An intelligent navigation system in an outdoor chemical plant dangerous environment based on visual simultaneous localization and mapping

Abstract

The application provides an intelligent navigation system based on visual simultaneous localization and mapping in an outdoor chemical plant dangerous environment, comprising a multi-source sensor system for collecting motion state data and surrounding environment feature data of a mobile robot in real time; a motor drive control module for driving the motion of a mobile robot chassis; a processor in communication connection with the multi-source sensor system and the motor drive control module, the processor being configured to execute the following control process: step 1, adaptive robust positioning; step 2, cross-modal risk modeling; step 3, risk perception path planning; and step 4, dynamic cooperative control. The application solves the problems that traditional navigation is prone to lose positioning in bad weather and cannot identify intangible chemical risks, and effectively improves the autonomous inspection safety of the mobile robot in a complex chemical environment. The application is suitable for a mobile robot platform.

Description

Technical Field

[0001] This invention belongs to the field of mobile robot navigation and path planning control technology, specifically relating to an intelligent navigation system for hazardous outdoor chemical plant environments based on visual synchronous localization and mapping. Background Technology

[0002] With the development of modern industry, the use of automated guided vehicles (AGVs) to replace manual labor in high-frequency chemical inspections has become a trend. However, in the complex environment of outdoor chemical plants, traditional autonomous navigation systems face severe challenges: First, extreme weather conditions (such as heavy fog and steam) can lead to severe degradation of visual perception, making traditional pure visual SLAM algorithms prone to catastrophic positioning drift or even system crashes when feature extraction is drastically reduced. Second, existing mainstream navigation cost maps are mostly based on "pure geometric measurements," only able to identify obstacles with physical collision volumes. For the "hidden chemical risks" such as frequent toxic gas leaks and high-temperature radiation from equipment in chemical plants, traditional frameworks are blinded by a lack of semantic cognitive ability. Finally, mainstream global path planning algorithms (such as the traditional A* algorithm) heavily rely on finding the shortest Euclidean geometric distance. This purely geometric greedy logic can blindly guide the robot through deadly chemical radiation zones in exchange for driving efficiency, leading to devastating hardware damage and potentially triggering secondary explosions. Summary of the Invention

[0003] Purpose of the Invention: The technical problem to be solved by this invention is to address the shortcomings of existing technologies by providing an intelligent navigation system for hazardous outdoor chemical plant environments based on Visual Simultaneous Localization and Mapping (VSLAM), comprising:

[0004] A multi-source sensor system is used to collect real-time motion state data and surrounding environment feature data of a mobile robot.

[0005] The motor drive control module is used to drive the movement of the mobile robot chassis.

[0006] A processor, communicatively connected to the multi-source sensor system and the motor drive control module, is configured to execute the following control flow:

[0007] Step 1, Perform spatial cognition and mapping localization based on adaptive multi-source tight coupling: Establish a nonholonomic constrained kinematic model of the mobile robot, fuse visual odometry, wheel odometry and inertial measurement unit data based on extended Kalman filter, and dynamically adjust the confidence level of visual observation using environmental degradation adaptive covariance reweighting mechanism to output a robust global pose estimate.

[0008] Step 2, construct a hierarchical hazard perception semantic cost map and perform cross-modal aggregation: on the basis of the global physical static base map, use the artificial potential field decay model to quantify the hidden chemical hazard sources into a continuous semantic risk layer, and use the local maximum aggregation strategy to deeply integrate physical obstacle constraints and semantic risk constraints;

[0009] Step 3: Execute a global replanning algorithm based on minimizing risk exposure integral: Improve the cost evaluation function of the traditional heuristic search algorithm by introducing the cumulative chemical exposure risk as a severe penalty, forcing the planning engine to actively abandon physical shortcuts across high-risk areas and plan a global safe topology trajectory.

[0010] Step 4: Perform local trajectory collaborative optimization combined with dynamic constraints: Generate control commands in the velocity sampling space, introduce local risk rejection functions for evaluation and screening, and coordinate the global path to achieve smooth and safe control of the robot.

[0011] The multi-source sensor system includes a wheeled odometer;

[0012] The processor is configured to perform kinematic state prediction based on discrete time series: receiving the chassis forward linear velocity from the wheel odometer in real time. With rotational angular velocity And based on the set discrete time step Calculate and update the theoretical pose of the mobile robot in the next moment;

[0013] The theoretical pose is used to provide kinematic deduction basis for the subsequent real-time positioning and control of the motor drive control module. The state prediction equation is as follows:

[0014] (1),

[0015] (2),

[0016] (3),

[0017] in, , Let x and y be the x and y coordinates of the mobile robot in the global coordinate system at time t, respectively. Let t be the heading angle. For discrete time steps, , , These are the x-coordinate, y-coordinate, and heading angle of the mobile robot at time t, respectively, in the previous time step.

[0018] In step 1, the general theoretical expression model of the prior state prediction and error covariance prediction matrix of the extended Kalman filter is defined as follows:

[0019] (4),

[0020] (5),

[0021] in, To control the input vector, This is the state transition function. Let F be the expected state vector at the current moment; F is the state transition function with respect to the state vector. The Jacobian matrix is ​​obtained by solving for Q, which is the system process noise covariance matrix, and T denotes the transpose. This is the expected state vector from the previous time step. This is the error covariance prediction matrix at the current time. This is the error covariance matrix of the previous time step.

[0022] In step 1, the environmental degradation adaptive covariance reweighting mechanism dynamically adjusts the diagonal elements of the error covariance matrix based on environmental characteristic quality: the environmental field of view abrupt change threshold coordinates are set as follows. When the robot's global lateral coordinates At that time, the number of feature points The estimation formula is:

[0023] (6),

[0024] in, This is the baseline number of feature points in the normal region. For linear decay rate, The lower limit of the feature point;

[0025] Subsequently, the diagonal elements of the visual observation noise covariance matrix were dynamically adjusted. :

[0026] (7),

[0027] in, Based on baseline observation noise, This is the sensitivity adjustment coefficient.

[0028] In step 2, the hierarchical hazard perception semantic cost map includes a static physical layer, a dynamic expansion layer, and a semantic risk layer; the cost of the dynamic expansion layer... Euclidean distance from the obstacle surface It decays exponentially:

[0029] (8),

[0030] in, To scale the decay factor, The distance from the current evaluation point to the surface of the physical obstacle is the Euclidean distance. Let e ​​be the safe inner radius of the mobile robot, and e be a natural constant.

[0031] When constructing the semantic risk layer, the system first builds a general physical foundation model of a multi-source risk field in a continuous space, for any evaluation point in the space. Its overall risk value It is composed of the superposition of the exponentially decaying potential fields of each hazard source: (9),

[0032] in, For the first The absolute spatial coordinates of a high-risk semantic detection source As assessment point With the A high-risk semantic detection source The Euclidean distance between them For the first The initial weighting coefficients for the inherent lethality of each hazard source. Let be the environmental degradation constant, and exp be the natural exponential function.

[0033] In step 2, discretization is performed based on the continuous risk field model. The semantic risk layer adopts an artificial potential field distance decay model: let the current evaluation grid coordinates be... The center of the hazard source is Define the fatal exposure radius With edge warning radius Set the core fatal zone base penalty cap to 253, and the semantic risk potential cost. The piecewise analytical expression of is:

[0034] (10),

[0035] Among them, the distance parameter ;

[0036] The system merges various layers into a global comprehensive cost map using a local maximum aggregation strategy. Achieving a unified cost metric for both physical constraints and chemical soft constraints:

[0037] (11),

[0038] in, To evaluate grid coordinates The static physical layer value at that location.

[0039] In step 3, the global reprogramming algorithm based on minimizing the risk exposure integral reconstructs the heuristic evaluation system of the A* heuristic search algorithm, and the reconstructed macroscopic multi-objective general evaluation model. Defined as:

[0040] (12),

[0041] in, For nodes The overall evaluation score Represents the physical path cost. This represents the cumulative risk penalty for nodes. For heuristic cost estimation; Increase the weight of overall risk sensitivity;

[0042] Combining the underlying discrete grid cost map, the macro model is specifically expanded into a reconstructed heuristic cost evaluation equation. Defined as: (13)

[0043] in, The cost is the mesh distance for physical translation. This represents the sum of the semantic risk values ​​of all rasters traversed on the search backtracking branch.

[0044] In step 4, the local trajectory cooperative optimization combined with dynamic constraints directly generates virtual linear velocity and angular velocity control pairs in the velocity sampling space through the local dynamic window method, and a multi-objective cooperative evaluation function. Reconstructed as:

[0045] (14)

[0046] in, For sampling linear velocity, For sampling angular velocity, For sampling rate pairs; A cost evaluation item to ensure the robot's heading conforms to the global topology trajectory; This is the positive translational velocity excitation term; To ensure the safe clearance assessment item for keeping the vehicle's physical geometric boundaries away from local obstacles; This is a risk penalty term for the local semantic potential field; This indicates the system adjustment weight for each evaluation sub-item.

[0047] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0048] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0049] The present invention has the following beneficial effects: This method improves the robustness of positioning under severe weather conditions: Through the adaptive covariance reweighting mechanism, the adverse effects of environmental degradation such as dense fog and strong backlight on visual feature matching can be effectively reduced, and the pose estimation is smoothly transferred to the inertial unit, which helps the system maintain a high global continuous positioning accuracy under complex interference.

[0050] Achieving cross-modal cognition of hidden dangers: Breaking through the limitations of traditional cost maps that mainly rely on the geometric boundaries of physical entities, by transforming models such as toxic gas leaks into continuous digital potential fields and using a maximum value strategy for layer aggregation, the navigation engine's semantic recognition capability for intangible chemical and thermodynamic risks is enhanced.

[0051] Facilitating proactive risk avoidance decisions with low exposure risk: The improved global heuristic operator effectively overcomes the limitations of greedy optimization based on pure Euclidean distance, enabling the planning engine to plan a safer continuous detour topology path by increasing some travel distance in the multi-objective trade-off between physical distance and chemical risk, thereby improving the overall safety of special robot inspections. Attached Figure Description

[0052] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0053] Figure 1 This is a schematic diagram of the overall architecture of a visual SLAM-based intelligent navigation system for hazardous environments.

[0054] Figure 2 A schematic diagram of the hierarchical danger perception semantic cost map architecture and cross-modal fusion mechanism.

[0055] Figure 3 This is a schematic diagram of the core algorithm flow and closed-loop feedback of the adaptive hazard perception navigation system. Detailed Implementation

[0056] This invention provides an intelligent navigation system for hazardous outdoor chemical plant environments based on visual simultaneous localization and mapping. The system is configured to execute the following control flow:

[0057] Step 1, Perform spatial cognition and mapping localization based on adaptive multi-source tight coupling: Establish a nonholonomic constrained kinematic model of the mobile robot, fuse visual odometry, wheel odometry and inertial measurement unit data based on extended Kalman filter, and dynamically adjust the confidence level of visual observation using environmental degradation adaptive covariance reweighting mechanism to output a robust global pose estimate.

[0058] Step 2, construct a hierarchical hazard perception semantic cost map and perform cross-modal aggregation: on the basis of the global physical static base map, use the artificial potential field decay model to quantify the hidden chemical hazard sources into a continuous semantic risk layer, and use the local maximum aggregation strategy to deeply integrate physical obstacle constraints and semantic risk constraints;

[0059] Step 3: Execute a global replanning algorithm based on minimizing risk exposure integral: Improve the cost evaluation function of the traditional heuristic search algorithm by introducing the cumulative chemical exposure risk as a severe penalty, forcing the planning engine to actively abandon physical shortcuts across high-risk areas and plan a global safe topology trajectory.

[0060] Step 4: Perform local trajectory collaborative optimization combined with dynamic constraints: Generate control commands in the velocity sampling space, introduce local risk rejection functions for evaluation and screening, and coordinate the global path to achieve smooth and safe control of the robot.

[0061] The system includes the following components:

[0062] High-performance embedded processors are used to run and implement adaptive multi-source fusion mapping algorithms, cost map construction algorithms, and global and local path planning algorithms.

[0063] A multi-source sensor system, including a depth camera, a 2D LiDAR, an inertial measurement unit, and a wheeled odometry, is used to provide real-time high-frequency kinematic state, 3D point cloud, and environmental data of the mobile robot;

[0064] The motor drive control module is used to receive and execute the linear velocity and angular velocity commands obtained in step 4 to drive the mobile robot chassis motor.

[0065] In step 1, the kinematic model is based on a discrete-time series, and the robot's state prediction equation is defined as:

[0066] (1),

[0067] (2),

[0068] (3),

[0069] in, Mobile robots The horizontal and vertical coordinates in the global coordinate system at any given time. For heading angle, The forward speed of the chassis. Angular velocity of rotation For time step, , , These represent the horizontal coordinate, vertical coordinate, and heading angle of the mobile robot at the previous moment, respectively. The main function of formulas (1), (2), and (3) is based on the linear velocity issued by the system. With angular velocity Instructions, at a given time interval The robot's theoretical pose for the next moment is calculated and updated internally, providing a basic kinematic inference for subsequent localization and control.

[0070] The general theoretical expression model for the prior state prediction and error covariance prediction matrix of the extended Kalman filter is defined as follows:

[0071] (4),

[0072] (5),

[0073] in, To control the input vector, This is the state transition function. Let F be the expected state vector at the current moment; F is the state transition function with respect to the state vector. The Jacobian matrix is ​​obtained by solving for Q, which is the system process noise covariance matrix, and T denotes the transpose. This is the expected state vector from the previous time step. This is the error covariance prediction matrix at the current time. Let be the error covariance matrix of the previous time step. The purpose of formulas (4) and (5) is to construct a recursive mathematical framework, which enables the system to continuously use new sensor observation data to correct prior predictions, thereby solving for a more accurate global pose in the multidimensional state space.

[0074] In step 1, the extended Kalman filter sets the observation constraint state of each sensor using a Boolean diagonal matrix during multi-source data assimilation. The observation vector of the adaptive visual odometry is defined as... The observation vector of the chassis inertial measurement unit is defined as follows: The system avoids observation divergence caused by wheel slippage by physically isolating the IMU from the absolute position observations.

[0075] In step 1, the adaptive covariance reweighting mechanism dynamically adjusts the diagonal elements of the error covariance matrix based on the environmental feature quality. When in the open-loop verification stage without external feature feedback, the number of effective feature points... Prediction is performed using a spatial distance attenuation model. The environmental view abrupt change threshold coordinates are set as follows: When the robot's global lateral coordinates When the number of feature points is estimated, the formula is:

[0076] (6),

[0077] in, This is the baseline number of feature points in the normal region. For linear decay rate, This represents the lower limit of the feature point.

[0078] Subsequently, the diagonal elements of the visual observation noise covariance matrix were dynamically adjusted. :

[0079] (7),

[0080] in, Based on baseline observation noise, This is the sensitivity adjustment coefficient. When visual degradation leads to... During decay, The weight of visual observation in system state updates is increased accordingly, thereby reducing the weight of visual observation. Formulas (6) and (7) are used to give the system the ability to self-diagnose perception degradation, actively reduce the trust weight of visual data in poor weather conditions, and effectively reduce the impact of erroneous observations on the positioning system.

[0081] In step 2, the hierarchical hazard perception semantic cost map includes a dynamic expansion layer and a semantic risk layer. The cost of the dynamic expansion layer... Euclidean distance from the obstacle surface It decays exponentially:

[0082] (8),

[0083] in, The scaling decay factor determines how steeply the penalty value decreases with distance; The physical boundaries of collisions were established based on the safe inner tangent radius of the mobile robot. is the Euclidean distance from the current evaluation point to the surface of the physical obstacle. The function of formula (8) is to generate a virtual buffer repulsion field around the physical obstacle to reduce the probability of physical collision when the robot is moving along the wall.

[0084] In step 2, when constructing the semantic risk layer, the system first builds a general multi-source risk field physical foundation model in a continuous space. For any evaluation point in the space... Its overall risk value It is composed of the superposition of the exponentially decaying potential fields of each hazard source: (9),

[0085] in, For the first The absolute spatial coordinates of a high-risk semantic detection source As assessment point With the A high-risk semantic detection source The Euclidean distance between them The initial weighting coefficients represent the inherent lethality of the hazard source. For environmental attenuation constant. Formula (9) provides a physical explanation for the spatial radiation threat of hazardous sources from a theoretical perspective, transforming the abstract leakage hazard into a continuously distributed risk potential field.

[0086] However, when transforming a continuous theoretical mathematical model into a two-dimensional discrete grid map (Costmap) required for the underlying navigation of a mobile robot, directly using theoretical superposition (summation) can easily lead to the overflow of cost values ​​at the boundaries of multiple hazard sources into the valid range of the underlying data structure (usually 0~253), thus creating "virtual dead ends" that hinder the robot's normal passage. Therefore, in the engineering implementation of this system, a dimensionality reduction strategy combining "single-source discrete mapping" and "multi-source local maximum aggregation" is adopted.

[0087] First, for any single, independent hazard source in the environment, the system uses an artificial potential field distance attenuation model for discretization mapping. Let the current evaluation grid coordinates be... The center of the hazard source is Define the fatal exposure radius With edge warning radius Its core fatal zone base penalty cap is set at 253, and semantic risk potential field cost. The piecewise analytical expression of is:

[0088] (10),

[0089] in, As assessment point absolute coordinates of the i-th high-risk semantic detection source space Euclidean distance, , The rounding down operation is used to prevent floating-point overflow and to match the discretization threshold of the local cost map. Formula (10) realizes the engineered dimensionality reduction mapping of the single-source theoretical risk model to the two-dimensional discrete grid space, providing a legitimate underlying data foundation for subsequent multi-source aggregation.

[0090] In the actual engineering deployment of the system, the potential field matrix generated by the semantic risk layer can be published through the nav_msgs / OccupancyGrid message bus of the Robot Operating System (ROS). Due to the limitations of the underlying data structure of this message format (int8, value range -128 to 127), the basic penalty upper limit of the core fatal zone is set to the maximum legal occupancy probability value of 100 (i.e., 100% occupancy) in the engineering implementation. When this cost value is received by the global cost map (Costmap2D) engine, the underlying matrix memory mapping mechanism automatically maps the input value of 100 to the internal absolute collision cost value of 254, linearly mapping 0~99 to 0~253. Therefore, through this dimensionality reduction mapping mechanism, this system effectively ensures a tight consistency between theoretical derivation (cost range 0~253) and physical bus transmission (probability range 0~100). It should be noted that the above-mentioned underlying mapping mechanism based on the ROS system is only a specific implementation example of the present invention. Those skilled in the art can adopt other equivalent data mapping rules according to the actual underlying communication framework. These modifications and substitutions are all within the protection scope of the present invention.

[0091] When M independent hidden risk sources exist simultaneously in the environment, to ensure the absolute boundedness of the underlying cost data and the safety of obstacle avoidance, the system adopts a "highest risk-dominated" mechanism instead of physical overlay. Let the center coordinates of the m-th risk source be... The semantic risk layer is in coordinates The ultimate value of the place Obtained by iterating through the global maximum value function:

[0092] (11),

[0093] Subsequently, the system merges the layers into a global comprehensive cost map using a local maximum aggregation strategy. Achieving a unified cost metric for both physical constraints and chemical soft constraints:

[0094] (12),

[0095] Formula (12) uses a local maximum aggregation strategy to aggregate the static physical layer. Dynamic expansion layer With semantic risk layer The cost value is compared and extracted across modes. Formulas (10), (11) and (12) work together to reasonably map the continuous theoretical mathematical model onto the two-dimensional occupied grid map, realizing the effective unification of physical constraints and chemical constraints in the underlying data structure.

[0096] In step 3, when calculating the cumulative chemical exposure risk penalty, the robot's current pose in the continuous physical space needs to be considered. The discretization is mapped to a one-dimensional memory array of the two-dimensional occupied raster map. Let the physical origin coordinates of the global cost map be... The raster resolution is If the horizontal grid width of the map is W, then the mapped memory index... The computational model is defined as follows:

[0097] (13),

[0098] In the formula, The floor operator is used to achieve unbiased quantization mapping from continuous space to discrete grid topology. The main function of formula (13) is to provide an accurate starting point index for the algorithm and build a bridge between the physical world and the algorithm's memory structure.

[0099] In step 3, based on the mapping index, the global path planner executes the improved risk-aware A* algorithm, whose reconstructed macroscopic multi-objective general evaluation model is defined as:

[0100] (14),

[0101] in For nodes The overall evaluation score Represents the physical path cost. To amplify the weight of overall risk sensitivity, This represents the cumulative risk penalty for nodes. This model uses heuristic cost estimation. Combining the underlying discrete grid cost map, the macroscopic model is specifically expanded into a reconstructed heuristic cost evaluation equation. Defined as: (15)

[0102] in, The cost is the mesh distance for physical translation. This represents the sum of the semantic risk values ​​of all rasters traversed on the search backtracking branch. To amplify the weight of overall risk sensitivity, The Euclidean distance to the target point is heuristically estimated. Formulas (14) and (15) optimize the logic of traditional algorithms that only pursue the physical shortest path. By increasing the mathematical penalty for crossing dangerous areas, the planning engine is prompted to reduce the search priority of high-risk routes, thereby outputting a globally safer detour topology trajectory.

[0103] In step 4, the Local Dynamic Window (DWA) method directly generates virtual linear velocity and angular velocity control pairs within the velocity sampling space, and its multi-objective collaborative evaluation function... Reconstructed as:

[0104] (16)

[0105] In the formula, For sampling rate pairs; A cost evaluation item to ensure the robot's heading conforms to the global topology trajectory; This is the positive translational velocity excitation term; To ensure the safe clearance assessment item for keeping the vehicle's physical geometric boundaries away from local obstacles; This is a newly added local semantic potential risk penalty term in this study; These correspond to the system adjustment weights of each evaluation sub-item. In this multi-objective game model, if the velocity sampling command in the micro-adjustment might enter a high-chemical-risk region, it will usually be due to... The increase in total score This reduces the execution priority of such high-risk instructions during the evaluation and screening phase. Formula (16) serves as a bottom-level constraint mechanism, effectively rejecting dynamic instructions that may enter areas with higher chemical risks during the micro-speed adjustment phase, so as to ensure that the final control law issued to the motor can balance driving smoothness and better risk avoidance safety.

[0106] In one specific embodiment of this method, the system includes:

[0107] I. System Composition

[0108] This embodiment provides an adaptive risk consistency intelligent navigation system, mainly mounted on an outdoor mobile robot (AGV) platform. The system includes the following components:

[0109] 1. Hardware environment:

[0110] The host computer computing hub adopts a high-performance embedded edge computing platform (such as an industrial control computer equipped with multi-core CPU and GPU computing power) to coordinate the operation of visual SLAM fusion mapping and multi-objective path planning algorithms with high computational complexity.

[0111] Multi-source sensor system: includes an industrial-grade high-fidelity depth camera (RGB-D), a two-dimensional single-line LiDAR, a high-frequency six-axis inertial measurement unit (IMU), and a chassis-mounted wheeled odometer encoder.

[0112] Lower-level actuator: Includes a motor drive control module and a differential chassis, used to receive linear velocity and angular velocity commands from the upper-level computer, and execute low-level PID control and motor drive.

[0113] 2. Software environment:

[0114] The Robot Operating System (ROS) is deployed on the Ubuntu operating system. All core algorithms are encapsulated as independent ROS node processes, and data flows between nodes are facilitated through high-frequency topics. The control loop frequency is typically set to greater than 10Hz to ensure real-time dynamic obstacle avoidance.

[0115] II. Specific Implementation Steps

[0116] This implementation details the workflow of an intelligent navigation and path planning system based on adaptive multi-source tight coupling and hierarchical hazard-aware semantic cost maps. After startup, the system executes the following steps sequentially, forming a real-time closed loop:

[0117] Step 1: System power-on and full initialization;

[0118] 1. Hardware Power-On and Self-Test: Close the main power supply of the mobile robot to power the host computer, slave computer motor drive modules, and multi-source sensor system. After the host computer operating system starts, it completes the interface initialization and communication link detection of the depth camera, LiDAR, and IMU to ensure normal acquisition of underlying data.

[0119] System node startup and parameter loading: In the ROS framework, the multi-source data assimilation node, cost map construction node, global path planning node, and local trajectory optimization node are sequentially awakened. The system synchronously loads preset physical and safety parameters, mainly including: the robot's physical shape contour and safe inner radius, maximum movement speed limit, lethal exposure radius and edge warning radius of hazardous sources such as toxic gases, and parameters such as feature point attenuation thresholds used to determine visual environment degradation.

[0120] Step 2: Multi-source sensor data acquisition and adaptive pose calculation;

[0121] 1. High-frequency acquisition of multi-source heterogeneous data: The system acquires environmental information and its own status in parallel. The depth camera front end outputs visual feature points and relative pose estimation; the lidar provides scanned point cloud slices of physical obstacles; the IMU and chassis encoder provide high-frequency angular velocity, linear acceleration, and wheel track estimation data.

[0122] 2. Environmental degradation detection and adaptive reweighting: The core node of perception fusion monitors the extraction and matching quality of effective feature points from the visual front end in real time. When the robot enters an area with poor weather conditions (such as dense fog in a chemical plant or strong backlight), causing the number of visual feature points to drop and approach the set lower threshold, the system will trigger an adaptive covariance reshaping mechanism in a timely manner.

[0123] 3. Optimal Global Pose Estimation Output: During multi-source data assimilation using the underlying Extended Kalman Filter (EKF), the system dynamically and significantly expands the noise covariance matrix of the visual observation data based on the aforementioned degradation state. This operation substantially reduces the weight of visual data affected by environmental interference at the underlying logic level, smoothly transferring the trust of system localization to the IMU and wheeled odometry, which are less affected by weather interference. Ultimately, it calculates the robot's continuous global absolute pose with strong anti-interference capability and low drift.

[0124] Step 3: Construction and cross-modal aggregation of hierarchical hazard perception semantic cost map;

[0125] 1. Construction of Basic Static and Dynamic Expansion Layers: Utilizing occupancy grid data output from LiDAR and SLAM systems, a static physical layer is generated in a 2D map to effectively mark the impassable boundaries of physical walls and facilities. Simultaneously, based on the robot's safe inner radius, a dynamic expansion buffer layer with exponentially decreasing cost is generated outwards to reduce the probability of collisions when the robot travels close to walls.

[0126] 2. Generation of Implicit Semantic Risk Layer: When the system obtains a chemical hazard source without a physical rigid entity (such as the coordinates of the center of a toxic gas leak) through external industrial IoT or visual semantic segmentation, it invokes an artificial potential field model. Centered on the leak coordinates, a higher risk cost is assigned within the lethal exposure radius, and a linearly smooth penalty cost gradient is generated within the edge warning radius. This explicitly transforms the abstract chemical hazard into a continuous digital potential field layer.

[0127] 3. Cross-modal Local Maximum Aggregation: The system traverses every raster pixel in the map, performing a bottom-up comparison of the aforementioned static physical layer, dynamic dilation layer, and semantic risk layer. At any coordinate, the system selects the maximum cost value from each layer and assigns it to the final global comprehensive master cost map. This strategy logically treats high-concentration "chemical soft constraints" as equivalent to "physical hard obstacles," effectively compensating for the cognitive blind spots that traditional navigation systems often struggle to identify intangible chemical risks.

[0128] Step 4: Global reprogramming based on minimizing the risk exposure integral;

[0129] 1. Coordinate Mapping and Search Initialization: Map the robot's current continuous physical coordinates to a memory array of a two-dimensional discrete grid map, which serves as the starting node for global path optimization and locks the task target node.

[0130] 2. Multi-objective topology optimization and risk penalty: The global planner activates an improved risk-aware A* algorithm to expand the search tree. In addition to the traditional physical grid movement distance and target estimated distance, the algorithm traverses all grids traversed by each virtual search branch in real time, accumulating the semantic risk value contained therein in the comprehensive cost map.

[0131] 3. Proactive Detour and Safe Path Generation: When the planner detects a path with a shorter physical distance attempting to cross a toxic gas field, its accumulated chemical risk penalty will increase significantly. This prompts the algorithm to automatically reduce the search priority of this high-risk branch and eliminate it. After a massive game among nodes, the algorithm usually outputs a globally safe detour path that compromises the physical travel distance but significantly reduces the risk of chemical exposure.

[0132] Step 5: Local dynamics co-optimization and underlying motor control;

[0133] Velocity space sampling and local evaluation: Guided by the global safe path, the local trajectory planner adopts the dynamic window method (DWA) and combines the robot's acceleration and deceleration physical limits to sample and generate multiple sets of virtual linear velocity and angular velocity control command pairs in a large number of samples within the current time window.

[0134] Local risk rejection mechanism: A multi-objective evaluation function comprehensively scores each set of sampled trajectories. In addition to rewarding heading alignment and physical safety clearance, the system introduces a strict local risk rejection mechanism. Any velocity sampling command that might enter a high-chemical-hazard zone during micro-adjustments will be assigned a low score and discarded.

[0135] Closed-loop control and task completion: The system selects the optimal speed command with the highest overall score and safety constraints, and sends it to the lower-level computer, where the PID controller drives the chassis motor to execute the action. The entire system continuously loops through steps two through five at a set frequency to achieve high-frequency feedback in dynamic environments until the robot safely reaches the final inspection target point.

[0136] III. Comparison of Specific Implementation Examples and Experimental Results

[0137] To explain in detail how this invention solves the technical problems of "location loss due to severe weather" and "blind penetration of hidden chemical risks" in real-world chemical plants, the following is a detailed explanation. Figures 1 to 3 This invention provides a complete specific embodiment and comparative experimental data to demonstrate the significant advantages of the present invention over the prior art.

[0138] 1. Specific steps and parameter derivations, with reference to the attached diagram.

[0139] (1) Known environment and system configuration parameter settings

[0140] Known environment and system configuration parameters:

[0141] Perception degradation parameter: Set the horizontal axis of the threshold for sudden changes in visibility due to dense fog. =10.0m; Number of normal feature points as reference =500; Characteristic linear decay rate =80; Lower limit of feature points =20; Basic observation noise =0.01; Visual sensitivity adjustment coefficient =5.0.

[0142] Chemical risk parameters: The system receives the coordinates of the toxic gas leak source ( = (20.0, 5.0); Blinding lethal exposure radius =1.0m; Edge warning radius =5.0m; the maximum base penalty is 253.

[0143] Global planning parameters: A* algorithm risk sensitivity amplification weights =2.0.

[0144] (2) Scenario assumptions and results obtained in each step

[0145] Combination Figure 3 The algorithm flow and closed-loop feedback loop, when the mobile robot enters the global coordinate system. In the dense fog region (15.0, 5.0), when the global planner evaluates the adjacent grid node (18.0, 5.0), the execution and results of each step of the system are as follows:

[0146] The result obtained in step 1 (adaptive fusion localization):

[0147] Robot x-axis =15.0> We have entered a dense fog zone. The system first estimates the number of currently valid feature points. Based on this, the system dynamically reshapes the diagonal elements of the visual observation noise covariance matrix. =max(0.01,5.0×1 / 100)=0.05;

[0148] The results show that the visual noise matrix value has significantly increased from the basic 0.01 to 0.05. The system automatically deprives the trusted weight of the disturbed visual data at the bottom layer and smoothly transfers the positioning control to the inertial unit.

[0149] The result obtained in step 2 (hierarchical cost map aggregation):

[0150] Assess the distance from node (18.0, 5.0) to the center of the toxic gas source. =2.0m, determine the distance to this node: 1.0 < A value less than 5.0 indicates that the node is not in the core critical zone, but is in the high-risk warning decay zone. Substituting... Figure 2 The piecewise decay formula for the semantic risk potential field, and its penalty cost: .

[0151] The results of the steps show that the invisible toxic gas with no physical collision volume was successfully transformed into a recognizable high-risk soft constraint cost of up to 189 in the bottom layer.

[0152] The result obtained in step 3 (risk perception global planning):

[0153] Set the physical geometry cost of this node =3.0, estimated heuristic distance =15.0. Substituting into the reconstructed evaluation equation, the total evaluation score of the node is 15.0. : =3.0 + 2.0 × 189 + 15.0 = 3.0 + 378.0 + 15.0 = 396.0;

[0154] The results show that the originally physically accessible node received a high score of 396.0 due to the high risk penalty, and was directly downgraded and eliminated from the algorithm's open list.

[0155] The final result: the planning engine actively abandoned this shortcut and output a global safe topology detour trajectory away from the center of the poison gas to the local planner (DWA), and the chassis motor performed safe avoidance accordingly.

[0156] 2. The real-world technical problems solved

[0157] Through the above specific implementation steps, this invention effectively addresses two major technical pain points in real-world factories: First, traditional AGVs are prone to "visual perception degradation and positioning drift" due to a significant reduction in feature points under extreme weather conditions (heavy fog, steam); second, traditional A* algorithms, when optimizing paths, rely excessively on the shortest physical distance, which can easily lead to "blindly penetrating high-risk areas" when facing dangers without physical boundaries, such as toxic gas leaks or high-temperature radiation.

[0158] 3. Superiority compared to existing technologies (experimental data comparison)

[0159] To quantify and verify the technical effectiveness of this invention, comparative experiments were conducted on the Gazebo 3D physics simulation platform. Traditional pure vision SLAM and traditional pure geometry-driven A* navigation algorithms were used as a control group of existing technologies, and independently repeated experiments were conducted to compare them with the intelligent navigation system experimental group of this invention.

[0160] First, experimental results demonstrate that the proposed technique exhibits significantly improved localization robustness compared to traditional visual SLAM. In a simulated high-concentration smoke-induced visual degradation scenario, the control group's traditional visual SLAM exhibited severe localization divergence, with a root mean square error (RMSE) of 3.229 m and a maximum local drift of 11.212 m. In contrast, the proposed invention, by introducing an adaptive reweighting mechanism, reduced the RMSE to 0.008 m and controlled the maximum local drift to 0.090 m. This indicates that the proposed system can maintain high-precision, stable, and continuous localization even under severe weather conditions.

[0161] Secondly, experimental results demonstrate that the proposed technology significantly improves safety and survival rate compared to traditional algorithms. In maze avoidance tasks involving multi-branch, hidden toxic gas leaks, the control group's traditional algorithm often directly plans routes through the toxic gas zone. Although its average geometric path length is only 5.500m, the average cumulative semantic risk cost is as high as 7150, resulting in an extremely low safe passage success rate (0%). In contrast, the algorithm of this invention can proactively plan detour trajectories that conform to the outer contour of the risk potential field. Although the average geometric path length is moderately increased to 11.511m, it successfully reduces the average cumulative chemical risk cost to 0, increasing the safe passage success rate to 98%.

[0162] The experimental data and results above demonstrate that the cross-modal cognition and multi-objective cooperative control technology proposed in this invention effectively overcomes the limitations of traditional industrial AGVs in perceptual interference and lack of semantic risk avoidance capabilities in harsh environments. Compared with existing technologies, it has substantial features and significant progress in improving the safety of high-risk inspections of special robots.

[0163] Figure 1This paper presents a flowchart of a closed-loop control system for intelligent navigation based on hazard perception. The scheme employs a multi-level planning and control framework, showcasing the entire closed-loop logic of the system, from environmental risk awareness and multi-objective optimization to dynamic trajectory execution. The diagram mainly includes: a layered cost map (including a semantic risk layer), a global path planner, a local trajectory planner, and a velocity control command output module. The specific workflow is as follows: the global path planner receives the task target point and current pose, and based on the comprehensive risk base provided by the layered cost map, plans a global safe topology path using the reconstructed multi-objective evaluation model. Subsequently, guided by the global path, the local trajectory planner, combined with real-time local cost updates, generates a final control command output including linear and angular velocities. After this command drives the robot to generate physical motion, the cost map is updated in a closed loop through sensor feedback, thus forming a safe navigation system capable of dynamically rejecting hidden hazards in real time.

[0164] Figure 2 This is a schematic diagram of a hierarchical hazard perception semantic cost map architecture. This functional architecture diagram illustrates how multi-dimensional environmental information is reduced in dimensionality and fused across modalities into a unified navigation constraint. The workflow begins by acquiring physical obstacle information from traditional physical sensors (LiDAR / depth cameras), constructing a "static physical layer" and a "dynamic expansion layer" as hard constraints. Simultaneously, the coordinates of hidden hazardous targets are obtained from industrial IoT or high-risk semantic detection sources, and transformed into a "semantic risk layer" as soft constraints through continuous risk gradients in an artificial potential field. These three layers are ultimately fused across modalities using a bottom-up local maximum aggregation strategy to form a global / local master cost map (comprehensive cost value 0~254), providing a unified cognitive foundation for subsequent safe path planning.

[0165] Figure 3 The core algorithm flow and closed-loop feedback loop of the adaptive hazard perception navigation system are demonstrated. It is mainly used for trajectory cooperative control of mobile robots in complex and harsh environments. The functions of each module are described below:

[0166] 1. Adaptive Fusion Positioning

[0167] Input sensor data from various hardware devices; the system detects that poor weather conditions have reduced the number of visual feature points ( During the decay process, the visual observation noise matrix is ​​appropriately increased through a dynamic covariance reweighting mechanism. This allows for the reasonable adjustment of the trust weights in multi-source fusion, resulting in the output of stable and reliable global pose and point cloud data to the next module.

[0168] 2. Layered cost map aggregation

[0169] Input the pose from the previous module and the coordinates of externally introduced hazards, and use the spatial exponential decay model to calculate the semantic risk cost at each coordinate point. Then, a local maximum traversal function is used to effectively merge the physical rigidity layer and the semantic risk layer, outputting the global comprehensive cost. ).

[0170] 3. Risk perception and overall planning

[0171] Combining the comprehensive cost map output by the pre-module, the equation is evaluated using the reconstructed and improved A-star heuristic search algorithm. The evaluation model incorporates the cumulative chemical exposure risk as a significant penalty, prompting the algorithm to reduce the search priority for traversing high-risk areas and output a globally safe topology path to the local planner.

[0172] 4. Local trajectory cooperative optimization

[0173] Guided by the global security path, the reconstructed local dynamic window method multi-objective evaluation function (G(v, The function evaluates the sampling velocity command. It introduces a risk rejection mechanism, effectively reducing the priority of commands that enter high-risk regions. The optimal linear velocity (v) and angular velocity (v) selected are... Not only are the actuator commands sent to the motor drive module, but the robot's state update after movement is also fed back to the perception fusion module, forming a complete data closed-loop control.

[0174] This invention provides an intelligent navigation system for hazardous outdoor chemical plant environments based on visual synchronous positioning and mapping. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. An intelligent navigation system for hazardous outdoor chemical plant environments based on visual simultaneous localization and mapping, characterized in that, include: A multi-source sensor system is used to collect real-time motion state data and surrounding environment feature data of a mobile robot. The motor drive control module is used to drive the movement of the mobile robot chassis. A processor, communicatively connected to the multi-source sensor system and the motor drive control module, is configured to execute the following control flow: Step 1, Perform spatial cognition and mapping localization based on adaptive multi-source tight coupling: Establish a nonholonomic constrained kinematic model of the mobile robot, fuse visual odometry, wheel odometry and inertial measurement unit data based on extended Kalman filter, and dynamically adjust the confidence level of visual observation using environmental degradation adaptive covariance reweighting mechanism to output a robust global pose estimate. Step 2, construct a hierarchical hazard perception semantic cost map and perform cross-modal aggregation: on the basis of the global physical static base map, use the artificial potential field decay model to quantify the hidden chemical hazard sources into a continuous semantic risk layer, and use the local maximum aggregation strategy to deeply integrate physical obstacle constraints and semantic risk constraints; Step 3: Execute a global replanning algorithm based on minimizing risk exposure integral: Improve the cost evaluation function of the traditional heuristic search algorithm by introducing the cumulative chemical exposure risk as a severe penalty, forcing the planning engine to actively abandon physical shortcuts across high-risk areas and plan a global safe topology trajectory. Step 4: Perform local trajectory collaborative optimization combined with dynamic constraints: Generate control commands in the velocity sampling space, introduce local risk rejection functions for evaluation and screening, and coordinate the global path to achieve smooth and safe control of the robot.

2. The system as described in claim 1, characterized in that, The multi-source sensor system includes a wheeled odometer; The processor is configured to perform kinematic state prediction based on discrete time series: receiving the chassis forward linear velocity from the wheel odometer in real time. With rotational angular velocity And based on the set discrete time step Calculate and update the theoretical pose of the mobile robot in the next moment; The theoretical pose is used to provide kinematic deduction basis for the subsequent real-time positioning and control of the motor drive control module. The state prediction equation is as follows: (1)， (2)， (3)， in, , Let x and y be the x and y coordinates of the mobile robot in the global coordinate system at time t, respectively. Let t be the heading angle. For discrete time steps, , , These are the x-coordinate, y-coordinate, and heading angle of the mobile robot at time t, respectively, in the previous time step.

3. The system as described in claim 2, characterized in that, In step 1, the general theoretical expression model of the prior state prediction and error covariance prediction matrix of the extended Kalman filter is defined as follows: (4)， (5)， in, To control the input vector, This is the state transition function. Let F be the expected state vector at the current moment; F is the state transition function with respect to the state vector. The Jacobian matrix is ​​obtained by solving for Q, which is the system process noise covariance matrix, and T denotes the transpose. This is the expected state vector from the previous time step. This is the error covariance prediction matrix at the current time. This is the error covariance matrix of the previous time step.

4. The system as described in claim 3, characterized in that, In step 1, the environmental degradation adaptive covariance reweighting mechanism dynamically adjusts the diagonal elements of the error covariance matrix based on environmental characteristic quality: the environmental field of view abrupt change threshold coordinates are set as follows. When the robot's global lateral coordinates At that time, the number of feature points The estimation formula is: (6)， in, This is the baseline number of feature points in the normal region. For linear decay rate, The lower limit of the feature point; Subsequently, the diagonal elements of the visual observation noise covariance matrix were dynamically adjusted. : (7)， in, Based on baseline observation noise, This is the sensitivity adjustment coefficient.

5. The system as described in claim 4, characterized in that, In step 2, the hierarchical hazard perception semantic cost map includes a static physical layer, a dynamic expansion layer, and a semantic risk layer; the cost of the dynamic expansion layer... Euclidean distance from the obstacle surface It decays exponentially: (8)， in, To scale the decay factor, The distance from the current evaluation point to the surface of the physical obstacle is the Euclidean distance. Let e ​​be the safe inner radius of the mobile robot, and e be a natural constant. When constructing the semantic risk layer, the system first builds a general multi-source risk field physical foundation model in a continuous space, for any evaluation point in the space. Global risk value It is composed of the superposition of the exponentially decaying potential fields of each hazard source: （9）， in, For the first The absolute spatial coordinates of a high-risk semantic detection source As assessment point With the A high-risk semantic detection source The Euclidean distance between them For the first The initial weighting coefficients for the inherent lethality of each hazard source. Let be the environmental degradation constant, and exp be the natural exponential function.

6. The system as described in claim 5, characterized in that, In step 2, discretization is performed based on the continuous risk field model. The semantic risk layer adopts an artificial potential field distance decay model: let the current evaluation grid coordinates be... The center of the hazard source is Define the fatal exposure radius With edge warning radius Set the core fatal zone base penalty cap to 253, and the semantic risk potential cost. The piecewise analytical expression of is: (10)， Among them, the distance parameter ; The system merges various layers into a global comprehensive cost map using a local maximum aggregation strategy. Achieving a unified cost metric for both physical constraints and chemical soft constraints: (11)， in, To evaluate grid coordinates The static physical layer value at that location.

7. The system as described in claim 6, characterized in that, In step 3, the global reprogramming algorithm based on minimizing the risk exposure integral reconstructs the heuristic evaluation system of the A* heuristic search algorithm, and the reconstructed macroscopic multi-objective general evaluation model Defined as: (12) ， in, For nodes The overall evaluation score Represents the physical path cost. This represents the cumulative risk penalty for nodes. For heuristic cost estimation; Increase the weight of overall risk sensitivity; Combining the underlying discrete grid cost map, the macro model is specifically expanded into a reconstructed heuristic cost evaluation equation. Defined as: （13）， in, The cost is the mesh distance for physical translation. This represents the sum of the semantic risk values ​​of all rasters traversed on the search backtracking branch.

8. The system as described in claim 7, characterized in that, In step 4, the local trajectory cooperative optimization combined with dynamic constraints directly generates virtual linear velocity and angular velocity control pairs in the velocity sampling space through the local dynamic window method, and a multi-objective cooperative evaluation function. Reconstructed as: （14）， in, For sampling linear velocity, For sampling angular velocity, For sampling rate pairs; A cost evaluation item to ensure the robot's heading conforms to the global topology trajectory; This is the positive translational velocity excitation term; To ensure the safe clearance assessment item for keeping the vehicle's physical geometric boundaries away from local obstacles; This is a risk penalty term for the local semantic potential field; This indicates the system adjustment weight for each evaluation sub-item.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 8.

Images