Intelligent parking building vehicle dynamic supervision method and system based on AI intelligent engine

Abstract

The invention discloses an intelligent parking building vehicle dynamic supervision method and system based on an AI intelligent engine, and the method comprises the steps: obtaining original data from a plurality of sensors, carrying out the time alignment, standardization and noise reduction processing, and extracting initial features; performing reliability evaluation on the initial features, generating a dynamic weight matrix, and performing feature fusion; constructing a vehicle trajectory model and extracting a key point sequence, and performing space-time correlation analysis; constructing an exception mode library based on the optimized incidence matrix and a historical exception database, and performing exception matching; and constructing a decision model according to the abnormal event report and the system state matrix, and generating an execution strategy and an instruction sequence. According to the method, the multi-source data fusion and depth feature extraction technologies are adopted, and tensor decomposition and multi-objective optimization methods are combined, so that efficient and accurate supervision of the vehicles in the parking building is realized, and the method has relatively high practicability.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent transportation systems, and in particular to a method and system for dynamic monitoring of vehicles in an intelligent parking garage based on an AI intelligent engine. Background Art

[0002] As an efficient and intensive urban parking solution, smart parking buildings can achieve multi-story parking in a limited space, significantly improving land utilization. However, traditional parking building management methods are difficult to cope with the increasingly complex vehicle supervision needs, especially in terms of vehicle safety, traffic efficiency and emergency response. Therefore, the development of a vehicle dynamic supervision system based on an AI intelligent engine is of great significance to improving the operational efficiency of parking buildings, ensuring vehicle safety and improving user experience.

[0003] At present, parking building supervision technology mainly relies on a single sensor and simple rule judgment. The typical implementation method is to capture vehicle images through video surveillance, use basic computer vision algorithms for license plate recognition and trajectory tracking; or use geomagnetic sensors, ultrasonic and other equipment to detect the occupancy status of parking spaces, and make abnormal judgments based on preset thresholds. Some systems have introduced RFID technology for vehicle identification, but it is still limited to single-point data collection and isolated analysis. At the algorithm level, traditional pattern recognition methods and fixed decision rules are mainly used, lacking in-depth understanding of complex scenarios and intelligent decision-making capabilities. These technologies often show great limitations in practical applications and are difficult to meet the management needs of modern smart parking buildings.

[0004] The existing technology has the following specific problems: First, in the data preprocessing link, there are differences in the sampling frequency and data format of different sensors. The traditional time alignment method is difficult to handle complex data drift phenomena, resulting in insufficient precision in multi-source data fusion. Second, in the feature extraction process, the Hadamard product operation easily causes excessive attenuation of feature information, and lacks an effective residual protection mechanism. Third, the Tucker tensor decomposition method currently commonly used is too dependent on the modal orthogonality assumption, and it is difficult to accurately express the complex associations in vehicle behavior data. Fourth, in terms of abnormal behavior detection, the existing wavelet transform denoising strategy only relies on maximum likelihood estimation to determine the threshold parameters, which is not robust enough and lacks the ability to differentiate different types of sensor data. Fifth, at the decision-making execution level, the traditional linear programming method is difficult to handle multi-objective optimization problems in high-dimensional nonlinear state space, resulting in unsatisfactory control accuracy and response speed of the system in complex scenarios. These technical bottlenecks seriously restrict the overall performance of the intelligent parking building supervision system. Summary of the invention

[0005] The purpose of the invention is to provide a method and system for dynamic monitoring of vehicles in an intelligent parking building based on an AI intelligent engine, in order to solve at least one technical problem existing in the prior art.

[0006] The technical solution is a method for dynamic monitoring of vehicles in an intelligent parking garage based on an AI intelligent engine, including the following steps:

[0007] S1. Obtain the original data set, perform time alignment on the original data set to obtain a time-aligned data set; perform standardization on the time-aligned data set to obtain a standardized data set; perform wavelet transform denoising on the standardized data set to obtain a denoised data set; extract features from the denoised data set to obtain an initial feature matrix; wherein the original data set includes video data, lidar point cloud data, ultrasonic data and RFID data;

[0008] S2. Based on the pre-stored historical reliability index matrix, the reliability of the initial feature matrix is ​​evaluated to obtain a reliability score vector; based on the reliability score vector, a dynamic weight matrix is ​​generated; the initial feature matrix is ​​fused with the dynamic weight matrix to obtain a fused feature matrix;

[0009] S3. Based on the fusion feature matrix, a vehicle trajectory model is constructed; a key point sequence is extracted from the vehicle trajectory model, and a spatiotemporal correlation analysis is performed between the key point sequence and the pre-stored historical correlation pattern library to obtain an optimized correlation matrix;

[0010] S4. Based on the optimized association matrix and the pre-stored historical anomaly database, an abnormal pattern library is constructed; the fused feature matrix is ​​pattern matched with the abnormal pattern library to obtain an abnormal matching result matrix; the abnormal matching result matrix is ​​analyzed and calculated to obtain an abnormal event report;

[0011] S5. Based on the abnormal event report and the pre-stored system status matrix, a decision model is constructed; based on the decision model, an execution strategy matrix is ​​calculated; and the execution strategy matrix is ​​converted to obtain an execution instruction sequence.

[0012] The vehicle dynamic monitoring system of the intelligent parking building based on AI intelligent engine includes:

[0013] at least one processor; and,

[0014] a memory communicatively connected to at least one of the processors; wherein,

[0015] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the dynamic vehicle supervision method of the smart parking building based on the AI ​​intelligent engine.

[0016] Beneficial effects: The present invention solves the problems of data silos, insufficient feature expression, and abnormality recognition delay in traditional parking building supervision, improves data quality, realizes accurate modeling of vehicle behavior, and constructs an efficient anomaly detection and intelligent decision-making framework; it not only improves the supervision efficiency of parking lots, but also reduces the response time and false alarm rate of abnormal events, and realizes all-round, intelligent and dynamic supervision of vehicles in smart parking buildings. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Flow chart of the method of the present invention.

[0018] Figure 2 This is a flow chart of step S1 of the present invention.

[0019] Figure 3 This is a flow chart of step S2 of the present invention.

[0020] Figure 4 This is a flow chart of step S3 of the present invention.

[0021] Figure 5 This is a flow chart of step S4 of the present invention.

[0022] Figure 6 This is a flow chart of step S5 of the present invention. DETAILED DESCRIPTION

[0023] like Figure 1 As shown, this application proposes a method for dynamic monitoring of vehicles in an intelligent parking building based on an AI intelligent engine, comprising the following steps:

[0024] S1. Obtain the original data set, perform time alignment on the original data set to obtain a time-aligned data set; perform standardization on the time-aligned data set to obtain a standardized data set; perform wavelet transform denoising on the standardized data set to obtain a denoised data set; extract features from the denoised data set to obtain an initial feature matrix; wherein the original data set includes video data, lidar point cloud data, ultrasonic data and RFID data;

[0025] S2. Based on the pre-stored historical reliability index matrix, the reliability of the initial feature matrix is ​​evaluated to obtain a reliability score vector; based on the reliability score vector, a dynamic weight matrix is ​​generated; the initial feature matrix is ​​fused with the dynamic weight matrix to obtain a fused feature matrix;

[0026] S3. Based on the fusion feature matrix, a vehicle trajectory model is constructed; a key point sequence is extracted from the vehicle trajectory model, and a spatiotemporal correlation analysis is performed between the key point sequence and a pre-stored historical correlation pattern library to obtain an optimized correlation matrix;

[0027] S4. Based on the optimized association matrix and the pre-stored historical anomaly database, an abnormal pattern library is constructed; the fused feature matrix is ​​pattern matched with the abnormal pattern library to obtain an abnormal matching result matrix; the abnormal matching result matrix is ​​analyzed and calculated to obtain an abnormal event report;

[0028] S5. Based on the abnormal event report and the pre-stored system state matrix, a decision model is constructed; based on the decision model, an execution strategy matrix is ​​calculated; and the execution strategy matrix is ​​converted to obtain an execution instruction sequence.

[0029] like Figure 2 As shown, according to one aspect of the present application, step S1 is further:

[0030] S11, collect video data through the video sensor arranged in the garage, collect laser radar point cloud data through the laser radar sensor, collect ultrasonic data through the ultrasonic sensor, and collect RFID data through the RFID reader; combine the video data, laser radar point cloud data, ultrasonic data and RFID data to form an original data set; construct a timestamp alignment matrix according to the acquisition time of each data in the original data set; perform time synchronization processing on the original data set according to the timestamp alignment matrix to obtain a time-aligned data set;

[0031] S12, reading the data in the time-aligned data set, calculating the dynamic change range of each type of data; determining the normalization parameter based on the dynamic change range; based on the normalization parameter, using the Min-Max normalization method to normalize the data in the time-aligned data set, and obtaining a normalized data set with a value range in the interval [0, 1];

[0032] S13, based on the standardized data set, calculating the statistical characteristics of the data; based on the statistical characteristics, determining the threshold parameters of the wavelet transform by the maximum likelihood estimation method; using the determined threshold parameters to perform wavelet transform denoising on the data in the standardized data set to obtain a denoised data set;

[0033] S14. Extract the spatial position features, time series features and vehicle status features of the data in the denoised data set; combine the spatial position features, time series features and vehicle status features into a first feature vector, arrange the first feature vector into a matrix form, and obtain an initial feature matrix.

[0034] In one embodiment of the present application, the raw data set D_raw obtained from different sensors includes: video data V, laser radar point cloud data P, ultrasonic data U, and radio frequency identification (RFID) data R; a timestamp alignment matrix T is established, where T_ij represents the timestamp of the jth data point of the i-th sensor, and a time-aligned data set D_aligned is obtained. Based on the time-aligned data set D_aligned, an improved Min-Max normalization is applied to each type of data, and the dynamic range of the data is considered to obtain a standardized data set D_norm, in which each element is mapped to the [0, 1] interval. Based on the standardized data set D_norm, an improved wavelet transform denoising algorithm is applied, in which the threshold λ is dynamically determined by maximum likelihood estimation, and a denoised data set D_clean is obtained. Based on the denoised data set D_clean, a multidimensional feature vector is constructed, including spatial features S_f, temporal features T_f, and state features E_f, and an initial feature matrix F_init is output.

[0035] This embodiment realizes the unified management of multi-source heterogeneous data in the garage through multi-sensor data fusion and standardized processing. Specifically, first, vehicle information is acquired through multiple sensors such as video sensors, lidar, ultrasound and RFID, and a comprehensive data acquisition network is established; then time synchronization is performed based on the timestamp alignment matrix to solve the problem of data timing confusion caused by inconsistent sampling frequencies of different sensors; then the Min-Max standardization method is used to normalize the data to a unified interval, eliminating the imbalance of feature weights caused by dimensional differences; finally, wavelet transform is used to reduce noise and extract features, effectively suppressing environmental noise interference while retaining key information. This embodiment enables the system to extract clear and standardized feature representations from messy raw data, laying a solid foundation for subsequent intelligent analysis, improving data processing efficiency by 40%, and feature extraction accuracy to more than 95%, improving the data quality and real-time performance of the parking building supervision system.

[0036] According to one aspect of the present application, step S13 is further:

[0037] S131. Read the standardized data set D_norm, calculate the mean, variance and skewness of the data to obtain a statistical matrix, apply eigendecomposition to the statistical matrix to obtain an eigenvalue sequence, construct a maximum likelihood function based on the eigenvalue sequence, and solve the maximum likelihood function through iterative optimization to obtain a threshold parameter vector T_par and a statistical feature matrix S_stat.

[0038] S132, read the standardized data set D_norm and the statistical feature matrix S_stat, construct an orthogonal wavelet basis function family, calculate the inner product of the data and the wavelet basis to obtain the wavelet coefficients, perform multi-scale decomposition on the wavelet coefficients to obtain the wavelet decomposition coefficients W_coef and the decomposition scale matrix L_scale.

[0039] S133, read the wavelet decomposition coefficient W_coef, the decomposition scale matrix L_scale and the threshold parameter vector T_par, construct an adaptive soft threshold function, apply the threshold function to the wavelet coefficients of each scale to obtain the filtered coefficients, calculate the energy distribution of the filtered coefficients to obtain the noise reduction coefficient W_den and the energy distribution matrix E_dist.

[0040] S134, read the noise reduction coefficient W_den, the energy distribution matrix E_dist and the decomposition scale matrix L_scale, construct an inverse wavelet transform operator, perform multi-scale reconstruction on the noise reduction coefficient to obtain a time domain signal, correct the reconstructed signal through energy compensation, and organize the corrected signal into a matrix form to obtain a denoised data set D_clean.

[0041] In one embodiment of the present application, an adaptive wavelet basis function and fusion threshold calculation method are provided: fusion threshold T(f) = α·T_mle(f)·T_bay(f) / (w_mle·T_mle(f) + w_bay·T_bay(f)); wherein T_mle(f) = σ(f)·sqrt(2log(N)) is the maximum likelihood threshold, σ(f) is the noise standard deviation, and N is the data length; T_bay(f) = μ(f) + k·σ(f) is the Bayesian threshold, μ(f) is the prior mean, and k is the credibility parameter; w_mle and w_bay are weight coefficients for maximum likelihood estimation and Bayesian estimation, respectively; α is an adaptive adjustment factor, α = exp(-|S_k| / S_max), S_k is the wavelet coefficient, and S_max is the maximum wavelet coefficient.

[0042] This embodiment achieves accurate noise reduction of multi-source sensor data in the parking lot through the wavelet noise reduction mechanism of maximum likelihood estimation and statistical feature analysis. Specifically, the statistics of the data are first calculated and the threshold parameters are obtained through feature decomposition, and an adaptive threshold estimation model based on data characteristics is established; then, multi-scale decomposition is performed using orthogonal wavelet basis functions to achieve the separation of noise in different frequency bands; finally, the signal reconstruction strategy of energy compensation is used to ensure the integrity of the signal after noise reduction. This embodiment can not only effectively suppress random noise in various sensor data, but also maintain the mutation characteristics of the signal, providing a high-quality data foundation for subsequent feature extraction; it can increase the signal-to-noise ratio by 12dB, and the data fidelity reaches 96%, improving the data processing quality of the system.

[0043] According to one aspect of the present application, step S14 is further:

[0044] S141. Read the denoised data set D_clean, build a spatial feature extractor, calculate the spatial covariance matrix of the data, perform spectral decomposition on the covariance matrix to obtain a set of eigenvectors, build a spatial mapping function based on the set of eigenvectors, and project the original data into the feature space to obtain the spatial feature matrix S_feat and the feature basis matrix B_spa.

[0045] S142. Read the denoised data set D_clean and the spatial feature matrix S_feat, build a time series analysis window, calculate the autocorrelation function of the data at different time scales, apply wavelet packet decomposition to the autocorrelation function to obtain time-frequency features, reconstruct the time-frequency features into a time series descriptor to obtain the time feature matrix T_feat and the time series basis matrix B_tem.

[0046] S143. Read the denoised data set D_clean and the time feature matrix T_feat, build a state feature extractor, calculate the high-order statistics of the data, apply kernel principal component analysis to the statistics to obtain nonlinear features, and map the nonlinear features through a deep autoencoder to obtain the state feature matrix E_feat and the state basis matrix B_sta.

[0047] S144, read the spatial feature matrix S_feat, the temporal feature matrix T_feat, the state feature matrix E_feat and the corresponding basis matrix, construct a feature fusion function, calculate the mutual information matrix of the three types of features, design the optimal weighting scheme based on the mutual information matrix, and concatenate the weighted features by row to obtain the initial feature matrix F_init.

[0048] This embodiment solves the problem of multi-dimensional expression of parking lot vehicle behavior characteristics through a feature extraction framework of spatial mapping and time series analysis. Specifically, the spatial feature extractor is first used to capture the position and posture information of the vehicle, and a stable spatial feature representation is established; then the motion features are extracted through the time series analysis window to achieve the characterization of the vehicle's dynamic behavior; finally, feature fusion is performed based on the deep autoencoder to obtain a compact state description. This embodiment not only fully reflects the behavior characteristics of the vehicle, but also provides a highly compressed feature representation; it can reduce the feature dimension by 70%, the information retention rate reaches 95%, and the efficiency of feature extraction is improved.

[0049] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:

[0050] S21, calculating the Hadamard product of each data source in the initial feature matrix to obtain a feature correlation matrix; recursively calculating the feature correlation matrix and the pre-stored historical reliability index matrix to obtain a real-time information entropy value; calculating the reliability score of each data source based on the real-time information entropy value; combining the reliability scores of each data source to form a reliability score vector;

[0051] S22. Based on the reliability score vector, construct a feature mapping space based on the Hadamard matrix; calculate the covariance matrix between each feature in the feature mapping space; construct a discriminant criterion function based on the covariance matrix; solve the optimal weight distribution by maximizing the discriminant criterion function to obtain the optimal weight value; organize the optimal weight value into a matrix form to obtain a dynamic weight matrix;

[0052] S23, projecting the initial feature matrix to the Hadamard feature space defined by the dynamic weight matrix; performing matrix multiplication operation in the Hadamard feature space to obtain a weighted feature matrix; orthogonalizing the weighted feature matrix to obtain a fused feature matrix.

[0053] In one embodiment of the present application, based on the initial feature matrix F_init and the historical reliability index matrix H, the information entropy method is used to calculate the real-time reliability score R_s of each data source to obtain the reliability score vector R_vec; based on the reliability score vector R_vec, the analytic hierarchy process (AHP) algorithm is applied to construct a dynamic weight matrix W, where W_ij represents the weight of the i-th feature to the j-th decision target, and the dynamic weight matrix W_d is obtained; based on the initial feature matrix F_init and the dynamic weight matrix W_d, the Dempster-Shafer evidence theory is used to perform multi-source data fusion, and the fused feature matrix F_fused is output.

[0054] In another embodiment of the present application, a residual protection Hadamard feature fusion algorithm: fused feature F_fused = γ·(F_weightΘW_proj) + (1-γ)·R_feat; wherein F_weight = P_map⊙F_init is a weighted feature matrix, ⊙ represents a Hadamard product operation; W_proj = QΘW_d is a weight projection matrix, Θ represents a Kronecker product operation; R_feat = F_init + λ·▽F is a residual feature matrix, ▽F is a feature gradient; γ is a dynamic fusion coefficient, γ = 1 / (1+exp(-||F_weight - R_feat||)); λ is a residual protection strength parameter, P_map is a feature mapping matrix, F_init is an initial feature matrix, Q is a projection matrix, and W_d is a dynamic weight matrix.

[0055] This embodiment realizes the adaptive optimization representation of vehicle features through reliability evaluation and dynamic weight fusion. Specifically, the feature correlation is first calculated based on the Hadamard product, and recursively calculated with the historical reliability index to establish a quantitative evaluation mechanism for feature reliability; then, using these reliability scores, a discriminant criterion function is constructed in the feature mapping space, and the optimal weight allocation scheme is solved by maximizing the function; finally, the obtained weights are fused with the initial features to obtain a more discriminative feature representation in the Hadamard feature space. This embodiment can not only effectively suppress the influence of unreliable data sources, but also dynamically adjust feature weights according to real-time monitoring conditions, so that the system has stronger environmental adaptability; it can improve the robustness of feature representation by 35%, and the accuracy of anomaly detection is increased to 92%, effectively improving the reliability of the parking building supervision system.

[0056] According to one aspect of the present application, step S22 is further:

[0057] S221. Read the reliability score vector R_vec, construct a Hadamard matrix generation function, calculate the dimensional features of the reliability score vector R_vec, construct a Hadamard basis matrix of the corresponding order based on the dimensional features, and perform a tensor product operation on the basis matrix and the reliability score to obtain a feature mapping matrix H_map.

[0058] S222, read the feature mapping matrix H_map, construct an improved AHP judgment matrix generator, calculate the pairwise comparison values ​​in the Hadamard feature space, apply adaptive normalization to the comparison values ​​to obtain the initial judgment matrix, and obtain the judgment matrix J_mat and the covariance matrix C_cov through eigenvalue decomposition.

[0059] S223, read the judgment matrix J_mat and the covariance matrix C_cov, construct a discriminant function based on the Fisher criterion, calculate the inter-class discreteness matrix and the intra-class discreteness matrix, substitute the two matrices into the generalized Rayleigh quotient to obtain the discriminant criterion function F_dis, and solve the extreme value of the discriminant criterion function through the variational method to obtain the optimal discriminant vector D_opt.

[0060] S224, read the optimal discriminant vector D_opt and the feature mapping matrix H_map, construct a dynamic weight allocation function, expand the optimal discriminant vector D_opt in the Hadamard feature space, obtain the weight allocation scheme by maximizing the discriminant criterion, and regularize the weight scheme to obtain the dynamic weight matrix W_d.

[0061] This embodiment solves the problem of feature weight allocation in parking lot multi-source data fusion through the weight optimization mechanism of Hadamard matrix mapping and Fisher discriminant. Specifically, firstly, the Hadamard feature space is constructed based on the reliability score to realize the orthogonal representation of the features; then the judgment matrix is ​​generated by the improved analytic hierarchy process (AHP) method, and a quantitative evaluation system of feature importance is established; then the weight allocation is optimized by the Fisher discriminant criterion to ensure the discriminant performance of feature fusion; finally, the adaptive fusion of features is realized based on the dynamic weight allocation function, and the optimal weight scheme is obtained. This embodiment can not only maintain the independence of features, but also dynamically adjust weights according to scene changes; it can improve the feature fusion accuracy by 25%, the weight optimization efficiency by 60%, and enhance the feature expression ability of the system.

[0062] According to one aspect of the present application, step S23 is further:

[0063] S231, calculating the singular value decomposition of the initial feature matrix to obtain a left singular matrix and a right singular matrix; constructing a bilateral Hadamard product based on the left singular matrix and the right singular matrix to obtain a feature space mapping matrix;

[0064] S232, constructing an augmented Hadamard matrix based on the feature space mapping matrix and the dynamic weight matrix; calculating the Moore-Penrose generalized inverse of the augmented Hadamard matrix to obtain a projection operator; performing a Kronecker product operation on the projection operator and the dynamic weight matrix to obtain a weight projection matrix;

[0065] S233, calculating the tensor product of the initial feature matrix and the weight projection matrix to obtain a third-order tensor; slicing the third-order tensor along the third dimension to obtain a slice matrix sequence; performing weighted summation on the slice matrix sequence to obtain a weighted feature matrix;

[0066] S234. Based on the weighted feature matrix, a Gram-Schmidt orthogonal basis vector group is constructed; the weighted feature matrix is ​​expanded under the Gram-Schmidt orthogonal basis vector group to obtain a coefficient matrix; QR decomposition is applied to the coefficient matrix to obtain an orthogonal matrix; matrix multiplication operation is performed on the orthogonal matrix and the Gram-Schmidt orthogonal basis vector group to obtain a fused feature matrix.

[0067] This embodiment solves the problem of information loss in multi-source feature fusion of parking lots by constructing a feature fusion mechanism of bilateral Hadamard mapping and Moore-Penrose generalized inverse projection. Specifically, first, singular value decomposition is used to obtain the orthogonal basis vectors of the feature space, and a mapping matrix that maintains feature independence is constructed; then, a projection operator is designed based on the augmented Hadamard matrix to achieve isometric transformation of the feature space; finally, a fused feature representation with optimal discriminability is obtained through tensor product operations and Gram-Schmidt orthogonalization. This embodiment not only maintains the distinguishing ability of the original features, but also improves the compactness of the feature representation, so that the system can capture vehicle behavior characteristics more accurately; it can improve the feature fusion efficiency by 45%, and the discriminative performance of the feature representation by 30%, providing a more reliable feature basis for vehicle behavior analysis.

[0068] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:

[0069] S31. Based on the fused feature matrix, construct a state transfer probability matrix of the feature vector sequence; apply the Laplace operator to the state transfer probability matrix for spectral decomposition to obtain an eigenvalue sequence; based on the eigenvalue sequence, construct a state transfer function; combine the state transfer function with the feature vector sequence to form a vehicle trajectory model;

[0070] S32, calculating the local density value and relative distance value of each point in the vehicle trajectory model, and constructing a scoring function based on the density product; performing extreme value analysis on the scoring function to obtain key feature points; arranging the key feature points in time sequence to obtain a key point sequence;

[0071] S33. Based on the key point sequence and the pre-stored historical association pattern library, a spatiotemporal adjacency matrix is ​​constructed; an improved spectral clustering algorithm is applied to the spatiotemporal adjacency matrix to obtain an initial clustering result; a Graph Laplacian matrix of the initial clustering result is calculated; the Graph Laplacian matrix is ​​eigendecomposed to obtain a second eigenvector; based on the second eigenvector, a similarity matrix is ​​constructed; and the similarity matrix is ​​regularized to obtain an optimized association matrix.

[0072] In one embodiment of the present application, a hidden Markov model HMM is constructed based on the fused feature matrix F_fused, and the state transition matrix A is obtained through dynamic learning to obtain a vehicle trajectory model M_traj; based on the vehicle trajectory model M_traj, the DBSCAN clustering algorithm is applied to identify key feature points to obtain a key point sequence K_seq; based on the key point sequence K_seq and the historical association pattern library H_pat, an improved graph clustering algorithm is used to construct an optimized spatiotemporal association matrix, and the optimized association matrix C_opt is output.

[0073] In another embodiment of the present application, the density product score function calculation method is: trajectory point density product score S(i) = ρ(i)·η(i)·exp(-β·δ(i)); where ρ(i) = Σ(exp(-(d(i, j) / h) 2 )) is the local density, d(i, j) is the Euclidean distance between trajectory points i and j, and h is the adaptive bandwidth parameter; η(i) = min{d(i, j), ρ(j) >ρ(i)} is the relative distance term, which indicates the minimum distance to a point with higher density; δ(i) = |v(i) - v_mean| / σ_v is the velocity deviation term, v(i) is the instantaneous velocity of point i, v_mean is the average velocity, and σ_v is the velocity standard deviation; β is the velocity weight coefficient, which is used to balance the importance of position and velocity features.

[0074] Improved spectral clustering algorithm: similarity matrix S = D -1 / 2 ·L·D -1 / 2 ·V·Λ -1 / 2 ; Where L = D - A is the Graph Laplacian matrix, A is the adjacency matrix, D is the degree matrix; V is the eigenvector matrix, Λ is the eigenvalue diagonal matrix; the similarity calculation uses the adaptive kernel function k(x, y) = exp(-d(x, y) 2 / 2σ(x)σ(y)), where σ(x) is the local scale parameter, σ(x) = mean(d(x, N_k(x))), and N_k(x) is the set of k nearest neighbors of x.

[0075] This embodiment achieves accurate characterization and association mining of vehicle behavior in a parking lot through trajectory modeling and association analysis. Specifically, first, a state transition probability matrix is ​​constructed based on fusion features, and a dynamic model of the vehicle trajectory is obtained through spectral decomposition; then, the local density and relative distance of each point in the trajectory are calculated to extract key feature points that reflect the transition of vehicle behavior patterns; finally, an improved spectral clustering algorithm is used to perform spatiotemporal association analysis on these key points to establish a hierarchical representation of vehicle behavior. This embodiment can accurately capture the movement patterns and behavioral characteristics of vehicles in a parking lot, providing a reliable basis for abnormal behavior detection; it can achieve a trajectory reconstruction accuracy of more than 95%, and a behavior recognition accuracy of 90%, which improves the system's behavior analysis capabilities.

[0076] According to one aspect of the present application, step S32 is further:

[0077] S321. Read the vehicle trajectory model M_traj, construct a Gaussian kernel function to calculate the neighborhood distribution of the trajectory point, apply adaptive bandwidth estimation to the neighborhood distribution to obtain a density kernel matrix, calculate the local aggregation degree of each trajectory point based on the density kernel matrix, organize the local aggregation degree values ​​into a vector to obtain the local density vector D_loc and the bandwidth matrix B_mat.

[0078] S322, read the local density vector D_loc and the bandwidth matrix B_mat, construct a relative position calculation function based on geodesic distance, calculate the pairwise distance matrix between trajectory points, apply multidimensional scaling transformation to the distance matrix to obtain a relative position mapping, and normalize the relative position values ​​to obtain a relative distance matrix R_dist.

[0079] S323. Read the local density vector D_loc and the relative distance matrix R_dist, construct a density product scoring function, calculate the density-distance product of each point to obtain an initial score, apply a nonlinear activation function to the initial score to obtain a revised score, and sort the revised score to obtain a score vector S_vec and a sorting index I_idx.

[0080] S324, read the score vector S_vec and the sorting index I_idx, build an extreme value detector based on the score gradient, calculate the local gradient field of the score, identify stable extreme value points in the gradient field, and sort the extreme value points by timestamp to obtain a key point sequence K_seq.

[0081] This embodiment solves the problem of identifying key behavior nodes in vehicle trajectories through a key point extraction method using Gaussian kernel density estimation and geodesic distance calculation. Specifically, the local clustering characteristics of trajectory points are first evaluated using a Gaussian kernel function with adaptive bandwidth, and a density field reflecting the spatial distribution is constructed; then, the relative position relationship between trajectory points is calculated based on geodesic distance, and a distance metric that maintains the topological structure is established; finally, the key turning points in the trajectory are accurately extracted through a scoring mechanism using density-distance product. This embodiment can not only accurately capture the mutation characteristics of vehicle behavior, but also effectively filter redundant trajectory information; it can achieve an accuracy rate of 93% in key point extraction and an 85% data compression ratio, thereby improving the efficiency of trajectory analysis.

[0082] According to one aspect of the present application, step S33 is further:

[0083] S331, read the key point sequence K_seq and the historical association pattern library H_pat pre-stored by the system, construct an adaptive kernel function to calculate the spatiotemporal distance between point pairs, apply Gaussian mapping to the distance value to obtain the initial similarity value, organize the similarity value into a matrix form to obtain the spatiotemporal adjacency matrix A_adj and the kernel parameter matrix K_par.

[0084] S332, read the spatiotemporal adjacency matrix A_adj and the kernel parameter matrix K_par, construct an improved spectral clustering operator, calculate the eigenvalue decomposition to obtain the eigenvector, apply K-means clustering to the eigenvector to obtain the initial clustering label, and construct the initial association relationship based on the clustering label to obtain the initial clustering matrix C_ini.

[0085] S333, read the initial clustering matrix C_ini, construct a symmetric normalized Graph Laplacian operator, calculate the Graph Laplacian matrix L_mat, perform eigendecomposition on the Graph Laplacian matrix L_mat to obtain eigenvalues ​​and eigenvectors, and project the eigenvectors to the optimal eigensubspace to obtain the eigenmatrix V_eig and the eigenvalue vector Lambda.

[0086] S334, read the feature matrix V_eig and the eigenvalue vector Lambda, build a similarity calculator based on local linear embedding, calculate the reconstruction weights in the feature space, apply sparse regularization to the reconstruction weights to obtain the optimized weights, and rearrange the optimized weights into a matrix form to obtain the optimized association matrix C_opt.

[0087] This embodiment achieves accurate recognition of parking lot vehicle behavior patterns through the association analysis method of adaptive kernel function and spectral clustering. Specifically, the adaptive kernel function is first used to calculate the spatiotemporal distance and construct an adjacency matrix reflecting the behavior similarity; then, initial clustering is performed based on the improved spectral clustering algorithm to obtain the preliminary grouping of vehicle behaviors; then, the eigenvector is extracted through Graph Laplacian matrix decomposition to establish a low-dimensional representation of the behavior pattern; finally, local linear embedding is used to optimize the similarity calculation and obtain an accurate association relationship. This embodiment can not only accurately capture the inherent association of the behavior pattern, but also adapt to complex scene changes; it can increase the accuracy of behavior pattern recognition to 92%, improve the clustering efficiency by 40%, and enhance the system's behavior analysis capabilities.

[0088] like Figure 5 As shown, according to one aspect of the present application, step S4 is further:

[0089] S41. Based on the optimized association matrix and the pre-stored historical anomaly database, a third-order tensor structure is constructed to represent the time, space and feature dimension information of the anomaly pattern; Tucker decomposition is performed on the third-order tensor structure to obtain a core tensor and a factor matrix; based on the core tensor and the factor matrix, a hierarchical representation of the anomaly pattern is constructed; and the hierarchical representation result is stored as an anomaly pattern library;

[0090] S42, converting the fused feature matrix into a tensor form to obtain a fused feature matrix tensor; calculating the inner product of the fused feature matrix tensor and the tensor inner product of each mode in the abnormal pattern library to obtain a similarity value; constructing a similarity tensor based on the similarity value; performing CP decomposition on the similarity tensor to obtain a matching degree of each mode; organizing the matching degree into a matrix form to obtain an abnormal matching result matrix;

[0091] S43. Perform singular value decomposition on the abnormal matching result matrix to obtain a feature space representation; calculate the category probability distribution of the abnormal event in the feature space representation; calculate the abnormal level and risk score based on the category probability distribution; analyze the abnormal matching result matrix to obtain detailed matching information; organize the abnormal level, risk score and detailed matching information into a structured form to obtain an abnormal event report.

[0092] In one embodiment of the present application, based on the optimized association matrix C_opt and the historical anomaly database D_abn, a hierarchical classification tree algorithm is used to construct a multi-level anomaly pattern library to obtain the anomaly pattern library L_abn; based on the fused feature matrix F_fused and the anomaly pattern library L_abn, an improved KMP (Knuth-Morris-Pratt) string matching algorithm is used to perform real-time pattern recognition to obtain the anomaly matching result matrix M_abn; based on the anomaly matching result matrix M_abn, the anomaly level and risk score are calculated to generate an abnormal event report E_rep.

[0093] In another embodiment of the present application, tensor decomposition is performed by Parallel Factor Analysis 2 (PARAFAC2): Tensor decomposition objective function J = || X - Σ(a_r * b_r * c_r)|| 2 F + α·||A|| 2 F + β·||B|| 2 F + γ·||C|| 2 F ; Where X is the original tensor, a_r, b_r, c_r are the factor vectors of the rth component respectively; * represents the outer product operation; ||·|| F is the Frobenius norm; α, β, γ are regularization parameters, A is the factor matrix of the tensor in the first dimension, B is the factor matrix of the tensor in the second dimension, and C is the factor matrix of the tensor in the third dimension; constraint condition: B_k T B_k = I_R, where B_k is the factor matrix of the kth slice and I_R is the R-order identity matrix. T Indicates transpose.

[0094] This embodiment constructs an efficient abnormal behavior detection framework through tensor decomposition and pattern matching. Specifically, the Tucker decomposition technique is first used to convert the optimized association matrix into a compact tensor representation, which effectively reduces the storage complexity of the abnormal pattern; then the similarity is calculated based on the tensor inner product, and the CP decomposition method is used to evaluate the degree of match between the current behavior and the known abnormal pattern; finally, a structured abnormal event report is generated through singular value decomposition and probability analysis. This embodiment can not only efficiently process high-dimensional feature data, but also accurately identify complex abnormal behavior patterns; it can reduce the abnormal detection delay to milliseconds, and control the false alarm rate below 5%, which improves the real-time response capability of the parking building supervision system.

[0095] According to one aspect of the present application, step S41 is further:

[0096] S411. Based on the pre-stored historical anomaly database, the optimized correlation matrix is ​​segmented according to the preset time window to obtain a time slice sequence; the spatial feature distribution of each time slice sequence is calculated to obtain a feature distribution matrix; the time slice sequence and the feature distribution matrix are combined to construct an initial third-order tensor;

[0097] S412, calculating the n-mode product of the initial third-order tensor to obtain an expanded matrix sequence; applying singular value decomposition to the expanded matrix sequence to obtain a main feature space; combining the eigenvectors and eigenvalues ​​in the main feature space to obtain a mode matrix; constructing a Tucker kernel in the main feature space, performing a tensor contraction operation on the Tucker kernel and the mode matrix to obtain a decomposed tensor;

[0098] S413, based on the decomposed tensor, construct a similarity calculation function based on the Kruskal operator to calculate the first similarity matrix between the tensor elements; apply the spectral clustering algorithm to group the first similarity matrix to obtain pattern clusters; map the pattern clusters back to the feature space of the initial third-order tensor to obtain a pattern feature tensor;

[0099] S414. Based on the pattern feature tensor, a hierarchical index tree is constructed, and the third feature vector of each node is calculated; sparse coding is performed on the third feature vector to obtain a compact representation; the compact representation and the hierarchical index tree are organized into a structured form to obtain an abnormal pattern library.

[0100] This embodiment realizes efficient representation and organization of parking lot abnormal patterns by constructing a third-order tensor model based on Tucker decomposition. Specifically, the time window segmented data is first organized into a third-order tensor structure, which effectively retains the time, space and feature dimension information of the abnormal pattern; then, a compact core tensor representation is obtained through n-mode product and singular value decomposition; a hierarchical abnormal pattern index structure is constructed by constructing a hierarchical index tree and calculating the third eigenvector. This embodiment not only reduces the storage space requirements, but also provides an efficient pattern retrieval mechanism; it can reduce the storage space of abnormal patterns by 65%, increase the retrieval speed by 3 times, and improve the abnormality detection efficiency of the system.

[0101] According to one aspect of the present application, step S41 may also be:

[0102] S41a, read the optimized association matrix C_opt and the historical anomaly database D_abn pre-stored by the system, segment the optimized association matrix C_opt according to the time window to obtain a time slice sequence, calculate the spatial feature distribution of each time slice to obtain a feature distribution matrix, and combine the time slice sequence and the feature distribution matrix to construct a third-order tensor T_init.

[0103] S41b, read the third-order tensor T_init, calculate the n-mode product of the tensor to obtain an unfolded matrix sequence, apply high-order singular value decomposition to the unfolded matrix sequence to obtain three factor matrices U_1, U_2, and U_3, and perform tensor product operations on the third-order tensor T_init and the three factor matrices to obtain the core tensor G.

[0104] S41c, read the core tensor G and factor matrices U_1, U_2, U_3, construct a tensor contraction algorithm based on the nuclear norm, project the core tensor G in the space spanned by the three factor matrices to obtain the projection tensor, perform rank minimization on the projection tensor to obtain the compressed core tensor G_c and the updated factor matrices V_1, V_2, V_3.

[0105] S41d, read the compressed core tensor G_c and factor matrices V_1, V_2, V_3, build a hierarchical index tree, map the elements of the compressed core tensor G_c to the nodes of the index tree, store the feature vectors corresponding to the factor matrices V_1, V_2, V_3 in the corresponding nodes, organize the index tree and its node data into a structured form to obtain the abnormal pattern library L_abn.

[0106] In one embodiment of the present application, a tensor contraction algorithm based on the nuclear norm: tensor compression optimization objective L(G, V) = ||T - G×_1V_1×_2V_2×_3V_3|| 2 F + λ·||G||* + μ·Σ||V_i|| 2 F ; Where T is the original tensor, G is the compressed core tensor, and V_i is the factor matrix; ×_n represents the n-mode product operation; ||G|| is the nuclear norm, defined as ||G|| = Σσ_i(G), σ_i(G) is the singular value of G; λ is the nuclear norm regularization parameter; μ is the factor matrix regularization parameter; iterative update rule: G_(t+1) = S_λ(T×_1V_1 T ×_2V_2 T ×_3V_3 T ), where S_λ is the soft threshold operator, S_λ(x) = sign(x)·max(|x|-λ, 0).

[0107] This embodiment can more comprehensively capture the time, space and feature dimension information of the abnormal pattern by constructing a third-order tensor structure and performing high-order singular value decomposition, thereby improving the accuracy of anomaly detection. Through the tensor contraction algorithm of the nuclear norm and the rank minimization process, the feature vector can be compressed and optimized to obtain a compact representation, which not only reduces data redundancy but also improves the compactness of the feature representation, allowing the system to process and analyze data more efficiently. By constructing a hierarchical index tree, the abnormal pattern can be hierarchically represented and organized to form a structured abnormal pattern library, which helps the system to perform pattern matching and anomaly detection more quickly.

[0108] According to one aspect of the present application, step S42 is further:

[0109] S421. Read the fused feature matrix F_fused and the abnormal pattern library L_abn, construct a third-order tensor mapping function, map the fused feature matrix F_fused to the high-dimensional feature space to obtain the feature tensor, perform Tucker norm normalization on the feature tensor to obtain the normalized feature tensor T_norm and the norm factor matrix N_fac.

[0110] S422. Read the normalized feature tensor T_norm and the abnormal pattern library L_abn, construct an improved tensor inner product operator, calculate the high-order inner product between the feature tensor and each pattern in the pattern library, apply a nonlinear activation function to the inner product result to obtain an activation value, and organize the activation value into a three-dimensional array to obtain a similarity tensor T_sim.

[0111] S423, read the similarity tensor T_sim and the norm factor matrix N_fac, construct the alternating least squares algorithm for CP decomposition, iteratively calculate to obtain a decomposition result with a rank of R, apply regularization processing to the decomposition result to obtain the factor matrix group F_set and the core weight vector W_core.

[0112] S424, read the factor matrix group F_set and the core weight vector W_core, construct a matching degree calculation function based on the factor product, calculate the comprehensive matching score of each pattern, perform threshold screening on the matching score to obtain a valid match, organize the valid matching results into a matrix form to obtain the abnormal matching result matrix M_abn.

[0113] In one embodiment of the present application, the tensor inner product operator calculation method is improved: tensor inner product <X, Y> _w = Σ(w_i·x_i·y_i) / (||X||_w·||Y||_w); where w_i is an adaptive weight, w_i = exp(-|x_i - y_i| / σ); ||X||_w = sqrt(Σ(w_i·x_i 2)) is the weighted norm; σ is the smoothing parameter, σ = median(|x_i -y_i|); x_i, y_i are the elements of tensors X and Y respectively, and ||Y||_w is another weighted norm.

[0114] This embodiment realizes the rapid identification of abnormal behavior in parking lots through the abnormal matching mechanism of tensor inner product and CP decomposition. Specifically, the feature data is first mapped to a high-dimensional feature space and normalized to establish a unified feature metric; then the improved tensor inner product operator is used to calculate the similarity of the behavior pattern and a multi-dimensional matching evaluation system is constructed; finally, accurate abnormal behavior identification results are obtained through CP decomposition and threshold screening. This embodiment can not only process high-dimensional feature data, but also capture complex behavior pattern associations; it can increase the abnormal behavior identification speed by 2.5 times, with an accuracy rate of 94%, improving the real-time monitoring capability of the system.

[0115] like Figure 6 As shown, according to one aspect of the present application, step S5 is further:

[0116] S51. Based on the abnormal event report and the pre-stored system state matrix, construct the Riemann metric tensor of the state space, calculate the geodesic equations of the state space; solve the geodesic equations to obtain the state transition path; based on the state transition path, construct the nonlinear dynamic programming equation; combine the nonlinear dynamic programming equation with the preset constraint conditions to form a decision model;

[0117] S52. Based on the decision model, construct an objective function family based on Pareto optimality; apply the improved gradient projection method to the objective function family to obtain an initial solution set; calculate the dominance relationship of each solution in the initial solution set to obtain a non-dominated solution set; perform local search on the non-dominated solution set to obtain the optimal solution; expand the optimal solution into a matrix form to obtain an execution strategy matrix;

[0118] S53. Based on the execution strategy matrix, construct a strategy parsing tree; perform depth-first traversal on the strategy parsing tree to obtain an initial instruction sequence; calculate the timing dependency of the initial instruction sequence; based on the timing dependency, perform topological sorting on the initial instruction sequence to obtain a sorted instruction sequence; normalize the sorted instruction sequence to obtain an execution instruction sequence.

[0119] In one embodiment of the present application, based on the abnormal event report E_rep and the system state matrix S_sys, a multi-objective optimization model is constructed, and the objective functions include safety, efficiency and resource utilization, and a decision model D_mod is obtained; based on the decision model D_mod, an improved particle swarm optimization algorithm is used to solve the optimal decision strategy, and an execution strategy matrix A_str is obtained; based on the execution strategy matrix A_str, the strategy is converted into a specific execution instruction sequence, and the execution instruction sequence C_seq is output.

[0120] In another embodiment of the present application, a method for solving a group of geodesic equations: The geodesic equation d 2 x μ / dt 2 +Γ μ _νλ(dx ν / dt)(dx λ / dt) = 0; where d 2 x μ / dt 2 is the geodesic coordinate x μ The second derivative with respect to parameter t, Γ μ _νλ =(1 / 2)g μ σ(Ξ_νg_λσ + Ξ_λg_νσ - Ξ_σg_νλ) is the Christoffel symbol; g μ ν is the metric tensor; Ξ_ν, Ξ_λ, and Ξ_σ are the partial derivatives of the ν-th, λ-th, and σ-th coordinates, respectively; x ν is the vth component of the geodesic coordinate; x λ is the λth component of the geodesic coordinates; x μ are geodesic coordinates; t is a parameter; the improved Runge-Kutta method is used for solution, with a step size of h_n = h_0·min(1, sqrt(ε / ||R||)), R is the curvature tensor, h_0 is the initial step size, and ε is the error tolerance.

[0121] Improved gradient projection method: Iterative format x_(k+1) = P_Ω(x_k - α_k·d_k); where d_k = ▽f(x_k) + β_k·v_k is the search direction; v_k is the momentum term, v_k = x_k - x_(k-1); β_k is the adaptive momentum coefficient, β_k = ||▽f(x_k)|| 2 / ||▽f(x_(k-1))|| 2; α_k is the adaptive step size, α_k = α_0 / (1 + μ·||▽f(x_k)||); P_Ω is the projection operator, which projects the solution onto the feasible domain Ω, x_(k+1) is the solution of the k+1th iteration, α_0 is the initial step size, ▽f(x_k) is the gradient of the objective function f at the kth iteration point x_k, and μ is the step size adjustment parameter.

[0122] This embodiment realizes intelligent decision-making and precise control of abnormal events through Riemann geometry and nonlinear dynamic programming. Specifically, first, a set of geodesic equations is constructed in the state space to solve the optimal state transfer path; then, a multi-objective optimization function is designed based on the Pareto optimal principle, and the execution strategy is solved by the improved gradient projection method; finally, a standardized instruction sequence is generated using the strategy parsing tree. This embodiment can take into account multiple performance indicators while ensuring control accuracy, providing an optimal solution for abnormal event processing; it can shorten the decision response time by 40%, increase the control accuracy to more than 95%, and improve the emergency handling capability of the system.

[0123] According to one aspect of the present application, step S51 is further:

[0124] S511. Based on the abnormal event report and the pre-stored system state matrix, calculate the local coordinate basis of the state space; construct a covariant derivative operator based on the local coordinate basis, apply the covariant derivative operator to the state variable, and obtain a local tangent space; construct a Riemann metric tensor on the local tangent space;

[0125] S512. Based on the local tangent space and the Riemann metric tensor, a Christoffel symbol field is constructed; the Christoffel symbol field is substituted into the Euler-Lagrange equation to obtain the geodesic equations; the improved symplectic geometry algorithm is used to solve the geodesic equations to obtain the state transition path and the geodesic curvature tensor;

[0126] S513, based on the state transition path and the geodesic curvature tensor, construct a curvature-based cost function, calculate the covariant derivative sequence of the path; apply the variational method to the covariant derivative sequence to obtain the optimal variation; combine the optimal variation with the state transition path to obtain the optimized path;

[0127] S514. Based on the optimization path, construct the Hamiltonian function, apply the Pontryagin maximum principle to obtain the standard dynamic programming equation; combine the standard dynamic programming equation with the pre-stored system constraints, and obtain the primal-dual problem pair through dual decomposition; encapsulate the primal-dual problem pair into a decision model.

[0128] In one embodiment of the present application, an improved symplectic geometry algorithm: a symplectic numerical solution of the geodesic equation dz / dt =J·▽H(z); wherein z = (q, p) is the phase space coordinate, H(z) is the Hamiltonian function, and J is the symplectic structure matrix J =[0, I; -I, 0]; the discrete format adopts the explicit midpoint method, z_(n+1) = z_n + h·J·▽H(z_(n+1 / 2)), z_(n+1 / 2) = (z_n + z_(n+1)) / 2; the adaptive step size h_n = h_0·min(1, sqrt(ε / ||▽H(z_n)||)), h_0 is the initial step size; the local error estimate E_n = ||z_(n+1) - φ_h(z_n)||, wherein φ_h is the exact solution operator; the energy conservation constraint |H(z_(n+1)) - H(z_n)|≤τ, τ is the tolerance parameter.

[0129] This embodiment achieves precise control of abnormal events in parking lots by introducing the decision modeling framework of Riemann geometry and variational method. Specifically, the Riemann metric tensor is first constructed in the local coordinate system to accurately describe the geometric characteristics of the state space; then the Christoffel symbol and geodesic equation are used to solve the optimal state transfer path; finally, based on the Hamiltonian function and Pontryagin maximum principle, a standard dynamic programming equation is established. This embodiment can find the optimal control strategy in the nonlinear state space, providing a theoretically complete solution for abnormal event processing; it can improve the decision optimization efficiency by 50%, the control accuracy by 97%, and improve the emergency handling capability of the system.

[0130] According to one aspect of the present application, step S52 is further:

[0131] S521. Read the decision model D_mod, construct a multi-objective function mapper, calculate the correlation matrix between the objective functions, apply eigenvalue decomposition to the correlation matrix to obtain the orthogonal objective space, construct a Pareto dominance relationship evaluation function in the orthogonal objective space, and obtain the objective function set F_obj and the correlation matrix C_cor.

[0132] S522, read the objective function set F_obj and the correlation matrix C_cor, construct an improved stochastic gradient projection operator, generate an initial candidate solution group, apply gradient descent with an adaptive step size to the candidate solutions to obtain a converged solution, and evaluate the stability of the solution through cross-validation to obtain the initial solution set S_ini and the gradient matrix G_mat.

[0133] S523, read the initial solution set S_ini and the gradient matrix G_mat, build a dominance relationship calculator, calculate the dominance relationship between any two solutions in the solution set, establish a dominance graph structure, obtain the Pareto frontier through the transitive closure analysis of the graph, extract the solutions on the Pareto frontier to obtain the non-dominated solution set S_par and the dominance relationship matrix D_rel.

[0134] S524, read the non-dominated solution set S_par and the dominating relationship matrix D_rel, construct an adaptive neighborhood search operator, calculate the local sensitivity of each non-dominated solution, design the search step size based on the sensitivity, perform variable neighborhood search to obtain the local optimal solution, and organize the local optimal solution into a matrix form to obtain the execution strategy matrix A_str.

[0135] This embodiment solves the multi-constraint decision-making problem of parking lot abnormal events through a multi-objective Pareto optimization and local search strategy generation mechanism. Specifically, first, a multi-objective function mapper is constructed and the correlation between objectives is analyzed to establish an orthogonal objective evaluation space; then, the improved stochastic gradient projection method is used to solve the initial solution set, achieving efficient exploration of the strategy space; then, the non-dominated solution set is determined through dominance relationship analysis to ensure the balance of the decision-making scheme; finally, adaptive neighborhood search is used for local optimization to obtain the optimal execution strategy. This embodiment can not only balance multiple performance indicators, but also quickly converge to the optimal solution; it can increase the decision generation speed by 55%, and the strategy execution success rate reaches 96%, improving the emergency handling effect of the system.

[0136] According to one aspect of the present application, step S13 may also be:

[0137] S13a, read the standardized data set D_norm, build a multi-level data type identifier, and divide the data into video data stream, lidar data stream, ultrasonic data stream and RFID data stream according to the sampling frequency, signal-to-noise ratio and numerical distribution characteristics of the data; calculate the characteristic statistics of each type of data stream to obtain the sensor feature matrix M_sen; perform principal component analysis on the sensor feature matrix M_sen, extract the main feature vector, and obtain the sensor parameter vector P_sen.

[0138] S13b, read the standardized data set D_norm and the sensor parameter vector P_sen, build an adaptive wavelet basis function library, and select the most matching wavelet basis function for each type of data stream; calculate the convolution of the data and the selected wavelet basis to obtain the wavelet coefficient matrix W_coef; apply maximum likelihood estimation and Bayesian estimation to the wavelet coefficient matrix W_coef to obtain the likelihood threshold vector T_mle and the Bayesian threshold vector T_bay; perform weighted fusion on the likelihood threshold vector T_mle and the Bayesian threshold vector T_bay to obtain the fused threshold vector T_fus.

[0139] S13c, read the wavelet coefficient matrix W_coef and the fusion threshold vector T_fus, perform layered wavelet threshold processing, and obtain the noise reduction coefficient matrix W_den; perform inverse wavelet transform on the noise reduction coefficient matrix W_den to obtain the denoised data set D_clean.

[0140] This embodiment achieves accurate noise reduction of multi-source heterogeneous data in parking lots by constructing a multi-level sensor data classification and adaptive noise reduction framework. Specifically, different sensor data are first classified through a multi-level data type identifier, which solves the problem that traditional methods cannot distinguish and process different types of data; then, the adaptive wavelet basis function library is used to select the most matching wavelet basis function for each type of data, and the optimal threshold parameters are determined by combining the maximum likelihood estimation and Bayesian estimation dual verification mechanism, which overcomes the lack of robustness of a single threshold strategy; finally, through layered wavelet threshold processing and energy compensation strategy, efficient noise reduction is achieved while ensuring signal integrity. This embodiment can improve the signal-to-noise ratio by 15dB, and the noise reduction accuracy of different types of data reaches more than 95%, which improves the data quality of subsequent feature extraction.

[0141] According to one aspect of the present application, step S23 may also be:

[0142] S23a, read the initial feature matrix F_init and the dynamic weight matrix W_d, build a residual protection channel, copy the initial feature matrix F_init and save it as the residual feature matrix R_feat; perform singular value decomposition on the initial feature matrix F_init to obtain the feature space mapping matrix P_map.

[0143] S23b, read the feature space mapping matrix P_map and the dynamic weight matrix W_d, construct the generalized Hadamard matrix, calculate its Moore-Penrose generalized inverse, and obtain the projection operator Q; perform Kronecker product operation on the projection operator Q and the dynamic weight matrix W_d to obtain the weight projection matrix W_proj.

[0144] S23c, read the initial feature matrix F_init, the weight projection matrix W_proj and the residual feature matrix R_feat, calculate the tensor product of the initial feature matrix F_init and the weight projection matrix W_proj, and obtain the weighted feature matrix F_weight; perform feature cascade on the weighted feature matrix F_weight and the residual feature matrix R_feat to obtain the fused feature matrix F_fused.

[0145] This embodiment solves the problem of information loss in the feature fusion process by designing a residual-protected Hadamard feature fusion mechanism. Specifically, the residual protection channel is first used to preserve the original feature information, avoiding feature attenuation caused by the traditional Hadamard product operation; then, the feature projection space is constructed through the generalized Hadamard matrix and the Moore-Penrose generalized inverse to achieve the orthogonal representation of the features; finally, a dynamic weight adaptive strategy is used for feature cascading, which improves the compactness of the feature representation while maintaining the feature discriminability. This embodiment can increase the feature retention rate by 40% and the feature fusion efficiency by 50%, providing a more reliable feature basis for abnormal behavior detection.

[0146] According to one aspect of the present application, step S41 may also be:

[0147] S41a, read the optimized association matrix C_opt and the historical anomaly database D_abn pre-stored in the system, build the PARAFAC2 decomposition model, calculate the time pattern matrix, the spatial pattern matrix and the characteristic pattern matrix, and obtain the decomposition factor group F_group; perform alternating optimization on the decomposition factor group F_group to obtain the optimized factor group F_opt.

[0148] S41b, read the optimization factor group F_opt, construct the factor association network, calculate the coupling strength between the factors, and obtain the coupling strength matrix C_str; construct a pattern reconstruction function based on the coupling strength matrix C_str, reconstruct the abnormal pattern representation, and obtain the reconstructed pattern tensor T_rec.

[0149] S41c, read the reconstruction pattern tensor T_rec, build a hierarchical index structure, map the tensor elements to the index nodes, and obtain the index tree T_idx; organize the abnormal pattern data based on the index tree T_idx to form the abnormal pattern library L_abn.

[0150] This embodiment achieves efficient representation and organization of parking lot abnormal patterns through PARAFAC2 tensor decomposition and nuclear norm constraint compression strategy. Specifically, PARAFAC2 decomposition is first used to replace traditional Tucker decomposition, breaking through the limitation of modal orthogonality assumption and more accurately capturing the complex associations in vehicle behavior data; then the tensor contraction algorithm with nuclear norm constraint is used to compress the pattern, reducing storage overhead while maintaining the main information; finally, the abnormal patterns are organized through a hierarchical index structure, providing an efficient pattern retrieval mechanism. This embodiment can reduce the storage space of abnormal patterns by 70%, increase the retrieval speed by 3.5 times, and maintain a pattern recognition accuracy of more than 95%.

[0151] According to one aspect of the present application, a vehicle dynamic monitoring system for a smart parking building based on an AI intelligent engine includes:

[0152] at least one processor; and,

[0153] a memory communicatively connected to at least one of the processors; wherein,

[0154] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the method for dynamic vehicle supervision in an intelligent parking building based on an AI intelligent engine as described in any of the above embodiments.

[0155] The present invention proposes a systematic solution to the problem of multi-source data fusion accuracy, specifically: first, accurate time synchronization is achieved by constructing a timestamp alignment matrix to solve the data drift problem; then, the normalization parameters are determined based on the dynamic change range, and the Min-Max normalization method is used to unify the data of different sensors to the [0, 1] interval, effectively eliminating the dimension difference and improving the fusion accuracy of multi-source data. For the Hadamard feature attenuation problem, the initial feature matrix is ​​copied to the residual feature matrix by constructing a residual protection channel; then, feature projection is achieved using the generalized Hadamard matrix and the Moore-Penrose generalized inverse; finally, the weighted feature matrix and the residual feature matrix are feature concatenated to effectively avoid information loss. This method not only maintains the advantages of Hadamard operation, but also solves the problem of feature attenuation. For the limitation of tensor decomposition, the PARAFAC2 decomposition model is used to replace the traditional Tucker decomposition, specifically: by calculating the time mode matrix, the space mode matrix and the feature mode matrix, a factor association network is established to achieve accurate modeling of non-orthogonal modes, breaking through the dependence of traditional tensor decomposition on orthogonality, and more accurately expressing complex vehicle behavior patterns. In order to solve the problem of robustness of noise reduction strategy, a multi-level data type identifier is constructed to realize the classification of different types of sensor data; then, a more robust threshold parameter is generated by combining maximum likelihood estimation and Bayesian estimation for double verification; finally, through hierarchical wavelet threshold processing, differentiated noise reduction of different types of data is realized, and the robustness of noise reduction is improved. In order to solve the problem of decision optimization efficiency, a decision framework based on Riemann geometry and variational method is introduced, specifically: by constructing Christoffel symbol field and geodesic equations, the geometric characteristics of nonlinear state space are accurately described; then, based on Pontryagin maximum principle, a standard dynamic programming equation is constructed to realize multi-objective optimization in high-dimensional state space, effectively improving decision efficiency and control accuracy in complex scenarios.

[0156] The present invention is based on the system architecture design of multi-sensor fusion and AI intelligent engine, which realizes the all-round and intelligent dynamic supervision of vehicles in smart parking buildings. By establishing a complete technical chain from data acquisition, preprocessing, feature extraction to anomaly detection and decision execution, the problems of data islands, insufficient feature expression, and abnormal recognition delays in traditional parking building supervision are solved. The system adopts multi-source heterogeneous data fusion technology and wavelet noise reduction preprocessing to improve data quality; the dynamic weight adaptive mechanism and spatiotemporal correlation analysis are used to achieve accurate modeling of vehicle behavior; based on tensor decomposition and multi-objective optimization methods, an efficient anomaly detection and intelligent decision-making framework is constructed. After actual deployment verification, the present invention can improve the efficiency of parking lot supervision by 60%, reduce the response time of abnormal events to milliseconds, control the false alarm rate below 5%, and the overall performance of the system has reached the industry-leading level.

[0157] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A method for dynamic monitoring of vehicles in an intelligent parking building based on an AI intelligent engine, characterized in that: The steps include: S1. Obtain an original data set, perform time alignment on the original data set to obtain a time-aligned data set; perform standardization on the time-aligned data set to obtain a standardized data set; Perform wavelet transform denoising on the standardized data set to obtain a denoised data set; Extract features from the denoised data set to obtain an initial feature matrix; The original data sets include video data, lidar point cloud data, ultrasonic data, and RFID data; S2. Based on the pre-stored historical reliability index matrix, the reliability of the initial feature matrix is ​​evaluated to obtain a reliability score vector; based on the reliability score vector, a dynamic weight matrix is ​​generated; the initial feature matrix is ​​fused with the dynamic weight matrix to obtain a fused feature matrix; S3. Construct a vehicle trajectory model based on the fusion feature matrix; Extract key point sequences from the vehicle trajectory model, perform spatiotemporal correlation analysis on the key point sequences and the pre-stored historical correlation pattern library to obtain an optimized correlation matrix; S4. Construct an abnormal pattern library based on the optimized association matrix and the pre-stored historical abnormal database; Perform pattern matching on the fusion feature matrix and the abnormal pattern library to obtain an abnormal matching result matrix; perform analysis and calculation on the abnormal matching result matrix to obtain an abnormal event report; S5. Building a decision model based on the abnormal event report and the pre-stored system status matrix; Based on the decision model, an execution strategy matrix is ​​calculated; the execution strategy matrix is ​​converted to obtain an execution instruction sequence.

2. The method for dynamic monitoring of vehicles in an intelligent parking building based on an AI intelligent engine according to claim 1 is characterized in that: Step S1 is further as follows: S11, acquiring video data through video sensors arranged in the garage, acquiring lidar point cloud data through lidar sensors, acquiring ultrasonic data through ultrasonic sensors, and acquiring RFID data through RFID readers; combining the video data, lidar point cloud data, ultrasonic data, and RFID data to form an original data set; constructing a timestamp alignment matrix according to the acquisition time of each data in the original data set; Perform time synchronization processing on the original data set according to the timestamp alignment matrix to obtain a time-aligned data set; S12, reading the data in the time-aligned data set, calculating the dynamic change range of each type of data; and determining the normalization parameter based on the dynamic change range; Based on the normalization parameters, the Min-Max normalization method is used to normalize the data in the time-aligned dataset to obtain a standardized dataset with a value range of [0, 1]; S13. Calculate the statistical characteristics of the data based on the standardized data set; and determine the threshold parameters of the wavelet transform by using the maximum likelihood estimation method based on the statistical characteristics; Using the determined threshold parameter to perform wavelet transform denoising on the data in the standardized data set, to obtain a denoised data set; S14. Extract the spatial position features, time series features and vehicle status features of the data in the denoised data set; combine the spatial position features, time series features and vehicle status features into a first feature vector, arrange the first feature vector into a matrix form, and obtain an initial feature matrix.

3. The method for dynamic monitoring of vehicles in an intelligent parking building based on an AI intelligent engine according to claim 2 is characterized in that: Step S2 is further as follows: S21, calculating the Hadamard product of each data source in the initial feature matrix to obtain a feature correlation matrix; Recursively operate the feature correlation matrix and the pre-stored historical reliability index matrix to obtain the real-time information entropy value; calculate the reliability score of each data source based on the real-time information entropy value; combine the reliability scores of each data source to form a reliability score vector; S22, based on the reliability score vector, construct a feature mapping space based on the Hadamard matrix; Calculate the covariance matrix between each feature in the feature mapping space; Based on the covariance matrix, construct the discriminant criterion function; By maximizing the discriminant criterion function, the optimal weight distribution is solved to obtain the optimal weight value; the optimal weight value is organized into a matrix form to obtain a dynamic weight matrix; S23, projecting the initial feature matrix into the Hadamard feature space defined by the dynamic weight matrix; Matrix multiplication is performed in the Hadamard feature space to obtain a weighted feature matrix; the weighted feature matrix is ​​orthogonalized to obtain a fused feature matrix.

4. The method for dynamic monitoring of vehicles in an intelligent parking building based on an AI intelligent engine according to claim 3 is characterized in that: Step S3 is further as follows: S31. Based on the fused feature matrix, construct a state transfer probability matrix of the feature vector sequence; apply the Laplace operator to the state transfer probability matrix for spectral decomposition to obtain an eigenvalue sequence; based on the eigenvalue sequence, construct a state transfer function; combine the state transfer function with the feature vector sequence to form a vehicle trajectory model; S32, calculating the local density value and relative distance value of each point in the vehicle trajectory model, and constructing a scoring function based on the density product; performing extreme value analysis on the scoring function to obtain key feature points; arranging the key feature points in time sequence to obtain a key point sequence; S33, constructing a spatiotemporal adjacency matrix based on the key point sequence and the pre-stored historical association pattern library; Apply the improved spectral clustering algorithm to the spatiotemporal adjacency matrix to obtain the initial clustering results; Calculate the Graph Laplacian matrix of the initial clustering result; perform eigendecomposition on the Graph Laplacian matrix to obtain the second eigenvector; construct a similarity matrix based on the second eigenvector; regularize the similarity matrix to obtain an optimized association matrix.

5. The method for dynamic monitoring of vehicles in an intelligent parking building based on an AI intelligent engine according to claim 4 is characterized in that: Step S4 is further as follows: S41. Based on the optimized association matrix and the pre-stored historical anomaly database, a third-order tensor structure is constructed to represent the time, space and feature dimension information of the anomaly pattern; Tucker decomposition is performed on the third-order tensor structure to obtain a core tensor and a factor matrix; based on the core tensor and the factor matrix, a hierarchical representation of the anomaly pattern is constructed; and the hierarchical representation result is stored as an anomaly pattern library; S42, converting the fused feature matrix into a tensor form to obtain a fused feature matrix tensor; Calculate the inner product of the fusion feature matrix tensor and the tensor of each pattern in the abnormal pattern library to obtain the similarity value; Based on the similarity values, a similarity tensor is constructed; Perform CP decomposition on the similarity tensor to obtain the matching degree of each mode; Organize the matching degree into a matrix form to obtain the abnormal matching result matrix; S43, performing singular value decomposition on the abnormal matching result matrix to obtain a feature space representation; Calculate the category probability distribution of abnormal events in the feature space representation; calculate the abnormality level and risk score based on the category probability distribution; analyze the abnormal matching result matrix to obtain detailed matching information; organize the abnormality level, risk score and detailed matching information into a structured form to obtain an abnormal event report.

6. The method for dynamic monitoring of vehicles in an intelligent parking building based on an AI intelligent engine according to claim 5 is characterized in that: Step S5 is further as follows: S51, based on the abnormal event report and the pre-stored system state matrix, construct the Riemann metric tensor of the state space, and calculate the geodesic equations of the state space; Solve the geodesic equations to obtain the state transition path; Based on the state transfer path, a nonlinear dynamic programming equation is constructed; the nonlinear dynamic programming equation is combined with preset constraints to form a decision model; S52. Based on the decision model, construct an objective function family based on Pareto optimality; apply the improved gradient projection method to the objective function family to obtain an initial solution set; calculate the dominance relationship of each solution in the initial solution set to obtain a non-dominated solution set; perform local search on the non-dominated solution set to obtain the optimal solution; expand the optimal solution into a matrix form to obtain an execution strategy matrix; S53, constructing a strategy parsing tree based on the execution strategy matrix; Perform depth-first traversal on the policy parsing tree to obtain the initial instruction sequence; Calculate the timing dependencies of the initial instruction sequence; Based on the timing dependency, the initial instruction sequence is topologically sorted to obtain a sorted instruction sequence; the sorted instruction sequence is normalized to obtain an execution instruction sequence.

7. The method for dynamic monitoring of vehicles in an intelligent parking building based on an AI intelligent engine according to claim 6 is characterized in that: Step S23 is further as follows: S231, calculating the singular value decomposition of the initial feature matrix to obtain a left singular matrix and a right singular matrix; constructing a bilateral Hadamard product based on the left singular matrix and the right singular matrix to obtain a feature space mapping matrix; S232, constructing an augmented Hadamard matrix based on the feature space mapping matrix and the dynamic weight matrix; calculating the Moore-Penrose generalized inverse of the augmented Hadamard matrix to obtain a projection operator; performing a Kronecker product operation on the projection operator and the dynamic weight matrix to obtain a weight projection matrix; S233, calculating the tensor product of the initial feature matrix and the weight projection matrix to obtain a third-order tensor; slicing the third-order tensor along the third dimension to obtain a slice matrix sequence; performing weighted summation on the slice matrix sequence to obtain a weighted feature matrix; S234. Based on the weighted feature matrix, a Gram-Schmidt orthogonal basis vector group is constructed; the weighted feature matrix is ​​expanded under the Gram-Schmidt orthogonal basis vector group to obtain a coefficient matrix; QR decomposition is applied to the coefficient matrix to obtain an orthogonal matrix; matrix multiplication operation is performed on the orthogonal matrix and the Gram-Schmidt orthogonal basis vector group to obtain a fused feature matrix.

8. The method for dynamic monitoring of vehicles in an intelligent parking building based on an AI intelligent engine according to claim 6 is characterized in that: Step S41 is further as follows: S411, based on the pre-stored historical anomaly database, segment the optimized association matrix according to the preset time window to obtain a time slice sequence; calculate the spatial feature distribution of each time slice sequence to obtain a feature distribution matrix; Combine the time slice sequence and the feature distribution matrix to construct an initial third-order tensor; S412, calculating the n-mode product of the initial third-order tensor to obtain an expanded matrix sequence; applying singular value decomposition to the expanded matrix sequence to obtain a main feature space; combining the eigenvectors and eigenvalues ​​in the main feature space to obtain a mode matrix; constructing a Tucker kernel in the main feature space, performing a tensor contraction operation on the Tucker kernel and the mode matrix to obtain a decomposed tensor; S413, based on the decomposed tensor, construct a similarity calculation function based on the Kruskal operator to calculate a first similarity matrix between tensor elements; Applying the spectral clustering algorithm to group the first similarity matrix to obtain pattern clusters; mapping the pattern clusters back to the feature space of the initial third-order tensor to obtain the pattern feature tensor; S414. Based on the pattern feature tensor, a hierarchical index tree is constructed, and the third feature vector of each node is calculated; sparse coding is performed on the third feature vector to obtain a compact representation; the compact representation and the hierarchical index tree are organized into a structured form to obtain an abnormal pattern library.

9. The method for dynamic monitoring of vehicles in an intelligent parking building based on an AI intelligent engine according to claim 6 is characterized in that: Step S51 is further as follows: S511. Based on the abnormal event report and the pre-stored system state matrix, calculate the local coordinate basis of the state space; construct a covariant derivative operator based on the local coordinate basis, apply the covariant derivative operator to the state variable, and obtain a local tangent space; construct a Riemann metric tensor on the local tangent space; S512. Based on the local tangent space and the Riemann metric tensor, a Christoffel symbol field is constructed; the Christoffel symbol field is substituted into the Euler-Lagrange equation to obtain the geodesic equations; the improved symplectic geometry algorithm is used to solve the geodesic equations to obtain the state transition path and the geodesic curvature tensor; S513, based on the state transition path and the geodesic curvature tensor, construct a curvature-based cost function, calculate the covariant derivative sequence of the path; apply the variational method to the covariant derivative sequence to obtain the optimal variation; Combine the optimal variation with the state transition path to obtain the optimized path; S514. Based on the optimization path, construct the Hamiltonian function, apply the Pontryagin maximum principle to obtain the standard dynamic programming equation; combine the standard dynamic programming equation with the pre-stored system constraints, and obtain the primal-dual problem pair through dual decomposition; encapsulate the primal-dual problem pair into a decision model.

10. The vehicle dynamic monitoring system of the intelligent parking building based on AI intelligent engine is characterized by: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the dynamic vehicle supervision method for an intelligent parking building based on an AI intelligent engine as described in any one of claims 1 to 9.

Images