JP2021174496A - Trajectory prediction method and system - Google Patents

Abstract

To provide a trajectory prediction method and a system that reduce an error between a predicted trajectory position and an actual trajectory position.SOLUTION: A trajectory prediction method according to one embodiment comprises the steps of: gridding and encoding a sub-trajectory of a predetermined length among a predicted trajectory to input to a long-term/short-term memory LSTM network; making a grid corresponding to a first top_N probability values output from the LSTM network into a candidate set of a prediction result; selecting a grid, in which an Euclidean distance from center coordinates to coordinates of a flag point of the sub-trajectory is between an upper threshold and a lower threshold, from center coordinates; and making the center coordinates of the grid having a highest selected probability value into a trajectory prediction result of the sub-trajectory. The LSTM network is obtained, based on historical trajectory data, by pre-training.SELECTED DRAWING: Figure 5

Description

The present specification relates to a technical field of trajectory detection, particularly a trajectory prediction method and system.

Hurricane movement is a very devastating natural phenomenon and often results in unpredictable losses of individuals and property. Therefore, trajectory prediction has become a hotspot for trajectory data research as a means of reducing hurricane losses.

In view of the above, one or more embodiments of the present invention provide a locus prediction method and system capable of reducing an error between a predicted locus position and an actual locus position.

One or more embodiments of the present specification provide a trajectory prediction method, and the trajectory prediction method is described.
A step of gridging and encoding sub-trajectories of a predetermined length among the predicted loci and inputting them to the long-term short-term memory RSTM network.
A step in which the grid corresponding to the first top_N probability values output from the LSTM network is used as a candidate set for the prediction result, and
A step of selecting a grid from the candidate set whose Euclidean distance from the center coordinates to the coordinates of the flag point of the sub locus is between the upper threshold and the lower threshold.
Including a step of using the center coordinates of the grid having the highest selected probability value as the trajectory prediction result of the sub-trajectory.
The LSTM network is obtained by pre-training based on historical trajectory data.

Preferably, after the step in which the grid corresponding to the first top_N probability values output from the LSTM network is set as a candidate set for the prediction result, the Euclidean distance from the center coordinates to the coordinates of the flag points of the sub-trajectories is When the grid between the upper limit threshold value and the lower limit threshold value does not exist in the candidate set, the step further includes a step in which the center coordinates of the grid including the flag point of the sub-trajectory are used as the locus prediction result of the sub-trajectory.

The length of the prediction target locus is n, the extraction interval of the sub loci is step = 1, the length of the extracted sub locus is k, and the number of the extracted sub loci is n-k.

Preferably, nk sub-trajectories of the locus are predicted by the LSTM network to obtain nk predicted locus coordinates, and then the coordinates are obtained.
A step of Kalman filtering the sequence consisting of the obtained nk predicted trajectory coordinates,
Further including a step of using the optimum estimation sequence obtained by filtering as the final prediction result of the trajectory.

Preferably, the step of gridging and encoding sub-trajectories of a predetermined length among the prediction target loci is
For any two locus segments in the training set, a step of calculating the spatiotemporal state distance between the two locus segments based on the time and spatial states of the two locus segments.
It includes a step of clustering the locus segments based on the calculated spatiotemporal state distance between the locus segments.

One or more embodiments of the present specification further provide a trajectory prediction system, which includes a grid encoding module, an LSTM network obtained by pre-training based on historical trajectory data, and a prediction module. , With
The grid encoding module grids and encodes sub-trajectories of a predetermined length among the predicted loci, outputs them, and inputs them to the LSTM network.
The prediction module uses a grid corresponding to the first top_N probability values output from the LSTM network as a candidate set for prediction results, and the Euclidean distance from the center coordinates to the coordinates of the flag points of the sub-trajectory is the upper limit. A grid between the value and the lower limit threshold is selected from the candidate set, and the center coordinate of the grid having the highest selected probability value is used as the locus prediction result of the sub locus.

One or more embodiments of the present specification further provide an electronic device, the electronic device comprising a memory, a processor, and a computer program stored in the memory and executable in the processor, wherein the processor is the program. Is executed, the locus prediction method is realized.

In the technical proposal of the present specification, sub-trajectories of a predetermined length among the prediction target loci are gridded and encoded, input to the long-term short-term memory LSTM network, and set to the first top_N probability values output from the LSTM network. The corresponding grid is used as a candidate set for the prediction result, and the grid whose Euclidean distance from the center coordinate to the coordinate of the flag point of the sub locus is between the upper limit threshold and the lower limit threshold is selected and selected. The center coordinate of the grid having the highest probability value is used as the locus prediction result of the sub locus. Compared to the conventional Brute Force LSTM locus prediction algorithm, the locus prediction algorithm according to the examples of the present specification does not use the only result output by the LSTM, but makes an appropriate prediction from the top_N candidate set output by the LSTM. By selecting the result, the predicted locus position becomes closer to the actual locus position, that is, the error between the predicted locus position and the actual locus position can be reduced.

Further, in the technical proposal according to one or more embodiments of the present specification, nk sub-trajectories of the locus are predicted by the LSTM network, nk predicted locus coordinates are obtained, and then nk obtained coordinates are obtained. The sequence consisting of the predicted locus coordinates can be Kalman-filtered, and the optimum estimation sequence obtained by filtering can be used as the final prediction result of the locus, whereby the error between the predicted locus position and the actual locus position can be further reduced. ..

In order to more clearly explain the examples of the present specification or the technical proposals of the prior art, the drawings necessary for the description of the examples or the prior art will be briefly described below, but of course, the drawings described below will be described. Only a few examples of this specification are included, and those skilled in the art can obtain other drawings based on these drawings without the need for creative effort.
It is a schematic diagram of the internal structure of the LSTM cell of the prior art. It is a schematic flowchart of the brute force LSTM algorithm of the prior art. It is a flowchart of the locus prediction method by one or more examples of this specification. It is a schematic diagram of the grid formation of the locus sequence according to one or more examples of this specification. It is a schematic diagram of the one-hot encoding for the grid locus according to one or more embodiments of the present specification. It is a flowchart of the method of predicting the next locus grid of a sub locus by an LSTM network according to one or more examples of this specification. It is a block diagram of the internal structure of the locus prediction system according to one or more examples of this specification. For LSTM network training rounds in three methods: the conventional brute force LSTM trajectory prediction algorithm according to one or more embodiments of the present specification, the ILSTM trajectory prediction algorithm according to the embodiments of the present specification, and the ILSTM-KF trajectory prediction algorithm. It is a schematic diagram of the change of RSME value accompanying. For LSTM network training rounds in three methods: the conventional brute force LSTM trajectory prediction algorithm according to one or more embodiments of the present specification, the ILSTM trajectory prediction algorithm according to the embodiments of the present specification, and the ILSTM-KF trajectory prediction algorithm. It is a schematic diagram of the change of RSME value accompanying. It is a schematic diagram of the hardware structure of the electronic device by one or more examples of this specification.

In order to further clarify the objectives, technical proposals and advantages of the present specification, one or more embodiments of the present specification will be described in more detail below with reference to the drawings with reference to specific examples. do.

Unless otherwise specified, technical or scientific terms used in the examples herein should have general meaning understood by those skilled in the art. The terms "first," "second," and similar terms used in this disclosure do not indicate order, quantity, or materiality, but are used solely to distinguish between different components. A term such as "provide" or "include" covers an object element or article whose subject element or article precedes the term is listed after the term and its equivalents. , Means that it does not exclude other elements or articles. Terms such as "connection" or "connection" are not limited to physical or mechanical connections, but may include electrical connections, both direct and indirect. "Upper", "lower", "left", "right", etc. indicate only the relative positional relationship, and when the absolute position of the explanation target changes, the corresponding relative positional relationship also changes accordingly.

With the spread and wide application of the artificial intelligence technology, currently, long term memory network; based on (L ong S hort- T erm M emory LSTM), through historical trajectory data, it is possible to train a model for predicting the trajectory Is.

A long-term short-term memory network (LSTM) is a time-recurrent neural network, which is specially designed to solve the problem of long-term dependence of general RNNs (recurrent neural networks). LSTMs are widely used in time series and have several uses for spatiotemporal trajectory data prediction. The internal structure of the LSTM cell is shown in FIG.

は現在の入力x_tと前時点の出力h_t-1に基づいて計算され、f_tの各次元の値は（０，１）の範囲内にあり、次に前時点のC_t-1にf_tベクトルをビット単位で乗算すると、f_tの値が０に近い次元では、情報は忘れられ、f_tの値が１に近い次元では、情報は保持される。 There are generally three types of LSTM gates at each sequence index position: forget gates, input gates and output gates. As the name implies, the forgetting gate controls whether or not it is forgotten, and in the LSTM, it controls whether or not the state of the hidden cell of the previous layer is forgotten with a certain probability. That is, the vector

Is calculated based on the current input x _{t and the output h t-1} at the previous time, and _{the value of each dimension of f t} is in the range of (0, 1), and then to C _{t-1 at the previous time.} multiplying f _t vector in units of bits, the dimensions close to the value of f _t is 0, information is forgotten, the value of f _t is the dimension closer to 1, the information is retained.

The input gate processes the input of the current sequence position. As shown in Equations 1 and 2, the input gate has a total of two steps. The sigma layer determines the information that needs to be updated, and the tanh layer creates a vector that maps the value to (-1,1), that is, alternative content for the update, and combines the two parts to form an input gate. Will be done.

Before the output gate, the cell state needs to be updated as shown in Equation 3, and the results of the previous forgetting gate and the input gate both affect the cell state.

The output gate determines what to output, which output is based on the state of the LSTM cell, but is the result of filtering. It is as shown in the formula 4 and the formula 5.

The conventional brute force LSTM locus prediction algorithm sequentially selects sub-trajectories seq of K locus points, grids and encodes them, inputs them into the LSTM cell network structure shown in FIG. It predicts the index of the grid.

The conventional brute force LSTM algorithm predicts the input trajectory sequence, and the algorithm flow is shown in FIG. This algorithm is divided into three modules: a data preprocessing module, a model training module, and a prediction module. The data processing module included gridding and encoding the locus data into an acceptable locus data vector for the LSTM, the model training module found the optimal LSTM prediction model, and the prediction module was trained. Predict the index of the next grid based on the LSTM model.

LSTMs based on neural networks have excellent functions for learning historical knowledge, but may not be adaptable to new knowledge. Hurricane trajectory data is a sample of relatively small data, and the important feature of hurricane trajectories is that in areas where hurricanes have occurred, past hurricanes may not have occurred or have occurred very few times. Therefore, the neural network may cause a large error in the predicted locus position from the actual locus position due to insufficient learning degree.

The technical proposals of one or more embodiments herein improve the conventional Bruteforce LSTM locus prediction algorithm, and the improved LSTM (ILSTM) locus prediction algorithm is part of the prediction module of the Bruteforce LSTM locus prediction algorithm. In the improved LSTM (ILSTM) algorithm, the trajectory data preprocessing part and the model training part are in agreement with the Brute Force LSTM trajectory prediction algorithm.

In the improved LSTM trajectory prediction, the appropriate prediction result is selected from the top_N candidate sets instead of having the LSTM output only one result. The flow of the improved LSTM (ILSTM) trajectory prediction algorithm according to the embodiment of the present specification is as follows. Let the grid corresponding to the first top_N probability values be the prediction result candidate set C = {S ₁ , S ₂ , L, S _{top_N} }.

The improved LSTM trajectory prediction avoids the special situation in the Brute Force LSTM prediction that the deviation of the prediction result is too large or the prediction result overlaps with the flag point of the sub-trajectory due to the restriction condition for the distance threshold, and the prediction result. By limiting to a specific reasonable range, the error between the predicted locus position and the actual locus position can be reduced.

Further, in one or more embodiments of the present specification, Kalman Filtering optimally estimates the trajectory coordinates predicted by the improved LSTM (ILSTM) trajectory prediction algorithm, thereby predicting the predicted trajectory. The error between the position and the actual trajectory position can be further reduced.

Hereinafter, technical proposals for one or more embodiments of the present specification will be described in detail with reference to the drawings.

The flow of the trajectory prediction method according to one or more embodiments of the present specification is specifically shown in FIG. 3 and includes the following steps.
Step S301: The trajectory data is preprocessed.
Specifically, an ordered sequence consisting of moving objects moving in geospatial space, their positions changing continuously over time, and m discrete position points and associated auxiliary information Defined as a trajectory, that is, as shown in Equation 6, a trajectory is an ordered set of points in multidimensional space.

In the equation, P _m indicates the spatial position information of the m-th locus point of the locus Trj, for example, latitude and longitude information P _m = (lon _m , lat _m ), where lon _m and lat _m are loci, respectively. longitude of the m-th trajectory points Trj indicates the latitude, I _m denotes the status information of the m-th path point trajectory Trj, hurricane data if example, the maximum sustained wind, hurricanes central minimum pressure, such as Indicates the status information of.

式中、kはサブ軌跡の長さを示す。 The sub-trajectory refers to an ordered set of locus points within a specific locus range and is shown in Equation 7.

In the equation, k indicates the length of the sub locus.

Data preprocessing is a necessary step before using raw data, and the data preprocessing module performs preprocessing on the raw data. Data preprocessing operations include operations such as data import, data filtering, and data missing value filling.

Raw trajectory data inevitably has data deficiencies and data errors, and the problems caused by the data deficiencies and errors directly affect the accuracy of the final result of the algorithm. Therefore, it is necessary to preprocess the data before it is actually used for training the algorithm module. Data preprocessing operations in this step include operations such as importing data into trajectory data, filtering data, and filling data missing values. The data preprocessing module imports and stores the raw data for training and the raw data for anomaly detection uploaded by the administrator into the database, and cleans the data and fills in the missing values through a series of database operations.

Based on the database-based system mode, the data in the Excel table is imported into the database during the data preprocessing stage, the imported data is read from the database, filtered, filled, and then back into the database. It will be remembered.

For example, in the case of a hurricane trajectory, the trajectory data may include information such as the latitude and longitude of the hurricane, the maximum sustained wind speed, the central minimum pressure, and the sampling time. In this step, data such as latitude and longitude, maximum sustained wind speed, and system status (Status of system) among the hurricane trajectory data can be preprocessed, and if there is a data loss in the data, the data loss value will be It is filled with the approximate value of the adjacent data. Duplicate data at the sampling point is also cleaned, and some outliers are changed.

Step S302: The locus data is gridded and encoded.

Specifically, the geographical space in which the movement target is located is divided into various spatial regions by a fixed grid (square, triangle, hexagon, etc.), and a single coordinate point of the trajectory is mapped to the spatial region, and the original The sequence of grid loci in which the locus Trj of is gridded is shown in Equation 8.

The structure of raw trajectory data is significantly different from that of text and images as spatiotemporal data, and therefore, a conventional deep learning model cannot be used directly. Therefore, in this step, the locus data is gridded into a data format that can be received by the neural network, that is, the representation format of the locus sequence is converted into a grid locus.

Also, since the locus data includes categorical data and numeric data, this step can also encode the grid locus based on the gridded locus, encoding various types of locus data. The process of encoding to be combined together can be divided into the following three steps.

1. 1. One-hot encoding is performed on the category data in the grid locus. For example, the grid index S _m in the grid locus can be One-hot encoded. One-hot encoding is the most common and basic encoding used to convert flags to vectors. First, map the category value to an integer value. Next, except for the index of the integer value, each integer value is represented as a binary vector, the rest are zero values, and the index position is made as 1. FIG. 4b shows a schematic diagram of the One-hot encoding of the grid locus.

2. Standardize the numeric data in the grid locus so that the numeric data range is between [01]. In order to eliminate the influence of dimensions between indicators, it is necessary to make data indicators comparable by performing data standardization. After the raw data has been processed by data standardization, each indicator has the same digits and is suitable for comprehensive comparative evaluation. The numeric data z in I _m, standardized using Equation 9 below, z 'is obtained.

3. 3. Connect category data and numeric data to form a complete locus vector V_Trj.

Step S303: The locus is predicted by the LSTM network obtained by pre-training based on the historical locus data.

と呼ばれる。 Specifically, among the prediction target trajectories, the grid-coded sub-trajectories are sequentially input to the LSTM network obtained by pre-training, and the position of the next locus grid of each sub-trajectory is predicted by the output of the LSTM network. NS. For example, in the case of a locus Trj of length n, each sub-locus sequence is extracted and gridded and encoded with step as the sub-locus extraction interval and k as the sub-locus length extraction. The next locus coordinates are predicted, and if step = 1, a sequence of nk sub-trajectories is predicted, so that nk predicted locus coordinates are generated and the sequence composed of these predicted coordinates is Predicted trajectory sequence

Is called.

Here, the flow of the method of predicting the position of the next locus grid of this sub locus by the LSTM network obtained by pre-training the sub locus of a predetermined length among the prediction target loci based on the historical locus data. Is specifically the flow of the method shown in FIG. 5, and includes the following substeps.

Sub-step S501: Among the prediction target loci, one sub-trajectory encoded in a grid is input to the LSTM network obtained by pre-training based on the historical locus data.

Sub-step S502: The grid corresponding to the first top_N probability values output from the LSTM network is used as a candidate set for the prediction result.

Sub-step S503: When a grid in which the Euclidean distance from the center coordinates to the coordinates of the flag point of the sub-trajectory is between the upper limit threshold value and the lower limit threshold value is searched in the candidate set and searched, the following sub-step S404 Is executed, and if the search is not performed, the following sub-step S405 is executed.

Sub-step S504: The center coordinates of the grid having the highest selected probability value are set as the locus prediction result of the sub-trajectory, that is, the position of the locus grid next to the predicted sub-trajectory.

Substep S505: The center coordinates of the grid including the flag point of the sub locus are set as the locus prediction result of the sub locus, that is, the position of the locus grid next to the predicted sub locus.

Definition of locus flag point Trj_flag:
The sub-trajectory flag point seq_flag in the first sub-trajectory sequence extracted from each locus is the locus flag point Trj_flag of the entire locus.

Step S304: Kalman filtering the prediction result.

Specifically, Kalman Filtering, which is a data processing technology that removes noise and restores actual data, is effectively used in many fields such as communication, navigation, guidance and control. Kalman Filtering optimally estimates the state of the system from the input / output observation data of the system. Optimal estimation can also be considered as a filtering process because the observed data includes the effects of noise and interference in the system.

Therefore, more preferably, in this step, Kalman filtering is performed on the sequence Z composed of the obtained nk predicted locus coordinates, and the optimal estimation sequence obtained by the filtering is the final of the Trj locus. Used as a prediction result, the locus prediction method using Kalman filtering is abbreviated as an ILSTM-KF locus prediction algorithm in the present specification.

The definition of the state vector of the motion system of the moving object X (m) is shown in Equation 10.

Kalman filtering requires the introduction of a discrete control process system. The system equation of state and the observation equation of the motion of the moving object are shown in Equations 11 and 12, respectively.

In the equation, X (m) represents the system state vector, and as shown in Equation 10, the state vector of the motion target at the m time point is represented, and A represents the transition mode of the motion state from the previous time point to the present time point. Represents the state transition matrix for, W (m) represents the system state noise of the model, its statistical characteristics are similar to white noise or Gaussian noise, and Z (m) is m in the predicted locus sequence Z. It represents the observation vector representing the second predicted value, where H is the observation matrix and V (m) is the observation noise.

For convenience of analysis, it is generally assumed that the system noise W (m) and the observed noise V (m) are Gaussian white noises that are independent of each other, and their covariances are Q and R, respectively, and their statistical characteristics. Are shown in Equations 13 and 14, respectively.

At the core of the Kalman filtering algorithm is to use a recursive algorithm to implement an estimation model for optimal state estimation and to update the current state variable estimates using previous and current observations. .. Use the system equation of state to estimate the state of the system at the next point in time. Assuming that the current state of the system is m, the current state can be estimated based on the previous state of the system according to the equation of state of the system, as shown in Equation 15.

In the equation, X (m | m-1) is the result of the previous state estimation, and X (m-1 | m-1) is the result of the optimum estimation of the previous state.

The update equation of the covariance corresponding to X (m | m-1) is shown in Equation 16.

In Equation 16, ρ (m | m-1) is the covariance corresponding to X (m | m-1), and ρ (m-1 | m-1) is X (m-1 | m-1). It is the covariance corresponding to 1), and A ^T is the transposed matrix of A.

Combining the locus coordinates Z (m) predicted by the improved RSTM locus prediction algorithm, that is, combining the measured value Z (m) and the estimated value X (m | m-1), as shown in Equation 17, the current The optimal estimate X (m | m) for the state m can be obtained.

As described above, the entire Kalman filtering process of the predicted trajectory sequence Z is completed. The role of Kalman filtering is to optimally estimate the locus coordinates predicted by the improved RSTM locus prediction algorithm.

The above LSTM network is obtained by pre-training based on historical trajectory data. For example, the LSTM network is trained using the hurricane data set from 1851 to 2015 as a training set, and this data includes 1780. There are book hurricane trajectories and 49139 sampling points, with a sampling time interval of 6 hours. The training method used may be similar to the method of training an LSTM network with a conventional brute force LSTM algorithm and will not be described in detail here.

According to the trajectory prediction method, one or more embodiments herein further provide a trajectory prediction system, the structure of which is based on grid encoding module 601 and historical trajectory data, as shown in FIG. The LSTM network 602 obtained by pre-training and the prediction module 603 are provided.

Among them, the grid encoding module 601 grids and encodes sub-trajectories of a predetermined length among the predicted loci, and inputs / outputs them to the LSTM network 602. That is, the grid encoding module 601 grids the locus data. Encode the grid locus based on the gridded grid locus.

The prediction module 603 uses the grid corresponding to the first top_N probability values output by the LSTM network 602 as a candidate set for the prediction result, and the Euclidean distance from the center coordinates to the coordinates of the flag point of the sub locus is the upper limit. A grid between the value and the lower limit threshold is selected from the candidate set, and the center coordinate of the grid having the highest selected probability value is used as the locus prediction result of the sub locus.

Further, the prediction module 603 also sets the grid corresponding to the first top_N probability values output from the LSTM network as a candidate set for the prediction result, and after the step, from the center coordinate to the flag point of the sub locus. If there is no grid in the candidate set whose Euclidean distance to the coordinates is between the upper and lower thresholds, the center coordinates of the grid including the flag points of the sub-trajectory are used as the locus prediction result of the sub-trajectory. do.

The length of the prediction target locus is n, the extraction interval of the sub locus is step = 1, the length of the extracted sub locus is k, the number of the extracted sub loci is nk, and the above. The prediction module 603 further predicts nk sub-trajectories of the locus with the RSTM network, obtains nk predicted locus coordinates, and then performs Kalman filtering of a sequence consisting of the obtained nk predicted locus coordinates. , The optimum estimation sequence obtained by filtering is used as the final prediction result of the locus.

In addition, the trajectory prediction system according to one or more embodiments herein may further include a data preprocessing module 604.

The data preprocessing module 604 performs data import, data filtering, and data missing value filling operations on the locus data, and inputs the data to the grid encoding module 601.

In addition, the trajectory prediction system according to one or more embodiments herein may further include a training module 605.

The training module 605 trains the LSTM network based on the historical trajectory data.

Specifically, the method of realizing the function of each module of the trajectory prediction system will not be described in detail here because the detailed method described in each step in the flow shown in FIG. 3 can be referred to.

The following experimental data use the 2016 hurricane as a validation set. The prediction errors when using the three methods of the conventional brute force LSTM locus prediction algorithm, the ILSTM locus prediction algorithm proposed in the examples of the present specification, and the ILSTM-KF locus prediction algorithm are compared.

Actual path sequence _{T = {Trj 1, Trj 2} , L, Trj N} and the estimated track sequence _{T '= {Trj' 1,} Trj '2, L, Trj' N} when assuming the actual and predicted trajectory point Using the geometrical space error of the locus point, the root mean square error RSME shown in Equation 20 is used as the prediction error, and the unit is degrees.

In the equation, n indicates the number of predicted loci, and m indicates the data of the points of the predicted loci included in each predicted locus.

The setting of the experimental parameter values is shown in Table 1, and the changes in the RSME values with the round of LSTM network training are shown in FIGS. 7 and 8. FIG. 7 shows that Q is [1,0,0,0; 0,1,0,0; 0,0,1,0; 0,0,0,1], and FIG. 8 shows Q. Is [2,0,0,0; 0,2,0,0; 0,0,2,0; 0,0,0,2], and Q is the covariance matrix of the system.

In FIGS. 7 and 8, the horizontal axis represents the training round of the neural network, and the vertical axis represents the RSEM value. The structure of the LSTM network shown in FIGS. 7 (a) and 8 (a) is 128 * 128 * 256, and the structure of the LSTM network shown in FIGS. 7 (b) and 8 (b) is 128 *. 256 * 512, the structure of the LSTM network shown in FIGS. 7 (c) and 8 (c) is 256 * 256, and the structure of the LSTM network shown in FIGS. 7 (d) and 8 (d) is It is 256 * 256 * 256.
As can be seen from FIGS. 7 and 8, the RSME value of the conventional Bruteforce LSTM trajectory prediction algorithm is large at first, and at this time, the learning degree of the LSTM network is low and the prediction trajectory error is large, while the present specification The RSEM values of the ILSTM locus prediction algorithm and the ILSTM-KF locus algorithm proposed in the examples are significantly lower than the RSME of the conventional Bruteforce LSTM locus prediction algorithm. As the number of rounds of LSTM network training increases, the RSEM value of the conventional brute force LSTM algorithm gradually decreases and gradually stabilizes, but the prediction error of the conventional brute force LSTM algorithm is ILSTM and ILSTM-KF. It was found that it was significantly higher than the prediction error of.

In order to better display the RSEM values of the above three methods, the RSME values of the three methods from round 71 to round 80 of the 128 * 256 * 512 LSTM network structure are shown below, that is, FIG. 7 ( In b) and FIG. 8 (b), the RSME values of rounds 71 to 80 are shown in Table 2 below. As can be seen from Table 2, the RSME value of each round of ILSTM-KF is slightly smaller than the value of ILSTM, indicating that the predictive effect of ILSTM-KF is the best of the three methods. There is.

In FIG. 7B, the conventional Bruteforce LSTM locus algorithm has an average prediction error of 2.1065 degrees for 10 rounds of rounds 71-80, and the improved LSTM locus prediction algorithm (ILSTM) has rounds 71-80. The average prediction error of 10 rounds is 1.4690, and the ILSTM-KF trajectory prediction algorithm has an average RSEM value of 1.4605 for 10 rounds of rounds 71-80.

In FIG. 8B, the conventional Brute Force LSTM locus prediction algorithm has an average RSEM value of 2.3865 degrees for 10 rounds of rounds 71-80, and the improved LSTM locus prediction algorithm (ILSTM) has rounds 71-80. The average RSEM value of 10 rounds is 1.5141, and the ILSTM-KF trajectory prediction algorithm has an average RSEM value of 1.5017 for 10 rounds of rounds 71 to 80. Therefore, the prediction error of ILSTM-KF. Was found to be the smallest. The prediction error of the conventional brute force LSTM trajectory prediction algorithm is about 2 degrees, which is the least effective. The prediction error of ILSTM-KF is slightly smaller than the prediction error of ILSTM. This indicates that the locus coordinates predicted by ILSTM-KF are more accurate than the locus coordinates predicted by ILSTM, and it is considered that ILSTM-KF has the best prediction effect.

The system of the above-described embodiment is for carrying out the corresponding method in the above-described embodiment, has a beneficial effect of the embodiment of the corresponding method, and is not described in detail here.

FIG. 10 shows a schematic diagram of a more specific hardware structure of an electronic device according to this embodiment, which device may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. .. Among them, the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040 are connected to each other so as to be able to communicate with each other in the device via the bus 1050.

The processor 1010 executes a related program and implements a trajectory prediction method according to an embodiment of the present specification, in order to carry out a general-purpose CPU (Central Processing Unit), a microprocessor, and an integrated circuit for a specific application (Application Special). It is realized in the form of an integrated CPU (ASIC), or one or more integrated circuits.

The memory 1020 can be implemented in the form of a ROM (Read Only Memory, read-only memory), a RAM (Random Access Memory, random access memory), a static storage device, and a dynamic storage device. The memory 1020 can store an operating system or other application program, and when the proposed technology according to the embodiments herein is implemented by software or firmware, the relevant program code is stored in the memory 1020 and the processor 1010. Called and executed by.

The input / output interface 1030 is used to connect input / output modules to input / output information. The input / output module may be arranged in the device as a component (not shown), or may be externally connected to the device to provide a corresponding function. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

The communication interface 1040 is used to connect a communication module (not shown) to realize communication between a device and another device. The communication module may communicate by a wired method (USB, network cable, etc.) or by a wireless method (mobile network, WIFI, Bluetooth (registered trademark), etc.).

Bus 1050 includes routes for transferring information between various components of the device (eg, processor 1010, memory 1020, input / output interface 1030, and communication interface 1040).

It should be noted that the device shows only the processor 1010, the memory 1020, the input / output interface 1030, the communication interface 1040, and the bus 1050, but in a particular implementation process, the device achieves normal operation. It may also include other required components. Further, as will be appreciated by those skilled in the art, the equipment may include only the components necessary to carry out the embodiments of the embodiments herein, rather than including all the components shown in the figures. ..

One or more embodiments of the present specification further provide a non-temporary computer-readable storage medium, wherein the non-temporary computer-readable storage medium can store computer instructions and is stored. The computer instruction is used to cause the computer to execute the trajectory prediction method of the above embodiment.

In the technical proposal according to one or more embodiments of the present specification, the sub-trajectories of a predetermined length among the predicted loci are gridded and encoded, input to the long-term short-term memory LSTM network, and first output from the LSTM network. The grid corresponding to the top_N probability values of It is selected from the candidate set, and the center coordinate of the grid having the highest selected probability value is used as the locus prediction result of the sub locus. Compared to the conventional Brute Force LSTM locus prediction algorithm, the locus prediction algorithm according to the examples of the present specification does not use the only result output by the LSTM, but makes an appropriate prediction from the top_N candidate set output by the LSTM. By selecting the result, the predicted locus position becomes closer to the actual locus position, that is, the error between the predicted locus position and the actual locus position can be reduced.

Further, in the technical proposal according to one or more embodiments of the present specification, nk sub-trajectories of the locus are predicted by the LSTM network to obtain nk predicted locus coordinates, and then nk are obtained. The sequence consisting of the predicted locus coordinates of the above can be Kalman filtered, and the optimum estimation sequence obtained by filtering can be used as the final prediction result of the locus, thereby further reducing the error between the predicted locus position and the actual locus position. can.

The computer-readable media of this embodiment include permanent and non-permanent, removable and non-removable media, and information can be stored by any method or technique. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), and electricity. Erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage Equipment, cassette-type magnetic tape, magnetic tape / magnetic disk storage equipment and other magnetic storage equipment, or other non-transmission media, but not limited to, used to store information accessible to computer equipment. ..

As will be appreciated by those skilled in the art, any description of the above embodiments is merely exemplary and is not intended to limit the scope of the present disclosure (including claims) to these examples. No. Based on the concepts of the examples herein, each of the above embodiments or the technical features of each embodiment may be combined, the steps may be performed in any order, and in different embodiments as described above. There are many changes and for brevity they are not described in detail.

Further, to simplify the description and discussion and not obscure the embodiments herein, the drawings provided are connected to known power sources for integrated circuit (IC) chips and other components. It may or may not be shown to be grounded. Further, in order to avoid difficulty in understanding the embodiments of the present specification, the devices may be shown in the form of block diagrams, which also detail the embodiments of the devices according to these block diagrams. However, take into account the fact that it relies heavily on the platform on which the examples herein are implemented (ie, these details are fully understandable to those skilled in the art). When specific details (eg, circuits) are described to illustrate exemplary embodiments herein, the examples herein are also in the absence of these specific details or changes in these specific details. It is clear to those skilled in the art that Therefore, these explanations should be considered exemplary rather than restrictive.

The technical proposals according to the present specification have been described with reference to specific examples, but those skilled in the art will be able to make many substitutions, modifications, and modifications to these examples based on the above description. Is clear. For example, other memory architectures (eg, dynamic RAM (DRAM)) can use the examples considered.

The examples herein are intended to include all such substitutions, modifications, and modifications that fall within the broad scope of the appended claims. Accordingly, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the examples herein should be included in the claims herein.

Claims (10)

It is a trajectory prediction method,
A step of gridging and encoding sub-trajectories of a predetermined length among the predicted loci and inputting them to the long-term short-term memory RSTM network.
A step in which the grid corresponding to the first top_N probability values output from the LSTM network is used as a candidate set for the prediction result, and
A step of selecting a grid from the candidate set whose Euclidean distance from the center coordinates to the coordinates of the flag point of the sub locus is between the upper threshold and the lower threshold.
Including a step of using the center coordinates of the grid having the highest selected probability value as the trajectory prediction result of the sub-trajectory.
The LSTM network is a locus prediction method characterized in that it is obtained by pre-training based on historical locus data. After the step, the grid corresponding to the first top_N probability values output from the LSTM network is set as a candidate set for the prediction result.
If there is no grid in the candidate set whose Euclidean distance from the center coordinates to the coordinates of the flag point of the sub locus is between the upper and lower thresholds, the center of the grid containing the flag point of the sub locus. The locus prediction method according to claim 1, further comprising a step in which the coordinates are used as the locus prediction result of the sub locus. The feature is that the length of the prediction target locus is n, the extraction interval of the sub locus is step = 1, the length of the extracted sub locus is k, and the number of the extracted sub loci is nk. The locus prediction method according to claim 1. After predicting the nk sub-trajectories with the LSTM network and obtaining nk predicted locus coordinates,
A step of Kalman filtering the sequence consisting of the obtained nk predicted trajectory coordinates,
The trajectory prediction method according to claim 3, further comprising a step of using the optimum estimation sequence obtained by filtering as the final prediction result of the trajectory. The step of gridging and encoding sub-trajectories of a predetermined length among the predicted loci is
For any two locus segments in the training set, a step of calculating the spatiotemporal state distance between the two locus segments based on the time and spatial states of the two locus segments.
The locus prediction method according to any one of claims 1 to 4, wherein the step of clustering the locus segments based on the calculated spatiotemporal state distance between the locus segments is included. It is a trajectory prediction system
It is equipped with a grid encoding module, an LSTM network obtained by pre-training based on historical trajectory data, and a prediction module.
The grid encoding module grids and encodes sub-trajectories of a predetermined length among the predicted loci, outputs them, and inputs them to the LSTM network.
The prediction module uses a grid corresponding to the first top_N probability values output from the LSTM network as a candidate set for prediction results, and the upper limit is the Euclidean distance from the center coordinates to the coordinates of the flag points of the sub-trajectory. A locus prediction system characterized in that a grid between a value and a lower limit threshold value is selected from the candidate set, and the center coordinates of the grid having the highest selected probability value are used as the locus prediction result of the sub locus. .. The prediction module further sets the grid corresponding to the first top_N probability values output from the RSTM network as a candidate set for the prediction result, and then moves from the center coordinates to the coordinates of the flag points of the sub locus. When the grid whose Euclidean distance is between the upper limit threshold value and the lower limit threshold value does not exist in the candidate set, the center coordinates of the grid including the flag point of the sub-trajectory are used as the locus prediction result of the sub-trajectory. 6. The trajectory prediction system according to claim 6. The length of the prediction target locus is n, the extraction interval of the sub locus is step = 1, the length of the extracted sub locus is k, and the number of extracted sub loci is nk.
The prediction module further predicts the nk sub-trajectories with the LSTM network, obtains nk predicted locus coordinates, and then performs Kalman filtering of the sequence consisting of the obtained nk predicted locus coordinates. The locus prediction system according to claim 6, wherein the optimum estimation sequence obtained by filtering is used as the final prediction result of the locus. The locus prediction system according to any one of claims 6 to 8, further comprising a training module for training the LSTM network based on historical locus data. It ’s an electronic device,
A memory, a processor, and a computer program stored in the memory and executed by the processor are provided, and the processor realizes the trajectory prediction method according to any one of claims 1 to 5 when executing the program. Electronics.

Images