Intelligent tracking and positioning system for container wharf hazardous chemical transport vehicle

Abstract

The invention discloses a container wharf hazardous chemical transport vehicle intelligent tracking and positioning system comprising a data acquisition module used for acquiring vehicle positioning, driving behaviors and environment state data; the data fusion and time synchronization module is used for performing time alignment on the multi-source data; the behavior modeling and trajectory prediction module is used for extracting driving behavior characteristics and predicting a driving trajectory; the risk assessment and anomaly detection module is used for calculating an anomaly score and predicting a driving risk; the security strategy optimization module is used for constructing a zero-sum game security control model and optimizing speed limit, path and early warning strategies; the remote control and intelligent intervention module is used for executing remote speed regulation, emergency braking, path optimization and alarm pushing in the intelligent dispatching system; and the cloud monitoring and data analysis module is used for storing a track, an environment, a driving score and a safety strategy and predicting a long-term risk trend. According to the invention, based on intelligent game and multi-modal fusion, accurate tracking and safety management and control of the hazardous chemical substance transport vehicle are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle safety monitoring, and particularly to an intelligent tracking and positioning system for dangerous chemical transportation vehicles in container terminals. Background Art

[0002] In modern logistics and traffic management, the safety and intelligent management of dangerous chemical transportation vehicles in container terminals have become important research directions. Due to the high-risk characteristics of dangerous chemical transportation, once an accident occurs, it may cause serious environmental pollution, property damage, and even casualties. Therefore, how to use advanced intelligent tracking and safety monitoring technologies to achieve high-precision positioning, real-time monitoring, and intelligent safety management of dangerous chemical transportation vehicles is an important research topic in the current industry. In the prior art, the tracking and management methods for dangerous chemical transportation vehicles mainly rely on traditional GNSS-RTK high-precision positioning and IMU inertial navigation, combined with sensors such as cameras and lidar for environmental perception, supplemented by manual intervention for risk management. However, these technologies still have many defects in practical applications and are difficult to meet the requirements of high safety, high real-time performance, and intelligent management.

[0003] The current tracking systems for dangerous chemical transportation vehicles usually rely on a single GNSS-RTK positioning technology. Although it has high positioning accuracy, it is prone to signal drift or loss in environments with severe signal occlusion (such as tunnels, densely packed container areas in docks, etc.). In addition, although the IMU inertial navigation system can provide short-term compensation in the event of GNSS failure, due to the large accumulation of errors in inertial navigation data over time, long-term reliance on inertial navigation data will lead to trajectory deviation and affect the positioning accuracy. To improve the positioning stability, some systems adopt a method of fusing GNSS-RTK and IMU, and optimize the data through Kalman filtering. However, traditional Kalman filtering methods are difficult to accurately compensate for non-linear errors in the face of complex environments and high-dynamic vehicles, resulting in a decrease in the accuracy of vehicle position data.

[0004] Existing environmental perception systems usually rely on cameras and lidar, and use deep learning methods such as object detection and image segmentation to identify the distribution of obstacles and road conditions in the dock area. However, due to the significant impact of light changes and bad weather (such as heavy fog, heavy rain, etc.) on camera data, the reliability of object detection decreases. And the point cloud data of lidar in complex environments is easily affected by noise interference, and the computational burden is heavy when performing large-scale environmental perception, making it difficult to meet the real-time requirements. In addition, traditional environmental monitoring methods mainly evaluate based on static data and lack the ability to predict dynamic changes in the environment, and cannot achieve early warning of potential risks during the transportation of dangerous chemicals.

[0005] In the aspect of driving behavior analysis and anomaly detection, existing technologies mainly adopt time series analysis methods based on gated recurrent units (GRUs) or long short-term memory networks (LSTMs) to classify the driving patterns of drivers. However, these methods only focus on the single historical trajectory data of vehicles, lack long-term modeling of drivers' behavior habits, and are difficult to identify subtle abnormal driving behaviors. In addition, most existing methods rely on fixed rules or statistical feature calculations to obtain anomaly behavior scores, lacking intelligent comparative learning of driving behavior patterns, resulting in low accuracy of anomaly behavior detection and being unable to meet the high requirements for driving safety in hazardous chemical transportation.

[0006] In the aspect of driving risk assessment and safety control strategies, traditional methods usually adopt fixed threshold judgment models, that is, fixed safety thresholds are set according to indicators such as vehicle speed, braking frequency, and number of hard accelerations. Once the set threshold is exceeded, a safety alarm is triggered. However, this method ignores the complexity of driving behaviors. For example, in the complex road conditions of a terminal, short-term braking and low-speed driving may be normal behaviors, while the same operations on a highway may mean danger. Existing rule- or threshold-based methods have poor adaptability in complex traffic environments and are prone to false alarms or missed alarms. In addition, some studies introduce Bayesian optimization or fuzzy logic reasoning to adjust the risk assessment model, but still fail to fully combine multi-modal data fusion and reinforcement learning methods, resulting in insufficient dynamic adjustment ability of driving risk assessment and being unable to adapt to the optimization of safety decisions in complex environments.

[0007] In the aspect of safety strategy optimization and game theory modeling, some studies use deep reinforcement learning methods to optimize driving behaviors, such as using Q-learning or deep reinforcement learning algorithms based on policy gradients (DQN, PPO, etc.) to optimize vehicle control strategies. However, most existing methods are for ordinary intelligent driving scenarios and do not fully consider the high risk and safety game requirements of hazardous chemical transportation vehicles. Since hazardous chemical transportation involves multiple factors such as drivers, supervision systems, and traffic environments, driving safety strategies should possess multi-agent collaborative decision-making capabilities. Traditional reinforcement learning models fail to introduce the zero-sum game idea and construct an effective game mechanism between drivers' behaviors and safety strategies, resulting in the lack of pertinence of safety strategies and being difficult to dynamically adjust speed limits, route optimization, and warning strategies in different environments. In addition, existing systems lack adaptive dynamic adjustment capabilities. Even if safety strategies are calculated, it is difficult to optimize and adjust them according to real-time road conditions, environmental monitoring data, and changes in drivers' behaviors.

[0008] Therefore, how to provide an intelligent tracking and positioning system for hazardous chemical transportation vehicles in container terminals is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0009] An object of the present invention is to provide an intelligent tracking and positioning system for hazardous chemical transportation vehicles in container terminals. The present invention makes full use of technologies such as spatio-temporal collaborative data fusion, temporal graph neural network, Bayesian dynamic autoregressive model, and zero-sum game reinforcement learning, and details algorithms for multi-modal data fusion, driving behavior modeling, dynamic risk assessment, and intelligent safety strategy optimization, with the advantages of high positioning accuracy, accurate driving behavior analysis, dynamically optimized risk prediction, and strong intelligent adaptability of safety control.

[0010] The intelligent tracking and positioning system for hazardous chemical transportation vehicles in container terminals according to an embodiment of the present invention includes:

[0011] A data acquisition module for collecting vehicle positioning, driving behavior, and environmental status data;

[0012] A data fusion and time synchronization module for time-aligning multi-source data;

[0013] A behavior modeling and trajectory prediction module for extracting driving behavior characteristics and predicting driving trajectories;

[0014] A risk assessment and anomaly detection module for calculating anomaly scores and predicting driving risks;

[0015] A safety strategy optimization module for constructing a zero-sum game safety control model and optimizing speed limit, path, and warning strategies;

[0016] A remote control and intelligent intervention module for remote speed regulation, emergency braking, and path optimization, and pushing alerts;

[0017] A cloud monitoring and data analysis module for storing trajectories, environments, driving scores, and safety strategies, and predicting long-term risk trends.

[0018] Optionally, the modules are implemented by the following methods:

[0019] S1. Collect GNSS-RTK positioning data, IMU inertial navigation data, camera image data, lidar scan data, and driving parameters, and perform time synchronization processing;

[0020] S2. Based on the collected GNSS-RTK positioning data and IMU inertial navigation data, use a non-linear motion prediction compensation method to correct the signal loss area and generate vehicle positioning data;

[0021] S3. Based on the camera image data and lidar scan data, use image segmentation technology to analyze terminal obstacles, driving routes, and environmental changes, and generate environmental safety assessment data;

[0022] S4. Use vehicle positioning data, environmental safety assessment data, and driving parameters to construct an implicit topological behavior analysis network, and adopt a temporal graph neural network to extract the driver's operation mode to generate driving behavior feature data;

[0023] S5. Based on the driving behavior feature data, vehicle positioning data, and environmental safety assessment data, use a multi-modal behavior fusion inference model to evaluate driving risks, combine dynamic Bayesian optimization to calculate anomaly scores, and generate driving risk assessment data;

[0024] S6. Based on the driving risk assessment data, vehicle positioning data, and environmental safety assessment data, construct a dynamic game safety strategy, and adopt a zero-sum game to adjust speed limits, routes, and warning strategies to generate safety control data;

[0025] S7. Based on the safety control data, perform real-time intervention, remotely control speed, braking, and path planning, and optimize transportation tasks through spatio-temporal prediction scheduling, and push alerts to the driver terminal and the supervision center.

[0026] Optionally, the S2 includes the following specific steps:

[0027] S21. Based on GNSS-RTK positioning data and IMU inertial navigation data, construct a multi-sensor spatio-temporal state fusion model. Let the state vector of the vehicle at time t be:

[0028] X t =[P t ,V t ,A t ,Θ t T ;

[0029] Among them, P t is the position vector, V t is the velocity vector, A t is the acceleration vector, and Θ t is the attitude angle vector;

[0030] S22. Use an adaptive non-linear filtering model to perform dynamic error correction on GNSS-RTK positioning data. Define the GNSS observation value as The measurement noise is e t , then the position update equation is:

[0031]

[0032] Among them, the covariance matrix Σ t is calculated by the Markov adaptive estimation method, and the update rule is:

[0033]

[0034] ​Among them, P t is the corrected vehicle position, e t is the GNSS measurement noise, α is the weight factor, Σ t-1 is the covariance matrix at the previous moment t-1, N is the size of the historical data window, and i is the index;

[0035] S23. Based on the IMU inertial navigation data, construct a multi-order dynamic prediction model. Let the angular velocity be ω t =(ω x, , ω y, , ω z,t ), the attitude angle be Θ t =(φ t , θ t , ψ t ), and the acceleration be A t =(a x, , a y, , a z, ). Then the state update equation is:

[0036]

[0037] Among them, Θ t+1 is the attitude angle of the vehicle at time t+1, P t+1 is the position of the vehicle at time t+1, V t+1 is the speed of the vehicle at time t+1, Θ t is the attitude angle at time t, ω t is the angular velocity at time t, V t is the speed vector at time t, P t is the position vector at time t, W t is the process noise, and Δt is the time step;

[0038] S24. Fuse the GNSS-RTK positioning data and the IMU inertial navigation data, and use adaptive dynamic Bayesian filtering for trajectory correction to obtain the state estimation result. The state transition model is:

[0039] X t =FX t-1 +GU t +W t ;

[0040] The observation equation is:

[0041] Y t =HX t +R t ;

[0042] Among them, X t is the vehicle state, F is the state transition matrix, Y tis the observed data vector, G is the control input matrix, and U t is the control input vector, H is the observation matrix, and R t is the observation noise, and X t-1 is the vehicle state at time t+1;

[0043] S25. Based on the state estimation results, combined with the Gaussian process regression-spatiotemporal attention network model, optimize the trajectory compensation, and define the trajectory compensation objective function:

[0044]

[0045] where is the optimized vehicle position, is the predicted vehicle position, T is the time window size, λ is the regularization coefficient, is the trajectory smoothing term, P is the vehicle trajectory position variable to be optimized, is to find the optimal trajectory compensation position

[0046] Optionally, the S3 includes the following specific steps:

[0047] S31. Based on the camera image data I t and the lidar point cloud data L t construct an environmental perception data set, and use multi-scale pyramid convolution transformation to extract features. The environmental feature matrix is expressed as:

[0048]

[0049] where D t is the environmental feature matrix, S is the number of different scale pyramid transformation layers used for the image data, Q is the number of different scale pyramid transformation layers used for the point cloud data, and W s is the convolution weight matrix of the image data at the s-th layer scale, and V q is the convolution weight matrix of the point cloud data at the q-th layer scale, and G s is the feature map of the image data at the s-th layer scale, and H q is the feature map of the point cloud data at the q-th layer scale;

[0050] S32. Perform image-point cloud joint correction on the environmental feature matrix. Let the camera internal parameter matrix be K, the external parameter matrix be R and T, then the projection transformation from the point cloud to the image is:

[0051] Z t = K(RL t + D);

[0052] where Z tis the target projection point in the image coordinates. The optimal projection parameter θ is solved based on the least - squares optimization, and the objective function is:

[0053]

[0054] where θ * is the optimal parameter, K is the internal parameter matrix of the camera, R is the rotation matrix, D is the translation matrix, T is the time step, is the projection point obtained after adjusting according to the optimized parameters, β is the regularization coefficient, is the gradient change of the trajectory, argmin θ is to solve the parameter θ that makes the objective function reach the minimum value;

[0055] S33. Based on the calibration result, an adaptive graph optimization is used to construct the three - dimensional obstacle distribution. Let the obstacle graph be represented as G(V, E), where the node V is the set of obstacle points and the edge E represents the adjacency relationship:

[0056] G * = argmin G ∑ (i,)∈E w ij ∥Z i - Z j ∥ 2 ;

[0057] where G * is the optimal obstacle graph, argmin G is to represent solving the optimal graph G that makes the objective function reach the minimum value * , Z i and Z j are two obstacle points in the three - dimensional point cloud data, w ij is the weight of the edge E in the graph;

[0058] S34. Based on the obstacle graph, an adaptive dynamic Bayesian filter is used to model the environmental state. Define the environmental state vector E t :

[0059]

[0060] E t+1 = FE t + W t ;

[0061] where Z t , V t , A t are the position information, velocity information, and acceleration information of the environmental obstacles respectively, W t is the process noise, E t+1 is the environmental state vector at time t + 1, and F is the environmental state transition matrix;

[0062] S35. Based on the environmental state, combine the Markov hidden variable prediction model to predict the environmental change trend, calculate the environmental safety risk score, and set the environmental safety assessment data as R t :

[0063]

[0064] Among them, N is the total number of environmental characteristic factors, β i is the weight parameter of the environmental characteristic factor, f i (E t ) is the environmental variable under the environmental state E t , η is the historical risk attenuation coefficient, R t-1 is the environmental safety assessment data at the previous moment, K t is the prediction error term, and i is the index.

[0065] Optionally, the S4 includes the following specific steps:

[0066] S41. Based on the vehicle positioning data, environmental safety assessment data, and driving parameters, construct a high-dimensional driving behavior state vector:

[0067] N t = [Z t , V t , A t , Θ t , C t T ;

[0068] Among them, Z t is the vehicle position, V t is the speed, A t is the acceleration, Θ t is the heading angle, C t is the environmental dynamic constraint factor, N t driving behavior input state vector;

[0069] S42. Adopt a multi-scale graph convolutional spatio-temporal network to model the driving behavior pattern. Let the behavior topology graph be represented as X(V, E, W), where V is the set of driving behavior nodes, E is the set of behavior time-series edges, and W is the edge weight matrix. The driving behavior state transition is described by the following update equation:

[0070]

[0071] Among them, H t is the driving behavior feature matrix at the current moment, σ is the non-linear activation function, W is the trainable parameter matrix, H t-1 is the driving behavior feature matrix at the previous moment t-1, Y is the number of layers of the graph convolutional network, β​y is the weighted coefficient of the graph convolutional layer at different scales, N y (H t ) is the driving behavior embedding feature calculated by the multi-scale graph convolutional network at the s-th layer, and y is the index;

[0072] S43. Based on the driving behavior characteristics, use contrast learning + Transformer attention mechanism for driving mode classification:

[0073]

[0074] Define the enhanced contrast loss function:

[0075]

[0076] Among them, is the driving behavior feature space, L is the contrast learning loss function, N is the number of training samples, ln is the logarithmic function, exp is the exponential function, is to measure and the calculation function of the similarity between is to measure and the calculation function of the similarity between is the i-th driving behavior feature vector, is the positive sample corresponding to the i-th feature, is the other driving mode features in the contrast sample set, j and t are indexes, γ is the regularization coefficient, T is the time step, is the driving behavior feature predicted at time t;

[0077] S44. Based on the driving mode classification result, use the Bayesian dynamic autoregressive model to calculate the driving anomaly score:

[0078]

[0079] Among them, B t is the driving anomaly score at the current moment, M is the total number of driving behavior feature factors, α k is the feature factor weight, g k (H t ) is the k-th driving behavior feature factor, δ is the risk adjustment parameter, B t-1 is the driving anomaly score at the previous moment t - 1, μ is the risk score threshold, K t is the prediction error term;

[0080] S45. Based on the driving anomaly score, use the temporal attention embedding model to generate driving behavior feature data. Let the driving behavior feature vector be:

[0081]

[0082] Among them, C t is the driving behavior characteristic data generated at the current moment t, O is the total number of environmental behavior influence factors, ω j is the weight of the environmental behavior factor, A j is the j-th environmental behavior influence factor, tanh is used to suppress the influence of extreme abnormal scores, π is used to control the influence degree of abnormal scores on driving behavior characteristics, ∈ t is the noise term, and j and k are indices.

[0083] Optionally, the S5 includes the following specific steps:

[0084] S51. Based on the driving behavior characteristic data, vehicle positioning data, and environmental safety assessment data, construct a driving risk assessment state vector, and set the state vector as:

[0085] Y t =[C t ,Z t ,R t T ;

[0086] Among them, C t is the generated driving behavior characteristic data, Z t is the vehicle positioning data, and R t is the generated environmental safety assessment data;

[0087] S52. Use a multi-modal behavior fusion inference model to calculate the driving risk score, and define the driving risk score function:

[0088]

[0089] Among them, D t is the driving risk score at the current moment t, M is the total number of driving behavior characteristic factors, α k is the weight of the characteristic factor, h k (Y t ) is the k-th driving behavior characteristic factor, β j is the weight of the environmental influence factor, f j (R t ) is the j-th environmental influence factor, ξ t is the prediction error term, U is the total number of environmental influence factors, and j and k are indices;

[0090] S53. Based on the driving risk score, construct a time series anomaly detection model, and define the risk state transition equation:

[0091] ​

[0092] Among them, K t is the prediction error term, D t+1 is the driving risk score at time t + 1, D t is the driving risk score at time t, ρ is the risk attenuation coefficient, T is the time window size, λ i is the weight coefficient of the i-th historical risk score within the time window, D t-i is the driving risk score within the past i time steps, and i is the index;

[0093] S54. Based on the risk state prediction result, adopt reinforcement learning to dynamically optimize the driving behavior risk assessment:

[0094]

[0095] Among them, X t+1 is the driving state at time t + 1, X t is the driving state at time t, A t is the decision-making action of the driver at time t, is the reward factor, π t is the driving behavior policy function, Q(X t , A t ) is the state-action value function of reinforcement learning, argmax x is to find the optimal driving state that maximizes the objective function, and x is the set of all possible driving states;

[0096] S55. Based on the optimized result of the driving behavior risk assessment, combine the multi-scale adaptive threshold algorithm to generate driving risk assessment data:

[0097]

[0098] Among them, is the driving risk assessment data, σ is the normalization function, M is the total number of driving risk characteristic factors, is the k-th driving risk score data, θ t is the dynamically adjusted risk threshold, ρ t is the prediction error term.

[0099] Optionally, the S6 includes the following specific steps:

[0100] S61. Based on the driving risk assessment data vehicle positioning data Z t and environmental safety assessment data R t , construct a dynamic game security policy state vector:

[0101]

[0102] S62. Construct a security control game model. Assume that the game participants include the driver P1 and the security control system P2, and define the game revenue function:

[0103]

[0104] Among them, U(G t , A t ) is the game revenue function of adopting the security control strategy A t under the current state G t . M is the total number of driving state characteristic factors, α k is the weight of the driving state characteristic factor, g k (G t ) is the k-th driving state characteristic factor, Z is the total number of security control strategy cost items, β j is the cost weight of the security control strategy, c j (A t ) is the j-th security control strategy cost item, and j and k are indices;

[0105] S63. Based on the game revenue function, use the zero-sum game reinforcement learning algorithm to calculate the optimal security strategy:

[0106]

[0107] Among them, is the optimal security strategy, E is the expected calculation, find the optimal security control strategy that maximizes the objective function E[U(G t , A t )]

[0108] S64. Based on the optimal security strategy, construct a multi-agent collaborative optimization framework. Assume that the agent set is P, then the joint optimization objective function is:

[0109] J∑ i∈P ω i U i (G t , A t );

[0110] Among them, J is the joint optimization objective function, ω i is the weight factor of agent i, U i (G t , A t ) is the revenue function of agent i after taking action A t under the state G t ;

[0111] S65. Based on the optimization results, dynamically adjust the safety control strategy in combination with the reinforcement learning model to generate safety policy control data:

[0112]

[0113] Among them, Q t is the generated safety control data, is the normalization function, is the k-th safety control strategy, ε is the control strategy update coefficient, A t-1 is the safety control strategy at the previous moment, σ t is the policy adjustment noise term.

[0114] The beneficial effects of the present invention are:

[0115] Through multi-modal data fusion and intelligent game optimization, the present invention realizes the intelligent tracking and safety control of dangerous chemical transportation vehicles in container terminals. Compared with the prior art, the present invention has made significant improvements in multiple key aspects. First, in terms of vehicle positioning, the spatio-temporal collaborative data fusion method of GNSS-RTK and IMU inertial navigation is adopted, effectively reducing the errors caused by signal loss or drift and improving the positioning accuracy in complex environments. Aiming at the problem that traditional Kalman filtering cannot adapt to non-linear errors, the present invention introduces non-linear motion prediction compensation and combines it with a multi-scale trajectory optimization model to achieve high-precision trajectory correction, enabling the vehicle to maintain stable positioning information even when the signal is blocked.

[0116] In terms of environmental perception, the present invention adopts the image-point cloud joint analysis technology and combines dynamic environmental modeling to accurately identify obstacles, road boundaries and dynamic traffic conditions in complex terminal scenarios. At the same time, the Markov hidden variable prediction model is introduced, which not only evaluates the current environment but also predicts the future environmental change trend, thus providing a more reliable safety warning ability during the transportation of dangerous chemicals. Compared with traditional static environmental analysis methods, the present invention has stronger environmental adaptability and prediction ability, avoiding potential safety hazards caused by sudden environmental changes.

[0117] In terms of driving behavior analysis, the present invention constructs a temporal graph neural network (T-GNN), combines contrastive learning and Transformer attention mechanism to achieve in-depth mining and classification of driver behavior patterns. Compared with existing time series analysis methods, the present invention not only focuses on single driving behaviors but also models long-term driving data to identify subtle abnormal behaviors, improving the accuracy of abnormal driving detection. At the same time, the Bayesian dynamic autoregressive model is combined to calculate the driving anomaly score. Compared with fixed rule or statistical feature analysis methods, the present invention can dynamically adapt to the behavior characteristics of different drivers, improve the accuracy of anomaly detection, and reduce the possibility of false alarms and missed detections.

[0118] In terms of risk assessment and safety strategy optimization, the present invention breaks through the limitations of traditional fixed-threshold safety control methods. It uses a multi-modal behavior integration reasoning model (MBIR) for risk assessment and combines zero-sum game reinforcement learning to optimize safety strategies, realizing intelligent dynamic adjustment of speed limits, route adjustments, and warning schemes. By constructing an adversarial game mechanism between the driver and the safety control system, the present invention can dynamically optimize safety control strategies in different environments, avoiding the adaptability problems caused by fixed strategies. At the same time, in the stage of safety strategy execution, the present invention introduces multi-agent reinforcement learning, enabling the safety control system to continuously optimize decisions based on real-time road conditions and historical driving behaviors, improving the ability to respond to emergencies.

[0119] In summary, the present invention not only solves the deficiencies of the prior art in terms of positioning accuracy, environmental perception, driving behavior analysis, and safety strategy optimization, but also improves in terms of real-time performance, adaptability, and safety. Through intelligent data fusion, in-depth behavior analysis, and reinforcement learning optimization, the present invention realizes all-round intelligent monitoring and safety control of hazardous chemical transport vehicles, effectively reducing safety risks during transportation, improving management efficiency, and being applicable to scenarios such as large container terminals, hazardous material logistics transportation, and intelligent traffic safety management. BRIEF DESCRIPTION OF THE DRAWINGS

[0120] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, but do not constitute a limitation to the present invention. In the drawings:

[0121] Figure 1 is a flowchart of the intelligent tracking and positioning system for hazardous chemical transport vehicles in a container terminal proposed by the present invention;

[0122] Figure 2 is a schematic diagram of the algorithm of the driving behavior modeling and trajectory prediction module of the intelligent tracking and positioning system for hazardous chemical transport vehicles in a container terminal proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0123] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0124] Referring to Figure 1-2 , the intelligent tracking and positioning system for hazardous chemical transport vehicles in a container terminal includes:

[0125] A data acquisition module for collecting vehicle positioning, driving behavior, and environmental state data;

[0126] A data fusion and time synchronization module for time alignment of multi-source data;

[0127] The behavior modeling and trajectory prediction module is used to extract driving behavior features and predict the driving trajectory;

[0128] The risk assessment and anomaly detection module is used to calculate the anomaly score and predict the driving risk;

[0129] The safety policy optimization module is used to build a zero-sum game safety control model and optimize the speed limit, path, and warning strategies;

[0130] The remote control and intelligent intervention module is used for remote speed regulation, emergency braking, and path optimization, and to push alerts;

[0131] The cloud monitoring and data analysis module is used to store the trajectory, environment, driving score, and safety policy, and to predict the long-term risk trend.

[0132] In this embodiment, the modules are implemented through the following methods:

[0133] S1. Collect GNSS-RTK positioning data, IMU inertial navigation data, camera image data, lidar scan data, and driving parameters, and perform time synchronization processing;

[0134] S2. Based on the collected GNSS-RTK positioning data and IMU inertial navigation data, use the non-linear motion prediction compensation method to correct the signal loss area and generate vehicle positioning data;

[0135] S3. Based on the camera image data and lidar scan data, use image segmentation technology to analyze the dock obstacles, driving route, and environmental changes, and generate environmental safety assessment data;

[0136] S4. Use the vehicle positioning data, environmental safety assessment data, and driving parameters to construct an implicit topological behavior analysis network, and use a temporal graph neural network to extract the driver's operation mode and generate driving behavior feature data;

[0137] S5. Based on the driving behavior feature data, vehicle positioning data, and environmental safety assessment data, use a multi-modal behavior fusion reasoning model to evaluate the driving risk, and combine dynamic Bayesian optimization to calculate the anomaly score and generate driving risk assessment data;

[0138] S6. Based on the driving risk assessment data, vehicle positioning data, and environmental safety assessment data, construct a dynamic game safety policy, and use a zero-sum game to adjust the speed limit, path, and warning strategies, and generate safety control data;

[0139] S7. Based on the safety control data, perform real-time intervention, remotely control the speed, braking, and path planning, and optimize the transportation task through spatio-temporal prediction scheduling, and push alerts to the driver terminal and the supervision center.

[0140] In this embodiment, S2 includes the following specific steps:

[0141] S21. Based on GNSS-RTK positioning data and IMU inertial navigation data, construct a multi-sensor spatio-temporal state fusion model. Let the state vector of the vehicle at time t be:

[0142] X t = [P t , V t , A t , Θ t T ;

[0143] where P t is the position vector, V t is the velocity vector, A t is the acceleration vector, and Θ t is the attitude angle vector;

[0144] S22. Use an adaptive non-linear filtering model to dynamically correct the GNSS-RTK positioning data. Define the GNSS observation value as The measurement noise is e t , then the position update equation is:

[0145]

[0146] where the covariance matrix Σ t is calculated by the Markov adaptive estimation method, and the update rule is:

[0147]

[0148] where P t is the corrected vehicle position, e t is the GNSS measurement noise, α is the weight factor, Σ t-1 is the covariance matrix at the previous time t-1, N is the size of the historical data window, and i is the index;

[0149] S23. Based on the IMU inertial navigation data, construct a multi-order dynamic prediction model. Let the angular velocity be ω t = (ω x, , ω y, , ω z,t ), the attitude angle be Θ t = (φ t , θ t , ψ t ), and the acceleration be A t = (a x, , a y, , a z, ), then the state update equation is:​

[0150]

[0151] where Θ t+1 is the attitude angle of the vehicle at time t + 1, P t+1 is the position of the vehicle at time t + 1, V t+1 is the speed of the vehicle at time t + 1, Θ t is the attitude angle at time t, ω t is the angular velocity at time t, V t is the velocity vector at time t, P t is the position vector at time t, W t is the process noise, and Δt is the time step;

[0152] S24. Fuse the GNSS-RTK positioning data and the IMU inertial navigation data, and use adaptive dynamic Bayesian filtering to correct the trajectory to obtain the state estimation result. The state transition model is:

[0153] X t = FX t-1 + GU t + W t ;

[0154] The observation equation is:

[0155] Y t = HX t + R t ;

[0156] where X t is the vehicle state, F is the state transition matrix, Y t is the observation data vector, G is the control input matrix, U t is the control input vector, H is the observation matrix, R t is the observation noise, and X t-1 is the vehicle state at time t + 1;

[0157] S25. Based on the state estimation result, combine the Gaussian process regression-spatiotemporal attention network model to optimize the trajectory compensation, and define the trajectory compensation objective function:

[0158]

[0159] where is the optimized vehicle position, is the predicted vehicle position, T is the time window size, λ is the regularization coefficient, is the trajectory smoothing term, P is the vehicle trajectory position variable to be optimized, is to find the optimal trajectory compensation position

[0160] In this embodiment, S3 includes the following specific steps:

[0161] S31. Based on the camera image data I t and the lidar point cloud data L t construct an environmental perception data set, and use multi-scale pyramid convolution transformation to extract features. The environmental feature matrix is expressed as:

[0162]

[0163] where D t is the environmental feature matrix, S is the number of layers of different scale pyramid transformations adopted for the image data, Q is the number of layers of different scale pyramid transformations adopted for the point cloud data, W s is the convolution weight matrix of the image data at the s-th layer scale, V q is the convolution weight matrix of the point cloud data at the q-th layer scale, G s is the feature map of the image data at the s-th layer scale, and H q is the feature map of the point cloud data at the q-th layer scale;

[0164] S32. Perform image-point cloud joint calibration on the environmental feature matrix. Let the camera internal parameter matrix be K, the external parameter matrix be R and T, then the projection transformation from the point cloud to the image is:

[0165] Z t = K(RL t + D);

[0166] where Z t is the target projection point in the image coordinates. Based on the least squares optimization, solve the optimal projection parameter θ, and the objective function is:

[0167]

[0168] where θ * is the optimal parameter, K is the internal parameter matrix of the camera, R is the rotation matrix, D is the translation matrix, T is the time step, is the projection point obtained after adjustment according to the optimization parameter, β is the regularization coefficient, is the gradient change of the trajectory, and argmin θ is to solve the parameter θ that makes the objective function reach the minimum value;

[0169] S33. Based on the calibration result, use adaptive graph optimization to construct a three-dimensional obstacle distribution. Let the obstacle graph be expressed as G(V, E), where the node V is the set of obstacle points and the edge E represents the adjacency relationship:

[0170] G * = argmin G∑ (i,)∈E w ij ∥Z i -Z j ∥ 2 ;

[0171] Among them, G * is the optimal obstacle map, and argmin G represents solving the optimal map G that makes the objective function reach the minimum value * , Z i and Z j are two obstacle points in the three-dimensional point cloud data, and w ij is the weight of the edge E in the graph;

[0172] S34. Based on the obstacle map, an adaptive dynamic Bayesian filter is used to model the environmental state, and the environmental state vector E t is defined as:

[0173]

[0174] E t+1 = FE t + W t ;

[0175] Among them, Z t , V t , A t are the position information, velocity information, and acceleration information of the environmental obstacle respectively, W t is the process noise, E t+1 is the environmental state vector at time t + 1, and F is the environmental state transition matrix;

[0176] S35. Based on the environmental state, a Markov hidden variable prediction model is combined to predict the environmental change trend, and the environmental safety risk score is calculated. Let the environmental safety assessment data be R t :

[0177]

[0178] Among them, N is the total number of environmental characteristic factors, β i is the weight parameter of the environmental characteristic factor, f i (E t ) is the environmental variable under the environmental state E t , η is the historical risk attenuation coefficient, R t-1 is the environmental safety assessment data at the previous moment, K t is the prediction error term, and i is the index.

[0179] In this embodiment, the S4 includes the following specific steps:

[0180] S41. Construct a high-dimensional driving behavior state vector based on vehicle positioning data, environmental safety assessment data, and driving parameters:

[0181] N t =[Z t ,V t ,A t ,Θ t ,C t T ;

[0182] where Z t is the vehicle position, V t is the speed, A t is the acceleration, Θ t is the heading angle, C t is the environmental dynamic constraint factor, and N t is the driving behavior input state vector;

[0183] S42. Use a multi-scale graph convolutional spatio-temporal network to model driving behavior patterns. Let the behavior topology graph be represented as X(V, E, W), where V is the set of driving behavior nodes, E is the set of behavior time-series edges, and W is the edge weight matrix. The driving behavior state transition is described by the following update equation:

[0184]

[0185] where H t is the driving behavior feature matrix at the current moment, σ is the non-linear activation function, W is the trainable parameter matrix, H t-1 is the driving behavior feature matrix at the previous moment t - 1, Y is the number of layers of the graph convolutional network, β y is the weighting coefficient of different-scale graph convolutional layers, N y (H t ) is the driving behavior embedding feature calculated by the s-th layer of the multi-scale graph convolutional network, and y is the index;

[0186] S43. Based on driving behavior features, use contrast learning + Transformer attention mechanism for driving mode classification:

[0187]

[0188] Define an enhanced contrast loss function:

[0189]

[0190] where is the driving behavior feature space, L is the contrast learning loss function, N is the number of training samples, ln is the logarithmic function, exp is the exponential function, is to measure and​ A calculation function for similarity between To measure and A calculation function for similarity between is the i-th driving behavior feature vector, is the positive sample corresponding to the i-th feature, is other driving mode features in the comparison sample set, j and t are indices, γ is the regularization coefficient, and T is the time step, is the driving behavior feature predicted at time t;

[0191] S44. Based on the driving mode classification result, use the Bayesian dynamic autoregressive model to calculate the driving anomaly score:

[0192]

[0193] where B t is the driving anomaly score at the current moment, M is the total number of driving behavior feature factors, α k is the feature factor weight, g k (H t ) is the k-th driving behavior feature factor, δ is the risk adjustment parameter, B t-1 is the driving anomaly score at the previous moment t-1, μ is the risk score threshold, K t is the prediction error term;

[0194] S45. Based on the driving anomaly score, use the time series attention embedding model to generate driving behavior feature data. Let the driving behavior feature vector be:

[0195]

[0196] where C t is the driving behavior feature data generated at the current moment t, O is the total number of environmental behavior impact factors, ω j is the environmental behavior factor weight, A j is the j-th environmental behavior impact factor, tanh is used to suppress the influence of extreme anomaly scores, π is used to control the influence degree of anomaly scores on driving behavior features, ∈ t is the noise term, and j and k are indices.

[0197] In this embodiment, S5 includes the following specific steps:

[0198] S51. Based on the driving behavior feature data, vehicle positioning data, and environmental safety assessment data, construct a driving risk assessment state vector. Let the state vector be:

[0199] Y t =[C t ,Zt ,R t T ;

[0200] Among them, C t is the generated driving behavior characteristic data, Z t is the vehicle positioning data, R t is the generated environmental safety assessment data;

[0201] S52. Calculate the driving risk score using a multi-modal behavior fusion inference model, and define the driving risk score function:

[0202]

[0203] Among them, D t is the driving risk score at the current moment t, M is the total number of driving behavior characteristic factors, α k is the characteristic factor weight, h k (Y t ) is the k-th driving behavior characteristic factor, β j is the environmental impact factor weight, f j (R t ) is the j-th environmental impact factor, ξ t is the prediction error term, U is the total number of environmental impact factors, and j and k are indices;

[0204] S53. Based on the driving risk score, construct a time series anomaly detection model, and define the risk state transition equation:

[0205]

[0206] Among them, K t is the prediction error term, D t+1 is the driving risk score at time t+1, D t is the driving risk score at time t, ρ is the risk attenuation coefficient, T is the time window size, λ i is the weight coefficient of the i-th historical risk score within the time window, D t-i is the driving risk score in the past i time steps, and i is the index;

[0207] S54. Based on the risk state prediction result, use reinforcement learning to dynamically optimize the driving behavior risk assessment:

[0208]

[0209] Among them, X t+1 is the driving state at time t+1, X t is the driving state at time t, A t is the decision-making action of the driver at time t, ​is the reward factor, π t is the driving behavior strategy function, Q(X t ,A t ) is the state-action value function of reinforcement learning, argmax x is to find the optimal driving state that maximizes the objective function, and x is the set of all possible driving states;

[0210] S55. Based on the optimization result of driving behavior risk assessment, combined with the multi-scale adaptive threshold algorithm, generate driving risk assessment data:

[0211]

[0212] Among them, is the driving risk assessment data, σ is the normalization function, M is the total number of driving risk characteristic factors, is the k-th driving risk score data, θ t is the dynamically adjusted risk threshold, ρ t is the prediction error term.

[0213] In this embodiment, the S6 includes the following specific steps:

[0214] S61. Based on the driving risk assessment data vehicle positioning data Z t and environmental safety assessment data R t , construct a dynamic game safety strategy state vector:

[0215]

[0216] S62. Construct a safety control game model. Assume that the game participants include the driver P1 and the safety control system P2, and define the game revenue function:

[0217]

[0218] Among them, U(G t ,A t ) is the game revenue function of taking the safety control strategy A t under the current state G t , M is the total number of driving state characteristic factors, α k is the weight of the driving state characteristic factor, g k (G t ) is the k-th driving state characteristic factor, Z is the total number of safety control strategy cost items, β j is the cost weight of the safety control strategy, c j (A t ) is the j-th safety control strategy cost item, and j and k are indexes;

[0219] S63. Calculate the optimal safety strategy using the zero-sum game reinforcement learning algorithm based on the game payoff function:

[0220]

[0221] Among them, is the optimal safety strategy, E is the expected calculation, Find the optimal safety control strategy that maximizes the objective function E[U(G t , A t )]

[0222] S64. Based on the optimal safety strategy, construct a multi-agent collaborative optimization framework. Let the set of agents be P, then the joint optimization objective function is:

[0223] J∑ i∈P ω i U i (G t , A t );

[0224] Among them, J is the joint optimization objective function, ω i is the weight factor of agent i, U i (G t , A t ) is the payoff function of agent i after taking action A t in state G t ;

[0225] S65. Based on the optimization results, dynamically adjust the safety control strategy in combination with the reinforcement learning model to generate safety strategy control data:

[0226]

[0227] t is the generated safety control data, is the normalization function, is the k-th safety control strategy, ε is the control strategy update coefficient, A t-1 is the safety control strategy at the previous moment, σ t is the policy adjustment noise term.

[0228] Example 1:

[0229] To verify the feasibility of the present invention in implementation, the present invention is applied to a large international container terminal. There are a large number of hazardous chemical transportation vehicles entering and leaving the terminal every day, and multiple high-risk scenarios are involved in the transportation process, such as complex road environments, adverse weather conditions, and irregular operations of drivers, etc. Traditional monitoring of hazardous chemical transportation mainly relies on manual supervision and safety strategies with fixed rules. However, due to the complex environment inside the terminal, there are a large number of container stacks, blind spots, and vehicle intersection sections. The traditional GPS positioning system is prone to a decrease in accuracy in signal-blocked areas, resulting in positioning drift. At the same time, due to the limited road space inside the terminal, drivers may exhibit abnormal behaviors such as sudden braking, speeding, and dangerous lane changes in complex environments. Traditional monitoring means are difficult to detect and intervene in a timely manner, increasing the accident risk. The present invention deploys a set of intelligent tracking and safety control system for hazardous chemical transportation vehicles in container terminals based on intelligent game and multi-modal data fusion, and realizes the full-range intelligent monitoring and safety control of hazardous chemical transportation vehicles through high-precision vehicle positioning, driving behavior analysis, environmental risk assessment, and intelligent safety strategy optimization.

[0230] First, the system installs sensing devices such as GNSS-RTK, IMU inertial navigation, cameras, lidar, gas sensors, and temperature and humidity sensors on hazardous chemical transportation vehicles to collect vehicle position, driving status, driver operation information, and surrounding environment data in real time. Through the spatio-temporal collaborative data fusion method, the GNSS-RTK and IMU data are optimized to improve the positioning accuracy, enabling the vehicle to maintain accurate positioning even in signal-blocked environments such as dense container areas and tunnels in the terminal. Compared with traditional methods, the average positioning error of the present invention in multi-blocked areas in the terminal is reduced from 2.8 meters to 0.6 meters, and the positioning loss rate in low-GNSS-signal areas is decreased from 12.4% to 2.1%, effectively improving the vehicle positioning stability in complex environments.

[0231] In terms of environmental perception, the system realizes the efficient recognition of the distribution of obstacles inside the terminal, road boundaries, and dynamic traffic conditions through image-point cloud joint analysis technology combined with dynamic environment modeling. In foggy weather, the recognition rate of obstacle detection by traditional cameras drops to 71.2%, while after the system integrates lidar point cloud data, the obstacle detection accuracy is increased to 94.5%, ensuring the environmental perception ability in complex weather environments. At the same time, the system combines the Markov hidden variable prediction model to predict the environmental change trend in the next 5 seconds to 30 seconds, enabling the system to give early warnings before the occurrence of danger and improving safety.

[0232] In terms of driving behavior analysis, the present invention uses a Temporal Graph Neural Network (T-GNN), combines contrastive learning with the Transformer attention mechanism, and models and analyzes driver behavior. After 30 consecutive days of operation monitoring, this system detected that the number of emergency brakes of hazardous chemical transport vehicles decreased by 32%, the speeding situation decreased by 28%, and the dangerous lane changes decreased by 45%. Compared with the traditional rule-based anomaly detection method, the accuracy of driving behavior anomaly detection of this system increased from 85.6% to 95.3%, effectively reducing the false alarm and missed alarm situations, so that abnormal driving behaviors can be more accurately identified and intervened.

[0233] In terms of risk assessment and safety strategy optimization, this system uses a Multi-modal Behavior Integration and Reasoning Model (MBIR) for risk assessment and combines zero-sum game reinforcement learning to dynamically optimize safety strategies. Experimental data shows that during peak hours, the safety intervention trigger rate of the traditional fixed rule strategy is as high as 65%, resulting in some normal driving behaviors being misjudged and affecting the transportation efficiency. However, based on the intelligent game mechanism, this system can dynamically adjust speed limits, route optimization, and warning strategies according to real-time road conditions, driving behaviors, and environmental monitoring data, reducing the safety intervention trigger rate during peak hours to 38%. While ensuring safety, the transportation efficiency is improved. In addition, after optimizing the speed limit strategy, the number of speeding alarms decreased by 41%. Through route optimization, the average transportation time was shortened by 7.8%, effectively improving the safety and timeliness of the transportation process.

[0234] Table 1 Safety Monitoring Data Table of Hazardous Chemical Transport Vehicles

[0235]

[0236]

[0237] The present invention has achieved significant optimization in the intelligent tracking and safety control of hazardous chemical transport vehicles. In terms of vehicle positioning accuracy, spatio-temporal collaborative data fusion is adopted, reducing the average positioning error in the dock occlusion area from 2.8 meters to 0.6 meters, and the positioning loss rate in the low GNSS signal area from 12.4% to 2.1%, greatly improving the positioning stability in complex environments. In terms of environmental perception ability, image-point cloud fusion and dynamic environment modeling are introduced, increasing the obstacle detection accuracy in foggy weather from 71.2% to 94.5%, and the environmental change prediction accuracy is increased to 93.2%, enhancing the adaptability of the system in complex environments.

[0238] In terms of driving behavior analysis, this system adopts the Temporal Graph Neural Network (T-GNN) and contrastive learning method, achieving a 32% reduction in hard braking, a 28% reduction in speeding, and a 45% reduction in dangerous lane changes. The accuracy of abnormal driving behavior detection has increased from 85.6% to 95.3%, reducing false alarms and missed reports. In terms of risk assessment and safety strategy optimization, zero-sum game reinforcement learning is used to dynamically adjust speed limits, route optimization, and warning strategies, resulting in a reduction of the safety intervention trigger rate during peak hours from 65% to 38%, a 41% reduction in speeding alarms, and a 7.8% reduction in transportation time, effectively improving transportation efficiency.

[0239] Finally, during the 6-month test, the system monitored 3,892 safety risks in total, successfully intervened 2,954 times, and the accident rate decreased from 0.87 times per 100,000 kilometers to 0.29 times, a 66.7% reduction, significantly enhancing the safety of hazardous chemical transportation.

[0240] As described above, it is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. Intelligent tracking and positioning system for hazardous chemicals transport vehicles at container terminals, characterized by: include: Data collection module, used to collect vehicle positioning, driving behavior and environmental status data; Data fusion and time synchronization module, used to align the time of multi-source data; Behavior modeling and trajectory prediction module, used to extract driving behavior features and predict driving trajectories; Risk assessment and anomaly detection module, used to calculate anomaly scores and predict driving risks; The safety strategy optimization module is used to build a zero-sum game safety control model and optimize the speed limit, path and warning strategies; Remote control and intelligent intervention module for remote speed regulation, emergency braking, route optimization, and push alerts; Cloud monitoring and data analysis module, used to store trajectories, environments, driving scores and safety strategies, and predict long-term risk trends.

2. The intelligent tracking and positioning system for hazardous chemicals transport vehicles at container terminals according to claim 1 is characterized in that: The modules are implemented in the following ways: S1, collect GNSS-RTK positioning data, IMU inertial navigation data, camera image data, lidar scanning data and driving parameters, and perform time synchronization processing; S2, based on the collected GNSS-RTK positioning data and IMU inertial navigation data, the nonlinear motion prediction compensation method is used to correct the signal loss area and generate vehicle positioning data; S3, based on camera image data and lidar scanning data, uses image segmentation technology to analyze dock obstacles, driving routes and environmental changes to generate environmental safety assessment data; S4. Using vehicle positioning data, environmental safety assessment data and driving parameters, an implicit topological behavior analysis network is constructed, and a time graph neural network is used to extract the driver's operation mode and generate driving behavior feature data; S5. Based on driving behavior feature data, vehicle positioning data and environmental safety assessment data, a multimodal behavior fusion reasoning model is used to assess driving risk, and dynamic Bayesian optimization is used to calculate abnormal scores to generate driving risk assessment data; S6. Based on driving risk assessment data, vehicle positioning data and environmental safety assessment data, a dynamic game safety strategy is constructed, and a zero-sum game is used to adjust the speed limit, path and warning strategy to generate safety control data; S7. Based on safety control data, perform real-time intervention, remotely control speed, braking and path planning, optimize transportation tasks through spatiotemporal prediction and scheduling, and push alarms to driver terminals and supervision centers.

3. The intelligent tracking and positioning system for hazardous chemicals transport vehicles at container terminals according to claim 2 is characterized in that: The S2 specifically includes: S21. Based on GNSS-RTK positioning data and IMU inertial navigation data, a multi-sensor spatiotemporal state fusion model is constructed. The state vector of the vehicle at time t is: X t =[P t ,V t ,A t ,I t ] T ; Among them, P t is the position vector, V t is the velocity vector, A t is the acceleration vector, Θ t is the attitude angle vector; S22, use the adaptive nonlinear filtering model to perform dynamic error correction on the GNSS-RTK positioning data, and define the GNSS observation value as The measurement noise is e t , then the position update equation is: Among them, the covariance matrix Σ t Calculated by the Markov adaptive estimation method, the update rule is: Among them, P t is the corrected vehicle position, e t is the GNSS measurement noise, α is the weight factor, Σ t-1 is the covariance matrix of the previous moment t-1, N is the size of the historical data window, and i is the index; S23. Based on IMU inertial navigation data, a multi-order dynamic prediction model is constructed, assuming that the angular velocity is ω t =(ω x,t ,ω y,t ,ω z,t ), the attitude angle is Θ t =(φ t ,θ t ,ψ t ), the acceleration is A t =(a x,t ,a y,t ,a z,t ), then the state update equation is: Among them, Θ t+1 is the attitude angle of the vehicle at time t+1, P t+1 is the position of the vehicle at time t+1, V t+1 is the speed of the vehicle at time t+1, Θ t is the attitude angle at time t, ω t is the angular velocity at time t, V t is the velocity vector at time t, P t is the position vector at time t, W t is the process noise, Δt is the time step; S24, integrating GNSS-RTK positioning data and IMU inertial navigation data, using adaptive dynamic Bayesian filtering for trajectory correction, and obtaining state estimation results. The state transfer model is: X t =FX t-1 +GU t +W t ; The observation equation is: Y t =HX t +R t ; Among them, X t is the vehicle state, F is the state transfer matrix, Y t is the observed data vector, G is the control input matrix, U t is the control input vector, H is the observation matrix, R t is the observation noise, X t-1 is the vehicle status at time t+1; S25. Based on the state estimation results, the trajectory compensation is optimized by combining the Gaussian process regression-spatial-temporal attention network model, and the trajectory compensation objective function is defined: in, For the optimized vehicle position, is the predicted vehicle position, T is the time window size, λ is the regularization coefficient, is the trajectory smoothing term, P is the vehicle trajectory position variable to be optimized, To find the optimal trajectory compensation position 4. The intelligent tracking and positioning system for hazardous chemicals transport vehicles at container terminals according to claim 2 is characterized in that: The S3 specifically includes: S31, based on camera image data I t and lidar point cloud data L t Construct an environmental perception dataset and use multi-scale pyramid convolution transform to extract features. The environmental feature matrix is ​​expressed as: Among them, D t is the environmental feature matrix, S is the number of pyramid transformation layers of different scales used for image data, Q is the number of pyramid transformation layers of different scales used for point cloud data, and W s is the convolution weight matrix of the image data at the sth level, V q is the convolution weight matrix of the point cloud data at the qth layer scale, G s is the feature map of the image data at the sth level, H q It is the feature map of point cloud data at the qth scale; S32, perform image-point cloud joint correction on the environmental feature matrix, assuming that the camera internal parameter matrix is ​​K, and the external parameter matrices are R and T, then the projection transformation from point cloud to image is: Z t =K(RL t +D); Among them, Z t is the target projection point in the image coordinates, and the optimal projection parameter θ is solved based on the least squares optimization. The objective function is: Among them, θ * is the optimal parameter, K is the intrinsic parameter matrix of the camera, R is the rotation matrix, D is the translation matrix, T is the time step, is the projection point obtained after adjusting the optimization parameters, β is the regularization coefficient, is the gradient change of the trajectory, argmin θ To solve the parameter θ that minimizes the objective function; S33. Based on the correction results, adaptive graph optimization is used to construct the three-dimensional obstacle distribution. The obstacle graph is represented by G = (V, E), where the node V is the obstacle point set and the edge E represents the adjacency relationship: G * =argmin G ∑ (i,j)∈E In ij ∥Z i -WITH j ∥ 2 ; Among them, G * is the optimal obstacle map, argmin G is the optimal graph G that is used to solve the objective function to the minimum value. * , Z i and Z j are two obstacle points in the 3D point cloud data, w ij is the weight of edge E in the graph; S34, based on the obstacle map, the adaptive dynamic Bayesian filter is used to model the environment state and define the environment state vector E t : HAVE BEEN t+1 =FE t +W t ; Among them, Z t ,V t ,A t are the position information, velocity information and acceleration information of environmental obstacles respectively, W t is the process noise, E t+1 is the environmental state vector at time t+1, and F is the environmental state transfer matrix; S35. Based on the environmental status, the environmental change trend is predicted by combining the Markov latent variable prediction model, and the environmental safety risk score is calculated. Suppose the environmental safety assessment data is R t : Among them, N is the total number of environmental characteristic factors, β i is the weight parameter of the environmental characteristic factor, f i (E t ) is the environmental state E t The environmental variables under the condition, η is the historical risk attenuation coefficient, R t-1 is the environmental safety assessment data at the previous moment, K t is the prediction error term and i is the index.

5. The intelligent tracking and positioning system for hazardous chemicals transport vehicles at container terminals according to claim 2 is characterized in that: The S4 specifically includes: S41. Construct a high-dimensional driving behavior state vector based on vehicle positioning data, environmental safety assessment data and driving parameters: N t =[Z t ,V t ,A t ,Θ t ,C t ] T ; Among them, Z t is the vehicle position, V t is the speed, A t is the acceleration, Θ t is the heading angle, C t is the environmental dynamic constraint factor, N t Driving behavior input state vector; S42. Use a multi-scale graph convolutional spatiotemporal network to model driving behavior patterns. Suppose the behavior topology graph is represented by X = (V, E, W), where V is the set of driving behavior nodes, E is the set of behavior temporal edges, and W is the edge weight matrix. The driving behavior state transition is described by the following update equation: Among them, H t is the driving behavior feature matrix at the current moment, σ is the nonlinear activation function, W is the trainable parameter matrix, H t-1 is the driving behavior feature matrix at the previous time t-1, Y is the number of layers of the graph convolutional network, β y is the weight coefficient of the convolutional layer of different scale graphs, N y (H t ) is the driving behavior embedding feature calculated by the s-th layer multi-scale graph convolutional network, and y is the index; S43. Based on driving behavior characteristics, driving mode classification is performed using contrastive learning + Transformer attention mechanism: Define the enhanced contrast loss function: in, is the driving behavior feature space, L is the contrastive learning loss function, N is the number of training samples, ln is the logarithmic function, exp is the exponential function, To measure and The calculation function of the similarity between To measure and The calculation function of the similarity between is the i-th driving behavior feature vector, is the positive sample corresponding to the i-th feature, To compare other driving mode features in the sample set, j and t are indexes, γ is the regularization coefficient, T is the time step, is the driving behavior characteristics predicted at time t; S44. Based on the driving mode classification results, the driving abnormality score is calculated using the Bayesian dynamic autoregressive model: Among them, B t is the driving abnormality score at the current moment, M is the total number of driving behavior characteristic factors, α k is the characteristic factor weight, g k (H t ) is the kth driving behavior characteristic factor, δ is the risk adjustment parameter, B t-1 is the driving abnormality score at the previous moment t-1, μ is the risk score threshold, K t is the prediction error term; S45. Based on the driving abnormality score, a temporal attention embedding model is used to generate driving behavior feature data. The driving behavior feature vector is set to: Among them, C t is the driving behavior characteristic data generated at the current time t, O is the total number of environmental behavior influencing factors, ω j is the environmental behavior factor weight, A j is the jth environmental behavior influencing factor, tanh is used to suppress the influence of extreme abnormal scores, π is the degree of influence of abnormal scores on driving behavior characteristics, ∈ t is the noise term, j and k are the indices.

6. The intelligent tracking and positioning system for hazardous chemicals transport vehicles at container terminals according to claim 2 is characterized in that: The S5 specifically includes: S51. Based on the driving behavior characteristic data, the vehicle positioning data and the environmental safety assessment data, a driving risk assessment state vector is constructed. The state vector is assumed to be: Y t =[C t ,Z t ,R t ] T ; Among them, C t is the generated driving behavior characteristic data, Z t is the vehicle positioning data, R t To assess the data for the generated environment safety; S52. Calculate the driving risk score using a multimodal behavior fusion reasoning model and define a driving risk score function: Among them, D t is the driving risk score at the current time t, M is the total number of driving behavior characteristic factors, α k is the characteristic factor weight, h k (Y t ) is the kth driving behavior characteristic factor, β j is the weight of environmental impact factor, f j (R t ) is the jth environmental impact factor, ξ t is the prediction error term, U is the total number of environmental impact factors, j and k are indexes; S53. Based on the driving risk score, a time series anomaly detection model is constructed and the risk state transfer equation is defined: Among them, K t is the prediction error term, D t+1 is the driving risk score at time t+1, D t is the driving risk score at time t, ρ is the risk attenuation coefficient, T is the time window size, λ i is the weight coefficient of the i-th historical risk score in the time window, D t-i is the driving risk score in the past i time steps, i is the index; S54. Based on the risk status prediction results, reinforcement learning is used to dynamically optimize the driving behavior risk assessment: Among them, X t+1 is the driving state at time t+1, X t is the driving state at time t, A t is the driver’s decision action at time t, is the reward factor, π t is the driving behavior strategy function, Q(X t ,A t ) is the state-action value function of reinforcement learning, argmax x To find the optimal driving state that maximizes the objective function, x is the set of all possible driving states; S55. Based on the driving behavior risk assessment optimization results, a multi-scale adaptive threshold algorithm is used to generate driving risk assessment data: in, is the driving risk assessment data, σ is the normalization function, M is the total number of driving risk characteristic factors, is the k-th driving risk score data, θ t is the dynamically adjusted risk threshold, ρ t is the prediction error term.

7. The intelligent tracking and positioning system for hazardous chemicals transport vehicles at container terminals according to claim 2 is characterized in that: The S6 specifically includes: S61. Based on driving risk assessment data Vehicle positioning data Z t and environmental safety assessment data R t , construct the dynamic game security strategy state vector: S62, construct a safety control game model, assume that the game participants include the driver P1 and the safety control system P2, and define the game profit function: Among them, U(G t ,A t ) is the current state G t Take security control strategy A t The game profit function, M is the total number of driving state characteristic factors, α k is the weight of the driving state characteristic factor, g k (G t ) is the kth driving state characteristic factor, Z is the total number of safety control strategy cost items, β j is the cost weight of the security control strategy, c j (A t ) is the j-th security control strategy cost item, j and k are indexes; S63. Based on the game profit function, the zero-sum game reinforcement learning algorithm is used to calculate the optimal security strategy: in, is the optimal security strategy, E is the expected calculation, Find the objective function E[U(G t ,A t )] The optimal safety control strategy that maximizes S64. Based on the optimal safety strategy, a multi-agent collaborative optimization framework is constructed. Assuming the set of agents is P, the joint optimization objective function is: J=∑ i∈P ω i U i (G t ,A t ); Among them, J is the joint optimization objective function, ω i is the weight factor of agent i, U i (G t ,A t ) is the agent i in state G t Next, take action A t The subsequent profit function; S65. Based on the optimization results, the security control strategy is dynamically adjusted in combination with the reinforcement learning model to generate security strategy control data: Among them, Q t To generate security control data, is the normalization function, is the kth safety control strategy, ε is the control strategy update coefficient, A t-1 is the safety control strategy at the previous moment, σ t Adjust the noise term for the policy.

Images