Intelligent driving-oriented multivariate measurement lightweight spatiotemporal redundancy data construction method

Abstract

The application discloses a kind of multivariate measurement lightweight spatiotemporal redundancy data construction methods for intelligent driving, belongs to intelligent driving and multi-sensor data fusion technical field.The application is first carried out multi-source perception data acquisition and spatiotemporal synchronization, then carries out spatiotemporal redundancy measurement and feature compression modeling, finally carries out lightweight spatiotemporal data structure construction: the application is realized efficient storage and quick search by indexing and hierarchical coding strategy, provides real-time support for subsequent fusion and fault-tolerant control.The application realizes the lightweight, high-reliability fusion of multi-source perception data by introducing physical consistency constraint and spatiotemporal redundancy measurement mechanism, effectively reduces data transmission and computing burden, improves the robustness and real-time response capability of intelligent driving system in complex environment.

Description

Technical Field

[0001] This invention belongs to the field of intelligent driving and multi-sensor data fusion technology, specifically involving a lightweight spatiotemporal redundancy data construction method for multivariable measurement in intelligent driving systems. It involves technologies such as synchronous acquisition of multi-source heterogeneous sensor data, physical consistency feature extraction, spatiotemporal redundancy measurement and compression modeling, lightweight data structure construction and fusion optimization, and is applicable to scenarios such as environmental perception, state estimation and fault-tolerant control of autonomous vehicles. Background Technology

[0002] With the rapid development of intelligent driving technology, vehicle perception systems typically integrate multiple types of sensors, including LiDAR, cameras, millimeter-wave radar, and inertial measurement units (IMUs). These sensors can perceive the surrounding environment from different modalities and dimensions, achieving high-precision measurements of vehicle position, speed, attitude, and obstacles. However, due to the large number of sensor types and varying measurement frequencies and accuracies, problems such as data redundancy, noise interference, synchronization deviations, and information inconsistencies can easily arise in complex road environments.

[0003] Existing multi-sensor data fusion methods typically rely on algorithms such as Kalman filtering, particle filtering, or deep neural networks. While these methods can improve sensing accuracy when achieving data fusion, they have the following shortcomings:

[0004] (1) The physical consistency constraints among multiple source measurements were not fully considered, which may cause the fusion results to deviate from the true physical laws;

[0005] (2) Multimodal data has high dimensionality and high redundancy. Direct fusion will increase the computational burden and storage pressure on the system, affecting real-time performance;

[0006] (3) Under complex working conditions, such as sensor obstruction, sudden noise changes or communication delays, traditional methods are not robust enough and it is difficult to achieve stable dynamic perception and real-time response.

[0007] Therefore, how to achieve lightweight, low-redundancy, and highly consistent representation of multi-source sensor data while maintaining high perception accuracy has become a key issue that intelligent driving systems urgently need to address.

[0008] To address the aforementioned problems, this invention proposes a lightweight spatiotemporal redundancy data construction method for multivariate measurement in intelligent driving. By introducing a physical consistency constraint model and a spatiotemporal redundancy measurement mechanism, it performs unified feature extraction, redundancy analysis, and lightweight modeling on multimodal sensor data, thereby significantly reducing data complexity and improving the real-time performance and robustness of the system while ensuring measurement accuracy. Summary of the Invention

[0009] The purpose of this invention is to overcome the problems of high redundancy, insufficient physical consistency, heavy computational burden and poor real-time performance of multi-source sensor data in existing intelligent driving systems, and to propose a lightweight spatiotemporal redundant data construction method for multivariate measurement in intelligent driving.

[0010] To achieve the above objectives, this invention provides a lightweight spatiotemporal redundancy data construction method for multivariate measurement in intelligent driving, comprising the following steps:

[0011] Step 1) Multi-source sensing data acquisition and spatiotemporal synchronization:

[0012] It collects environmental perception data from various types of sensors, including lidar, cameras, millimeter-wave radar, and inertial measurement units.

[0013] By unifying timestamps and spatial coordinate transformation matrices, spatiotemporal alignment of multi-source data is achieved. Noise filtering, scale normalization, and missing value compensation are then performed to obtain a multimodal preprocessed dataset with unified spatiotemporal resolution.

[0014] Step 2) Key feature extraction under physical consistency constraints:

[0015] Based on the kinematic and dynamic model of the intelligent driving system, a multi-dimensional feature space including physical variables such as position, velocity, acceleration, and attitude is defined, and a mapping model between the measurements of each sensor and the real physical variables is established.

[0016] By constructing a physical consistency error function and minimizing this error, common-mode features that satisfy physical coupling relationships are extracted, noise and spurious signals are eliminated, and a stable feature set that conforms to physical laws is formed.

[0017] Step 3) Construction of the spatiotemporal redundancy measurement model:

[0018] For synchronous observation data from multiple time points and multiple perspectives, we calculate the mutual information and similarity indices between features and construct a spatiotemporal redundancy matrix to quantify the correlation between different features.

[0019] This matrix can reflect the redundancy distribution of multimodal features in the temporal and spatial dimensions, providing a basis for subsequent feature compression and lightweight representation.

[0020] Step 4) Adaptive feature compression and redundancy removal modeling:

[0021] Based on the analysis results of the spatiotemporal redundancy matrix, an adaptive linear mapping algorithm or a multi-source subspace compression method is used to reduce the dimensionality and aggregate highly correlated features, and extract the spatiotemporal evolution-sensitive principal feature components.

[0022] By constructing a mapping matrix, redundant features can be compressed and represented, forming a lightweight feature set, which effectively reduces data dimensionality and computational complexity.

[0023] Step 5) Construction of lightweight spatiotemporal data structure:

[0024] Design a hierarchical, lightweight spatiotemporal data structure to organize and encode compressed features in layers according to time, space, and physical quantities.

[0025] This structure supports fast indexing, retrieval, and updating, enabling efficient management and access to multi-scale spatiotemporal information, and ensuring the real-time performance and scalability of the data structure.

[0026] Step 6) Fusion Output and Real-time Support:

[0027] Features that satisfy physical consistency and redundancy constraints are weighted and fused to output joint measurement results.

[0028] Based on the fusion results, the physical model parameters are dynamically and adaptively updated to achieve online optimization and self-adjustment of the perception system, thereby improving the perception accuracy and system robustness of intelligent driving vehicles under complex conditions.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] (1) Enhanced physical consistency: By introducing a physical constraint model in the multi-source fusion process, the consistency of measurement results from different sensors in physical quantities such as position, velocity, and acceleration is ensured, thereby improving data credibility. (2) Spatiotemporal redundancy suppression: By constructing a spatiotemporal redundancy metric matrix and an adaptive feature compression algorithm, the redundant information between highly correlated features is effectively reduced, significantly reducing data dimensionality and storage volume. (3) Lightweight data representation: A hierarchical spatiotemporal data structure design is proposed to achieve lightweight representation and efficient organization of high-dimensional perception features, facilitating rapid retrieval and real-time updates. (4) Improved real-time performance: By compressing and mapping spatiotemporal features, the computational complexity and communication burden of the system are reduced, enabling the intelligent driving system to still have high real-time response capability under complex working conditions. (5) Improved robustness and adaptability: By combining a dynamic parameter adjustment mechanism and online model update capability, the system can still maintain stable and reliable perception and fusion performance even when sensors are abnormal, the environment changes, or the data is incomplete.

[0031] In summary, the lightweight spatiotemporal redundancy data construction method for multivariate measurement provided by this invention can achieve efficient fusion and lightweight modeling of multi-source perception data, providing highly reliable, low-redundancy, and high-real-time perception data support for intelligent driving systems. Attached Figure Description

[0032] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0033] This invention proposes a lightweight spatiotemporal redundancy data construction method for multivariate measurements in intelligent driving, enabling highly reliable and low-redundancy data fusion and feature compression representation in multi-sensor environments. This method significantly improves the robustness and real-time performance of intelligent driving systems by constructing a physical consistency constraint model, a spatiotemporal redundancy metric matrix, and a hierarchical lightweight data structure. The specific implementation steps are as follows.

[0034] Step 1) Multi-source sensing data acquisition and spatiotemporal synchronization: Environmental measurement data from various types of sensors, including LiDAR, cameras, millimeter-wave radar, and inertial measurement units, are acquired. Spatial coordinate alignment and temporal synchronization are performed on different data sources. Spatiotemporal registration of multi-source data is achieved using a unified timestamp and spatial transformation matrix. Noise filtering, scale normalization, and missing value compensation are then applied to obtain a multimodal preprocessed dataset with unified spatiotemporal resolution, providing consistent input for subsequent feature extraction.

[0035] (1.1) Spatial coordinate alignment

[0036] In intelligent driving systems, environmental perception information is acquired through various sensors, including LiDAR, cameras, millimeter-wave radar, and inertial measurement units (IMUs). Let D be the data collected by the i-th sensor at time t. i (t), whose corresponding local spatial coordinates are P. i (t). To achieve a unified spatial representation of multi-source sensor data, it is necessary to establish an extrinsic parameter transformation model from each sensor coordinate system to the vehicle's global coordinate system:

[0037]

[0038] Where: P i ′(t): Spatial coordinates (3×1 vector) of the i-th sensor in the vehicle's global coordinate system; P i (t): Local coordinates (3×1 vector) of the i-th sensor in its own coordinate system; R i : The rotation matrix (3×3 matrix) from the i-th sensor coordinate system to the vehicle global coordinate system, representing the change in attitude direction; T i Translation vector (3×1 vector) represents the offset of the sensor's installation position in the vehicle coordinate system.

[0039] (1.2) Timestamp Alignment

[0040] After completing the spatial coordinate transformation, a unified timestamp synchronization is required to ensure data alignment across different sensors in the time dimension. Let the system define the unified timestamp as T. s , m The adjacent sampling times of the i-th sensor are t and t respectively. i , k and t i , k+1 The corresponding sampling data is D i (t i , k ) and D i (t i , k+1 Synchronized data at a unified timestamp is obtained through linear interpolation.

[0041]

[0042] Where: D i (T s , m ): Interpolated data of the i-th sensor at a uniform timestamp; D i (t i , k D i (t i , k+1 ): The raw sampling data of the i-th sensor at two adjacent time points; w: the time interpolation weighting coefficient, calculated as: w = (t i , k+1 - T s , m ) / (t i , k+1 - t i , k ).

[0043] To eliminate sampling noise and amplitude differences between different sensors, the synchronized data is standardized:

[0044]

[0045] in: i Standardized sensor data; μ i σ: The mean of the sampled data from the i-th sensor; i : Standard deviation of the data sampled by the i-th sensor.

[0046] (1.3) Missing data compensation

[0047] When some sensors experience short-term data loss, a Kalman filter can be used for state prediction and data compensation to maintain the integrity of the time series. Let the system state-space model be:

[0048]

[0049]

[0050] in: (t|t-1): The predicted state vector (n×1 vector) at time t; (t-1|t-1): The optimal state estimate at the previous time t-1; A: State transition matrix (n×n), representing the evolution of the system state over time; B: Control input matrix (n×m), used to describe the influence of external control input on the system state; u(t-1): Control input vector (m×1); H: Observation matrix (p×n), mapping the system state to observable outputs; i (t): The predicted measurement value of the i-th sensor obtained based on the predicted state.

[0051] After spatial coordinate transformation, time synchronization, and data preprocessing, a multi-source synchronized dataset with unified spatiotemporal resolution is obtained:

[0052]

[0053] Among them, D n ′ represents the nth sensor data after spatial registration, temporal alignment, and noise processing.

[0054] Step 2) Key Feature Extraction under Physical Consistency Constraints: Based on the kinematic and dynamic model of the intelligent driving system, a multi-dimensional feature space including physical variables such as position, velocity, acceleration, and attitude is defined. A mapping relationship is established between the measurements of each sensor and the actual physical variables. A physical consistency error function is constructed, and constraint optimization is performed on the observation data from different sensors. Common-mode features that satisfy the physical coupling law are extracted, and spurious and noise interference signals are filtered out.

[0055] After achieving spatiotemporal synchronization of multi-source sensing data, a feature extraction method based on physical consistency constraints is constructed to ensure the synergy and interpretability of multi-sensor data at the physical level. This method, based on vehicle kinematics and dynamics constraints, extracts common-mode features from multi-dimensional physical variables such as spatial position, velocity, and acceleration for subsequent fusion analysis and anomaly detection.

[0056] (2.1) Constructing a kinematic model

[0057] First, let P be the global coordinates of the i-th sensor at time t.i ′(t) = [x i (t), y i (t), z i (t)] T .

[0058] Based on time series data, the velocity and acceleration characteristics of each sensor can be calculated:

[0059]

[0060]

[0061] Where: v i (t): The velocity vector (3×1) of the i-th sensor at time t; a i (t): The acceleration vector (3×1) of the i-th sensor at time t; Δt: The continuous sampling interval; P i ′(t): The spatial position of the i-th sensor in the global coordinate system.

[0062] To ensure the consistency of the motion characteristics of different sensors with the overall motion of the vehicle, a physical consistency constraint is introduced. Let the reference position of the vehicle center be P_c(t), then the physical consistency error of the i-th sensor relative to the vehicle center is:

[0063]

[0064] Where: E i (t): Spatial physical consistency error of the i-th sensor; R_c(t): Vehicle attitude rotation matrix at time t (3×3); T_c(t): Translation vector of the vehicle center (3×1); L i : The installation position vector of the i-th sensor relative to the center of the vehicle (constant, 3×1); ‖·‖2: L2 norm, used to represent spatial distance error.

[0065] By minimizing the physical consistency error of all sensors, a globally consistent set of motion estimation parameters can be obtained:

[0066]

[0067] Solving this optimization problem can constrain the spatial coordination relationship between different sensors and ensure the geometric consistency of the fused features.

[0068] (2.1) Constructing a dynamic model

[0069] Furthermore, dynamic consistency constraints are introduced at the velocity and acceleration levels. According to the vehicle dynamics equations, we have:

[0070]

[0071] Where: a_c(t): theoretical acceleration vector at the center of the vehicle; F(t): vehicle driving force (including motor output, tire friction, etc.); R(t): rolling resistance; D(t): air resistance; m: vehicle mass.

[0072] To ensure that the acceleration measured by the multi-source sensors is consistent with the theoretical dynamic acceleration, a dynamic consistency error term is defined:

[0073]

[0074] in: (t) = (1 / n) Σ (i=1) n a i (t): The average acceleration measured by each sensor.

[0075] Finally, by jointly optimizing the spatial consistency error E i (t) and the dynamic consistency error E_d(t) are used to obtain the key feature set F_phy under the physical consistency constraint:

[0076]

[0077] Where: F_phy: the set of key physical features extracted under physical consistency constraints; λ: weighting coefficient, used to balance the influence of spatial and dynamic consistency constraints.

[0078] This step integrates kinematic and dynamic constraints to achieve physical consistency correction and common-mode feature extraction among multi-source sensor data, providing basic feature support for subsequent anomaly detection and fault-tolerant control.

[0079] Step 3) Construction of the Spatiotemporal Redundancy Measurement Model: Based on multi-source synchronous observation data, the mutual information and similarity of features under different time windows and spatial perspectives are calculated to construct a spatiotemporal redundancy measurement matrix. Through statistical dependency analysis between features, the redundancy degree of each feature component is quantified, highly correlated feature regions are identified, and a spatiotemporal redundancy description model reflecting the correlation of features at multiple times and from multiple perspectives is established, providing a data basis for subsequent feature compression.

[0080] After extracting the physical consistency features of multi-source data, a spatiotemporal redundancy measurement model is constructed to achieve adaptive fault tolerance and information compensation for multiple sensors in intelligent driving systems. This model is used to measure the degree of information redundancy of different sensors in spatial layout and temporal response, thereby determining their reliability weight in data fusion and fault compensation.

[0081] (3.1) Calculation of spatial redundancy

[0082] Let the first one after coordinate unification be... i Each sensor at time t The global position vector is P i ′(t) = [x i (t), y i (t), z i (t)] T Then any two sensors i and j The spatial distance between them is:

[0083]

[0084] Define spatial redundancy R s (i, j) is a weighted function of positional proximity and field of view intersection:

[0085]

[0086] Where: d ij :sensor i and j Euclidean distance; d0: spatial scale constant, used to normalize the distance effect; θ ij : Field of view direction vector n of the two sensors i n j The angle between them, cos(θ) ij ) = (n i ·n j ) / (‖n i |||n j ||); exp(-d ij / d0): Indicates the spatial redundancy enhancement effect between sensors located close to each other.

[0087] When the two sensors are close in position and have the same field of view, R s When the value of (i, j) is close to 1, it indicates that there is high spatial redundancy; conversely, when the value is close to 0, it indicates that the information complementarity between the two is strong.

[0088] (3.2) Calculation of time redundancy

[0089] Let the first i The characteristic signal output by each sensor in the time series is s i (t), which is related to the sensor j The time lag cross-correlation function is defined as:

[0090]

[0091] Where: C ij(τ): the first i and j Normalized cross-correlation function of each sensor under time lag τ; T: calculation window duration; μ i μ j : The mean of the corresponding signal; σ i σ j : Standard deviation of the corresponding signal.

[0092] Temporal redundancy is defined as the maximum value of the cross-correlation function at τ=0:

[0093]

[0094] Where, τ max This represents the maximum allowable time delay. When the signals from the two sensors exhibit consistent trends and good time synchronization, R... t (i, j) approaches 1.

[0095] (3.3) Calculation of spatiotemporal redundancy

[0096] By jointly weighting spatial redundancy and temporal redundancy, we obtain the first... i and j Overall spatiotemporal redundancy of the sensors:

[0097]

[0098] Where: R_st(i, j): spatiotemporal redundancy comprehensive index; α, β: weight coefficients, satisfying α + β = 1, used to adjust the relative importance of spatial and temporal characteristics.

[0099] Further define the first i Global redundancy index of a single sensor:

[0100]

[0101] Where: R i :sensor i Average spatiotemporal redundancy compared to other sensors in the system; n: total number of sensors in the system.

[0102] When R i When R is large, it indicates that the sensor information can be well compensated by other sensors, possessing high redundancy characteristics; when R... i When the value is relatively small, it indicates that the information is more independent and it is a key sensing node.

[0103] (3.4) Spatiotemporal redundancy matrix and redundancy graph model

[0104] Construct a spatiotemporal redundancy matrix from the redundancy measurement results of all sensors:

[0105]

[0106] This matrix can be mapped to an undirected weighted graph G = (V, E, W).

[0107] Where: V = {v1, v2, …, v n}: Represents the set of sensor nodes; E = {(v i , v j )}: Represents a sensor pair with redundant relationships; W = {R_st(i, j)}: Represents the weight of the edge, i.e., the spatiotemporal redundancy strength.

[0108] The redundancy centrality index can be calculated based on the graph model:

[0109]

[0110] C i The larger the value, the better the sensor. i The stronger the redundancy and correlation in a system, the higher its information independence, and the more important it is to monitor its health status.

[0111] Finally, the spatiotemporal redundancy measurement model outputs: the inter-sensor redundancy matrix R_ST; and the global redundancy index R_ST for each sensor. i Redundancy centrality distribution C i .

[0112] This model can quantify the spatiotemporal correlation characteristics of multi-source sensors, providing a weight basis for subsequent anomaly compensation and fault-tolerant control, and realizing redundant adaptive scheduling and data reliability enhancement of intelligent driving systems under complex working conditions.

[0113] Step 4) Adaptive Feature Compression and Redundancy Removal Modeling: Based on the spatiotemporal redundancy measurement results, an adaptive compression algorithm or multi-source subspace mapping method is used to reduce the dimensionality and aggregate highly correlated features, while retaining the main feature components sensitive to spatiotemporal evolution. By constructing a feature mapping matrix, redundancy removal and compact representation of multi-source data are achieved, forming a low-dimensional, structured, lightweight spatiotemporal feature set, thereby improving system computational efficiency and transmission performance.

[0114] After obtaining the spatiotemporal redundancy measurement model, to reduce the redundancy of multi-source sensing data and improve the model's computational efficiency and feature discrimination capability, an adaptive feature compression and de-redundancy modeling method based on redundancy sensing is constructed. This method achieves adaptive compression of multimodal features through feature contribution analysis, principal component mapping, and weighted dimensionality reduction, retaining key representational information while eliminating redundant feature dimensions.

[0115] (4.1) Construction of the feature correlation matrix

[0116] Let the feature vector of the i-th sensor obtained under a uniform spatiotemporal resolution be:

[0117]

[0118] Where: f ik : The k-th feature component of the i-th sensor; m: the feature dimension of each sensor.

[0119] Combine the feature vectors of all sensors into a single feature matrix:

[0120]

[0121] Calculate the correlation matrix between features based on this matrix:

[0122]

[0123] Where: C: Multi-source feature covariance matrix (m×m); μ: Feature mean vector, μ = (1 / n)Σ (i=1) n f i ;(F-μ): The centered feature matrix.

[0124] (4.2) Feature contribution and redundancy weight analysis

[0125] To measure the contribution of each feature to the overall information, a feature energy distribution function is introduced:

[0126]

[0127] Where: λ k E: The k-th eigenvalue of matrix C; k The energy percentage of the k-th feature dimension reflects its information contribution.

[0128] Combined with the global spatiotemporal redundancy index R obtained in the previous stage i Define the comprehensive redundancy weight for each feature dimension:

[0129]

[0130] Where: ω k : Redundancy weights of feature k; γ1, γ2: Coefficients balancing feature energy and system redundancy effects, satisfying γ1 + γ2 = 1; Σ (i=1) n R i / n: System average redundancy.

[0131] higher ω k This indicates that the feature has strong redundancy, which can be weakened during the compression stage.

[0132] (4.3) Adaptive feature compression mapping

[0133] Based on the characteristic energy and redundant weights, an adaptive dimensionality reduction transformation matrix W_d ∈ ^{m×r}, its objective function is:

[0134]

[0135] Where: Tr(·): matrix trace operator; η: regularization parameter, used to control the redundancy removal intensity; ‖W_d(k,:)‖2²: represents the energy of the k-th row in the dimension reduction matrix.

[0136] This optimization problem can be solved by generalized eigenvalue decomposition, yielding a compressed mapping matrix W_d that retains feature dimensions with high information content and low redundancy.

[0137] The new features after feature compression are represented as follows:

[0138]

[0139] Where: F′ ∈ ^{r×n}: Compact feature matrix after dimensionality reduction; r < m: Number of principal feature dimensions retained after dimensionality reduction.

[0140] (4.4) Redundant feature removal and remodeling

[0141] To maintain the interpretability of the deredundant multi-source features at the physical level, a remapping model for compressed features is established based on the physical consistency constraints in step 2:

[0142]

[0143] Where: Φ(F′): Feature representation after physical consistency constraint mapping; A_p: Physical constraint linear transformation matrix (r×r); b_p: Offset vector (r×1), used to correct the constant terms lost during compression.

[0144] Optimize A_p and b_p by minimizing the compression reconstruction error:

[0145]

[0146] Where: ||·||_F: Frobenius norm; λ_p: regularization term balancing physical consistency and reconstruction error; I: identity matrix.

[0147] (4.5) Output Results and Model Significance

[0148] Finally, the redundant-free adaptive feature set is obtained:

[0149] F_opt = Φ(F′)

[0150] This feature set simultaneously satisfies the following properties:

[0151] Information integrity: Preserve the main energy components in the original data;

[0152] Redundancy suppression: Significantly reduces redundant information among multiple sensor sources;

[0153] Physical consistency: The compression results still conform to the vehicle motion and sensor geometry constraints.

[0154] Through the above adaptive feature compression and redundancy removal modeling process, data dimensionality can be reduced and computational efficiency can be improved while ensuring feature validity. This also provides simpler and more robust feature inputs for subsequent anomaly detection and fault tolerance control.

[0155] Step 5) Lightweight Spatiotemporal Data Structure Construction: Design a hierarchical spatiotemporal data structure, encode the feature set after redundancy removal according to time, space, and physical quantity hierarchy, and construct a lightweight feature map that can be quickly indexed and updated. Through hierarchical encoding and index management mechanisms, achieve efficient organization and storage of data, and ensure the traceability and recallability of features at different time scales and spatial ranges.

[0156] After completing adaptive feature compression and redundancy removal modeling, to achieve real-time perception and efficient computation of the intelligent driving system in resource-constrained environments, a lightweight spatiotemporal data structure supporting fast access, dynamic updates, and spatial consistency needs to be constructed. This structure achieves structured organization and dynamic maintenance of data through hierarchical indexing, block-level compression, and time synchronization mechanisms, significantly reducing storage redundancy and improving spatiotemporal query efficiency.

[0157] (5.1) Spatiotemporal data representation model

[0158] Suppose that the fused multi-source sensing data in the system, after feature compression, is represented as follows:

[0159]

[0160] Where: f i (t k ): The i-th sensor at time t k The compressed feature vector (r×1); n: number of sensors; T: time series length; F_opt: multi-source spatiotemporal feature set after redundancy removal.

[0161] To efficiently manage these feature data, a lightweight spatiotemporal tensor structure is introduced. ∈ ^{r×n×T}:

[0162]

[0163] in, The three dimensions correspond to “feature dimension, sensor number, and time step”, which can simultaneously express spatial distribution and temporal evolution characteristics.

[0164] (5.2) Block-level index and hierarchical storage structure

[0165] To reduce the complexity of real-time data access, the tensor structure is divided into block-level units along both temporal and spatial dimensions:

[0166]

[0167] Where: S_p: the p-th spatial sensor subset; T_q: the q-th time segment; B_{p,q}: ​​the corresponding spatiotemporal data block, serving as the smallest access unit.

[0168] Define a block-level index table I_B to support fast retrieval:

[0169]

[0170] Where: Addr_{p,q}: ​​the starting address of the block in the storage medium; Size_{p,q}: ​​the size of the block; T_start, T_end: time range boundaries, used for time period retrieval.

[0171] This architecture supports O(1) level block location and cache prefetching, significantly improving real-time access efficiency.

[0172] (5.3) Lightweight compression based on sparse coding

[0173] To further reduce data volume while preserving key spatiotemporal variation features, sparse coding compression is performed on each block B_{p,q}:

[0174]

[0175] Where: Φ ∈ ^{r×m}: Sparse basis matrix, composed of Discrete Cosine Transform (DCT) or principal component bases; α_{p,q} ∈ ^{m×1}: A sparse coefficient vector that satisfies ||α_{p,q}||0 m; _{p,q}: ​​The compressed data block representation.

[0176] The sparsity coefficients are obtained by minimizing the reconstruction error:

[0177]

[0178] Where: λ: sparsity regularization coefficient, which controls the trade-off between compression ratio and information fidelity.

[0179] This compression method reduces the average storage size of spacetime blocks by about 60%–80% while maintaining the main dynamic information without distortion.

[0180] (5.4) Spatiotemporal incremental update mechanism

[0181] To support continuous updates in the real-time sensing system, an incremental update model based on a time-sliding window is defined. Let the time window length be ΔT, then the following is executed when new data arrives:

[0182]

[0183] in: (t + ΔT): The updated spatiotemporal tensor; F_new(t + ΔT): The newly added compressed feature data; γ ∈ (0,1): The updated weight coefficient, which determines the fusion ratio of historical and new data.

[0184] This mechanism ensures that the spatiotemporal data structure maintains continuity and real-time performance within the sliding time window.

[0185] (5.5) Lightweight Index Mapping and Access Optimization

[0186] To further accelerate retrieval, a lightweight index mapping function is constructed:

[0187]

[0188] Where: M(i, k): the block number corresponding to the i-th sensor at time step k; p, q: spatial and temporal partition indices, respectively; Q: the total number of time blocks.

[0189] This mapping enables a fast mapping from the sensor-time two-dimensional index to block-level storage units, allowing the system to access only the corresponding block when querying a specific spatiotemporal region, thereby significantly reducing data I / O overhead.

[0190] (5.6) Output and Application

[0191] Ultimately, the resulting lightweight spatiotemporal data structure is represented as follows:

[0192]

[0193] in: : Main storage tensor; I_B: Block-level index table; Φ, α_{p,q}: ​​Sparse compression parameters; M: Fast access map table.

[0194] Step 6) Fusion Output and Real-time Support: Weighted fusion and reconstruction are performed on feature data with good physical consistency and low redundancy to generate joint measurement results. Based on the fusion output, the physical constraint model is dynamically updated and its parameters are adaptively adjusted to achieve online optimization of the model and adaptive improvement of system performance, thereby maintaining high robustness and real-time response capability in complex driving environments.

[0195] Based on the construction of a lightweight spatiotemporal data structure, a fusion output and real-time support mechanism is established to realize real-time perception, decision-making, and fault-tolerant control of intelligent driving systems. This mechanism comprehensively utilizes multi-source compression features, spatiotemporal redundancy, and physical consistency constraints to achieve dynamic fusion of multi-level information and real-time response to control commands, thereby ensuring the stability and safety of the system in complex environments.

[0196] (6.1) Multi-source feature fusion model

[0197] Let f be the feature of the i-th sensor obtained after feature compression and redundancy removal. i ′(t), whose corresponding spatiotemporal redundancy weight is w i If (t), then the multi-source fusion feature F_fuse(t) is expressed as:

[0198]

[0199] Where: F_fuse(t): the fused system-aware features (r×1 vector); f i ′(t): The compressed feature output of the i-th sensor; w i (t): Fusion weight, dynamically adjusted based on the spatiotemporal redundancy index and data confidence, defined as:

[0200]

[0201] Where: R i The global redundancy index obtained in step 3; C i Redundancy centrality index of sensors.

[0202] This weighting method ensures that highly independent and high-confidence sensors have a greater weight in the fusion process, while the weight of highly redundant sensors is automatically reduced, thereby achieving adaptive weighted fusion.

[0203] (6.2) Fusion optimization under physical consistency constraints

[0204] To ensure that the fusion results conform to the laws of vehicle motion and dynamics, a physical consistency constraint term is introduced, and the fusion optimization objective function is defined as follows:

[0205]

[0206] Where: Φ_phy(·): physical consistency mapping function (established by step 2); F_opt: adaptive compression feature set; E_d(t): dynamic consistency error (i.e., acceleration consistency error term); λ_c: constraint weight coefficient.

[0207] By minimizing J, the physical feasibility of the fusion output can be corrected, ensuring that the fusion result maintains the integrity of multi-source information while conforming to vehicle dynamics constraints.

[0208] (6.3) Real-time support and system feedback mechanism

[0209] To support real-time control decision-making, the fused feature F_fuse(t) is mapped to the control layer to generate dynamic state estimates and fault-tolerant control commands. The system state estimate is defined as:

[0210]

[0211] in: (t): System state estimate (position, attitude, velocity, etc.) at time t; A_c: Mapping matrix from system characteristics to state; B_c: Control input mapping matrix; u_ref(t): Reference control input signal (e.g., desired steering angle or acceleration).

[0212] Combining the lightweight data structure from step 5, the fused features can be quickly accessed and updated through the index table I_B and the mapping function M(i,k), thereby completing the perception-decision closed loop within milliseconds.

[0213] (6.4) Real-time fusion feedback and fault-tolerant correction

[0214] During system operation, the deviation between the current fused features and the predicted state is calculated in real time:

[0215]

[0216] Where: x_pred(t): the state vector predicted by the state-space model; Δx(t): the fusion prediction bias, used for fault tolerance correction.

[0217] If ||Δx(t)||2 > ε_th (preset threshold), then the fault tolerance compensation mechanism is triggered, and the state is corrected through Kalman update:

[0218]

[0219] Where: K_t: adaptive Kalman gain matrix; H: observation matrix; (t|t-1): Predicted state vector; (t|t): The corrected optimal estimated state.

[0220] This mechanism ensures that when individual sensors drift, experience sudden noise changes, or suffer short-term failures, the system can still maintain stable output with the support of redundant information.

[0221] (6.5) Integrated Output and System Interface

[0222] The real-time support results of the fusion output include: F_fuse(t): multimodal fusion sensing features under unified physical constraints; (t|t): System state estimate after fault-tolerant correction; u_adj(t): Adaptive control output based on deviation correction, defined as:

[0223]

[0224] Where K_f is the feedback gain matrix.

[0225] Ultimately, the system forms an integrated closed loop of "data fusion → state estimation → fault-tolerant control", providing real-time perception support and highly reliable control input for the autonomous driving decision-making layer.

Claims

1. A method for constructing multi-variable measurement lightweight spatiotemporal redundant data for intelligent driving, characterized in that, Includes the following steps: Step 1) Multi-source sensing data acquisition and spatiotemporal synchronization: Collect environmental measurement data from multiple types of sensors, synchronize the time and spatial coordinates of different data sources, use a unified timestamp and spatial transformation matrix to achieve spatiotemporal registration of multi-source data, and obtain a multimodal preprocessed dataset with unified spatiotemporal resolution through noise filtering, scale normalization and missing value compensation, so as to provide consistent input for feature extraction. Step 2) Key feature extraction under physical consistency constraints: Based on the kinematic and dynamic model of the intelligent driving system, a multi-dimensional feature space containing multiple physical variables is defined, the mapping relationship between the measurement quantities of each sensor and the real physical variables is established, a physical consistency error function is constructed, the observation data of different sensors are constrained and optimized, common mode features that satisfy the physical coupling law are extracted, and false and noise interference signals are filtered out. Step 3) Construction of the spatiotemporal redundancy measurement model: Based on multi-source synchronous observation data, the mutual information and similarity of features under different time windows and spatial perspectives are calculated to construct a spatiotemporal redundancy measurement matrix; through statistical dependency analysis between features, the redundancy degree of each feature component is quantified, highly correlated feature regions are identified, and a spatiotemporal redundancy description model is established to provide data basis for feature compression. Step 4) Adaptive feature compression and redundancy removal modeling: Based on the spatiotemporal redundancy measurement results, an adaptive compression algorithm or multi-source subspace mapping method is used to reduce the dimensionality and aggregate features whose covariance index exceeds a preset threshold, while retaining the main feature components that are sensitive to spatiotemporal evolution; by constructing a feature mapping matrix, the redundancy removal and compact representation of multi-source data are achieved, forming a lightweight spatiotemporal feature set; Step 5) Lightweight spatiotemporal data structure construction: Design a hierarchical spatiotemporal data structure, encode the feature set after redundancy removal according to time, space and physical quantity hierarchy, construct a lightweight feature map, and realize the organization and storage of data through hierarchical encoding and index management mechanism, so that the features have traceability and callability in different time scales and spatial ranges; Step 6) Fusion Output and Real-time Support: Weighted fusion and reconstruction are performed on feature data whose physical consistency constraint error is less than a preset threshold and whose redundancy evaluation index is lower than a preset threshold to generate joint measurement results. The physical constraint model is dynamically updated and its parameters are adaptively adjusted based on the fusion output to optimize the model.

2. The intelligent driving-oriented multivariate measurement lightweight spatiotemporal redundant data construction method according to claim 1, characterized in that, The specific method in step 1) is as follows: In intelligent driving systems, environmental data is collected using various types of sensors, including LiDAR, cameras, millimeter-wave radar, and inertial measurement units (IMUs), denoted as follows: LiDAR point cloud data , image data , radar range-velocity data and IMU inertial data ; To realize the unified space-time alignment of multiple sensors, a unified timestamp is adopted for different data sources with a spatial coordinate transformation matrix with a translation vector The following synchronization is completed: wherein, represents the original coordinate collected by the i-th sensor, is the synchronized coordinate in the global coordinate system, and the synchronized coordinate is formed by time interpolation and space mapping to form a comparable multi-modal synchronized dataset .

3. The intelligent driving-oriented multivariate measurement lightweight spatiotemporal redundant data construction method according to claim 1, characterized in that, The specific method in step 2) is as follows: Based on the physical consistency theory, the coupling constraint models of position , velocity and acceleration are established: The multi-source sensor measurements are denoted as a consistency error function is defined as By minimizing the physical constraints, combined with convolutional neural network (CNN) and statistical modeling, a stable and physically reasonable feature set is extracted , and abnormal and false measurements are removed.

4. The intelligent driving-oriented multivariate measurement lightweight spatiotemporal redundant data construction method according to claim 1, characterized in that, The specific method in step 3) is as follows: To analyze the redundancy relationship between multi-time and multi-view features, a spatio-temporal redundancy matrix is constructed whose elements are defined as wherein, denotes the mutual information between the i, j features, is the entropy of the features; This matrix is ​​used to measure the degree of redundancy between different times and different sensor perspectives, providing a quantitative basis for subsequent compression modeling.

5. The intelligent driving-oriented multivariate measurement lightweight spatiotemporal redundancy data construction method according to claim 1, characterized in that, The specific method in step 4) is as follows: According to the analysis result of the redundancy matrix , a feature correlation threshold value is defined , when , it is determined that the features are highly redundant An adaptive compression algorithm is used to perform linear dimensionality reduction and aggregation on highly correlated features: wherein, is a compression mapping matrix, is a light feature set after de-redundancy, retaining the most representative components of the spatio-temporal evolution.

6. The intelligent driving-oriented multivariate measurement lightweight spatiotemporal redundant data construction method according to claim 1, characterized in that, The specific method in step 5) is as follows: Designing hierarchical spatio-temporal data structures wherein each layer corresponds to different scale spatio-temporal feature representation; mapping the compressed features to a light-weight feature map: where the mapping function implements the transformation from the high-dimensional feature space to the sparse representation space; Through hierarchical indexing and encoding strategies, data can be quickly retrieved spatially and updated efficiently over time.

7. The intelligent driving-oriented multivariate measurement lightweight spatiotemporal redundant data construction method according to claim 1, characterized in that, The specific method in the step 6) is: Lightweight structure As the unified data input of intelligent driving perception module, it is called by subsequent multi-source data fusion, state estimation and fault-tolerant control modules. Define the fusion output function: wherein, represents a fusion operation function, including weight update, confidence calculation and real-time state estimation; through the above steps, the system has high robustness and real-time response ability under complex working conditions.

Images