Trajectory prediction method based on future low frequency and historical high frequency heterogenous decoupling

Abstract

The application discloses a trajectory prediction method based on future low frequency and historical high frequency heterogeneous decoupling, and belongs to the technical field of automatic driving and intelligent traffic trajectory prediction. The method first acquires historical trajectories of traffic participants, scene global features and interaction features, and generates a rough predicted future trajectory based on the features. A heterogeneous frequency division method is adopted to extract low frequency trend features from the rough predicted trajectory and high frequency detail features from the measured historical trajectory. Trend prediction branches and detail compensation branches are respectively constructed, and multi-modal predicted trajectories are output by relying on trend maintaining type heterogeneous bridging coupling. Low frequency reliability evaluation, high frequency noise gating, adaptive frequency band division and heterogeneous consistency training can be optionally configured. The application overcomes the problems of trajectory prediction over-smoothing or detail distortion caused by existing homogeneous frequency division and fixed frequency band, and takes into account long-term trend accuracy and local detail fitting capability, and can be applied to automatic driving, intelligent robots and low-altitude aircraft trajectory prediction scenes.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving and intelligent transportation trajectory prediction technology, and in particular to a multimodal trajectory prediction method based on heterogeneous decoupling of future low-frequency trends and historical high-frequency details. Background Technology

[0002] In autonomous driving trajectory prediction tasks, the goal is to predict one or more possible trajectories at several future moments based on the historical motion states of traffic participants, road geometry information, surrounding environment, and interaction relationships. Existing methods typically model historical trajectories directly in the time domain, extracting historical motion features through recurrent neural networks, convolutional neural networks, graph neural networks, or Transformers, and then outputting the future trajectory by a decoder.

[0003] The trajectory of traffic participants is essentially a time-series signal, which can be further analyzed from the perspective of the frequency domain. Generally speaking, the low-frequency components in the trajectory often correspond to the overall movement trend, long-term direction, destination bias, and macro-level behavioral intentions, while the high-frequency components correspond more to local curvature changes, short-term maneuvers, slight acceleration and deceleration, local avoidance, and fine-grained dynamic fluctuations.

[0004] While existing technologies include methods for incorporating frequency domain modeling, the following problems still exist:

[0005] First, extracting both high and low frequencies from the same source signal means that the high-frequency details in the coarse future trajectory prediction are often unreliable. If used directly, they will bring false local fluctuations into the final prediction result.

[0006] Second, the high-frequency components in the historical trajectory reflect the target's recent local maneuvering habits, short-term motion continuity, and dynamic detail patterns. This information itself is stable and truly observed. If it is completely ignored, the final prediction result will be too smooth and it will be difficult to accurately express the fine-grained trajectory changes when turning, changing lanes, or avoiding obstacles.

[0007] Third, existing models typically lack a concept of "heterogeneous decoupling," meaning they do not clearly distinguish which frequency bands are suitable for extraction from coarse future predictions and which are more suitable for extraction from historical observations. As a result, the models cannot fully utilize reliable trends in coarse future predictions, nor can they effectively utilize the true details in historical trajectories.

[0008] Fourth, fixed high and low frequency division thresholds are often difficult to adapt to different sampling frequencies, different scene complexities, and different motion types. The spectral structure of straight-moving scenes, turning scenes, acceleration and deceleration scenes, and even pedestrians and vehicles may be different, and fixed frequency band boundaries are prone to causing insufficient generalization problems.

[0009] Therefore, a new trajectory prediction method is needed to explicitly extract and decouple low-frequency trends and high-frequency details from different sources during the future trajectory prediction process. This method utilizes both the relatively reliable low-frequency trends in the coarse future prediction and the stable high-frequency details in the historical real trajectory, thereby improving the accuracy of long-term trend prediction and the ability to fit local details, and enhancing the robustness and interpretability of the model in complex traffic scenarios. Summary of the Invention

[0010] To address the aforementioned technical problems, this invention provides a trajectory prediction method based on the decoupling of future low-frequency and historical high-frequency sources, comprising the following steps:

[0011] S1. Obtain the historical trajectories of traffic participants and extract global scene features and detailed interaction features of traffic participants;

[0012] S2. Input the global features of the scene and the detailed features of traffic participants' interactions into the coarse prediction network to generate the first stage of coarse prediction of future trajectories;

[0013] S3. Perform frequency domain transformation on the coarsely predicted future trajectory in the first stage, and extract future low-frequency trend features through low-frequency masking; the future low-frequency trend features are used to characterize future long-term trends and macro-behavioral intentions.

[0014] S4. Perform frequency domain transformation on the historical trajectory, and extract historical high-frequency detail features through high-frequency masking; the historical high-frequency detail features are used to characterize the local dynamic patterns in the real historical trajectory.

[0015] S5. Construct a trend prediction branch and a detail compensation branch using future low-frequency trend features and historical high-frequency detail features respectively. The trend prediction branch integrates future low-frequency trend features, scene global features, and interaction detail features to output a multimodal global trend trajectory and modal probability. The detail compensation branch integrates historical high-frequency detail features, scene global features, and interaction detail features to output a multimodal detail compensation trajectory.

[0016] S6. By using a trend-preserving heterogeneous bridging coupling mechanism, the global trend trajectory and the detailed compensation trajectory are coupled to obtain the final future prediction trajectory.

[0017] Furthermore, in step S2, the method for generating the first-stage coarse prediction of the future trajectory is as follows:

[0018] Based on scene global features and interactive details Through coarse prediction network Generate the first stage coarse prediction of future trajectory :

[0019]

[0020]

[0021] in, Predict the length for the future.

[0022] Furthermore, in step S3, the method for extracting future low-frequency trend features is as follows:

[0023] The first stage provides a rough prediction of the future trajectory. Perform a frequency domain transformation along the time dimension to obtain the coarsely predicted future spectrum. Represented as:

[0024]

[0025] in, Represents the Discrete Fourier Transform;

[0026] Using low-frequency masks Extracting future low-frequency spectrum , Indicates frequency index; uses fixed frequency band boundaries Define low-frequency mask as follows:

[0027]

[0028] Then the future low-frequency spectrum Represented as:

[0029]

[0030] in, This represents element-wise multiplication;

[0031] The low-frequency transform network is denoted as follows: [The network is then mapped in the complex domain to represent the low-frequency spectrum]. ,but:

[0032]

[0033]

[0034] in, This indicates the low-frequency mapping result; and These represent taking the real and imaginary parts of a complex number, respectively. , Let be the learnable weight matrices for the real and imaginary parts, respectively. , These are the learnable biases for the real and imaginary parts, respectively. For activation functions;

[0035] The low-frequency mapping result is inversely transformed back to the time domain, and then subjected to pooling or projection operations to obtain future low-frequency trend features. :

[0036]

[0037]

[0038] in, This represents the inverse discrete Fourier transform. This indicates a pooling or projection operation. For the number of traffic participants, Indicates the dimension of the hidden feature layer.

[0039] Furthermore, step S3 introduces a trend reliability assessment factor to assess the reliability of future low-frequency trend characteristics, as follows:

[0040] First, calculate the proportion of low-frequency energy in the coarse prediction of the future trajectory in the first stage:

[0041]

[0042] in, This indicates the proportion of low-frequency energy in the rough prediction of future trajectories. Indicates frequency index, Indicates a low-frequency mask. This indicates that the coarse prediction of the future trajectory is obtained by performing a Discrete Fourier Transform along the time dimension at the frequency index. Complex spectral components at that location, To prevent extremely small constants with a denominator of zero;

[0043] Combined with global scene features After pooling, the reliability coefficient of the constructed trend is:

[0044]

[0045]

[0046] in, This represents the pooling result of global features of the scene. This indicates a pooling operation. Indicates feature splicing, and These are the learnable weight matrix and bias of the reliability estimation module, respectively. For activation functions;

[0047] Finally, the trend reliability coefficient is used to analyze the characteristics of future low-frequency trends. Perform adaptive weighting:

[0048]

[0049] in, The weighted future low-frequency trend characteristics, This indicates element-wise multiplication.

[0050] Furthermore, in step S4, the method for extracting historical high-frequency detail features is as follows:

[0051] The true trajectory of history Perform a discrete Fourier transform to obtain the historical spectrum. Represented as:

[0052]

[0053] in, Represents the Discrete Fourier Transform;

[0054] Define high-frequency mask :

[0055]

[0056] in, Indicates a low-frequency mask;

[0057] Based on length High-frequency mask, extracting high-frequency spectrum from historical spectrum. :

[0058]

[0059] in, This represents element-wise multiplication;

[0060] A high-frequency noise gating mechanism is introduced, as follows:

[0061] First, calculate the average high-frequency energy at each frequency position in the historical trajectory spectrum and the average energy of its neighborhood to determine the spectral abrupt change ratio;

[0062] Calculate the average high-frequency energy at each frequency position:

[0063]

[0064] in, Indicates the first The frequency component represents the high-frequency energy across all targets and all input channels. For the number of traffic participants, As a dimension of historical trajectory features, Indicates the first The traffic participant's number After the historical trajectory feature channels are subjected to discrete Fourier transform, the frequency index is... Complex spectral components at the location;

[0065] Secondly, a neighborhood is constructed for each frequency location. Calculate the average energy of the neighborhood :

[0066]

[0067] in, Indicates the first The average high-frequency energy of each frequency, among which ;

[0068] Spectral mutation ratio Defined as:

[0069]

[0070] in, To prevent extremely small constants with a denominator of zero;

[0071] A high-frequency noise gating factor is defined based on the aforementioned spectral mutation ratio. :

[0072]

[0073] in, For activation function, For scaling parameters, For threshold parameters;

[0074] Based on the aforementioned high-frequency noise gating factor, the high-frequency spectrum is suppressed or preserved, resulting in the gated historical high-frequency spectrum. Represented as:

[0075]

[0076] Perform an inverse transform on the gated historical high-frequency spectrum to obtain the historical high-frequency trajectory signal. :

[0077]

[0078] in, This represents the inverse discrete Fourier transform. This indicates taking the real part of a complex number;

[0079] High-frequency feature extraction network Obtain historical high-frequency detailed features :

[0080]

[0081]

[0082] in, It is a multilayer perceptron. For the number of traffic participants, Indicates the dimension of the hidden feature layer.

[0083] In another preferred embodiment, the frequency domain transformation in steps S3 and S4 employs an adaptive frequency band boundary, which is generated in the following manner:

[0084] Let the current spectrum for which the frequency band boundary needs to be calculated be represented as: , Take historical spectrum Or a rough prediction of the future spectrum Or the fused spectrum of the two, defined In frequency index The spectral energy obtained by averaging along all traffic participants and input channels at the point of origin. for:

[0085]

[0086] in, For the number of traffic participants, Indicates spectrum input The number of channels, when hour, , For historical trajectory feature dimension; when hour, ;when When it is an intermediate characteristic spectrum, This corresponds to the number of feature channels; Indicates the current spectrum The Middle The traffic participant's number Each input channel in frequency index Complex spectral components at that location, Indicates frequency index;

[0087] Normalize the spectral energy:

[0088]

[0089] in, This represents the normalized spectral energy. To prevent extremely small constants with a denominator of zero, This represents the summation frequency index used to iterate through all frequency components;

[0090] Recalculate the cumulative energy distribution :

[0091]

[0092] Then, based on at least one of the following: global scene features, interaction detail features, target speed, or uncertainty, the low-frequency proportion of the target is generated. :

[0093] ( )

[0094] in, For scene global features, For interactive details, This represents the average velocity statistic of targets within the scene. This represents a statistic indicating uncertainty. and For learnable weight matrix and bias;

[0095] This yields the adaptive frequency band boundary coefficient. :

[0096]

[0097] in, This represents the sharpening factor for selecting the frequency band boundary; represents the normalized exponential function;

[0098] Constructing a low-frequency mask :

[0099]

[0100] and high-frequency mask :

[0101]

[0102] The high- and low-frequency division changes dynamically with the scene and trajectory structure, where This is the frequency band boundary smoothing coefficient.

[0103] Furthermore, in step S5, after obtaining the weighted future low-frequency trend characteristics... and historical high-frequency details Then, a heterogeneous decoupled prediction structure is constructed, including a trend prediction branch and a detail compensation branch;

[0104] The future low-frequency trend features are obtained by fusion and weighting of the trend prediction branch. Scene global features Interactive details To obtain trend representation :

[0105]

[0106] in, for, A multi-head attention mechanism module for trend representation is used to complete feature fusion;

[0107] Output A global trend trajectory and its modal probabilities :

[0108]

[0109] in, , , For trajectory generation networks, For the number of traffic participants, For predicting future length, The number of trajectory modes;

[0110] The detail compensation branch fuses historical high-frequency detail features. Scene global features Interactive details , to obtain detailed representation :

[0111]

[0112] in, A multi-head attention mechanism module for detailed representation is used to complete feature fusion;

[0113] Output Detailed compensation trajectory :

[0114]

[0115] in, , A trajectory generation network.

[0116] Furthermore, in step S6, the trend-maintaining heterogeneous bridging coupling mechanism is as follows:

[0117] For the Each modality, taking the endpoint of the trend trajectory. :

[0118]

[0119] in, For predicting future length, Indicates the first Global trend trajectory of each modality;

[0120] Injecting the endpoint of the trend trajectory into the detailed compensation branch as conditional information yields the first... Detailed compensation trajectory for each modality :

[0121]

[0122] in, For trajectory generation networks, Indicates splicing, Detailed representation of the compensation branch;

[0123] Define with prediction time step Decreasing detail injection weights :

[0124]

[0125] in, , These are the attenuation control parameters;

[0126] Then time step The Final trajectory of each mode Represented as:

[0127]

[0128] in, For time step The Global trend trajectory of each modality For time step The Modal detail compensation trajectory;

[0129] For each mode Final trajectory The output is:

[0130]

[0131] The final future predicted trajectory is obtained by superimposing the final trajectories of K modes modally. Modal probability:

[0132]

[0133]

[0134] in, These represent the probabilities of the corresponding final trajectories.

[0135] In another preferred embodiment, in step S6, directly set:

[0136]

[0137] in, This represents the detailed compensated trajectory after time step t has been zeroed out at the endpoint. and The time step t and the endpoint are respectively the first and second steps. Modal detail compensation trajectory;

[0138] Therefore:

[0139]

[0140] in, This represents the end-point detail compensation trajectory after end-point zeroing correction;

[0141] For each mode Final trajectory The output is:

[0142]

[0143] in, Indicates the first Global trend trajectory of each modality This represents the detailed compensation trajectory after zeroing correction at the endpoint.

[0144] The beneficial effects of this invention are:

[0145] 1. Improve the long-term accuracy and local realism of trajectory prediction.

[0146] This invention proposes for the first time the concept of "heterogeneous frequency band decoupling": low-frequency trends are derived from coarse future predictions to ensure stable and reliable macroscopic trends; high-frequency details are derived from historical real trajectories to provide truly observed local maneuver information. The two complement each other, ensuring that the final trajectory is both accurate in direction and free from systematic deviations, while also possessing natural curvature changes and short-term dynamics, overcoming the problems of excessive smoothness or the coexistence of spurious fluctuations in traditional methods.

[0147] 2. Enhance robustness in complex scenarios

[0148] By introducing a reliability assessment of future low-frequency trends, the model can adaptively adjust its confidence in trend features based on scenario uncertainties, preventing erroneous trends from being amplified. Simultaneously, a historical high-frequency detail noise gating mechanism effectively suppresses isolated spikes such as sensor noise and positioning jitter, preserving continuous, true high-frequency motion information. Both mechanisms together enhance the system's adaptability to noise, complex interactions, and highly uncertain scenarios.

[0149] 3. Enhance generalization capabilities across scenarios and sports types.

[0150] Existing fixed frequency band thresholds are difficult to adapt to different sampling frequencies, scene complexity, and motion types. This invention employs an adaptive frequency band boundary generation method, which dynamically determines the frequency band division based on the cumulative energy of the spectrum and scene semantics. This enables the model to automatically adapt to different scenarios such as straight-ahead driving, turning, pedestrians, and vehicles, significantly enhancing generalization performance.

[0151] 4. Maintain consistency between macro-level behavioral intentions and the final destination.

[0152] This invention designs a trend-preserving heterogeneous bridging coupling mechanism, which injects the trend endpoint into the detailed branch as conditional information. Combined with time-varying weights or endpoint zeroing processing, it ensures that high-frequency detail compensation does not change the macroscopic endpoint direction expressed by the trend trajectory, so that the prediction results have both rich detail and logical consistency.

[0153] In summary, this invention has made significant progress in trajectory prediction accuracy, robustness, generalization ability, and interpretability, and is particularly suitable for applications such as autonomous driving, intelligent robots, and trajectory inference of low-altitude aircraft. Attached Figure Description

[0154] Figure 1 This is a schematic diagram of the overall process of the method of the present invention;

[0155] Figure 2 A schematic diagram of the overall framework for decoupling future low-frequency trends from historical high-frequency details;

[0156] Figure 3 A schematic diagram illustrating the process for extracting future low-frequency trends;

[0157] Figure 4 This is a schematic diagram of the historical high-frequency detail extraction and noise gating process. Detailed Implementation

[0158] like Figure 1-4 As shown, this invention provides a trajectory prediction method based on the decoupling of future low-frequency and historical high-frequency sources, comprising the following steps:

[0159] S1. Obtain the historical trajectories of traffic participants and extract global scene features and detailed interaction features of traffic participants;

[0160] S2. Input the global features of the scene and the detailed features of the interaction into the coarse prediction network to generate the first stage of coarse prediction of the future trajectory;

[0161] S3. Perform frequency domain transformation on the coarsely predicted future trajectory in the first stage, and extract future low-frequency trend features through low-frequency masking; the future low-frequency trend features are used to characterize future long-term trends and macro-behavioral intentions.

[0162] S4. Perform frequency domain transformation on the historical trajectory, and extract historical high-frequency detail features through high-frequency masking; the historical high-frequency detail features are used to characterize the local dynamic patterns in the real historical trajectory.

[0163] S5. Construct a trend prediction branch and a detail compensation branch using future low-frequency trend features and historical high-frequency detail features respectively. The trend prediction branch integrates future low-frequency trend features, scene global features, and interaction detail features to output a multimodal global trend trajectory and modal probability. The detail compensation branch integrates historical high-frequency detail features, scene global features, and interaction detail features to output a multimodal detail compensation trajectory.

[0164] S6. By using a trend-preserving heterogeneous bridging coupling mechanism, the global trend trajectory and the detailed compensation trajectory are coupled to obtain the final future prediction trajectory.

[0165] Assume there are a total of Each traffic participant, with a historical observation length of [number] times. The predicted length in the future is Output is required. One possible future modality.

[0166] Furthermore, in step S1 of this embodiment, the first... Historical trajectory of a traffic participant Represented as:

[0167]

[0168] in: For the first The two-dimensional position vectors of traffic participants at time t, t∈( , , ), For the length of historical observation, This refers to the current observation time, which is the last moment of the historical trajectory observation window. Represented as:

[0169]

[0170] in, These are the coordinates in the x and y directions, respectively. Indicates transpose;

[0171] Organize the historical trajectories of all traffic participants into a historical trajectory tensor :

[0172]

[0173] in, For the number of traffic participants, For time step features;

[0174] In other implementations, position, velocity, heading angle, category information, etc., can be combined to form a higher-dimensional input. It can be greater than 2.

[0175] Let the global scene features obtained by the scene coding network be:

[0176]

[0177] Let the interaction detail features obtained by the interaction coding network be:

[0178]

[0179] in, For scene global features, For interactive details, This refers to the implicit space dimension. and It can be obtained from any scene encoder, such as an encoding network built based on map features, neighbor vehicle interaction features, and relative pose features. In a preferred embodiment, This mainly represents the relationship between the target and the map, road topology, and global scene semantics. It mainly represents the interaction between the target and surrounding traffic participants.

[0180] Furthermore, in step S2 of this embodiment, the method for generating the first-stage coarse prediction of the future trajectory is as follows:

[0181] Based on scene global features and interactive details Through coarse prediction network Generate the first stage coarse prediction of future trajectory :

[0182]

[0183]

[0184] in, Predicting the length for the future; It can be a multilayer perceptron, a temporal convolutional network, a Transformer decoder, or other trajectory generation modules.

[0185] The first stage, which roughly predicts the future trajectory, is used to provide a general direction and prior macroscopic behavior of the future trajectory. In this invention, it is not required that... The goal is not to achieve absolute precision in local details, but rather to use this as the primary source of future low-frequency trends. This is the first core judgment of this invention: most trajectory prediction methods stop here, and their results can be used as the final trajectory output. The low-frequency part of the coarse prediction of future trajectories is usually more reliable than the high-frequency part.

[0186] Furthermore, in step S3 of this embodiment, the method for extracting future low-frequency trend features is as follows:

[0187] The first stage provides a rough prediction of the future trajectory. Perform a frequency domain transformation along the time dimension to obtain the coarsely predicted future spectrum. Represented as:

[0188]

[0189] in, Represents the Discrete Fourier Transform;

[0190] Using low-frequency masks Extracting future low-frequency spectrum , Represents the frequency index; in this embodiment, a fixed frequency band boundary is used. Define low-frequency mask as follows:

[0191]

[0192] Then the future low-frequency spectrum Represented as:

[0193]

[0194] in, This represents element-wise multiplication;

[0195] Complex domain mapping is performed on the future low-frequency spectrum to enhance its expressive power; the low-frequency transform network is denoted as... ,but:

[0196]

[0197]

[0198] in, This indicates the low-frequency mapping result; and These represent taking the real and imaginary parts of a complex number, respectively. , Let be the learnable weight matrices for the real and imaginary parts, respectively. , These are the learnable biases for the real and imaginary parts, respectively. For activation functions;

[0199] The low-frequency mapping result is inversely transformed back to the time domain, and then subjected to pooling or projection operations to obtain future low-frequency trend features. :

[0200]

[0201]

[0202] in, This represents the inverse discrete Fourier transform. This indicates a pooling operation. For the number of traffic participants, Indicates the dimension of the hidden feature layer;

[0203] The aforementioned future low-frequency trend characteristics mainly reflect future long-term trends, overall direction, and macro-level behavioral intentions.

[0204] While the low-frequency components of future coarse-predicted trajectories are generally reliable, their low-frequency trends can still be misleading in scenarios with strong interaction and high uncertainty. To improve robustness, this embodiment further introduces a trend reliability assessment factor to evaluate the reliability of future low-frequency trend characteristics, as follows:

[0205] First, calculate the proportion of low-frequency energy in the coarse prediction of the future trajectory in the first stage:

[0206]

[0207] in, This indicates the proportion of low-frequency energy in the rough prediction of future trajectories. Indicates frequency index, Indicates a low-frequency mask. This indicates that the coarse prediction of the future trajectory is obtained by performing a Discrete Fourier Transform along the time dimension at the frequency index. Complex spectral components at that location, To prevent extremely small constants with a denominator of zero;

[0208] Combined with global scene features After pooling, the reliability coefficient of the constructed trend is:

[0209]

[0210]

[0211] in, This represents the pooling result of global features of the scene. This indicates a pooling operation. Indicates feature splicing, and These are the learnable weight matrix and bias of the reliability estimation module, respectively. For activation functions; preferably, It can be designed as a scalar or a vector; when it is a scalar, it represents the confidence level of the overall trend for each target; when it is a vector, it represents the confidence level of the trend feature in each dimension.

[0212] Finally, the trend reliability coefficient is used to analyze the characteristics of future low-frequency trends. Perform adaptive weighting:

[0213]

[0214] in, The weighted future low-frequency trend characteristics, This represents element-wise multiplication;

[0215] Thus, in scenarios where the coarse prediction of future trends is relatively stable, the model will place more trust in low-frequency future trends; in scenarios where the coarse prediction is unstable, the model will adaptively reduce its impact.

[0216] Furthermore, in step S4 of this embodiment, the method for extracting historical high-frequency detail features is as follows:

[0217] The true trajectory of history Perform a discrete Fourier transform to obtain the historical spectrum. Represented as:

[0218]

[0219] in, Represents the Discrete Fourier Transform;

[0220] Define high-frequency mask :

[0221]

[0222] in, Indicates a low-frequency mask;

[0223] Based on length High-frequency mask, extracting high-frequency spectrum from historical spectrum. :

[0224]

[0225] in, This represents element-wise multiplication;

[0226] Due to the influence of the spectral bias principle, deep neural networks typically tend to prioritize fitting low-frequency, smooth motion trends during training, while their learning ability for high-frequency local changes, short-term micro-movements, and fine-grained dynamic patterns is relatively weak. Therefore, relying solely on coarse future predictions is insufficient to reliably and stably recover high-frequency details in future trajectories. In contrast, historical real trajectories, directly derived from observed motion processes, can more realistically reflect the curvature changes, velocity perturbations, and short-term interactive responses of the agent within a local time range, exhibiting higher reliability and continuity. However, historical high-frequency components may also contain isolated noise spikes caused by sensing errors, trajectory sampling errors, or anomalous observations. To address this, this embodiment introduces a high-frequency noise gating mechanism. By identifying and suppressing anomalous high-frequency spikes through spectral mutation ratios, it retains true high-frequency details with neighborhood continuity, thereby enhancing the auxiliary role of historical high-frequency features in fine-grained prediction of future trajectories. The method is as follows:

[0227] First, calculate the average high-frequency energy at each frequency position in the historical trajectory spectrum and the average energy of its neighborhood to determine the spectral abrupt change ratio;

[0228] Calculate the average high-frequency energy at each frequency position:

[0229]

[0230] in, Indicates the first The average high-frequency energy of each frequency component across all targets and all input channels. For the number of traffic participants, As a dimension of historical trajectory features, Indicates the first The traffic participant's number After the historical trajectory feature channels are subjected to discrete Fourier transform, the frequency index is... Complex spectral components at the location;

[0232] Secondly, a neighborhood is constructed for each frequency location. Calculate the average energy of the neighborhood (Excluding the current frequency point) ):

[0233]

[0234] in, Indicates the first The average high-frequency energy of each frequency, among which ;

[0235] When a certain frequency location is an isolated noise spike, there is usually ;

[0236] Spectral mutation ratio Defined as:

[0237]

[0238] in, To prevent extremely small constants with a denominator of zero;

[0239] If the energy at a certain frequency is much higher than the average value of its neighborhood, then the corresponding A larger value usually means that the high-frequency component may be an isolated noise spike.

[0240] A high-frequency noise gating factor is defined based on the aforementioned spectral mutation ratio. :

[0241]

[0242] in, For activation function, For scaling parameters, For threshold parameters;

[0243] Based on the aforementioned high-frequency noise gating factor, the high-frequency spectrum is suppressed or preserved. Thus, the gated historical high-frequency spectrum... Represented as:

[0244]

[0245] In this way, isolated high-frequency spikes are more likely to be judged as noise and suppressed, while high-frequency details with structural continuity are preserved.

[0246] Perform an inverse transform on the gated historical high-frequency spectrum to obtain the historical high-frequency trajectory signal. :

[0247]

[0248] in, This represents the inverse discrete Fourier transform. This indicates taking the real part of a complex number;

[0249] High-frequency feature extraction network Obtain historical high-frequency detailed features :

[0250]

[0251]

[0252] in, It can be a multilayer perceptron. For the number of traffic participants, Indicates the dimension of the hidden feature layer;

[0253] These high-frequency detail features primarily express local curvature changes, short-term micro-motions, and fine-grained dynamic patterns in the real historical trajectory.

[0254] Furthermore, in step S5 of this embodiment, after obtaining the weighted future low-frequency trend characteristics... and historical high-frequency details Then, a heterogeneous decoupled prediction structure is constructed, including a trend prediction branch and a detail compensation branch;

[0255] The future low-frequency trend features are obtained by fusion and weighting of the trend prediction branch. Scene global features Interactive details To obtain trend representation :

[0256]

[0257] in, for, A multi-head attention mechanism module for trend representation is used to complete feature fusion;

[0258] Output A global trend trajectory and its modal probabilities :

[0259]

[0260] in, , , For multilayer perceptrons, temporal convolutional networks, Transformer decoders, or other trajectory generation networks, For the number of traffic participants, For predicting future length, To predict the number of trajectory modes;

[0261] The detail compensation branch fuses historical high-frequency detail features. Scene global features Interactive details , to obtain detailed representation :

[0262]

[0263] in, A multi-head attention mechanism module for detailed representation is used to complete feature fusion;

[0264] Output Detailed compensation trajectory :

[0265]

[0266] in, , It is used for multilayer perceptrons, temporal convolutional networks, Transformer decoders, or other trajectory generation networks.

[0267] The It is more appropriate to understand it as "local detail residuals" rather than the complete trajectory.

[0268] Furthermore, in step S6 of this embodiment, in order to ensure that high-frequency detail compensation does not disrupt the overall macroscopic trend, a trend-preserving heterogeneous bridging coupling mechanism is introduced.

[0269] For the Each modality, taking the endpoint of the trend trajectory. :

[0270]

[0271] in, For predicting future length, Indicates the first Global trend trajectory of each modality;

[0272] Injecting the endpoint of the trend trajectory into the detailed compensation branch as conditional information yields the first... Detailed compensation trajectory for each modality :

[0273]

[0274] in, For trajectory generation networks, Indicates splicing, This provides a detailed representation of the detail compensation branch; through this design, the detail compensation branch can explicitly perceive the macroscopic endpoint constraints of the corresponding trend mode.

[0275] Define with prediction time step Decreasing detail injection weights :

[0276]

[0277] in, , These are the attenuation control parameters;

[0278] Then time step The Final trajectory of each mode Represented as:

[0279]

[0280] in, For time step The A modal global trend trajectory point For time step The Modal detail compensation trajectory;

[0281] In this way, the detailed compensation plays a greater role in the early stages of prediction; as the prediction approaches its end, the detailed compensation gradually decreases, thereby preventing the detailed branches from excessively altering the trend endpoint.

[0282] For each mode Final trajectory The output is:

[0283]

[0284] In another preferred embodiment, let:

[0285]

[0286] in, This represents the detailed compensated trajectory after time step t has been zeroed out at the endpoint. and The time step t and the endpoint are respectively the first and second steps. Modal detail compensation trajectory;

[0287] Therefore:

[0288]

[0289] in, This represents the end-point detail compensation trajectory after end-point zeroing correction;

[0290] For each mode Final trajectory The output is:

[0291]

[0292] in, Indicates the first Global trend trajectory of each modality This represents the detailed compensated trajectory after zeroing correction at the endpoint.

[0293] This form mathematically guarantees that the final trajectory endpoint will not deviate from the endpoint given by the trend branch.

[0294] The final future predicted trajectory is obtained by superimposing the final trajectories of K modes modally. Modal probability:

[0295]

[0296]

[0297] in, These represent the probabilities of the corresponding final trajectories.

[0298] Example 2

[0299] This embodiment provides a trajectory prediction method based on the decoupling of future low-frequency and historical high-frequency sources. Building upon Embodiment 1, in step S3, to avoid insufficient generalization when using fixed frequency band boundaries across different scenarios, target types, and datasets, this embodiment further provides a preferred implementation method for adaptive frequency band boundaries.

[0300] This embodiment preferably does not use a fixed frequency band boundary. Instead, it dynamically generates frequency band boundaries based on the accumulated energy of the trajectory spectrum and scene characteristics; let the current spectrum for which the frequency band boundaries need to be calculated be represented as... , Take historical spectrum Or a rough prediction of the future spectrum Or the fused spectrum of the two, defined Spectral energy for:

[0301]

[0302] in, For the number of traffic participants, Indicates spectrum input The number of channels, when hour, , For historical trajectory feature dimension; when hour, ;when When it is an intermediate characteristic spectrum, This corresponds to the number of feature channels; Indicates the current spectrum The Middle The traffic participant's number Each input channel in frequency index Complex spectral components at that location, Indicates frequency index;

[0303] Normalize the spectral energy:

[0304]

[0305] in, This represents the normalized spectral energy. To prevent extremely small constants with a denominator of zero, This represents the summation frequency index used to iterate through all frequency components;

[0306] Recalculate the cumulative energy distribution :

[0307]

[0308] Then, based on at least one of the following: global scene features, interaction detail features, target speed, or uncertainty, the low-frequency proportion of the target is generated. :

[0309] ( )

[0310] in, For scene global features, For interactive details, This represents the average velocity statistic of targets within the scene. This represents a statistic indicating uncertainty. and For learnable weight matrix and bias;

[0311] This yields the adaptive frequency band boundary coefficient. :

[0312]

[0313] in, This represents the sharpening factor for selecting the frequency band boundary; represents the normalized exponential function;

[0314] Constructing a low-frequency mask :

[0315]

[0316] and high-frequency mask :

[0317]

[0318] This method allows the high- and low-frequency division to dynamically change with the scene and trajectory structure, where... This is the frequency band boundary smoothing coefficient.

[0319] Example 3

[0320] This embodiment provides a trajectory prediction method based on heterogeneous decoupling of future low-frequency and historical high-frequency sources. Building upon Embodiment 1 or 2, to clarify the learning responsibilities of trend branches and detail branches, a heterogeneous consistency training objective is introduced during the training phase, setting the true future trajectory as... for:

[0321]

[0322] The winner-takes-all approach is used to select the prediction mode that best matches the actual future trajectory.

[0323]

[0324] in, This represents the predicted mode that best matches the actual future trajectory, i.e., the optimal mode. Indicates the first The final trajectory of each mode;

[0325] Low-frequency trend consistency loss for:

[0326]

[0327] in, This represents the global trend trajectory of the predictive mode that best approximates the actual future trajectory. These are the weighting coefficients. For low-frequency masking, The loss is a discrete Fourier transform; on the one hand, it constrains the trend trajectory to be close to the real future trajectory in the time domain, and on the other hand, it constrains its low-frequency spectrum to be consistent with the low-frequency spectrum of the real trajectory.

[0328] Define the true local detail residual for:

[0329]

[0330] High-frequency detail consistency loss for:

[0331]

[0332] in, For the detailed compensation trajectory of the optimal modality, These are the weighting coefficients. It is a high-frequency mask.

[0333] To prevent severe conflicts between the detailed compensation branch and the trend prediction branch, and to increase bridging consistency loss. for:

[0334]

[0335] in, Indicates the first The final trajectory endpoint of the (best) prediction mode at the Fth prediction time step; Indicates the endpoint of the global trend trajectory;

[0336] When using the strict endpoint preservation form, the endpoint compensation range is equivalently constrained:

[0337]

[0338] in, This indicates the end-point details to compensate for constraint losses. Indicates the first The amount of detail compensation for each prediction mode at the last prediction time step;

[0339] Final training total loss Represented as:

[0340]

[0341] in, For modal classification loss, These are the weighting coefficients for the corresponding loss terms.

Claims

1. A trajectory prediction method based on the decoupling of future low-frequency and historical high-frequency sources, characterized in that, Includes the following steps: S1. Obtain the historical trajectories of traffic participants and extract global scene features and detailed interaction features of traffic participants; S2. Input the global features of the scene and the detailed features of traffic participants' interactions into the coarse prediction network to generate the first stage of coarse prediction of future trajectories; S3. Perform frequency domain transformation on the first stage coarse prediction of the future trajectory, and extract the future low-frequency trend features through low-frequency masking. S4. Perform frequency domain transformation on the historical trajectory and extract historical high-frequency detail features through high-frequency masking; S5. Construct a trend prediction branch and a detail compensation branch using future low-frequency trend features and historical high-frequency detail features respectively. The trend prediction branch integrates future low-frequency trend features, scene global features, and interaction detail features to output a multimodal global trend trajectory and modal probability. The detail compensation branch integrates historical high-frequency detail features, scene global features, and interaction detail features to output a multimodal detail compensation trajectory. S6. By using a trend-preserving heterogeneous bridging coupling mechanism, the global trend trajectory and the detailed compensation trajectory are coupled to obtain the final future prediction trajectory.

2. The trajectory prediction method based on the decoupling of future low-frequency and historical high-frequency sources according to claim 1, characterized in that: In step S2, the method for generating the first-stage coarse prediction of the future trajectory is as follows: Based on scene global features and interactive details Through coarse prediction network Generate the first stage coarse prediction of future trajectory : in, Predict the length for the future.

3. The trajectory prediction method based on the decoupling of future low-frequency and historical high-frequency sources according to claim 1, characterized in that: In step S3, the method for extracting future low-frequency trend features is as follows: Rough prediction of future trajectory in the first stage Perform a frequency domain transformation along the time dimension to obtain the coarsely predicted future spectrum. Represented as: in, Represents the Discrete Fourier Transform; Using low-frequency masks Extracting future low-frequency spectrum , Indicates frequency index; uses fixed frequency band boundaries Define low-frequency mask as follows: Then the future low-frequency spectrum Represented as: in, This represents element-wise multiplication; The low-frequency transform network is denoted as follows: [The network is then mapped in the complex domain to represent the low-frequency spectrum]. ,but: in, This indicates the low-frequency mapping result; and These represent taking the real and imaginary parts of a complex number, respectively. , Let be the learnable weight matrices for the real and imaginary parts, respectively. , These are the learnable biases for the real and imaginary parts, respectively. For activation functions; The low-frequency mapping result is inversely transformed back to the time domain, and then subjected to pooling or projection operations to obtain future low-frequency trend features. : in, This represents the inverse discrete Fourier transform. This indicates a pooling operation. For the number of traffic participants, Indicates the dimension of the hidden feature layer.

4. The trajectory prediction method based on the decoupling of future low-frequency and historical high-frequency sources according to claim 3, characterized in that: A trend reliability assessment factor is introduced to assess the reliability of future low-frequency trend characteristics. The steps are as follows: First, calculate the proportion of low-frequency energy in the coarse prediction of the future trajectory in the first stage: in, This indicates the proportion of low-frequency energy in the rough prediction of future trajectories. Indicates frequency index, Indicates a low-frequency mask. This indicates that the coarse prediction of the future trajectory, after undergoing a discrete Fourier transform along the time dimension, is indexed at the frequency. Complex spectral components at that location, To prevent extremely small constants with a denominator of zero; Combined with global scene features After pooling, the reliability coefficient of the constructed trend is: in, This represents the pooling result of global features of the scene. This indicates a pooling operation. Indicates feature splicing, and These are the learnable weight matrix and bias of the reliability estimation module, respectively. For activation functions; Finally, the trend reliability coefficient is used to analyze the characteristics of future low-frequency trends. Perform adaptive weighting: in, The weighted future low-frequency trend characteristics, This indicates element-wise multiplication.

5. The trajectory prediction method based on the decoupling of future low-frequency and historical high-frequency sources according to claim 1, characterized in that: In step S4, the method for extracting historical high-frequency detail features is as follows: The true trajectory of history Perform a discrete Fourier transform to obtain the historical spectrum. Represented as: in, Represents the Discrete Fourier Transform; Define high-frequency mask : in, Indicates a low-frequency mask; Based on length High-frequency mask, extracting high-frequency spectrum from historical spectrum. : in, This represents element-wise multiplication; A high-frequency noise gating mechanism is introduced, as follows: First, calculate the average high-frequency energy at each frequency position in the historical trajectory spectrum and the average energy of its neighborhood to determine the spectral abrupt change ratio; Calculate the average high-frequency energy at each frequency position: in, Indicates the first The average high-frequency energy of each frequency component across all targets and all input channels. For the number of traffic participants, As a dimension of historical trajectory features, Indicates the first The traffic participant's number After the historical trajectory feature channels are subjected to discrete Fourier transform, the frequency index is... Complex spectral components at the location; Secondly, a neighborhood is constructed for each frequency location. Calculate the average energy of the neighborhood : in, Indicates the first The average high-frequency energy of each frequency, among which ; Spectral mutation ratio Defined as: in, To prevent extremely small constants with a denominator of zero; A high-frequency noise gating factor is defined based on the aforementioned spectral mutation ratio. : in, For activation function, For scaling parameters, For threshold parameters; Based on the aforementioned high-frequency noise gating factor, the high-frequency spectrum is suppressed or preserved, resulting in the gated historical high-frequency spectrum. Represented as: Perform an inverse transform on the gated historical high-frequency spectrum to obtain the historical high-frequency trajectory signal. : in, This represents the inverse discrete Fourier transform. This indicates taking the real part of a complex number; High-frequency feature extraction network Obtain historical high-frequency detailed features : in, It is a multilayer perceptron. For the number of traffic participants, Indicates the dimension of the hidden feature layer.

6. The trajectory prediction method based on heterogeneous decoupling of future low-frequency and historical high-frequency sources according to claim 1, characterized in that: The frequency domain transformation in steps S3 and S4 employs an adaptive frequency band boundary, which is generated in the following manner: Let the current spectrum for which the frequency band boundary needs to be calculated be represented as: , Take historical spectrum Or a rough prediction of the future spectrum Or the fused spectrum of the two, defined In frequency index The spectral energy obtained by averaging along all traffic participants and input channels at the point of origin. for: in, For the number of traffic participants, Indicates spectrum input The number of channels, when hour, , For historical trajectory feature dimension; when hour, ;when When it is an intermediate characteristic spectrum, This corresponds to the number of feature channels; Indicates the current spectrum The Middle The traffic participant's number Each input channel in frequency index Complex spectral components at that location, Indicates frequency index; Normalize the spectral energy: in, This represents the normalized spectral energy. To prevent extremely small constants with a denominator of zero, This represents the summation frequency index used to iterate through all frequency components; Then calculate the cumulative energy distribution : Then, based on at least one of the following: global scene features, interaction detail features, target speed, or uncertainty, the low-frequency proportion of the target is generated. : ( ) in, For scene global features, For interactive details, This represents the average velocity statistic of targets within the scene. This represents a statistic indicating uncertainty. and For learnable weight matrix and bias; This yields the adaptive frequency band boundary coefficient. : in, This represents the sharpening factor for selecting the frequency band boundary; This represents the normalized exponential function; Constructing a low-frequency mask : and high-frequency mask : The high- and low-frequency division changes dynamically with the scene and trajectory structure, where This is the frequency band boundary smoothing coefficient.

7. The trajectory prediction method based on heterogeneous decoupling of future low-frequency and historical high-frequency sources according to claim 1, characterized in that: In step S5, after obtaining the weighted future low-frequency trend characteristics... and historical high-frequency details Then, a heterogeneous decoupled prediction structure is constructed, including a trend prediction branch and a detail compensation branch; The future low-frequency trend features are obtained by fusion and weighting of the trend prediction branch. Scene global features Interactive details To obtain trend representation : in, for, A multi-head attention mechanism module for trend representation is used to complete feature fusion; Output A global trend trajectory and its modal probabilities : in, , , For trajectory generation networks, For the number of traffic participants, For predicting future length, The number of trajectory modes; The detail compensation branch fuses historical high-frequency detail features. Scene global features Interactive details , to obtain detailed representation : in, A multi-head attention mechanism module for detailed representation is used to complete feature fusion; Output Detailed compensation trajectory : in, , A trajectory generation network.

8. The trajectory prediction method based on the decoupling of future low-frequency and historical high-frequency sources according to claim 1, characterized in that: In step S6, the trend-maintaining heterogeneous bridging coupling mechanism is as follows: For the Each modality, taking the endpoint of the trend trajectory. : in, For predicting future length, Indicates the first Global trend trajectory of each modality; Injecting the endpoint of the trend trajectory into the detailed compensation branch as conditional information yields the first... Detailed compensation trajectory for each modality : in, For trajectory generation networks, Indicates splicing, Detailed representation of the compensation branch; Define with prediction time step Decreasing detail injection weights : in, , These are the attenuation control parameters; Then time step The Final trajectory of each mode Represented as: in, For time steps The Global trend trajectory of each modality For time steps The Modal detail compensation trajectory; For each mode Final trajectory The output is: The final future predicted trajectory is obtained by superimposing the final trajectories of K modes modally. Modal probability: in, These represent the probabilities of the corresponding final trajectories.

9. The trajectory prediction method based on the decoupling of future low-frequency and historical high-frequency sources according to claim 1, characterized in that: In step S6, directly set: in, This represents the detailed compensated trajectory after time step t has been zeroed out at the endpoint. and The time step t and the endpoint are respectively the first and second steps. Modal detail compensation trajectory; Therefore: in, This represents the end-point detail compensation trajectory after end-point zeroing correction; For each mode Final trajectory The output is: in, Indicates the first Global trend trajectory of each modality This represents the detailed compensation trajectory after zeroing correction at the endpoint.

Images