High-precision densification method for sparse sampling data of driving track

Abstract

The invention provides a high-precision densification method for sparse sampling data of a driving track. The high-precision densification method comprises the following steps: S1, extracting mixed Fourier coding features; s2, sparse trajectory coding; s3, performing multi-sensor fusion; s4, road network constraint trajectory prediction; s5, fusing and smoothing multiple frames of data; s6, carrying out dynamic binding and geographical correction; s7, dense semantic map assistance; s8, generating a parallel track; s9, lightweight model deployment; visual semantic information and L i DAR point cloud data are combined, through a BEV space multi-sensor fusion technology, the context perception capability of a complex traffic scene is enhanced, dynamic semantic modeling under an open road is supported, and through road network constrained space-time diagram convolution and mixed Fourier coding, space-time dependence characteristics of a track are captured, so that the real-time performance of the track is improved. The problem of continuity deficiency caused by sparse sampling is solved, high-frequency noise is dynamically suppressed, the transverse positioning error is low, and accurate alignment of a track, a lane line and an obstacle is realized by using a dense semantic map.

Description

TECHNICAL FIELD

[0001] The present application relates to a densification method, in particular to a high-precision densification method for sparse sampling data of driving track, and belongs to the technical field of data processing. BACKGROUND

[0002] In modern traffic monitoring and data analysis, efficient acquisition and utilization of driving track is an important means to reduce traffic accidents, optimize traffic flow management and improve road safety. Traditional vehicle tracking methods usually rely on sampling of each frame of image or GPS data point per second. However, due to the limitations or cost factors of equipment, the sampling points are often sparse, which cannot comprehensively and accurately restore the driving track of the vehicle. Such sparse sampling not only limits the accurate monitoring of traffic conditions, but also affects the application range of subsequent data analysis techniques such as vehicle path planning and travel mode recognition.

[0003] Currently, the densification methods for sparse sampling data mainly include three types: interpolation-based method, dynamic programming-based method and machine learning-based method. The interpolation-based method is relatively direct, which fills in the data gaps through simple linear or polynomial interpolation. However, this method cannot well maintain the dynamic characteristics of the original track. The dynamic programming-based method attempts to expand data while ensuring the smoothness and continuity of the track. However, this method is usually complex and requires higher density data, increasing the complexity of algorithm implementation and processing. The machine learning-based method predicts the missing points in the sparse track by training a model. This technology is more advanced, but when dealing with high-dimensional and large amounts of data, it often requires a long training time and high data preprocessing requirements.

[0004] To solve the above technical problems, a high-precision densification method for sparse sampling data of driving track is proposed. SUMMARY

[0005] Therefore, the present application provides a high-precision densification method for sparse sampling data of driving track to solve or alleviate the technical problems in the prior art, and at least provides a beneficial choice.

[0006] The technical solution of the present application is as follows: a high-precision densification method for sparse sampling data of driving track, comprising the following steps:

[0007] S1, mixed Fourier coding feature extraction;

[0008] S2, sparse track coding;

[0009] S3, multi-sensor fusion;

[0010] S4, road network constraint track prediction;

[0011] S5, multi-frame data fusion and smoothing;

[0012] S6, dynamic binding and geographical correction;

[0013] S7, dense semantic map assistance;

[0014] S8, parallelized trajectory generation;

[0015] S9. Lightweight model deployment.

[0016] Further preferably, in said S1, a hybrid Fourier encoder is used to perform frequency domain feature analysis on historical trajectory points, and the attention mechanism is combined to extract lane line geometry and dynamic interaction features to solve the problem of characterizing behavioral patterns such as turning and speed mutations implied in sparse trajectory points.

[0017] Further preferably, in S2, a hybrid Fourier encoder is used to extract frequency domain features of sparse trajectories, and key behavior patterns such as turning and acceleration are captured through an attention mechanism to solve the problem of lack of behavior continuity caused by low-frequency sampling.

[0018] Further preferably, in the S3, the LiDAR point cloud and visual semantic information are fused through the bird's-eye view (BEV) to construct a multimodal space-time field including lane line types and obstacle positions, thereby improving the context perception capability of trajectory prediction.

[0019] Further preferably, in S4, a lane-level topology map is constructed based on the high-precision map, and a multimodal candidate trajectory is generated through a CVAE (conditional variational autoencoder) model.

[0020] Further preferably, in said S5, based on the time series update algorithm, the lane line POI point pool is dynamically updated, and multi-frame data fusion and B-spline interpolation are used to achieve trajectory smoothing, suppress high-frequency noise and retain key turning points.

[0021] Further preferably, in said S6, based on the high-precision map road network data, a hidden Markov model (HMM) is used to dynamically match trajectory points and road sections, and the search radius is adaptively adjusted to 20m, which effectively corrects GPS drift and improves geographic accuracy.

[0022] Further preferably, in the S7, a dense semantic map is generated by combining 3D Gaussian splattering technology, and accurate alignment of the trajectory with the lane line and obstacles is achieved through multi-channel optimization (appearance, geometry, semantics).

[0023] Further preferably, in S8, CUDA is used to accelerate the calculation of horizontal and vertical trajectory merging, and a graphics memory pre-allocation strategy is used to process 10^5 level trajectory points, so that the calculation time is reduced to within 50ms.

[0024] Further preferably, in the S9, the trajectory generation model is quantized to FP16 precision by TensorRT, the edge inference delay is ≤50ms, and the real-time processing of 5-second interval sparse sampling data on a vehicle-mounted platform is supported.

[0025] The embodiments of the present application have the following advantages due to the adoption of the above technical solutions:

[0026] First, the present application combines visual semantic information with LiDAR point cloud data, enhances the context perception ability of complex traffic scenes through multi-sensor fusion technology in BEV space, supports dynamic semantic modeling under open roads, captures the spatio-temporal dependence features of trajectories through spatio-temporal graph convolution and hybrid Fourier encoding under road network constraints, solves the continuity loss problem caused by sparse sampling, dynamically suppresses high-frequency noise, and has low horizontal positioning error.

[0027] Second, the present application realizes accurate alignment of trajectories, lane lines and obstacles by using dense semantic maps, supports real-time processing of 5-second interval sparse data, and meets the resource constraints of vehicle-mounted platforms.

[0028] The above summary is for the purpose of description only and is not intended to limit in any way. In addition to the illustrative aspects, embodiments and features described above, further aspects, embodiments and features will become apparent to those skilled in the art from the following detailed description, the accompanying drawings and the following claims. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0030] Figure 1 Flow chart of the high-precision densification method for the sparse sampling data of the driving trajectory of the present application. DETAILED DESCRIPTION

[0031] In the following, only some exemplary embodiments are simply described. As those skilled in the art can recognize, the described embodiments can be modified in various different ways without departing from the spirit or scope of the present application. Therefore, the drawings and description are considered to be exemplary in nature rather than limiting.

[0032] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0033] Embodiment one

[0034] As Figure 1As shown, the embodiment of the present application provides a high-precision densification method for sparse sampling data of driving track, which comprises the following steps:

[0035] S1, mixed Fourier coding feature extraction, a mixed Fourier encoder is used to analyze the frequency domain features of historical track points, and the lane line geometry and dynamic interaction features are extracted by combining the attention mechanism, so that the representation problem of behavior modes such as turning and speed mutation implied in sparse track points is solved;

[0036] S2, sparse trajectory coding, a mixed Fourier encoder is used to extract the frequency domain features of sparse trajectories, and the attention mechanism is used to capture key behavior modes such as turning and acceleration, so as to solve the problem of lack of behavior continuity caused by low-frequency sampling;

[0037] S3, multi-sensor fusion, LiDAR point cloud and visual semantic information are fused through bird's eye view (BEV), a multi-modal spatio-temporal field containing lane line types and obstacle positions is constructed, and the context perception ability of trajectory prediction is improved;

[0038] S4, road network constraint trajectory prediction, a lane-level topology graph is constructed based on a high-precision map, and a multi-modal candidate trajectory is generated by a CVAE (conditional variational autoencoder) model;

[0039] S5, multi-frame data fusion and smoothing, based on a time series updating algorithm, the lane line POI point pool is dynamically updated, multi-frame data fusion and B-spline interpolation are used to realize trajectory smoothing, high-frequency noise is suppressed, and key turning points are preserved;

[0040] S6, dynamic binding road and geographical correction, based on high-precision map road network data, a hidden Markov model (HMM) is used to dynamically match trajectory points and road segments, the search radius is self-adaptively adjusted to 20m, GPS drift is effectively corrected, and geographical accuracy is improved;

[0041] S7, dense semantic map assistance, a dense semantic map is generated by combining 3D Gaussian splashing technology, and the trajectory is accurately aligned with the lane line and obstacle through multi-channel optimization (appearance, geometry, and semantics);

[0042] S8, parallel trajectory generation, CUDA is used to accelerate horizontal and vertical trajectory merging calculation, and 10^5 order trajectory points are processed through memory pre-allocation strategy, and the calculation time is reduced to within 50ms;

[0043] S9, lightweight model deployment, the trajectory generation model is quantized to FP16 precision by using TensorRT, the edge inference delay is ≤50ms, and the vehicle-mounted platform can support real-time processing of sparse sampling data with an interval of 5 seconds.

[0044] Embodiment two

[0045] The embodiment of the application also provides a high-precision densification method for sparse sampling data of a driving track, including the following steps:

[0046] S1, multi-modal perception and feature modeling, a visual-language-action large model (MindVLA) is used to analyze traffic sign text (such as a tidal lane rule), multi-sensor data (LiDAR point cloud+visual semantics) in a BEV space is combined, a dynamic semantic field is generated, open scene track context understanding is supported, a hybrid Fourier encoder is used to extract frequency domain features of sparse tracks, an attention mechanism is used to capture key behavior patterns such as turning and accelerating, and a behavior continuity loss problem caused by low-frequency sampling is solved;

[0047] S2, spatio-temporal joint modeling and track generation, a lane-level topology graph is constructed based on a high-precision map, a CVAE (conditional variational autoencoder) model is used to generate multi-modal candidate tracks, transverse displacement error is less than or equal to 0.3m, longitudinal speed error is less than 5%, a double-layer optimization strategy is designed, a cubic B-spline curve interpolation is used in the geometric layer to generate a smooth candidate track, and curvature continuity meets vehicle dynamics constraints; the semantic layer is combined with dynamic traffic rules (such as a variable lane sign) to correct the track, and the Fréchet distance is used to select an optimal solution;

[0048] S3, anti-interference and robustness enhancement, an improved Sage-Husa Kalman filter is used to dynamically adjust noise covariance, a sliding window (window size 5-15 points) is combined to realize local track correction, high-frequency noise suppression rate is improved by 40%, a hidden Markov model (HMM) is used to match track points with high-precision map lane lines in real time, a search radius (5-20m) is adaptively adjusted, and track distortion caused by GPS drift is solved;

[0049] S4, efficient calculation and deployment, CUDA is used to accelerate horizontal and vertical track merging calculation, a memory pre-allocation strategy is used to process 10^5 order track points, calculation time is reduced to less than or equal to 50ms, a CVAE model is quantized to FP16 precision by using TensorRT, edge computing platform (such as Jetson AGX) inference delay is less than or equal to 30ms, and real-time track completion is supported.

[0050] Embodiment three

[0051] The embodiment of the application also provides a high-precision densification method for sparse sampling data of a driving track, including the following steps:

[0052] S1, dynamic semantic binding and federated learning fusion, real-time semantic field construction, fusion of multi-car LiDAR point cloud and visual features using BEV feature transmission technology (256 times compression rate), generation of dynamic 4D semantic field, support for open text query of lane line type, traffic signs, etc., integration of edge vehicle data through a federated learning framework, privacy protection collaborative training of multi-car trajectory features using homomorphic encryption technology, data sharing bandwidth demand reduced to 0.32MB / s, semantic guided trajectory generation, analysis of dynamic traffic rules (such as tidal lane identification) using MindVLA multi-modal large model, generation of a candidate trajectory set with semantic constraints, prediction of multi-modal trajectory distribution using a Transformer decoder, covering turning and avoidance behavior patterns using a Gaussian Mixture Model (GMM), and a long-term prediction error reduction of 12.3%;

[0053] S2, cooperative double-layer optimization of vehicle and road, joint optimization of signal and trajectory, establishment of a double-layer optimization model, dynamic adjustment of signal timing in the upper layer to minimize vehicle delay (double-ring barrier structure optimization phase period), optimization of CAV (connected and automated vehicle) arrival time and fuel consumption based on economic driving strategy in the lower layer, HDV (human-driven vehicle) constraint through a car-following model, real-time inference on the edge, deployment of a lightweight trajectory optimization model (TensorRT quantization to FP16 precision), and support for on-board computing platforms (such as Jetson AGXXavier) to complete trajectory interpolation and smoothing within 50ms;

[0054] S3, anti-interference and robustness enhancement, spatio-temporal anti-noise filtering, dynamic adjustment of process noise covariance using an improved Sage-Husa adaptive Kalman filter, dynamic lane binding compensation based on a multi-precision high-precision map (lateral error ≤0.3m), dynamic adjustment of road matching search radius (20m) using a Hidden Markov Model (HMM), and correction of GPS drift based on lane line geometric similarity;

[0055] S4, performance verification and deployment, performance verification based on dense accuracy, real-time performance, anti-interference capability, and multi-car collaboration efficiency.

[0056] Embodiment Four

[0057] The application also provides a high-precision densification method for sparse sampling data of driving trajectories, including the following steps:

[0058] S1, dynamic semantic field construction and multi-modal fusion, 4D spatiotemporal semantic modeling, using dynamic three-dimensional Gaussian splash technology, combining the 4D labeling scheme of BEV perception, fusing sparse trajectory points with real-time camera and LiDAR data to generate a spatiotemporal field containing semantic information such as lane line types and traffic signs. Through state-variable network compression of semantic feature dimensions, open text query and dynamic scene understanding are realized, multi-modal data alignment is achieved, and hardware clock synchronization (PTP protocol) is used to align multi-sensor data. LiDAR point cloud is voxelized (voxel_size = 0.1m), and is fused with visual features through a 256-fold compression rate BEV feature transmission technology;

[0059] S2, cooperative prediction and trajectory generation, multi-vehicle cooperative prediction framework, introducing a multi-modal prediction decoder based on the Transformer architecture, combining Gaussian Mixture Model (GMM) to generate diverse trajectory hypotheses, dynamically integrating surrounding vehicle prediction results through attention mechanism, 5-second long-time prediction error reduced by 12.3%, spatiotemporal joint trajectory optimization, coarse-grained sampling in longitudinal acceleration and lateral displacement dimensions, retaining key turning points and speed mutation points, filtering physically feasible trajectories through Conditional Variational Autoencoder (CVAE) combined with vehicle dynamics constraints (maximum lateral acceleration ≤ 2.5m / s 2 );

[0060] S3, high-precision map matching and noise-resistant processing, dynamic lane binding compensation, integrating multi-precision high-precision map data, using an improved Hidden Markov Model (HMM) for map matching, dynamically adjusting the search radius to 20m, combining lane line geometry matching algorithm to correct GPS drift points, lateral positioning error ≤ 0.3m, adaptive noise-resistant filtering based on Sage-Husa adaptive Kalman filter, dynamically adjusting process noise covariance Q, eliminating high-frequency noise through sliding window (5-15 points) B-spline fitting, trajectory smoothness improved by 40%.

[0061] In one embodiment, in the tidal lane area, real-time analysis of dynamic signs is performed through the VLA model, and the trajectory deviation rate is reduced from 12% to 3.7%; multi-vehicle cooperative prediction makes the intersection collision warning time advance by more than 2 seconds.

[0062] In work: mixed Fourier coding feature extraction, the history trajectory point is analyzed in frequency domain feature by using mixed Fourier encoder, the lane line geometry and dynamic interaction feature is extracted by combining attention mechanism, the representation problem of behavior mode such as steering and speed mutation hidden in sparse trajectory point is solved, sparse trajectory coding, the frequency domain feature of sparse trajectory is extracted by using mixed Fourier encoder, the key behavior mode such as steering and acceleration is captured through attention mechanism, the behavior continuity loss problem caused by low-frequency sampling is solved, multi-sensor fusion, the multi-modal space-time field containing lane line type and obstacle position is constructed by fusing LiDAR point cloud and visual semantic information through bird's eye view (BEV), the context perception ability of trajectory prediction is improved, road network constraint trajectory prediction, the lane level topology graph is constructed based on high-precision map, the multi-modal candidate trajectory is generated through CVAE (conditional variational autoencoder) model, multi-frame data fusion and smoothing, based on time series updating algorithm, the lane line POI point pool is dynamically updated, trajectory smoothing is realized by using multi-frame data fusion and B-spline interpolation, high-frequency noise is suppressed and key steering points are reserved, dynamic road binding and geographic rectification, based on high-precision map road network data, the trajectory point and road segment are dynamically matched by using hidden Markov model (HMM), the search radius is adaptively adjusted to 20m, the GPS drift is effectively corrected and the geographic accuracy is improved, dense semantic map auxiliary, the dense semantic map is generated by combining 3D Gaussian splashing technology, the trajectory and lane line and obstacle are accurately aligned through multi-channel optimization (appearance, geometry, semantic), parallel trajectory generation, the horizontal and vertical trajectory merging calculation is accelerated by using CUDA, the 10^5 order trajectory points are processed through memory pre-allocation strategy, the calculation time is reduced to within 50ms, lightweight model deployment, the trajectory generation model is quantized to FP16 precision by using TensorRT, the edge inference delay is less than or equal to 50ms, the real-time processing of 5-second interval sparse sampling data on vehicle-mounted platform is supported.

[0063] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any skilled person in the art can easily think of various changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A high-precision densification method for sparsely sampled vehicle trajectory data, characterized by: The following steps are involved: S1, hybrid Fourier coding feature extraction; S2, sparse trajectory coding; S3, multi-sensor fusion; S4, road network constraint trajectory prediction; S5, multi-frame data fusion and smoothing; S6, dynamic binding and geographical correction; S7, dense semantic map assistance; S8, parallelized trajectory generation; S9. Lightweight model deployment.

2. The high-precision densification method for sparsely sampled vehicle trajectory data according to claim 1, characterized in that: In S1, a hybrid Fourier encoder is used to perform frequency domain feature analysis on historical trajectory points, and the attention mechanism is combined to extract lane line geometry and dynamic interaction features to solve the problem of representing behavioral patterns such as turning and speed mutations implied in sparse trajectory points.

3. The high-precision densification method for sparsely sampled vehicle trajectory data according to claim 1, characterized in that: In the S2, a hybrid Fourier encoder is used to extract the frequency domain features of sparse trajectories, and the key behavior patterns such as turning and acceleration are captured through the attention mechanism to solve the problem of lack of behavior continuity caused by low-frequency sampling.

4. The high-precision densification method for sparsely sampled vehicle trajectory data according to claim 1, characterized in that: In the S3, the LiDAR point cloud and visual semantic information are fused through the bird's-eye view (BEV) to construct a multimodal spatiotemporal field including lane line types and obstacle positions, thereby improving the context perception capability of trajectory prediction.

5. The high-precision densification method for sparsely sampled vehicle trajectory data according to claim 1, characterized in that: In the S4, a lane-level topology map is constructed based on the high-precision map, and a multimodal candidate trajectory is generated through a CVAE (conditional variational autoencoder) model.

6. The high-precision densification method for sparsely sampled vehicle trajectory data according to claim 1, characterized in that: In the S5, based on the time series update algorithm, the lane line POI point pool is dynamically updated, and multi-frame data fusion and B-spline interpolation are used to achieve trajectory smoothing, suppress high-frequency noise and retain key turning points.

7. The high-precision densification method for sparsely sampled vehicle trajectory data according to claim 1, characterized in that: In the S6, based on high-precision map road network data, a hidden Markov model (HMM) is used to dynamically match trajectory points and road sections, and the search radius is adaptively adjusted to 20m, effectively correcting GPS drift and improving geographic accuracy.

8. The high-precision densification method for sparsely sampled vehicle trajectory data according to claim 1, characterized in that: In the S7, a dense semantic map is generated by combining 3D Gaussian splashing technology, and precise alignment of trajectories with lane lines and obstacles is achieved through multi-channel optimization (appearance, geometry, and semantics).

9. The high-precision densification method for sparsely sampled vehicle trajectory data according to claim 1, characterized in that: In the S8, CUDA is used to accelerate the calculation of horizontal and vertical trajectory merging, and a graphics memory pre-allocation strategy is used to process 10^5 trajectory points, reducing the calculation time to within 50ms.

10. The high-precision densification method for sparsely sampled vehicle trajectory data according to claim 1, characterized in that: In the S9, TensorRT is used to quantize the trajectory generation model to FP16 precision, and the edge inference delay is ≤50ms, supporting the vehicle platform to process sparse sampling data at 5-second intervals in real time.

Images