Vehicle surrounding environment target trajectory prediction method, system, device and medium

Abstract

According to the vehicle surrounding environment target trajectory prediction method, system and device and the medium provided by the invention, the dynamic characteristics of the trajectory are captured by integrating time-frequency analysis, so that the prediction accuracy is improved; capturing a global driving trend and local motion details of the vehicle through current information and scene information of the vehicle and historical track information; an encoder-decoder structure aims at generating wavelet coefficients, a wavelet attention mechanism is integrated in a decoder, and a specific scale mode in historical trajectory data is utilized, so that the capability of learning driving modes of different scales by a model is enhanced, and the prediction performance is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field related to target detection, and in particular relates to a method, system, device and medium for predicting target trajectory in the surrounding environment of a vehicle. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Vehicle movement patterns captured as a series of data points such as locations, timestamps, etc. have become increasingly common due to the popularity of navigation tools and location-recording technologies such as smartphones. These patterns are now considered a rich data source for intelligent transportation networks, providing crucial insights for practical applications such as movement prediction, anomaly detection, trajectory clustering, trajectory similarity measurement, and travel time estimation. The increasing availability and application of vehicle movement data has prompted the creation of models to effectively learn and apply this data.

[0004] To improve the performance of practical applications, it is necessary to develop a trajectory learning model that can effectively extract rich semantic information from these trajectories, including the vehicle's motion behavior and environmental information. In addition, efficient extraction of this information is also necessary to reduce the computational burden and response time. However, there are still many challenges to achieve these goals.

[0005] First, it is challenging to extract continuous motion patterns from irregular and discrete trajectory points. The motion of a vehicle is inherently fluid in both spatial and temporal dimensions. However, this motion is captured using discrete data points. The relationships between these points may not fully capture the fluidity of the vehicle motion. Traditional sequential analysis models such as RNNs and Transformers focus only on the connections between these discrete points, which limits their ability to accurately reproduce the dynamics of vehicle motion patterns. Neural ODEs and custom-designed kernels combined with Transformer models explicitly extract the continuity of vehicle motion from intermittently recorded data points. While these methods show improved performance, they also incur significant computational overhead that may hinder the efficiency of the model. Neural ODEs require the analysis of integral equations, which is computationally expensive even with modern hardware, while Transformers involve quadratic computational complexity. Therefore, there is currently no cost-effective solution to accurately model continuous motion from discrete trajectory data points.

[0006] Second, capturing the underlying patterns of vehicle trajectories is a challenge. As mentioned above, short-term trajectory prediction is essentially a time series prediction problem, which is achieved by modeling the complex nonlinear transition patterns of driving trajectories (multiple interrelated dynamic attributes at each time instant). Although temporal modeling has been used to capture the autoregressive nature of dynamic attributes, it is still a challenging task to study the underlying driving patterns in sufficient detail. Generally speaking, the main dynamic characteristics come from the vehicle's maneuvers during the straight and turning phases and the operation of the intended drive. Considering the safety of the vehicle and the comfort of the driver, the intensity of the maneuvering control is restricted to obtain limited significance on the temporal trajectory series, resulting in the inability of traditional models to capture such maneuvering patterns. A practical solution is to utilize frequency domain information to capture informative patterns from time-frequency features to support the trajectory prediction task. In the field of TSF, frequency domain analysis is used to decompose time series to facilitate deep reasoning.

[0007] Considering the time series nature of trajectory prediction tasks, frequency information also involves trajectory attributes. For example, during driving, the position and direction always change along the overall evolution from the starting point to the end point, so the coordinates of the driving trajectory can be regarded as a partial description of the global driving trend. Driven by driving intentions (going straight, turning, turning around, etc.), as local details of vehicle motion, the time series of position, orientation, and speed will also change accordingly. However, current methods mainly focus on time domain modeling, and time-frequency analysis is still a virgin work in trajectory prediction research, without fine decomposition of frequency details.

[0008] Therefore, how to conduct time-frequency analysis to capture the potential patterns of driving trajectories and thus achieve trajectory prediction tasks from a more refined perspective is a problem that needs to be solved at present. Summary of the invention

[0009] To overcome the above-mentioned deficiencies of the prior art, the present invention provides a method, system, device and medium for predicting the trajectory of a target in the surrounding environment of a vehicle, which combines wavelet analysis and an encoder-decoder to capture the overall driving pattern and fine-grained vehicle motion details, thereby improving the accuracy of the prediction.

[0010] In order to achieve the above object, the present invention adopts the following technical solution:

[0011] In a first aspect, the present invention provides a method for predicting a target trajectory in a vehicle surrounding environment, comprising:

[0012] Obtain historical trajectory data, current vehicle information, and current vehicle scene information;

[0013] Using a target vehicle encoder and a map encoder, respectively encoding the vehicle information at the current moment and the vehicle scene information at the current moment to obtain a vehicle coding feature and a scene coding feature;

[0014] Performing discrete wavelet transform on the historical trajectory data to decompose the obtained wavelet coefficients of different time scales and frequencies;

[0015] The high-dimensional trajectory embedding features of the historical trajectory data are extracted by using an input embedding network, and the high-dimensional trajectory embedding features are enhanced and convolved by using a wavelet attention module to generate context perception of the wavelet coefficients, and the vehicle coding features, the scene coding features, and the context perception of the wavelet coefficients are respectively input into different decoders corresponding to different wavelet transform coefficients to obtain different sub-band time-frequency attributes of the trajectory sequence;

[0016] The inverse discrete wavelet transform is used to inversely transform the time-frequency attributes of different sub-bands of the obtained trajectory sequence to predict the trajectory of the vehicle at the next moment.

[0017] In a second aspect, the present invention provides a vehicle surrounding environment target trajectory prediction system, comprising:

[0018] An acquisition module is configured to: acquire historical trajectory data, current vehicle information, and current vehicle scene information;

[0019] The encoding module is configured to: respectively encode the vehicle information at the current moment and the vehicle scene information at the current moment using a target vehicle encoder and a map encoder to obtain a vehicle encoding feature and a scene encoding feature;

[0020] A wavelet transform module is configured to: perform discrete wavelet transform on the historical trajectory data to decompose the obtained wavelet coefficients of different time scales and frequencies;

[0021] An embedding module is configured to: extract high-dimensional trajectory embedding features of the historical trajectory data using an input embedding network, perform enhancement and convolution operations on the high-dimensional trajectory embedding features using a wavelet attention module to generate context perception of the wavelet coefficients, and input the vehicle coding features, the scene coding features, and the context perception of the wavelet coefficients into different decoders corresponding to different wavelet transform coefficients, respectively, to obtain different sub-band time-frequency attributes of the trajectory sequence;

[0022] The prediction module is configured to: use inverse discrete wavelet transform to inversely transform the time-frequency attributes of different sub-bands of the obtained trajectory sequence, and predict the trajectory of the vehicle at the next moment.

[0023] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in the first aspect is performed.

[0024] In a fourth aspect, the present invention provides a computer-readable storage medium for storing computer instructions, wherein when the computer instructions are executed by a processor, the method described in the first aspect is performed.

[0025] One or more of the above technical solutions have the following beneficial effects:

[0026] In the present invention, the dynamic characteristics of the trajectory are captured by integrating time-frequency analysis, thereby improving the accuracy of prediction; the global driving trend and local motion details of the vehicle are captured through the vehicle's current information and scene information, as well as historical trajectory information; its encoder-decoder structure is designed to generate wavelet coefficients, and the wavelet attention mechanism is integrated in the decoder, which utilizes specific scale patterns in historical trajectory data to enhance the model's ability to learn driving patterns of different scales and improve prediction performance.

[0027] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0029] Figure 1 This is an overall schematic diagram of the WMTP framework in Embodiment 1 of the present invention;

[0030] Figure 2 This is a schematic diagram of the structure of an encoder in Embodiment 1 of the present invention;

[0031] Figure 3 This is a schematic diagram of the structure of a decoder in Embodiment 1 of the present invention;

[0032] Figure 4 Schematic diagram of the wavelet attention mechanism processing process in the first embodiment of the present invention;

[0033] Figure 5 (a)-(h) are descriptive analysis diagrams of the Argoverse dataset motion prediction in Example 1 of the present invention. DETAILED DESCRIPTION

[0034] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0035] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0036] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0037] Embodiment 1

[0038] This embodiment discloses a method for predicting a target trajectory in a vehicle surrounding environment, including:

[0039] Obtain historical trajectory data, current vehicle information, and current vehicle scene information;

[0040] Using a target vehicle encoder and a map encoder, respectively encoding the vehicle information at the current moment and the vehicle scene information at the current moment to obtain a vehicle coding feature and a scene coding feature;

[0041] Performing discrete wavelet transform on the historical trajectory data to decompose the obtained wavelet coefficients of different time scales and frequencies;

[0042] The high-dimensional trajectory embedding features of the historical trajectory data are extracted by using an input embedding network, and the high-dimensional trajectory embedding features are enhanced and convolved by using a wavelet attention module to generate context perception of the wavelet coefficients, and the vehicle coding features, the scene coding features, and the context perception of the wavelet coefficients are respectively input into different decoders corresponding to different wavelet transform coefficients to obtain different sub-band time-frequency attributes of the trajectory sequence;

[0043] The inverse discrete wavelet transform is used to inversely transform the time-frequency attributes of different sub-bands of the obtained trajectory sequence to predict the trajectory of the vehicle at the next moment.

[0044] This embodiment proposes a new framework called Wavelet Mamba Trajectory Prediction, or WMTP, which combines wavelet analysis with a Mamba backbone network to capture both overall driving patterns and fine-grained vehicle motion details. This analysis method uses discrete wavelet transform (DWT) to decompose trajectory data into wavelet coefficients representing different time scales and frequencies; these coefficients can then be used to reconstruct the trajectory through an inverse wavelet transform (IDWT) process. In terms of trajectory prediction, the main goal is to extract the most effective wavelet coefficients from the input data so that the inverse wavelet transform (IDWT) can accurately predict subsequent trajectory points.

[0045] In this embodiment, discrete wavelet transform DWT is introduced into time-frequency analysis. The filter bank derived from wavelet transform can effectively capture the overall motion trend and intricate motion details of the driving path. Specifically, the wavelet coefficients of each terminal node in the filter bank are called wavelet transform components WTC, which are arranged in order of increasing frequency. For example, in the three-level DWT, WTC 0 corresponds to the coefficients of the highest level of approximation, while the set {WTC 4-i} i∈(0,3)∩i∈Z Represents the detail coefficient of level i. Through wavelet multi-resolution analysis, WTC 0 It can describe the overall trend of time series gradual change, and {WTC 4-i} i∈(0,3)∩i∈Z Represents fine-grained dynamics characterized by rapid fluctuations.

[0046] Given that wavelet analysis includes both broader dynamics and more detailed dynamics, the time-frequency properties of trajectories revealed by wavelets can be used to reveal the intrinsic properties of driving dynamics. Based on this insight, this embodiment introduces the wavelet-mamba trajectory predictor framework, namely WMTP, which aims to distinguish the subtle differences between global driving patterns and local movements from the perspective of time-frequency, thereby improving the recognition ability of driving behavior and the accuracy of the prediction model.

[0047] State space models (SSMs) are a family of models that perform continuous autoregression on data using unobserved states. A prominent example is the S4 model, whose autoregressive process is defined as follows:

[0048]

[0049] y t =Ch t +Dx t (2)

[0050] Among them, x t is the input signal, h t represents the potential state, y t is the output signal. Parameters C and D are the matrix coefficients. and It is derived from zero-order hold (ZOH) and Euler's method, which discretizes continuous systems. Although the S4 model provides an efficient framework for sequence modeling, there are ongoing efforts to extend its applicability to longer and more complex sequences.

[0051] Based on the S4 model, the S5 model integrates advanced features such as parallel scanning and MIMO-SSM, which enhances the ability of SSM to model sequences. Currently, a data-driven SSM layer called Mamba is introduced, which parameterizes the matrices Δ, B, and C based on sequence data, thereby improving sequence capture and modeling of extended sequences. The design of the Mamba module allows linear scaling with sequence length, providing efficient memory usage and fast inference.

[0052] This example develops a motion prediction model based on SSM and further investigates its potential in time series prediction tasks. For trajectory prediction tasks, the Wavelet-Mamba trajectory predictor WMTP method is centered on predicting wavelet coefficients, which enables the execution of an inverse discrete wavelet transform IDWT algorithm that reconstructs the original trajectory sequence and predicts the position of subsequent time intervals.

[0053] To achieve this, this embodiment constructs a motion prediction model with an encoder-decoder structure, such as Figure 1 As shown. The motion prediction model consists of a trajectory embedding layer, a pair of encoders and a set of decoders aligned with wavelet elements, followed by a series of inverse discrete wavelet transforms (IDWT). Among them, a pair of encoders is responsible for extracting surrounding vehicle and map features and splicing the outputs; the trajectory embedding layer is responsible for embedding historical trajectories; the outputs of the encoder and embedding layers are used as inputs to the decoder, and the decoder predicts the output wavelet components. Subsequently, the IDWT synthesizes the wavelet elements into a series of trajectories covering historical and upcoming motions, and uses the mean square error of the reconstructed wavelet components as the loss function to optimize the motion prediction model parameters. Unlike traditional trajectory prediction techniques, the WMTP framework proposed in this embodiment aims not only to predict the trajectory of the next time step, but also to reconstruct past trajectory sequences. This dual function allows a comprehensive understanding of the multi-scale characteristics of motion patterns, thereby improving the accuracy of predictions.

[0054] In the motion prediction model proposed in this embodiment, the historical trajectory data It is initially converted into a high-dimensional vector through the input embedding network, denoted as I in formula (3). At the same time, the encoder extracts Agent and Map features and concatenates the extracted features, denoted as G in formula (4). The concatenation result G is passed to the decoder together with the high-dimensional vector of the historical trajectory, denoted as H in formula (5). This latent representation encapsulates the temporal dynamics of the trajectory, which helps the subsequent prediction step and enables learning of various driving behaviors.

[0055]

[0056] G=Cat(Encoder(A),Encoder(M′)) (4)

[0057] H=Gather(H′,G) (5)

[0058] Among them, A and M′ represent vehicle status information and scene information respectively. Cat(·) represents the concatenation operation, Gather(·) represents the collection of input information, and H′ represents the output of the input embedding network. represents a historical trajectory point, N represents the Nth trajectory point, M represents the selected M trajectory points, G represents the concatenated output of the two encoders, and H represents the input of the decoder.

[0059] The architecture of the motion prediction model proposed in this embodiment combines multiple decoders to facilitate in-depth feature extraction at various driving behavior scales. In the context of L-level wavelet decomposition, the specific task of each decoder is to generate different sub-band time-frequency attributes of the trajectory sequence, denoted as Q i , represents the i-th wavelet transform coefficient WTC i , specifically:

[0060] Q i =Decoder i (H),i=0,1,...,L (6)

[0061] Among them, Q i represents the i-th wavelet coefficient, and L represents the wavelet decomposition level.

[0062] To improve the learning ability of the model, a novel wavelet attention mechanism is integrated in the decoder. This mechanism, together with the historical trajectory embeddings, generates scale-specific features, which are subsequently exploited by a Mamba-inspired module to grasp temporal dependencies.

[0063] Finally, the expected trajectory points are derived through the inverse discrete wavelet transform (IDWT) process. The basic feature of this process, which combines the stacked wavelet transform coefficients WTC from various decoders, is:

[0064]

[0065] in, Represents M historical trajectory points, namely the NMth to the N-1th, a total of M.

[0066] In order to update the model parameters, the wavelet component MSE loss method is introduced, which measures the difference between the predicted and actual wavelet component values, to improve the MRA capability and achieve model convergence.

[0067] The following is a detailed description of each part involved in this embodiment:

[0068] Input trajectory embedding: Each point in a trajectory is represented by a distinct vector in a three-dimensional (3D) space defined by its position and velocity. In order to effectively capture the latent features of the trajectory for further analysis by subsequent layers, these vectors must be transformed into a high-dimensional feature space. To this end, this embodiment develops an input embedding network that uses a linear mapping to enhance the representation of the trajectory sequence as follows:

[0069]

[0070] Where I∈R M×D Represents the input trajectory sequence A high-dimensional representation of . The matrix W i1 ∈R d ×(D / / 2) and W i2 ∈R (D / / 2)×D It is used to linearly project the low-dimensional attribute vector of the trajectory point into the high-dimensional feature space, and d represents the order of wavelet decomposition. The ReLU activation function is represented by σ(·) and is used to enhance the model's ability to capture nonlinear relationships. M represents the count of historical trajectory points, while D represents the dimension of the feature space, and N refers to the Nth point to be predicted.

[0071] The input embedding network is able to cleverly capture the intrinsic characteristics of trajectory points, including the attribute dynamics of individual points. Nevertheless, for accurate trajectory prediction over a sequence of trajectories, high-level semantic feature extraction is of great importance. To address this issue, this embodiment utilizes a Mamba-inspired encoder to extract features, resulting in dynamic, task-specific, feature-rich scene embeddings. This approach enables complex analysis of the interactions occurring in a trajectory sequence.

[0072] like Figure 2 As shown in the figure, the encoder has two encoders: the target vehicle encoder and the map encoder. The target vehicle encoder and the map encoder are each composed of multiple Mamba blocks. The outputs of the two are spliced ​​together and used as the input of the downstream decoder. The input of the Agent Encoder is the state of the vehicle, such as speed, direction, position, etc. The input of the map encoder is the map or environmental features, such as lane lines. The task of the target vehicle encoder is to encode information related to the movement of the target vehicle, such as speed and position. At the same time, the map encoder is responsible for encoding scene information. The resulting hidden state generated by the encoder is regarded as a high-level abstract semantic feature, which is embedded as follows:

[0073] J,h e =Mamba(F,h 0 ) (9)

[0074] in, are initialized to zero, representing the initial hidden state. The input of the Mamba module is composed of F∈R M×D Indicates that h e represents the hidden state of the output, F represents the state information or scene information of the vehicle, and J represents the output of the Mamba block.

[0075] The output of the Mamba module is divided into two parts: J∈R M×D and Variable S 1 Corresponds to the number of layers in the recurrent structure.

[0076] The encoder of this embodiment uses Mamba blocks to extract high-level trajectory embeddings rich in temporal details, providing a comprehensive description of the input trajectory sequence. In the method of this embodiment, a key prerequisite for trajectory prediction is to predict the wavelet coefficients required for IDWT execution. Over-reliance on traditional wavelet feature engineering, such as simply decomposing the sequence and feeding it partially into the model, cannot fully utilize the multi-resolution capabilities of wavelet analysis.

[0077] Figure 3 This is the decoder structure diagram in this embodiment. The wavelet transform coefficient WTC, as a time-frequency representation of the driving mode, provides a series of dynamic features of different resolutions. In order to cleverly analyze the trajectory embedding and conduct in-depth analysis of the overall trend and details of the driving behavior, a wavelet attention mechanism is integrated in each decoder to extract hierarchical features from the trajectory data. The components generated by the decoder illustrate the overall driving trend and local motion details. The output of the encoder splicing is used as hidden state initialization, and the historical trajectory embedding is used as the input of the Mamba block after passing through the wavelet attention module. LN is the LayerNorm layer and FC is the fully connected layer.

[0078] The motion prediction model of this embodiment includes multiple decoders, so that various decoders can grasp the time-frequency characteristics of trajectory attributes at different scales. The purpose of the decoder is to obtain wavelet components. According to the preset wavelet decomposition level, the number of decoders and the level are the same, and the wavelet components of the same level are generated. For example, if the preset decomposition level is 2, the two decoders need to output two wavelet components of different frequency bands. Each decoder includes several cascaded Mamba blocks, and each Mamba block is connected to the LayerNorm layer and the fully connected layer in sequence. The output of the fully connected layer after each Mamba block will be used as the output of the decoder together;

[0079] like Figure 4 As shown in Figure 2, wavelet attention mechanism: Initially, there is an enhancement stage in which the trajectory embeddings derived from the input embedding network are modulated by attention coefficients to generate enhanced, scale-specific features, i.e., refined trajectory embeddings. These attention coefficients (score i,i=0,1,…,M-1) are derived from trajectory embeddings and used to assess the importance of past trajectory positions. Subsequently, the second stage involves convolution operations to refine these enhanced embeddings into context embeddings. These are then used to generate wavelet components and are synchronized to match the time scale of their respective wavelet components.

[0080] Specifically, the architecture of the Watt module consists of two stages: the enhancement stage and the convolution stage. The trajectory embeddings derived from the encoder are enhanced by the attention scores to produce refined scale-oriented features, fundamentally enriching the trajectory embeddings. Subsequently, the WAtt module applies convolution operations on these enhanced embeddings to generate context-aware embeddings of WTC. Mathematically, the enhancement process of the WAtt module can be expressed as follows:

[0081] I p =σ(I·W p1 )·W p2 (10)

[0082] E=γ(W s2 ·σ(W s1 I p )) (11)

[0083] I h =I+Diag(E·I) (12)

[0084] Among them, W p1 ∈R D×(D / / 2) and W p2 ∈R (D / / 2)×1 Used to linearly project trajectory embeddings to aggregate features I p ∈R M×1 In. W s1 ∈R (M / / 2)×M , W s2 ∈R M×(M / / 2) Used to evaluate the importance of features within trajectory embedding. The ReLU activation function is represented by σ, while the Sigmoid activation function is represented by γ. The feature E activated by Sigmoid = R M×1 Acts as an attention score, measuring the relevance of past trajectory points to upcoming trajectory points. h =R M×D It is achieved by embedding the original embedding I with the attention weighted embedding E = R M×1 The operator Diag(·) constructs a diagonal matrix from the input elements.

[0085] Once the refined trajectory embedding is obtained, the subsequent task converts it into a contextual embedding of WTC. To achieve this, the WAtt module applies a 1D convolution operation to align the enhanced trajectory embeddings, resulting in a highly correlated contextual embedding C∈Rh×D , the corresponding WTC obtained by discrete wavelet transform of these embeddings and historical trajectories maintain the same time dimension h. This process can be mathematically expressed as follows:

[0086] C = σ(Conv1d(I h )) (13)

[0087] For the time dimension of C, the length of the L-level WTC derived from the discrete wavelet transform DWT is typically reduced to about half the length of the original series due to downsampling. In addition, the DWT typically involves shifting the wavelet function to convolve the time series. When the filter approaches the boundary of a finite signal, the convolution operation requires values ​​beyond the signal limit, which can be addressed by signal extension techniques. Therefore, both the choice of wavelet and the extension method affect the exact length of the WTC. Choosing a symmetric extension is a common choice to guarantee signal continuity at the boundary. With this approach, the length of the WTC for the L-level DWT in a convolution-based algorithm can be derived using the following equation:

[0088]

[0089] b L =b L-1 (16)

[0090] Where L is the level of wavelet analysis, l is half the length of the wavelet filter selected in the DWT procedure, and M is the number of historical trajectory points. i} i∈[0,L)∩i∈Z Indicates the duration of high frequency WTC L-i , b i Indicates the duration of high frequency. L Indicates the duration of low frequency WTC 0 . Operator Gives the largest integer less than or equal to the input.

[0091] After obtaining the contextual embeddings from the WAtt module, the decoder’s Mamba module continues to refine these embeddings to produce wavelet embeddings; these are then linearly transformed to obtain WTCs:

[0092] H w =Mamba(G,C) (17)

[0093] Q=FC(LN(H w )) (18)

[0094] Among them, H w Represents the output of the first Mamba block.

[0095] The Mamba block is preprocessed using the encoder’s long-term memory to provide it with prior knowledge of historical trajectory sequences and enhance its ability to exploit scale-specific features. The resulting output feature H w ∈R H×D It acts as a wavelet embedding, and then passes through the LayerNorm operation LN(·) and linear projection FC(·), and finally derives the required wavelet coefficients Q∈R h×d .

[0096] In this embodiment, the inverse discrete wavelet transform (IDWT) module performs an inverse transform on the output of each decoder, i.e., the WTCs of each attribute, so as to reconstruct the historical trajectory sequence and predict the trajectory attributes at the next moment. Converted to a trajectory attribute series of size d×(M+1).

[0097] Specifically, for each attribute, the filter sequentially merges low-frequency and high-frequency coefficient pairs until the time domain series of the attribute is reconstructed. Mathematically, for the j-th attribute, the time series is derived as follows:

[0098]

[0099] Among them, all matrix indices start from zero. At the end of the time series, in order to account for signal expansion, Trim(·) is used to eliminate redundant segments, leaving only the initial M+1 elements along the time axis. It is worth noting that the first M elements correspond to the reconstructed historical trajectory sequence, while the last element represents the expected vehicle state at the next time interval.

[0100] In order to enable various decoders to grasp the time-frequency characteristics of trajectory attributes at different scales, this embodiment constructs a wavelet loss function. This function summarizes the mean square error of the predicted WTC and guides the parameter update of the network, as shown below:

[0101]

[0102] Where L represents the depth of wavelet analysis, h L-k Represents the kth wavelet transform coefficient WTC k The length of g=6 specifies the attribute count considered in this embodiment.

[0103] In formula (21), corresponds to the i-th element of the k-th WTC predicted by the (k+1)-th decoder for the j-th attribute, and Represents the actual value used for comparison.

[0104] Trajectory prediction of objects in the environment surrounding an autonomous vehicle is crucial for autonomous vehicles. Learning from trajectories is of great significance for extracting valuable semantic data to improve the accuracy of subsequent tasks. However, the motion of vehicles is continuous in space and time, and it is challenging to effectively extract data from scattered trajectory data points. In addition, identifying subtle patterns in vehicle trajectories is another difficulty, which are not obvious in the original data and are difficult to distinguish from complex scene details related to specific areas and road features. To address these problems, this embodiment proposes a new framework called Wavelet-Mamba Trajectory Prediction (WMTP), which incorporates wavelet analysis through state-space modeling to capture the overall trend of driving patterns and the details of vehicle motion. The method uses discrete wavelet transform (DWT) to decompose trajectory data into wavelet coefficients representing different time scales and frequencies, and then uses these coefficients to reconstruct the trajectory through the inverse wavelet transform (IDWT) process. By designing an encoder-decoder neural architecture, latent temporal features are learned from the input trajectory sequence and projected into the wavelet domain, so that wavelet coefficient estimates of historical and future trajectories are generated using decoders with different scale orientations. Through the inverse wavelet transform (IDWT) module, these coefficients are used to reconstruct the trajectory, which improves the prediction accuracy.

[0105] The following is an experimental description of the method proposed in this embodiment:

[0106] A. Experimental Setup

[0107] Datasets: The motion prediction datasets of Argoverse 1 and Argoverse 2 are used to evaluate the performance of the methods. Argoverse 1 is divided into 205,942 training sequences, 39,472 validation sequences, and 78,143 test sequences, each captured at a frequency of 10 Hz. The goal is to predict the trajectory for the next 3 seconds using 2 seconds of past data. Argoverse 2 consists of 200,000 training, 25,000 validation, and 25,000 test sequences, also sampled at 10 Hz. It differs in that it uses a 5-second historical window to predict the motion for the next 6 seconds. Both datasets provide high-definition maps.

[0108] Metrics: We mainly use popular standard metrics in the field of multimodal trajectory prediction, including the minimum average displacement error (minADE k ), minimum final displacement error (minFDE k ) and bit error rate (MR k ). These metrics evaluate the accuracy of the best predicted trajectory of a single target vehicle among K possible hypotheses compared to the actual trajectory. minADE k Measures the average Euclidean distance between the predicted path and the ground truth, while minFDEk Focus only on endpoint differences. MR k Indicates minFDE k The proportion of sequences exceeding the 2-meter threshold.

[0109] Parameter settings: The model is configured with a latent vector dimension D of 128, 8 layers for both encoder and decoder, 2 wavelet analysis levels, and 2 decoder modules. The multi-mode decoder is set to k=6 output modes, which is standard practice. The Adam optimizer is used with a batch size of 16, and the learning rate starts at 1e-3 and gradually decreases to 1e-5 after 150 iterations.

[0110] like Figure 5 As shown, the model algorithm predicts the realistic, reasonable, and accurate trajectory prediction capabilities of all target vehicles in the scene on the Argoverse 1 and 2 datasets.

[0111] In order to further study the performance contribution of the method proposed in this embodiment, especially the impact of different wavelet analysis depths and wavelet attention modules on trajectory prediction, an ablation experiment was conducted. The configuration is summarized as follows:

[0112] A1: Level 1 WMTP - We explored the impact of a more in-depth wavelet analysis on the prediction accuracy, keeping the other parameters the same as the basic MTP model.

[0113] A2: 3-level WMTP-Extended Analysis, this configuration increases the wavelet depth to 3 while keeping other hyperparameters the same.

[0114] B1: WMTP without WAtt - This variant omits the wavelet attention module WAtt to evaluate its contribution and the value of time-frequency analysis in trajectory prediction. It uses autoregressive inference to generate trajectories.

[0115] B2: 2-level WMTP without WAtt - This setup mirrors B1, but with a two-level wavelet analysis.

[0116] B3: 3-level WMTP without WAtt - Similar to B2, but with a three-level wavelet analysis, this configuration further examines the role of the WAtt module.

[0117] Table 3: Experimental results of the Argoverse 1 motion prediction dataset test segmentation ablation study

[0118] Models WAtt minADE_6 minFDE_6 MR_6 MDE WMTP w / 0.78 1.16 12.2 0.3727 A1 w / 0.80 1.17 12.3 0.3855 A2 w / 0.85 1.27 12.3 0.5271 B1 w / o 0.88 1.25 12.4 0.5090 B2 w / o 0.85 1.21 12.3 0.4642 B3 w / o 0.90 1.28 12.5 0.5186

[0119] Table 3 lists the experimental results of the ablation study and draws the following conclusions:

[0120] 1. Experimental results on the depth of wavelet analysis show that the proposed method maintains comparable performance at different analysis levels. Specifically, 2-level wavelet analysis leads to the strongest trajectory prediction, while 1-level and 3-level analysis lead to degraded predictions. WMTP outperforms the A1 framework on all metrics, which may be due to the high-frequency wavelet transform coefficients (WTC) capturing more detailed dynamics, thus facilitating fine-grained trajectory modeling at the decoder. However, the lack of the wavelet attention (WAtt) module in B2 is also an improvement over B1, suggesting that higher wavelet levels may be beneficial in the absence of WAtt.

[0121] The superior performance of WMTP compared to A1 may be attributed to more WTCs and richer details of trajectory dynamics. In contrast, B2 and B3, which lack WAtt, perform poorly, indicating that excessive wavelet depth may be harmful without WAtt compensation. This is mainly due to the shorter WTC length, as shown in Equation (14). The minimum WTC length of A2 and B3 is only 2, which hinders the decoder's ability to identify trajectory evolution.

[0122] 2. Experimental results consistently show that the WAtt module improves the WTC prediction accuracy by capturing scale-oriented features and enriching the contextual representation of cross-scale trajectory sequences. The WAtt module reduces the prediction error of all metrics at the equivalent wavelet level.

[0123] The performance degradation observed in the 3-level wavelet analysis in A2 is related to the impaired concise temporal modeling of WTC. The WAtt module is able to enhance the correlation between historical trajectory data and future motion at a specific scale, enabling the WMTP framework to predict trajectory patterns more accurately. As a result, it surpasses B3 in all the proposed position component metrics.

[0124] 3. As shown in Table 3, although the wavelet-based framework proposed in this embodiment has a simple encoder-decoder structure, it is good at trajectory prediction even without the WAtt module. This is mainly due to the time-frequency analysis of the framework, which cleverly captures the global trends and local nuances of various driving modes. At the same time, the WMTP framework uses deep feature extraction to identify multi-resolution vehicle motion features. The MDE indicator in Table 3 shows that the model based on time-frequency analysis in this embodiment can predict driving trajectories with high accuracy, highlighting its robustness and practicality. These findings further verify the effectiveness of time-frequency analysis in trajectory prediction tasks.

[0125] This paper introduces a new time-frequency analysis method to predict driving trajectories, providing a new perspective for improving the modeling of trajectory patterns. The newly developed framework, called WMTP, emphasizes the importance of time-frequency analysis in predicting trajectories and overcomes the limitations of traditional techniques that have difficulty capturing a wide range of subtle trajectory features. The framework uses discrete wavelet transform to establish a general time-frequency model, which is designed to refine wavelet coefficients, which are crucial for reconstructing past trajectories and predicting future positions. The encoder-decoder neural network architecture uses historical trajectory data to generate wavelet coefficients, which are used for trajectory prediction by the inverse discrete wavelet transform (IDWT). In addition, the framework introduces a wavelet attention mechanism in the neural network to identify and emphasize features at different scales, thereby improving the learning efficiency of the model.

[0126] Embodiment 2

[0127] The purpose of this embodiment is to provide a vehicle surrounding environment target trajectory prediction system, including:

[0128] An acquisition module is configured to: acquire historical trajectory data, current vehicle information, and current vehicle scene information;

[0129] The encoding module is configured to: respectively encode the vehicle information at the current moment and the vehicle scene information at the current moment using a target vehicle encoder and a map encoder to obtain a vehicle encoding feature and a scene encoding feature;

[0130] A wavelet transform module is configured to: perform discrete wavelet transform on the historical trajectory data to decompose the obtained wavelet coefficients of different time scales and frequencies;

[0131] An embedding module is configured to: extract high-dimensional trajectory embedding features of the historical trajectory data using an input embedding network, perform enhancement and convolution operations on the high-dimensional trajectory embedding features using a wavelet attention module to generate context perception of the wavelet coefficients, and input the vehicle coding features, the scene coding features, and the context perception of the wavelet coefficients into different decoders corresponding to different wavelet transform coefficients, respectively, to obtain different sub-band time-frequency attributes of the trajectory sequence;

[0132] The prediction module is configured to: use inverse discrete wavelet transform to inversely transform the time-frequency attributes of different sub-bands of the obtained trajectory sequence, and predict the trajectory of the vehicle at the next moment.

[0133] In further embodiments, there is also provided:

[0134] An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in Embodiment 1 is performed. For the sake of brevity, no further description is given here.

[0135] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0136] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0137] A computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the method described in embodiment 1 is completed.

[0138] The method in the first embodiment can be directly embodied as a hardware processor, or a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it will not be described in detail here.

[0139] A computer program product includes a computer program, and when the computer program is executed by a processor, the method described in the first embodiment is implemented.

[0140] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer executable instructions, such as instructions included in a program module, which are executed in a device on a real or virtual processor of the target to perform the process / method as described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform specific tasks or implement specific abstract data types. In various embodiments, the functions of program modules can be combined or divided between program modules as needed. Machine executable instructions for program modules can be executed in local or distributed devices. In distributed devices, program modules can be located in local and remote storage media.

[0141] The computer program code for implementing the method of the present invention can be written in one or more programming languages. These computer program codes can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the computer or other programmable data processing device, causes the function / operation specified in the flow chart and / or block diagram to be implemented. The program code can be executed completely on a computer, partially on a computer, as an independent software package, partially on a computer and partially on a remote computer or completely on a remote computer or server.

[0142] In the context of the present invention, computer program codes or related data may be carried by any appropriate carrier to enable a device, apparatus or processor to perform the various processes and operations described above. Examples of carriers include signals, computer readable media, etc. Examples of signals may include electrical, optical, radio, acoustic or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0143] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0144] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A method for predicting target trajectory in vehicle surrounding environment, characterized in that: include: Obtain historical trajectory data, current vehicle information, and current vehicle scene information; Using a target vehicle encoder and a map encoder, respectively encode the vehicle information at the current moment and the vehicle scene information at the current moment to obtain a vehicle coding feature and a scene coding feature; Performing discrete wavelet transform on the historical trajectory data to decompose the obtained wavelet coefficients of different time scales and frequencies; The high-dimensional trajectory embedding features of the historical trajectory data are extracted by using an input embedding network, and the high-dimensional trajectory embedding features are enhanced and convolved by using a wavelet attention module to generate context perception of the wavelet coefficients, and the vehicle coding features, the scene coding features, and the context perception of the wavelet coefficients are respectively input into different decoders corresponding to different wavelet transform coefficients to obtain different sub-band time-frequency attributes of the trajectory sequence; The inverse discrete wavelet transform is used to inversely transform the time-frequency attributes of different sub-bands of the obtained trajectory sequence to predict the trajectory of the vehicle at the next moment.

2. A method for predicting a target trajectory in a vehicle surrounding environment as claimed in claim 1, characterized in that: The high-dimensional vector of the historical trajectory data is extracted as follows: Where I∈R M×D Represents the input history trajectory sequence The high-dimensional trajectory embedding feature representation of i1 ∈R d ×(D / / 2) and the matrix W i2 ∈R (D / / 2)×D It is used to linearly project the low-dimensional attribute vector of the trajectory point into the high-dimensional feature space; σ(·) represents the ReLU activation function; M represents the count of the historical trajectory points, and D represents the dimension of the feature space.

3. A method for predicting a target trajectory in a vehicle surrounding environment as claimed in claim 1, characterized in that: The target vehicle encoder and the map encoder respectively use a plurality of sequentially connected Mamba blocks to encode the vehicle information at the current moment and the vehicle scene information at the current moment to obtain vehicle coding features and scene coding features.

4. A method for predicting a target trajectory in a vehicle surrounding environment as claimed in claim 1, characterized in that: Each decoder consists of multiple cascaded Mamba blocks. Each Mamba block is sequentially connected to the LayerNorm layer and the fully connected layer to process the output features of the Mamba block. The output of the fully connected layer corresponding to each Mamba block is used as the different sub-band time-frequency attributes of the trajectory sequence output by the corresponding encoder.

5. A method for predicting a target trajectory in a vehicle surrounding environment as claimed in claim 1 or 4, characterized in that: The context-aware time dimension of the wavelet transform coefficients is determined as follows: b L =b L-1 , Where L is the level of wavelet analysis, l is half the length of the wavelet filter selected in the wavelet transform, M is the number of historical trajectory points, {b i}i∈[0,L)∩i∈Z Indicates the duration of high frequency, operator Indicates the largest integer less than or equal to the input, d i Represents the duration of high frequencies.

6. A method for predicting a target trajectory in a vehicle surrounding environment as claimed in claim 1, characterized in that: The high-dimensional trajectory embedding features of the historical trajectory data are extracted using the input embedding network, and the high-dimensional trajectory embedding features are enhanced and convolved using the wavelet attention module to generate the context perception of the wavelet coefficients, specifically: AND p =σ(I ​​W p1 )·IN p2 , E=γ(W s2 ·s ( W s1 ·I p )), I h =I+Diag(E·I), C=σ(Conv1d(I h )), Among them, W p1 ∈R D×(D / / 2) and W p2 ∈R (D / / 2)×1 Respectively represent the linear projection matrix, W s1 ∈R (M / / 2)×M , W s2 ∈R M×(M / / 2) It is used to evaluate the importance of features within trajectory embedding, σ(·) represents the ReLU activation function; M represents the count of historical trajectory points, D represents the dimension of the feature space, I represents the high-dimensional trajectory embedding feature representation of the historical trajectory sequence, and Conv1d represents a one-dimensional convolution.

7. A method for predicting a target trajectory in a vehicle surrounding environment as claimed in claim 1, characterized in that: For different sub-band time-frequency attributes of each trajectory sequence, the low-frequency and high-frequency coefficient pairs are merged in sequence until the time domain sequence of the attribute is reconstructed.

8. A vehicle surrounding environment target trajectory prediction system, characterized in that: include: An acquisition module is configured to: acquire historical trajectory data, current vehicle information, and current vehicle scene information; The encoding module is configured to: respectively encode the vehicle information at the current moment and the vehicle scene information at the current moment using a target vehicle encoder and a map encoder to obtain a vehicle encoding feature and a scene encoding feature; A wavelet transform module is configured to: perform discrete wavelet transform on the historical trajectory data to decompose the obtained wavelet coefficients of different time scales and frequencies; An embedding module is configured to: extract high-dimensional trajectory embedding features of the historical trajectory data using an input embedding network, perform enhancement and convolution operations on the high-dimensional trajectory embedding features using a wavelet attention module to generate context perception of the wavelet coefficients, and input the vehicle coding features, the scene coding features, and the context perception of the wavelet coefficients into different decoders corresponding to different wavelet transform coefficients, respectively, to obtain different sub-band time-frequency attributes of the trajectory sequence; The prediction module is configured to: use inverse discrete wavelet transform to inversely transform the time-frequency attributes of different sub-bands of the obtained trajectory sequence, and predict the trajectory of the vehicle at the next moment.

9. An electronic device, characterized in that: The method comprises a memory and a processor and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method according to any one of claims 1 to 7 is completed.

10. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the method described in any one of claims 1 to 7.

Images