028] S02.02.02, the constructed true predicted track sequence only includes the future latitude and longitude information of the target.
[0029] Furthermore, in step S03, constructing a multi-possibility track prediction model includes the following sub-steps:
[0030] S03.01, the multi-possible track prediction model based on GAN includes two adversarial networks: track generator and track discriminator; the two adversarial networks share a track encoder;
[0031] S03.02, the track encoder Enc is used to globally encode the track information of the input target and surrounding targets, the surrounding routes and the connection relationship between the routes, and its output is represented by Enc(X);
[0032] S03.03, the track generator G is used to predict the track based on the historical observation track generation model, which consists of a track encoder Enc and a track decoder Dec. The output of the track generator G is represented by Y G = Dec(Enc(X));
[0033] S03.04, the track discriminator D is used to distinguish the real track from the model-generated track. It consists of a track encoder Enc and a binary classification MLP. The input of the track discriminator D is Among them, Y i is the result of splicing the historical observation track sequence and the real prediction track sequence. is the concatenation result of the historical observed track sequence and the model predicted track sequence; the output of the track discriminator D is:
[0034] Y D =MLP(Enc(X D ))
[0035] S03.05, the loss function is defined as:
[0036] L=min G max D L GAN (G,D)+λ1L1(G)+λ2L2(G)+λ3L3(G)
[0037] Among them, L GAN (G, D) is a universal joint adversarial loss function used to train the two adversarial networks, track generator G and track discriminator D; λ1L1(G)+λ2L2(G)+λ3L3(G) is the loss function for training the track generator, and λ1, λ2, λ3 are used to adjust the weights of the loss functions of each sub-module in the overall loss function.
[0038] Furthermore, the specific encoding steps of the track encoder Enc are as follows:
[0039] S03.02.01, first use the same Transformer Encoder to encode the track point sequences generated by the current target and the surrounding targets respectively, and then use the global graph GNN G1 Encode the interaction relationship between different track point sequences to obtain the current target track point sequence code EncTr and the global track point sequence information code EncEnvTr;
[0040] S03.02.02, using the hierarchical graph neural network VectorNet to encode key routes and their connection information: first use the subgraph network GNN L For each critical route R i Encode and then use the global graph GNN G2 Encode the connection and interaction information between different routes, and output the global route information code EncEnvR in the current scenario;
[0041] S03.02.03, connect the coding information in step S03.02.01 and step S03.02.02, fuse the coding information through the MLP model, and obtain the total output Enc(X) = MLP([EncTr, EncEnvTr, EncEnvR]) of the track encoder.
[0042] Furthermore, the specific steps of decoding and generating the future track in the track generator include:
[0043] S03.03.01, taking the latest track point of the current target as the center, divide the surrounding space into grids as candidate destinations, and predict the destination based on the candidate destinations; the input of this step is the track encoder output Enc(X), the architecture used is the MLP model, and the loss function is:
[0044] L1(G)=L cls (tar)+L reg (tar)
[0045] Among them, L cls (tar) is the logarithmic loss function, which constrains the model to select the location closest to the true destination from the candidate destinations. reg (tar) is the Huber loss function, which is used to perform regression prediction on the distance between the candidate destination and the real destination, thereby realizing the prediction of the real destination;
[0046] S03.03.02, using the Transformer Decoder model, taking the track encoder output Enc(X) and the target position Loc as input, outputs the most likely future track trPred; the loss function of this step is:
[0047] L2(G)=L reg (trPred)
[0048] Among them, L reg (trPred) is the Huber loss function, which is used to perform regression prediction on the future track. The prediction result of this step is related to the predicted destination in step S03.03.01. The real destination will be brought in for training during the training process.
[0049] S03.03.03, in predicting multiple future tracks, each track is screened and scored through the MLP model, the possibility of each track is predicted, and the tracks with the highest probability are selected as the model output; the loss function of this step is:
[0050] L3(G)=L cls (trPred)
[0051] Among them, L cls (trPred) is a logarithmic loss function, which is used to select the track closest to the true predicted track from multiple candidate tracks.
[0052] Furthermore, in step S04, training the multi-possibility track prediction model includes the following sub-steps:
[0053] S04.01: Read the historical observation track sequence and the real prediction track sequence from the track prediction data set;
[0054] S04.02: Input the historical observation track sequence into the track generator to obtain the model predicted track sequence;
[0055] S04.03: During the training process, the lengths of the real predicted track sequence and the model predicted track sequence are randomly sampled at each iteration; the historical observed track sequence and the real predicted track sequence cut according to the randomly sampled length are concatenated to obtain Y i , the historical observation track sequence is concatenated with the model prediction track sequence after being truncated by length to obtain Y i and Input into the track discriminator to calculate the score results of the two track sequences;
[0056] S04.04: Substitute the model prediction results of the track sequence of the track generator and the scoring results of the two track sequences of the track discriminator into the loss function to calculate the loss, select the optimization algorithm for gradient descent, and update the weight parameters of the model;
[0057] S04.05: Repeat steps S04.02-S04.04 until the conditions for completing model training are met, and a trained multi-possibility track prediction model is obtained.
[0058] Furthermore, in step S05, applying the multi-possibility track prediction model includes the following sub-steps:
[0059] S05.01: perform data preprocessing on the observed track sequence of the target being tracked in real time;
[0060] S05.02: Input the target observation track sequence into the track generator in the trained multi-possibility track prediction model to obtain multiple model predicted tracks and the confidence of each track;
[0061] S05.03: For each model-predicted track, cut it into different lengths, concatenate it with the observed track sequence, and input it into the track discriminator in the trained multi-possibility track prediction model to obtain the credibility score result;
[0062] S05.04: Combining step S05.02 and step S05.03, the trained multi-possibility track prediction model adaptively outputs all possible indefinite-length tracks within the credibility threshold range and their respective confidence levels and credibility levels for the current prediction target.
[0063] The present invention also provides a multi-possibility track prediction system for implementing the multi-possibility track prediction method; the multi-possibility track prediction system comprises:
[0064] An extraction module is used to read the historical track data in the historical track database, perform data preprocessing on the historical track data, and extract key routes and connection relationships between routes in the historical track data;
[0065] A data set construction module is used to divide the historical track data in the historical track database into a track observation sequence and a real predicted track sequence, and to construct a track prediction data set in combination with the key routes and the connection relationship between the routes;
[0066] Model building module, used to build a multi-possible trajectory prediction model based on GAN;
[0067] A model training module, used for performing multiple rounds of adversarial training on a GAN-based multi-possible track prediction model using the track prediction dataset through a custom loss function;
[0068] The model application module is used to use the trained multi-possible track prediction model to make real-time predictions of the target's future track, output multiple future tracks of indefinite length, and output the confidence and credibility for each predicted track.
[0069] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:
[0070] 1. Output multiple possible future tracks of the target and their respective confidence levels: The present invention can not only output the best predicted track of the target at present, but also output other possible tracks of the target, and output the confidence level for each possible track, which helps analysts to fully consider all possibilities of the target's activities.
[0071] 2. Evaluate the credibility of the target predicted track output by the model: The track discriminator in the GAN-based multi-possible track prediction model proposed in the present invention evaluates and analyzes multiple possible tracks of the output target in the future, and gives the credibility of each output track, which can be used as an auxiliary basis for decision-making and enhance the practical application value of the model.
[0072] 3. Adaptively predict the target in the long and short time: The present invention uses a track identifier to evaluate the output track of the track prediction model, and selects the longest track that meets the credibility requirements as the predicted track of the target, thereby realizing adaptive long and short time prediction of the target, which can improve the accuracy of the prediction and the system performance, and provide better services for analysis and decision makers.
[0073] 4. The present invention can be used for the prediction of the activity tracks of moving targets such as airplanes, ships, cars, pedestrians, etc., the auxiliary identification of moving targets, etc. It can also be used for abnormal state detection and early warning, etc., and has strong practical engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS
[0074] Figure 1 The present invention provides a flowchart of a method for predicting multiple possible flight paths.
[0075] Figure 2 The figure is an architecture diagram of a multi-possibility track prediction model provided in an embodiment of the present invention.
[0076] Figure 3 The diagram is an architecture diagram of the track encoder Enc module provided in an embodiment of the present invention.
[0077] Figure 4 This is a flow chart of a track decoder Dec module provided in an embodiment of the present invention.
[0078] Figure 5 The present invention provides an architecture diagram of a multi-possibility trajectory prediction system. DETAILED DESCRIPTION
[0079] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.
[0080] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0081] Example
[0082] like Figure 1 As shown, this embodiment proposes a multi-possibility track prediction method, comprising the following steps:
[0083] S01, extracting key routes and connection relationships between routes: reading historical track data in the historical track database, performing data preprocessing on the historical track data, and extracting key routes and connection relationships between routes in the historical track data through algorithms such as track compression and track clustering.
[0084] S02, constructing a track prediction data set: dividing the historical track data in the historical track database into a historical observation track sequence and a real prediction track sequence by time equidistant sampling and other methods, and constructing a track prediction data set in combination with the key routes and the connection relationship between the routes.
[0085] S03, construct a multi-possible track prediction model: construct a multi-possible track prediction model based on GAN (Generative Adversarial Network), which specifically includes a track encoder, a track generator, and a track discriminator.
[0086] S04, training a multi-possible track prediction model: using a custom loss function and utilizing the track prediction dataset to perform multiple rounds of adversarial training on the GAN-based multi-possible track prediction model.
[0087] S05, applying a multi-possibility trajectory prediction model: using the trained multi-possibility trajectory prediction model, the future trajectory of the target is predicted in real time, multiple future trajectories of indefinite length are output, and the confidence and credibility of each predicted trajectory are output respectively.
[0088] In step S01, extracting key routes and the connection relationship between routes includes the following sub-steps:
[0089] S01.01, read massive historical track data from the historical track database; each track is represented as a sequence of track points:
[0090] setP={P1,P2,…,P I}
[0091] Among them, the track point P i =(lon i ,lat i ,alt i ,speed i ,heading i ,t i ), i = 1, 2, ..., I, I is the number of track points in the track point sequence; track point P i Including longitude lon, latitude lat, altitude alt, speed speed, heading heading and time t.
[0092] S01.02, track data cleaning: first, invalid values and abnormal values in the historical track data are eliminated to ensure the accuracy and completeness of the historical track data, and then the Kalman filter algorithm is used to smooth the track data to reduce and remove the noise data in the historical track data.
[0093] S01.03, track compression: In this embodiment, the Douglas-Peucker algorithm is used to implement track compression; through track compression, the historical track data is divided into different straight track segments as candidate key routes.
[0094] S01.04, extract key routes. In this embodiment, the DBSCAN clustering algorithm is used to cluster the candidate key routes, and the clustering results are output as key routes. In particular, in the process of clustering the candidate key routes, this embodiment performs equidistant interpolation on the candidate key routes according to the set distance threshold based on the head and tail positions, and represents each route as a coordinate sequence setCoor={(lon1,lat1,alt1),(lon2,lat2,alt2),…(lon Nr ,lat Nr ,alt Nr )}, Nr represents the number of coordinate points after interpolation, and then the DTW distance is used to measure the sequence similarity between different routes.
[0095] S01.05, extract key routes and the connection relationship between routes; based on the key routes, traverse the compressed tracks in the historical track database again, match the historical compressed tracks to key routes, count and supplement the attribute characteristics PropR of the key routes (such as historical occurrence frequency, degree of deviation of the matching track relative to the key route, etc.), and the historical connection status ConnR with other routes (such as whether there is a connection, the order of connection, the frequency of historical connection, etc.), thus, each key route is finally represented as R = {setCoor, PropR, ConnR}.
[0096] In step S02, constructing a track prediction data set includes the following sub-steps:
[0097] S02.01, time equidistant sampling: the historical track data after data cleaning in step S01.02 is sampled equidistantly at a fixed time interval T, which is conducive to better capturing the track correlation and change trend of the previous and next time points, and improving the accuracy and stability of the prediction.
[0098] S02.02, based on the equidistant sampling data, by generating and dividing the historical observation track sequence and the real prediction track sequence, and combining the key routes and the connection relationship between the routes, construct a track prediction data set. Specifically, the present invention adopts asymmetric historical observation track sequence and real prediction track sequence:
[0099] S02.02.01, the purpose of the present invention is to make full use of the track information of the current target and other surrounding targets and the surrounding key route information. When constructing a historical observation track sequence, according to the different types of targets, the activity time threshold and distance threshold of other targets that have an impact on the activity trend of the current target, as well as the distance threshold of the surrounding key routes are selected, and the tracks of the surrounding targets and the surrounding key routes of the current target are obtained according to the threshold screening. Among them, the track information includes longitude, latitude, altitude, speed and heading, etc., and the route information includes the attribute characteristics of the route and the connection relationship between routes.
[0100] S02.02.02. The real predicted track sequence constructed by the present invention only includes the future longitude and latitude information of the target, thereby reducing the output variables of the track prediction model, which is beneficial to reducing the instability of the model and improving the convergence of the model.
[0101] In step S03, constructing a multi-possibility track prediction model includes the following sub-steps:
[0102] S03.01, Multi-possible trajectory prediction model architecture based on GAN: Figure 2As shown, the multi-possibility track prediction model based on GAN in this embodiment mainly includes two adversarial networks, a track generator and a track discriminator. In particular, since both the track generator and the track discriminator involve encoding the input track data, this embodiment proposes to let the two adversarial networks share the track encoder, including its network architecture and model parameters, which is conducive to reducing model training parameters and accelerating model training reasoning.
[0103] S03.02, track encoder Enc: The track encoder Enc is used to globally encode the track information of the input target and surrounding targets, the surrounding routes and the connection relationship between the routes, and its output is expressed as Enc(X). Figure 3 As shown in the figure, assuming that the observed track sequence of the current target is Tr, and there are N surrounding active targets: setTr = {Tr1, Tr2, ..., Tr N}, M surrounding key routes: setR = {R1, R2, ..., R M}, the specific encoding steps are as follows:
[0104] S03.02.01, Track point sequence encoding: In this embodiment, the same Transformer Encoder is first used to encode the track point sequences generated by the current target and the surrounding targets, and then the global graph GNN is used to encode the track point sequences generated by the current target and the surrounding targets. G1 The interactive relationship between different track point sequences is encoded to obtain the current target track point sequence encoding EncTr and the global track point sequence information encoding EncEnvTr.
[0105] S03.02.02, key route encoding. In this embodiment, a hierarchical graph neural network VectorNet is used to encode key routes and their connection information: first, a subgraph network GNN is used L For each critical route R i Encode and then use the global graph GNN G2 Encode the connection and interaction information between different routes, and output the global route information code EncEnvR in the current scenario;
[0106] S03.02.03, connect the coding information in step S03.02.01 and step S03.02.02, fuse the coding information through the MLP model, and obtain the total output Enc(X) = MLP([EncTr, EncEnvTr, EncEnvR]) of the track encoder.
[0107] S03.03, track generator G: The track generator G is used to predict the track based on the historical observation track generation model. It consists of a track encoder Enc and a track decoder Dec. The output of the track generator G is represented by Y G= Dec(Enc(X)). The specific steps of decoding and generating the future track include:
[0108] S03.03.01, discretize and predict the destination: take the latest track point of the current target as the center, divide the surrounding space into grids, and use them as candidate destinations. The destination is predicted based on the candidate destinations. The input of this step is the track encoder output Enc(X), the architecture used is the MLP model, and the loss function is:
[0109] L1(G)=L cls (tar)+L reg (tar)
[0110] Among them, L cls (tar) is the logarithmic loss function, which constrains the model to select the location closest to the true destination from the candidate destinations. reg (tar) is the Huber loss function, which is used to perform regression prediction on the distance between the candidate destination and the actual destination, thereby realizing the prediction of the actual destination.
[0111] S03.03.02, if Figure 4 As shown in the figure, based on the predicted destination, the future track of the decoded target is further predicted; this step is implemented by the Transformer Decoder model, which takes the track encoder output Enc(X) and the target position Loc as input and outputs the most likely future track trPred under this condition. The loss function of this step is:
[0112] L2(G)=L reg (trPred)
[0113] Among them, L reg (trPred) is the Huber loss function, which is used to perform regression prediction on the future track. The prediction result of this step is related to the predicted destination of step S03.03.01. To improve the model training effect, the real destination will be brought in for training during the training process.
[0114] S03.03.03, in predicting multiple future tracks, each track is screened and scored through the MLP model, the possibility of each track is predicted, and the tracks with the highest probability are selected as the model output. The loss function of this step is:
[0115] L3(G)=L cls (trPred)
[0116] Among them, L cls (trPred) is a logarithmic loss function, which is used to select the track closest to the true predicted track from multiple candidate tracks.
[0117] S03.04, track discriminator D: The track discriminator D is used to distinguish the real track from the model-generated track. It consists of a track encoder Enc and a binary classification MLP. The input of the track discriminator D is Among them, Y i is the result of splicing the historical observation track sequence and the real prediction track sequence. It is the concatenation result of the historical observed track sequence and the model predicted track sequence. The output of the track discriminator D is:
[0118] Y D =MLP(Enc(X D ))
[0119] S03.05, loss function: The loss function of the variable duration track prediction model in this embodiment is defined as:
[0120] L=min G max D L GAN (G,D)+λ1L1(G)+λ2L2(G)+λ3L3(G)
[0121] Among them, L GAN (G, D) is used as a general joint adversarial loss function to train the two adversarial networks of track generator G and track discriminator D. λ1L1(G)+λ2L2(G)+λ3L3(G) is the loss function for training the track generator, which is conducive to generating multiple possible tracks and predicting the confidence of each track. λ1,λ2,λ3 are used to adjust the weight of the loss function of each submodule in the overall loss function.
[0122] In step S04, the multi-possible track prediction model is trained: the track prediction data set is used to perform multiple rounds of adversarial training on the multi-possible track prediction model based on GAN. The multi-possible track prediction model based on GAN includes two adversarial networks, a track generator and a track discriminator. Therefore, an alternating iterative training method is adopted. Each iteration is trained in a single way in which the track generator and the track discriminator are alternately executed. The specific training steps are as follows:
[0123] S04.01: Read the historical observation track sequence and the real prediction track sequence from the track prediction data set;
[0124] S04.02: Input the historical observation track sequence into the track generator to obtain the model predicted track sequence;
[0125] S04.03: In order to achieve the purpose of identifying the authenticity of the indefinite length track, the present invention proposes that in the training process, the lengths of the real predicted track and the model predicted track are randomly sampled in each iteration. The historical observed track sequence and the real predicted track sequence cut according to the randomly sampled length are concatenated to obtain Y i , the historical observation track sequence is concatenated with the model prediction track sequence after being truncated by length to obtain Y i and Input into the track discriminator to calculate the score results of the two track sequences;
[0126] S04.04: Substitute the model prediction results of the track sequence of the track generator and the scoring results of the two track sequences of the track discriminator into the loss function to calculate the loss, select a suitable optimization algorithm for gradient descent, and update the weight parameters of the model;
[0127] S04.05: Repeat steps S04.02-S04.04 until the conditions for completing model training are met, such as reaching the maximum number of iterations or terminating model training early because the model loss no longer decreases after several iterations, to obtain a trained multi-possibility track prediction model.
[0128] In step S05, a multi-possibility track prediction model is applied: the future track of the target is predicted in real time using the trained multi-possibility track prediction model, and multiple future tracks of indefinite length are outputted, and the confidence and credibility are outputted for each predicted track respectively, which specifically includes the following sub-steps:
[0129] S05.01: Data preprocessing of the observed track sequence of the real-time tracked target, including outlier removal and time equidistant sampling;
[0130] S05.02: Input the target observation track sequence into the track generator in the trained multi-possibility track prediction model to obtain multiple model predicted tracks and the confidence of each track;
[0131] S05.03: For each model-predicted track, cut it into different lengths, concatenate it with the observed track sequence, and input it into the track discriminator in the trained multi-possibility track prediction model to obtain the credibility score result;
[0132] S05.04: Combining step S05.02 and step S05.03, the trained multi-possibility track prediction model adaptively outputs all possible indefinite-length tracks within the credibility threshold range and their respective confidence levels and credibility levels for the current prediction target.
[0133] Corresponding to the above-mentioned multi-possibility track prediction method, this embodiment further provides a multi-possibility track prediction system, including:
[0134] An extraction module is used to read the historical track data in the historical track database, perform data preprocessing on the historical track data, and extract key routes and connection relationships between routes in the historical track data;
[0135] A data set construction module is used to divide the historical track data in the historical track database into a track observation sequence and a real predicted track sequence, and to construct a track prediction data set in combination with the key routes and the connection relationship between the routes;
[0136] Model building module, used to build a multi-possible trajectory prediction model based on GAN;
[0137] A model training module, used for performing multiple rounds of adversarial training on a GAN-based multi-possible track prediction model using the track prediction dataset through a custom loss function;
[0138] The model application module is used to use the trained multi-possible track prediction model to make real-time predictions on the target's future track, output multiple future tracks of indefinite length, and output the confidence and credibility of each predicted track.
[0139] The working principle of each module refers to the aforementioned multi-possibility track prediction method and will not be repeated here.
[0140] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A multi-possibility track prediction method, characterized in that: The steps include: S01, extracting key routes and connection relationships between routes: reading historical track data in the historical track database, performing data preprocessing on the historical track data, and extracting key routes and connection relationships between routes in the historical track data; S02, constructing a track prediction data set: dividing the historical track data in the historical track database into a track observation sequence and a real predicted track sequence, and constructing a track prediction data set in combination with the key routes and the connection relationship between the routes; S03, construct a multi-possible trajectory prediction model: construct a multi-possible trajectory prediction model based on GAN; S04, training a multi-possible track prediction model: performing multiple rounds of adversarial training on the GAN-based multi-possible track prediction model using the track prediction dataset through a custom loss function; S05, applying a multi-possibility trajectory prediction model: using the trained multi-possibility trajectory prediction model, the future trajectory of the target is predicted in real time, multiple future trajectories of indefinite length are output, and the confidence and credibility of each predicted trajectory are output respectively.
2. The multi-possible track prediction method according to claim 1, characterized in that: In step S01, extracting key routes and the connection relationship between routes includes the following sub-steps: S01.01, read historical track data from the historical track database; each track is represented by a track point sequence, and the track points in the track point sequence include longitude, latitude, altitude, speed, heading and time; S01.02, cleaning track data: first remove invalid values and abnormal values in historical track data, and then use Kalman filter algorithm to smooth the track data; S01.03, Track Compression: Through track compression, the historical track data is divided into different straight track segments as candidate key routes; S01.04, extracting key routes: clustering the candidate key routes using a clustering algorithm, and outputting the clustering results as key routes; in the process of clustering the candidate key routes, equidistant interpolation is performed on the candidate key routes according to the set distance threshold based on the head and tail positions, and each route is represented as a coordinate sequence setCoor; S01.05, extract key routes and connection relationships between routes: based on the key routes, traverse the candidate key routes in the historical track database again, perform key route matching on the candidate key routes, count and supplement the attribute characteristics PropR of the key routes and the historical connection status ConnR with other routes, so that each key route is finally represented as R = {setCoor, PropR, ConnR}.
3. The multi-possible track prediction method according to claim 2, characterized in that: In step S02, constructing a track prediction data set includes the following sub-steps: S02.01, performing equidistant sampling on the historical track data after data cleaning in step S01.02 at a fixed time interval T; S02.02, based on the equidistant sampling data, construct a track prediction data set by generating and dividing the historical observation track sequence and the real prediction track sequence, and combining the key routes and the connection relationship between the routes.
4. The multi-possible track prediction method according to claim 4, characterized in that: Using asymmetric historical observation track sequences and true prediction track sequences: S02.02.01, when constructing a historical observation track sequence, according to the different types of targets, select the activity time threshold and distance threshold of other targets that have an impact on the activity trend of the current target, as well as the distance threshold of the surrounding key routes, and obtain the surrounding target tracks and surrounding key routes of the current target based on the threshold screening; wherein the track information includes longitude, latitude, altitude, speed and heading, and the route information includes the attribute characteristics of the route and the connection relationship between routes; S02.02.02, the constructed true predicted track sequence only includes the future latitude and longitude information of the target.
5. The multi-possible track prediction method according to claim 1, characterized in that: In step S03, constructing a multi-possibility track prediction model includes the following sub-steps: S03.01, the multi-possible track prediction model based on GAN includes two adversarial networks: track generator and track discriminator; the two adversarial networks share a track encoder; S03.02, track encoder En c It is used to globally encode the track information of the input target and surrounding targets, the surrounding routes and the connection relationship between routes, and its output is expressed as Enc(X); S03.03, the track generator G is used to generate a model to predict the track based on the historical observation track, and the track encoder En c and track decoder De c The output of the track generator G is represented by Y G = Dec(Enc(X)); S03.04, the track discriminator D is used to distinguish the real track from the model-generated track. It consists of a track encoder Enc and a binary classification MLP. The input of the track discriminator D is Among them, Y i is the result of splicing the historical observation track sequence and the real prediction track sequence. is the concatenation result of the historical observed track sequence and the model predicted track sequence; the output of the track discriminator D is: YD=MLP(Enc(X D )) S03.05, the loss function is defined as: L=min G max D L GAN (G,D)+λ1L1(G)+λ2L2(G)+λ3L3(G) Among them, L GAN (G, D) is used as a universal joint adversarial loss function to train the two adversarial networks, track generator G and track discriminator D; λ1L1(G)+λ2L2(G)+λ3L3(G) is the loss function for training the track generator, and λ1, λ2, and λ3 are used to adjust the weights of the loss functions of each sub-module in the overall loss function.
6. The multi-possible track prediction method according to claim 5, characterized in that: Track encoderEn c The specific encoding steps are as follows: S03.02.01, first use the same Transformer Encoder to encode the track point sequences generated by the current target and the surrounding targets respectively, and then use the global graph GNN G1 Encode the interaction relationship between different track point sequences to obtain the current target track point sequence code EncTr and the global track point sequence information code EncEnvT r ; S03.02.02, using the hierarchical graph neural network VectorNet to encode key routes and their connection information: first use the subgraph network GNN L For each critical route R i Encode and then use the global graph GNN G2 Encode the connection and interaction information between different routes, and output the global route information code EncEnvR in the current scenario; S03.02.03, connect the coding information in step S03.02.01 and step S03.02.02, fuse the coding information through the MLP model, and obtain the total output Enc(X) = MLP([EncTr, EncEnvTr, EncEnvR]) of the track encoder.
7. The multi-possible track prediction method according to claim 5, characterized in that: The specific steps of decoding and generating future tracks in the track generator include: S03.03.01, taking the latest track point of the current target as the center, divide the surrounding space into grids as candidate destinations, and predict the destination based on the candidate destinations; the input of this step is the track encoder output Enc(X), the architecture used is the MLP model, and the loss function is: L1(G)=L cls (tar)+L reg (tar) Among them, L cls (tar) is the logarithmic loss function, which constrains the model to select the location closest to the true destination from the candidate destinations. reg (tar) is the Huber loss function, which is used to perform regression prediction on the distance between the candidate destination and the real destination, thereby realizing the prediction of the real destination; S03.03.02, using the Transformer Decoder model, taking the track encoder output Enc(X) and the target position Loc as input, outputs the most likely future track trPred; the loss function of this step is: L2(G)=L reg (trPred) Among them, L reg (trPred) is the Huber loss function, which is used to perform regression prediction on the future track. The prediction result of this step is related to the predicted destination in step S03.03.
01. The real destination will be brought in for training during the training process. S03.03.03, in predicting multiple future tracks, each track is screened and scored through the MLP model, the possibility of each track is predicted, and the tracks with the highest probability are selected as the model output; the loss function of this step is: L3(G)=L cls (trPred) Among them, L cls (trPred) is a logarithmic loss function, which is used to select the track closest to the true predicted track from multiple candidate tracks.
8. The multi-possible track prediction method according to claim 5, characterized in that: In step S04, training a multi-possibility track prediction model includes the following sub-steps: S04.01: Read the historical observation track sequence and the real prediction track sequence from the track prediction data set; S04.02: Input the historical observation track sequence into the track generator to obtain the model predicted track sequence; S04.03: During the training process, the lengths of the real predicted track sequence and the model predicted track sequence are randomly sampled at each iteration; the historical observed track sequence and the real predicted track sequence cut according to the randomly sampled length are concatenated to obtain Y i , the historical observation track sequence is concatenated with the model prediction track sequence after being truncated by length to obtain Yi and Input into the track discriminator to calculate the score results of the two track sequences; S04.04: Substitute the model prediction results of the track sequence of the track generator and the scoring results of the two track sequences of the track discriminator into the loss function to calculate the loss, select the optimization algorithm for gradient descent, and update the weight parameters of the model; S04.05: Repeat steps S04.02-S04.04 until the conditions for completing model training are met, and a trained multi-possibility track prediction model is obtained.
9. The multi-possible track prediction method according to claim 1, characterized in that: In step S0S, applying the multi-possibility track prediction model includes the following sub-steps: S0S.01: Data preprocessing of the observed track sequence of the real-time tracked target; S0S.02: Input the target observation track sequence into the track generator in the trained multi-possibility track prediction model to obtain multiple model predicted tracks and the confidence of each track; S0S.03: For each model-predicted track, cut it into different lengths, concatenate it with the observed track sequence, and input it into the track discriminator in the trained multi-possible track prediction model to obtain the credibility score result; S0S.04: Combining steps S05.02 and S05.03, the trained multi-possibility track prediction model adaptively outputs all possible indefinite-length tracks within the credibility threshold range and their respective confidence levels and credibility levels for the current prediction target.
10. A multi-possibility track prediction system, characterized in that: Used to implement the multi-possibility track prediction method as described in any one of claims 1 to 9; The multi-possibility track prediction system comprises: An extraction module is used to read the historical track data in the historical track database, perform data preprocessing on the historical track data, and extract key routes and connection relationships between routes in the historical track data; A data set construction module is used to divide the historical track data in the historical track database into a track observation sequence and a real predicted track sequence, and to construct a track prediction data set in combination with the key routes and the connection relationship between the routes; Model building module, used to build a multi-possible trajectory prediction model based on GAN; A model training module, used for performing multiple rounds of adversarial training on a GAN-based multi-possible track prediction model using the track prediction dataset through a custom loss function; The model application module is used to use the trained multi-possible track prediction model to make real-time predictions on the target's future track, output multiple future tracks of indefinite length, and output the confidence and credibility of each predicted track.