Safety analysis method and system for complex driving scenarios of warehouse robots

Abstract

The present invention discloses a safety analysis method and system for complex driving scenarios of AGVs. The method comprises the following steps: obtaining driving monitoring information, perceiving the driving environment of a target AGV based on the driving monitoring information, and obtaining environmental perception information; performing dynamic obstacle detection based on the environmental perception information, detecting dynamic obstacles in the target AGV's real-time driving environment, and predicting the motion trajectory of the detected dynamic obstacles to obtain trajectory prediction information; performing scenario risk modeling based on the environmental perception information and motion trajectory information, and analyzing the driving risk of the target AGV following a predetermined driving trajectory; and, when a driving risk is detected within the predetermined driving trajectory, adjusting the driving trajectory by combining the environmental perception information and motion trajectory prediction information, generating a driving strategy for driving assistance. This method improves the accuracy and intelligence of driving safety analysis for AGVs in complex dynamic environments.

Description

Technical Field

[0001] The present invention relates to the technical field of warehouse robots, and in particular to a safety analysis method and system applied to complex driving scenarios of warehouse robots. Background Art

[0002] With technological advancements, AGV warehouse robots have been widely used in smart warehousing, manufacturing logistics, and industrial automated transportation systems. The safety of their driving paths directly affects overall operational efficiency and equipment stability. However, in complex driving scenarios, such as dense dynamic obstacles, mixed traffic of humans and robots, narrow passages, and changing lighting environments, the safe driving of warehouse robots faces many challenges. Traditional path planning and safety control methods typically rely on a single sensor (such as a lidar or visual camera) for environmental perception and adjust the path through preset rules or simple optimization algorithms. However, such methods have limited adaptability to the environment and struggle to accurately predict driving risks in real time. This is especially true when responding to emergencies (such as the sudden appearance of obstacles, uneven ground, or accumulated system errors). These methods can be delayed in response and have unreasonable path planning, potentially leading to collisions, stagnation, and even equipment damage.

[0003] Therefore, how to improve the safe driving ability of warehouse robots in complex environments and ensure the real-time response and driving safety of warehouse robots in the event of emergencies is an urgent problem to be solved. Summary of the Invention

[0004] The present invention overcomes the shortcomings of the prior art and provides a safety analysis method and system for complex driving scenarios of storage robots.

[0005] To achieve the above objectives, the present invention provides a first aspect of a safety analysis method for a warehouse robot in a complex driving scenario, comprising:

[0006] Acquiring driving monitoring information based on a driving monitoring module set on the target warehouse robot, and sensing the driving environment of the target warehouse robot according to the driving monitoring information to obtain environmental perception information;

[0007] Performing dynamic obstacle detection based on the environmental perception information to detect dynamic obstacles in the real-time driving environment of the target warehouse robot, and predicting the motion trajectory of the detected dynamic obstacles to obtain trajectory prediction information;

[0008] Performing scenario risk modeling based on the environmental perception information and the motion trajectory information, analyzing the driving risk of the target warehouse robot following the predetermined driving trajectory, and obtaining driving risk analysis information;

[0009] When a driving risk is detected in the planned driving trajectory, the driving trajectory is adjusted by combining environmental perception information and action trajectory prediction information, and a driving strategy is generated for driving assistance.

[0010] In this solution, the driving monitoring module set on the target warehouse robot obtains driving monitoring information, and perceives the driving environment of the target warehouse robot according to the driving monitoring information to obtain environmental perception information, which specifically includes:

[0011] A driving monitoring module is set on the target storage robot. The driving monitoring module is composed of a laser monitoring array and a visual monitoring array. The driving environment is monitored by the set driving monitoring module to obtain driving monitoring information, including laser driving monitoring information and visual driving monitoring information;

[0012] Extracting laser driving monitoring information based on the driving monitoring information, performing feature extraction on the laser driving monitoring information to obtain a laser environment detection feature signal of the target warehouse robot in the driving area;

[0013] Extracting the local maximum and local minimum of the laser environment detection characteristic signal, and performing signal interpolation processing based on the local maximum and local minimum to construct the upper envelope and lower envelope of the laser environment detection characteristic signal;

[0014] Preset an envelope threshold interval, calculate the average value of the upper envelope and the lower envelope to obtain an envelope mean, and if the envelope mean is within the envelope threshold interval, remove the local signal trend corresponding to the envelope mean from the laser environment detection characteristic signal;

[0015] Repeat the envelope operation judgment step to obtain the remaining laser environment detection feature signals of the laser environment detection feature signals, perform voxel processing based on the remaining laser environment detection feature signals, and finally obtain the driving environment point cloud features of the target warehouse robot in the driving area.

[0016] In this solution, the driving monitoring module set on the target warehouse robot obtains driving monitoring information, and perceives the driving environment of the target warehouse robot according to the driving monitoring information to obtain environmental perception information, and further includes:

[0017] Establishing an environmental perception model based on BEVFormer, using the driving monitoring information to obtain visual driving monitoring information, and importing the visual driving monitoring information into the visual perception model for analysis;

[0018] Extract multi-view image features from the input visual driving monitoring information through a set feature extraction layer, generate a learnable query vector on a preset BEV grid, and project the multi-view image features into the BEV space using a deformable attention mechanism;

[0019] Obtain the point cloud features of the target warehouse robot's driving environment, obtain visual driving monitoring image features through the BEV space, perform feature alignment processing with the driving environment point cloud features, and fuse the aligned features to obtain multimodal driving monitoring features;

[0020] An environmental semantic analysis is performed based on the multimodal driving monitoring features to analyze the semantics of the real-time driving environment of the target warehouse robot, and the passable areas and obstacles are calibrated to obtain environmental perception information.

[0021] In this solution, the dynamic obstacle detection is performed based on the environmental perception information, the dynamic obstacles in the real-time driving environment of the target warehouse robot are detected, and the motion trajectory of the detected dynamic obstacles is predicted to obtain trajectory prediction information, which specifically includes:

[0022] Obtaining environmental perception information, building a driving obstacle analysis model based on the YOLOv5 framework, inputting the environmental perception information into the driving obstacle analysis model to perform dynamic obstacle analysis and obtain dynamic obstacle detection information;

[0023] Target locking is performed based on the dynamic obstacle detection information, and a target detection frame is generated. Position features and category features of each dynamic obstacle are extracted, and a unique data label defining each locked target is generated.

[0024] Use the extended Kalman filter to estimate the target's motion state, calculate the state transfer matrix of the locked target through first-order approximation, and generate the predicted motion state at the next moment based on the state transfer matrix;

[0025] Obtaining the observed motion state of the locked target based on the cosine metric algorithm, extracting the initial position of the locked target through the dynamic obstacle detection information, and generating observation trajectory information in combination with the observed motion state of the locked target;

[0026] Importing the observed trajectory information as training data into a Bayesian inference network for training to obtain a transient posterior distribution, and inferring the next motion state of the locked target in the Bayesian inference network through the transient posterior distribution to generate an inferred motion state at the next moment;

[0027] Extract the next moment inferred motion state with the highest inference probability value, calculate the deviation between the next moment predicted motion state and the next moment predicted motion state to obtain a deviation value, and compare the calculated deviation value with a preset threshold value;

[0028] If the deviation value is greater than a preset threshold, the next moment predicted motion state is corrected by weighted fusion based on the inferred motion state at the next moment, and trajectory prediction information is generated by the corrected next moment predicted motion state.

[0029] In this solution, the scenario risk modeling is performed based on the environmental perception information and trajectory prediction information, and the driving risk of the target warehouse robot following the predetermined driving trajectory is analyzed to obtain driving risk analysis information, which specifically includes:

[0030] Acquiring driving monitoring information and real-time transportation information of a target warehouse robot, calculating dynamic obstacle state characteristics based on the driving monitoring information, and extracting static obstacle state characteristics, extracting real-time transportation characteristics of the target warehouse robot from the real-time transportation information, including transportation type, transportation speed, and transportation status, and combining the dynamic obstacle state characteristics, the static obstacle state characteristics, and the real-time transportation characteristics to generate first characteristic information;

[0031] Obtaining environmental perception information and trajectory prediction information, introducing a Monte Carlo algorithm to perform risk simulation based on the first feature information and the trajectory prediction information, initializing parameters using the first feature information, and obtaining a trajectory prediction distribution for each dynamic obstacle based on the trajectory prediction information;

[0032] The risk diffusion range of each predicted trajectory is calculated based on the predicted trajectory distribution of each dynamic obstacle, and the risk diffusion range of all predicted trajectories is obtained by taking the weighted average of the risk diffusion ranges of all predicted trajectories.

[0033] The Gaussian diffusion model is used to describe the risk impact range of static obstacles. The risk diffusion range of dynamic obstacles is superimposed on the risk diffusion range of obstacles to obtain the real-time driving environment risk field of the target warehouse robot.

[0034] A risk heat map is generated through the real-time driving environment risk field, and the real-time driving trend characteristics of the target storage robot are obtained based on the driving monitoring information. The driving risk is analyzed in combination with the risk heat map to obtain driving risk analysis information.

[0035] In this solution, when a driving risk is detected in the predetermined driving trajectory, the driving trajectory is adjusted by combining the environmental perception information and the trajectory prediction information, and a driving strategy is generated to provide driving assistance, specifically including:

[0036] Acquiring driving risk analysis information, extracting driving risk characteristics of the target warehouse robot according to the current driving trend through the driving risk analysis information, performing a driving safety analysis based on the extracted driving risk characteristics, and obtaining safety analysis information;

[0037] If the safety analysis information indicates that the target warehouse robot has a safety risk according to the current driving trend, the driving trajectory is adjusted based on the environmental perception information and the action trajectory prediction information;

[0038] The target warehouse robot's driving environment is gridded, and the cuckoo search algorithm is used to adjust the driving trajectory. Constraints and objective functions are generated based on environmental perception information and action trajectory prediction information. The adjusted path is initialized to generate the initial path population.

[0039] A penalty term is set based on the path curvature, the fitness value of each path in the initial path population is calculated, paths with fitness values ​​greater than a preset threshold are selected as elite paths, a path replacement operation is performed using the elite paths, and an iterative analysis is performed to obtain an initial optimal path set;

[0040] The tabu search algorithm is introduced for improvement. The population is initialized based on the initial optimal path set and a tabu table is constructed. A neighborhood solution is generated based on the neighborhood movement rule. The optimal solution is selected based on the generated neighborhood solutions, and the tabu table is updated.

[0041] Perform iterative search and update the taboo table to obtain the optimized individuals after the search to update the population. When the number of iterations is reached or the stopping criteria are met, the final population is obtained. The optimal solution is output through the final population to obtain the adjusted path information. Based on the adjusted path information, a driving strategy is generated to provide driving assistance.

[0042] A second aspect of the present invention provides a safety analysis system for complex driving scenarios of a warehouse robot. The system includes: a memory and a processor. The memory contains a safety analysis method program for complex driving scenarios of a warehouse robot. When the safety analysis method program for complex driving scenarios of a warehouse robot is executed by the processor, the following steps are implemented:

[0043] Acquiring driving monitoring information based on a driving monitoring module set on the target warehouse robot, and sensing the driving environment of the target warehouse robot according to the driving monitoring information to obtain environmental perception information;

[0044] Performing dynamic obstacle detection based on the environmental perception information to detect dynamic obstacles in the real-time driving environment of the target warehouse robot, and predicting the motion trajectory of the detected dynamic obstacles to obtain trajectory prediction information;

[0045] Performing scenario risk modeling based on the environmental perception information and the motion trajectory information, analyzing the driving risk of the target warehouse robot following the predetermined driving trajectory, and obtaining driving risk analysis information;

[0046] When a driving risk is detected in the planned driving trajectory, the driving trajectory is adjusted by combining environmental perception information and action trajectory prediction information, and a driving strategy is generated for driving assistance.

[0047] The present invention discloses a safety analysis method and system for complex driving scenarios of AGVs. The method comprises the following steps: obtaining driving monitoring information, perceiving the driving environment of a target AGV based on the driving monitoring information, and obtaining environmental perception information; performing dynamic obstacle detection based on the environmental perception information, detecting dynamic obstacles in the target AGV's real-time driving environment, and predicting the motion trajectory of the detected dynamic obstacles to obtain trajectory prediction information; performing scenario risk modeling based on the environmental perception information and motion trajectory information, and analyzing the driving risk of the target AGV following a predetermined driving trajectory; and, when a driving risk is detected within the predetermined driving trajectory, adjusting the driving trajectory by combining the environmental perception information and motion trajectory prediction information, generating a driving strategy for driving assistance. This method improves the accuracy and intelligence of driving safety analysis for AGVs in complex dynamic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments or exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or exemplary descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained according to these drawings without paying any creative work.

[0049] Figure 1 A flow chart of a safety analysis method for a complex driving scenario of a warehouse robot provided by one embodiment of the present invention;

[0050] Figure 2 A flowchart of a method for assisting a warehouse robot in driving in complex scenarios, provided by one embodiment of the present invention;

[0051] Figure 3 A block diagram of a safety analysis system for complex driving scenarios of a warehouse robot, provided by one embodiment of the present invention;

[0052] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0053] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0054] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0055] Figure 1 A flow chart of a safety analysis method for a complex driving scenario of a warehouse robot provided by one embodiment of the present invention;

[0056] like Figure 1 As shown, the present invention provides a flow chart of a safety analysis method applied to complex driving scenarios of a warehouse robot, including:

[0057] S102, acquiring driving monitoring information based on a driving monitoring module set on the target warehouse robot, and sensing the driving environment of the target warehouse robot according to the driving monitoring information to obtain environmental perception information;

[0058] S104, performing dynamic obstacle detection based on the environmental perception information, detecting dynamic obstacles in the real-time driving environment of the target warehouse robot, and predicting the motion trajectory of the detected dynamic obstacles to obtain trajectory prediction information;

[0059] S106, performing scenario risk modeling based on the environmental perception information and the motion trajectory information, analyzing the driving risk of the target warehouse robot following the predetermined driving trajectory, and obtaining driving risk analysis information;

[0060] S108: When a driving risk is detected in the predetermined driving trajectory, the driving trajectory is adjusted by combining the environmental perception information and the action trajectory prediction information, and a driving strategy is generated to provide driving assistance.

[0061] Furthermore, in a preferred embodiment of the present invention, the driving monitoring module set on the target warehouse robot obtains driving monitoring information, and perceives the driving environment of the target warehouse robot according to the driving monitoring information to obtain environmental perception information, which specifically includes:

[0062] A driving monitoring module is set on the target storage robot. The driving monitoring module is composed of a laser monitoring array and a visual monitoring array. The driving environment is monitored by the set driving monitoring module to obtain driving monitoring information, including laser driving monitoring information and visual driving monitoring information;

[0063] Extracting laser driving monitoring information based on the driving monitoring information, performing feature extraction on the laser driving monitoring information to obtain a laser environment detection feature signal of the target warehouse robot in the driving area;

[0064] Extracting the local maximum and local minimum of the laser environment detection characteristic signal, and performing signal interpolation processing based on the local maximum and local minimum to construct the upper envelope and lower envelope of the laser environment detection characteristic signal;

[0065] Preset an envelope threshold interval, calculate the average value of the upper envelope and the lower envelope to obtain an envelope mean, and if the envelope mean is within the envelope threshold interval, remove the local signal trend corresponding to the envelope mean from the laser environment detection characteristic signal;

[0066] Repeat the envelope operation judgment step to obtain the remaining laser environment detection feature signals of the laser environment detection feature signals, perform voxel processing based on the remaining laser environment detection feature signals, and finally obtain the driving environment point cloud features of the target warehouse robot in the driving area.

[0067] It should be noted that a driving monitoring module is installed on the target warehouse robot. This module consists of a laser monitoring array and a visual monitoring array. The laser monitoring array acquires distance information about the target warehouse robot's surroundings by emitting laser beams and receiving reflected signals, while the visual monitoring array captures images of the environment using a camera. The driving monitoring module monitors the driving environment in real time, acquiring both laser and visual driving monitoring information. Laser driving monitoring information primarily includes distance data between the target warehouse robot and surrounding obstacles, while visual driving monitoring information includes environmental features such as texture, color, and shape. Based on this driving monitoring information, a laser environment detection feature signal is obtained within the target warehouse robot's driving area. This laser environment detection feature signal is generated by processing the laser reflection signal and reflects the geometric structure and obstacle distribution of the target warehouse robot's surroundings. The local maxima and local minima of the laser environment detection feature signal are further analyzed and extracted. Local maxima represent peaks in the laser reflection signal, typically corresponding to obstacle surfaces, while local minima represent valleys in the signal, typically corresponding to free space or low-reflection areas. The laser environment detection characteristic signal is interpolated using local maxima and local minima to construct the upper and lower envelopes of the signal. The upper envelope is a smooth curve generated by connecting local maxima, while the lower envelope is a smooth curve generated by connecting local minima. These two envelopes reflect the overall trend and fluctuation range of the laser environment detection characteristic signal. A preset envelope threshold interval is used to determine signal validity. If the envelope mean falls within the preset envelope threshold interval, the local signal trend corresponding to the envelope mean is removed from the laser environment detection characteristic signal, eliminating low-frequency trend components and retaining high-frequency details. Ultimately, the remaining laser environment detection characteristic signal reflects the high-frequency details of the target warehouse robot's surroundings. The remaining laser environment detection characteristic signal is then voxelized. Voxelization is the process of discretizing continuous signal data into a three-dimensional voxel grid, where each voxel contains signal data within a certain range. Through voxelization, a point cloud feature of the target warehouse robot's driving area is obtained, accurately describing the geometric structure of the target warehouse robot's surroundings.

[0068] Furthermore, in a preferred embodiment of the present invention, the driving monitoring module set on the target warehouse robot obtains driving monitoring information, and perceives the driving environment of the target warehouse robot according to the driving monitoring information to obtain environmental perception information, further comprising:

[0069] Establishing an environmental perception model based on BEVFormer, using the driving monitoring information to obtain visual driving monitoring information, and importing the visual driving monitoring information into the visual perception model for analysis;

[0070] Extract multi-view image features from the input visual driving monitoring information through a set feature extraction layer, generate a learnable query vector on a preset BEV grid, and project the multi-view image features into the BEV space using a deformable attention mechanism;

[0071] Obtain the point cloud features of the target warehouse robot's driving environment, obtain visual driving monitoring image features through the BEV space, perform feature alignment processing with the driving environment point cloud features, and fuse the aligned features to obtain multimodal driving monitoring features;

[0072] An environmental semantic analysis is performed based on the multimodal driving monitoring features to analyze the semantics of the real-time driving environment of the target warehouse robot, and the passable areas and obstacles are calibrated to obtain environmental perception information.

[0073] It should be noted that an environmental perception model is established based on the BEVFormer, and visual driving monitoring information is imported into the visual perception model for analysis. Visual driving monitoring information primarily consists of environmental image data captured by cameras, which provides rich texture, color, and shape information. A predefined feature extraction layer extracts multi-view image features from the input visual driving monitoring information. This feature extraction layer typically employs a convolutional neural network (CNN) or Transformer architecture, which can extract high-dimensional feature representations from multi-view images and capture key information about the environment. A learnable query vector is generated on a pre-set BEV grid to guide the feature projection process. A deformable attention mechanism is used to project the multi-view image features into the BEV space. This deformable attention mechanism adaptively adjusts the weights of feature projections to ensure effective integration of features from different viewpoints in the BEV space. This transforms the visual driving monitoring information into a feature representation in the BEV space, which more intuitively reflects the global structure of the target warehouse robot's surroundings. Subsequently, point cloud features of the target warehouse robot's driving environment are obtained. Visual driving monitoring image features are then acquired in the BEV space and aligned with the driving environment point cloud features. The purpose of feature alignment is to spatially align visual features and point cloud features to ensure that the two can be effectively integrated. The aligned features are fused through a feature fusion module to generate multimodal driving monitoring features. Multimodal feature fusion usually uses weighted summation or convolution operations, which can combine the advantages of visual features and point cloud features to provide a more comprehensive environmental representation. Environmental semantic analysis is performed based on the multimodal driving monitoring features to analyze the semantic information of the real-time driving environment of the target warehouse robot. The semantic analysis module usually uses a semantic segmentation network, such as FCN, U-Net, etc., which can classify different areas in the environment into passable areas and passable obstacles. Through semantic analysis, the passable areas and obstacle positions around the target warehouse robot are calibrated, and environmental perception information is obtained, which provides an important basis for subsequent path planning and obstacle avoidance decisions.

[0074] Furthermore, in a preferred embodiment of the present invention, the dynamic obstacle detection based on the environmental perception information is performed to detect dynamic obstacles in the real-time driving environment of the target warehouse robot, and the motion trajectory of the detected dynamic obstacles is predicted to obtain trajectory prediction information, which specifically includes:

[0075] Obtaining environmental perception information, building a driving obstacle analysis model based on the YOLOv5 framework, inputting the environmental perception information into the driving obstacle analysis model to perform dynamic obstacle analysis and obtain dynamic obstacle detection information;

[0076] Target locking is performed based on the dynamic obstacle detection information, and a target detection frame is generated. Position features and category features of each dynamic obstacle are extracted, and a unique data label defining each locked target is generated.

[0077] Use the extended Kalman filter to estimate the target's motion state, calculate the state transfer matrix of the locked target through first-order approximation, and generate the predicted motion state at the next moment based on the state transfer matrix;

[0078] Obtaining the observed motion state of the locked target based on the cosine metric algorithm, extracting the initial position of the locked target through the dynamic obstacle detection information, and generating observation trajectory information in combination with the observed motion state of the locked target;

[0079] Importing the observed trajectory information as training data into a Bayesian inference network for training to obtain a transient posterior distribution, and inferring the next motion state of the locked target in the Bayesian inference network through the transient posterior distribution to generate an inferred motion state at the next moment;

[0080] Extract the next moment inferred motion state with the highest inference probability value, calculate the deviation between the next moment predicted motion state and the next moment predicted motion state to obtain a deviation value, and compare the calculated deviation value with a preset threshold value;

[0081] If the deviation value is greater than a preset threshold, the next moment predicted motion state is corrected by weighted fusion based on the inferred motion state at the next moment, and trajectory prediction information is generated by the corrected next moment predicted motion state.

[0082] It should be noted that the obstacle analysis model is built based on the YOLOv5 framework. Environmental perception information is input into the model for dynamic obstacle analysis, generating dynamic obstacle detection information. The YOLOv5 model processes this input environmental perception information using a convolutional neural network. It can quickly detect dynamic obstacles around the target warehouse robot and generate target detection boxes. Each target detection box corresponds to a dynamic obstacle and contains its location and category information. Further processing of the dynamic obstacle detection information extracts the location and category features of each dynamic obstacle, generating a unique data label for each locked target. These data labels uniquely identify each dynamic obstacle, facilitating subsequent tracking and analysis. An extended Kalman filter (EKF) is used to estimate the motion state of locked dynamic obstacles. The EKF uses a first-order approximation to calculate the state transition matrix of the locked target. This matrix describes the target's motion relationship between the current moment and the next moment. Based on the state transition matrix, a predicted motion state for the next moment is generated. The predicted motion state includes information such as the target's position, velocity, and acceleration, reflecting the target's future motion trend. To further improve prediction accuracy, the cosine metric algorithm is used to obtain the observed motion state of the locked target. This algorithm calculates the similarity between targets in different frames and can determine whether two locked targets are the same. The initial position and observed motion state from the dynamic obstacle detection information are combined to generate observed trajectory information. This observed trajectory information reflects the actual motion of the target and provides an important basis for subsequent state estimation.

[0083] It should be noted that the observed trajectory information is imported as training data into the Bayesian inference network for training, obtaining the transient posterior distribution. The Bayesian inference network models the target's motion state using a probabilistic model, which can capture the uncertainty and dynamic changes in the target's motion. Using the transient posterior distribution, the Bayesian inference network infers the next motion state of the locked target, generating an inferred motion state at the next moment. This inferred motion state reflects the target's possible position and velocity at the next moment and has high prediction accuracy. The inferred motion state at the next moment with the highest inference probability is extracted and the deviation is calculated from the next moment predicted motion state generated by the extended Kalman filter to obtain a deviation value. The deviation value reflects the difference between the predicted and inferred states and is used to evaluate the accuracy of the prediction. The calculated deviation value is compared with a preset threshold. If the deviation value exceeds the preset threshold, it indicates a significant difference between the predicted and inferred states, requiring correction of the predicted state. Based on the inferred motion state at the next moment, a weighted fusion approach is used to correct the predicted motion state at the next moment. By combining the advantages of the predicted and inferred states, weighted fusion can generate a more accurate corrected predicted motion state. The corrected predicted motion state for the next moment is combined with the observed motion state of the target at the previous moment to form a single trajectory prediction result. After multiple recursive analyses, the single trajectory prediction results are integrated according to the time sequence to generate trajectory prediction information. This trajectory prediction information reflects the target's future motion trajectory and provides an important basis for path planning and obstacle avoidance decisions for the target warehouse robot, ensuring its safe operation in complex environments.

[0084] It should be noted that the motion state modeling process of dynamic obstacles is based on the following parameters to construct the state vector , specifically:

[0085] ;

[0086] in, , is the two-dimensional plane coordinate, i.e. the position of the dynamic obstacle, , is the velocity of the dynamic obstacle, , is the acceleration of the dynamic obstacle, and T is the period.

[0087] In the Bayesian network modeling process, the input target is the historical observation sequence output by the extended Kalman filter (EKF), and the output target is the state at the next moment The conditional probability distribution P. The Bayesian neural network includes an output layer, an output layer, and a hidden layer. The input layer is used to input the historical observation sequence, and the output layer is used to generate the state The parameters of the restored Gaussian distribution, including the mean and covariance matrix, are constructed using a three-layer fully connected network. The weights in each layer follow a Gaussian prior distribution. During network training, observation trajectory fragments are extracted from historical observation data as input, and Gaussian noise is added to simulate sensor errors. The loss function is to maximize the Evidence Lower Bound (ELBO). The Adam optimizer is used with sliding sampling over a time window. The batch size is set to 64, and training is terminated if the validation set loss does not decrease after 10 consecutive rounds.

[0088] Furthermore, in a preferred embodiment of the present invention, the scenario risk modeling is performed based on the environmental perception information and trajectory prediction information, and the driving risk of the target warehouse robot following the predetermined driving trajectory is analyzed to obtain driving risk analysis information, which specifically includes:

[0089] Acquiring driving monitoring information and real-time transportation information of a target warehouse robot, calculating dynamic obstacle state characteristics based on the driving monitoring information, and extracting static obstacle state characteristics, extracting real-time transportation characteristics of the target warehouse robot from the real-time transportation information, including transportation type, transportation speed, and transportation status, and combining the dynamic obstacle state characteristics, the static obstacle state characteristics, and the real-time transportation characteristics to generate first characteristic information;

[0090] Obtaining environmental perception information and trajectory prediction information, introducing a Monte Carlo algorithm to perform risk simulation based on the first feature information and the trajectory prediction information, initializing parameters using the first feature information, and obtaining a trajectory prediction distribution for each dynamic obstacle based on the trajectory prediction information;

[0091] The risk diffusion range of each predicted trajectory is calculated based on the predicted trajectory distribution of each dynamic obstacle, and the risk diffusion range of all predicted trajectories is obtained by taking the weighted average of the risk diffusion ranges of all predicted trajectories.

[0092] The Gaussian diffusion model is used to describe the risk impact range of static obstacles. The risk diffusion range of dynamic obstacles is superimposed on the risk diffusion range of obstacles to obtain the real-time driving environment risk field of the target warehouse robot.

[0093] A risk heat map is generated through the real-time driving environment risk field, and the real-time driving trend characteristics of the target storage robot are obtained based on the driving monitoring information. The driving risk is analyzed in combination with the risk heat map to obtain driving risk analysis information.

[0094] It should be noted that after obtaining driving monitoring information and the real-time transport information of the target warehouse robot, the state features of dynamic obstacles, including position, speed, and direction of movement, are calculated based on the driving monitoring information. The state features of static obstacles, such as position, shape, and size, are also extracted. The real-time transport information is used to extract the real-time transport features of the target warehouse robot, including transport type, transport speed, and transport status. The transport type reflects the attributes of the AGV warehouse robot's current transport task, the transport speed describes the AGV warehouse robot's real-time motion status, and the transport status includes information such as load, battery charge, and operating mode. The dynamic obstacle state features, static obstacle state features, and real-time transport features are combined to generate first feature information. The state features of dynamic obstacles can be calculated using the YOLOv5 model during object detection by detecting previous and subsequent frames. The state features of static obstacles can be extracted through a combination of lidar signal processing (geometric features) and the BEVFormer visual model (semantic features).

[0095] It should be noted that a Monte Carlo algorithm is used to combine the first feature information and trajectory prediction information for risk simulation. Parameters are initialized using the first feature information, including the initial position, velocity, and direction of motion of dynamic obstacles, as well as the distribution of static obstacles. Based on the trajectory prediction information, the predicted trajectory distribution of each dynamic obstacle is obtained. This trajectory prediction distribution describes the possible position and trajectory of a dynamic obstacle over a period of time, reflecting the uncertainty of its motion. The risk diffusion range of each predicted trajectory is calculated using the predicted trajectory distribution of each dynamic obstacle. The risk diffusion range describes the impact area of ​​a dynamic obstacle on the surrounding environment and is typically modeled using a Gaussian distribution. A weighted average of the risk diffusion ranges of all predicted trajectories is taken to obtain the comprehensive risk diffusion range of the dynamic obstacle. This weighted average takes into account the probability weights of different predicted trajectories, ensuring the accuracy of the risk diffusion range. A Gaussian diffusion model is used to describe the risk impact range of static obstacles. The risk impact range of a static obstacle is typically related to its shape and size, and the Gaussian diffusion model effectively describes the attenuation characteristics of its risk distribution. The risk diffusion range of the dynamic obstacle is superimposed with the risk diffusion range of the static obstacle to obtain the real-time risk field of the target warehouse robot's driving environment. The real-time driving environment risk field integrates the risk impacts of dynamic and static obstacles, comprehensively reflecting the risk distribution of the AGV warehouse robot's surroundings. This field generates a risk heat map, which uses a color gradient to represent the risk level of different areas. Subsequently, based on driving monitoring information, the real-time driving trend characteristics of the target warehouse robot are obtained, including current speed, acceleration, and direction of movement. Driving risk is analyzed in conjunction with the risk heat map to generate driving risk analysis information. This driving risk analysis information includes the risk level and potential collision probability along the current driving path.

[0096] It should be noted that when using the Monte Carlo algorithm for risk simulation, input parameters are first initialized based on the first feature information and the predicted trajectory information. Dynamic object state characteristics, including real-time velocity v, acceleration a, and direction of motion θ, are derived from the first feature information. The predicted trajectory distribution for dynamic targets is also derived from the predicted trajectory information. Simultaneously, static obstacle state characteristics, including geometric position, size, and category, are obtained based on the first feature information. These are combined with real-time transportation characteristics to generate the initial input parameters for the Monte Carlo algorithm. During the simulation, simulation parameters are set based on the input first feature information and predicted trajectory information. For dynamic obstacles, M possible trajectories are sampled from the predicted trajectory distribution, resulting in M ​​Monte Carlo trials. Since static obstacles are inherently fixed in position, only the risk impact radius needs to be described using a Gaussian model to determine their risk impact range. The risk value decays exponentially with distance, forming a smooth risk gradient. During the simulation, the probability of spatial overlap between the M sampled trajectories and the target warehouse robot is calculated. The risk impact area of ​​a dynamic obstacle is defined as a circular area with a radius of R, centered at the current position. If the target warehouse robot's path point falls within R, it is marked as a risk event, and the number of occurrences is counted. Subsequently, the probability of each predicted trajectory output by the Bayesian network is used as a weight, and the weighted average of the risk diffusion range of the M sampled trajectories is taken to obtain a dynamic risk field. Finally, the dynamic risk field is superimposed with the static risk field to generate a global environmental risk, thereby avoiding the risk of collisions in the target warehouse robot's driving environment, which could affect the safety of driving and cargo.

[0097] Among them, the weighted average of the risk diffusion range of M sampling trajectories can be expressed as:

[0098] ;

[0099] Where M is the number of sampled trajectories, is the probability estimate of the occurrence of the j-th trajectory, is the instantaneous risk value of the j-th trajectory.

[0100] The instantaneous risk value of the j-th trajectory is calculated as follows:

[0101] ;

[0102] in, is the safety factor, is the obstacle speed, is the prediction time window.

[0103] Among them, the description of the static risk field using the Gaussian model is represented by a two-dimensional Gaussian function, as follows:

[0104] ;

[0105] in,( , ) is the center position of the static obstacle, is the standard deviation of the Gaussian distribution in the x and y directions to control the diffusion range of the risk field, is the maximum risk value at the obstacle center, and the risk interval is set to (0,1).

[0106] The risk field intensity of static obstacles decays with distance and can be expressed as:

[0107] ;

[0108] in, is the Euclidean distance from the robot to the static obstacle, Determined by the obstacle type, for example, the shelf can be set as =1.

[0109] Furthermore, in a preferred embodiment of the present invention, when a driving risk is detected in a predetermined driving trajectory, the driving trajectory is adjusted by combining environmental perception information and trajectory prediction information to generate a driving strategy for driving assistance, specifically including:

[0110] Acquiring driving risk analysis information, extracting driving risk characteristics of the target warehouse robot according to the current driving trend through the driving risk analysis information, performing a driving safety analysis based on the extracted driving risk characteristics, and obtaining safety analysis information;

[0111] If the safety analysis information indicates that the target warehouse robot has a safety risk according to the current driving trend, the driving trajectory is adjusted based on the environmental perception information and the action trajectory prediction information;

[0112] The target warehouse robot's driving environment is gridded, and the cuckoo search algorithm is used to adjust the driving trajectory. Constraints and objective functions are generated based on environmental perception information and action trajectory prediction information. The adjusted path is initialized to generate the initial path population.

[0113] A penalty term is set based on the path curvature, the fitness value of each path in the initial path population is calculated, paths with fitness values ​​greater than a preset threshold are selected as elite paths, a path replacement operation is performed using the elite paths, and an iterative analysis is performed to obtain an initial optimal path set;

[0114] The tabu search algorithm is introduced for improvement. The population is initialized based on the initial optimal path set and a tabu table is constructed. A neighborhood solution is generated based on the neighborhood movement rule. The optimal solution is selected based on the generated neighborhood solutions, and the tabu table is updated.

[0115] Perform iterative search and update the taboo table to obtain the optimized individuals after the search to update the population. When the number of iterations is reached or the stopping criteria are met, the final population is obtained. The optimal solution is output through the final population to obtain the adjusted path information. Based on the adjusted path information, a driving strategy is generated to provide driving assistance.

[0116] It should be noted that driving risk analysis information is used to extract the target warehouse robot's driving risk characteristics based on its current driving trend. Driving risk characteristics include the risk level on the current path, potential collision probability, and the movement trend of dynamic obstacles. Driving safety analysis can be performed using methods such as preset safety level thresholds to obtain safety analysis information. This safety analysis information reflects the safety status of the target warehouse robot under the current driving trend, including whether there is a collision risk and whether the path needs to be adjusted. If the safety analysis information indicates that the target warehouse robot faces a safety risk based on the current driving trend, the driving trajectory is adjusted based on environmental perception information and motion trajectory prediction information.

[0117] It should be noted that the target warehouse robot's driving environment is gridded, divided into several grid cells. Each grid cell contains information about the environmental risk level and obstacles. A cuckoo search algorithm is used to adjust the driving trajectory. Constraints and an objective function are generated based on environmental perception and trajectory prediction. Constraints include obstacle avoidance and risk constraints, while the objective function comprehensively considers factors such as path length and risk. The adjusted path is initialized to generate an initial path population. The initial path population consists of multiple candidate paths, each represented as a node sequence and obtained through random generation or mutation of existing paths. A penalty term is set based on path curvature, and the fitness value of each path in the initial path population is calculated. The fitness value reflects the overall performance of the path, including factors such as path length, risk value, and smoothness. Paths with fitness values ​​greater than a preset threshold are selected as elite paths. Path replacement is performed using these elite paths, replacing paths with lower fitness in the population. An iterative analysis is performed, continuously optimizing the path population through the Lévy flight mechanism and random replacement operations, to obtain the initial optimal path set. The initial optimal path set contains multiple candidate paths with high performance. Subsequently, the cuckoo search algorithm is improved using a tabu search algorithm. A population is initialized based on the initial optimal path set, and a tabu table is constructed. The tabu table records recently searched paths to avoid repeated searches. Neighborhood solutions are generated based on neighborhood movement rules. These solutions are obtained by making local adjustments to the current path, such as inserting new nodes, deleting nodes, or adjusting node positions. The optimal solution is selected from the generated neighborhood solutions, and the tabu table is updated. An iterative search is performed and the tabu table is updated to obtain the optimized individuals after the search, and the population is updated. When the number of iterations is reached or the stopping criterion is met, the final population is obtained. This final population contains candidate paths that have been optimized multiple times. The optimal solution is output by comparing fitness values, resulting in adjusted path information. Finally, a driving strategy is generated based on the adjusted path information for driving assistance. The driving strategy includes path tracking control, speed adjustment, and obstacle avoidance, ensuring that the target warehouse robot can safely and efficiently travel along the optimized path. By real-time monitoring of driving status and environmental changes, driving strategies are dynamically adjusted to ensure the safety and adaptability of target warehouse robots in complex environments.

[0118] It should be noted that the penalty term based on path curvature is specifically set as follows: a path consists of a series of consecutive points, and the curvature of each path point reflects the degree of turn at that location. For a path point P, its curvature C can be approximately calculated by the inverse radius of the arc formed by the three adjacent points. The greater the curvature, the sharper the turn. If the curvature C of a path point exceeds the robot's maximum allowable curvature (determined by the robot's kinematic constraints), a penalty is applied to that point. The penalty value is proportional to the degree of the excess.

[0119] The penalty term can be expressed as: ;

[0120] in, is the curvature penalty weight coefficient, which is used to control the penalty intensity. The curvature penalty term of the entire path is the accumulation of the penalty values ​​of all points. is the curvature of path point i, is the maximum allowable curvature of the robot.

[0121] Furthermore, the fitness value is calculated by the set objective function, where the objective function is composed of path length, environmental risk value and path curvature penalty term. The specific expression is as follows:

[0122] ;

[0123] in, is a preset weight coefficient used to balance the priorities of path length, safety, and smoothness. is the total length of the path, which is calculated by accumulating the Euclidean distance of each segment of the path. is the integral value of the environmental risk field that the path passes through, which is calculated by accumulating or integrating the risk values ​​of the path points in the risk heat map. It is the path curvature penalty term, which accumulates penalty values ​​for path points that exceed the maximum allowable curvature of the robot.

[0124] Figure 2 A flowchart of a method for assisting a warehouse robot in driving in complex scenarios, provided by one embodiment of the present invention;

[0125] like Figure 2 As shown, the present invention provides a flow chart of an assisting method for a warehouse robot to drive in a complex scene, including:

[0126] S202, acquiring driving monitoring information based on a set driving monitoring module, and sensing the driving environment of the target storage robot according to the driving monitoring information to obtain environmental perception information;

[0127] S204, performing dynamic obstacle detection to detect dynamic obstacles in the real-time driving environment of the target storage robot, and performing motion trajectory prediction on the detected dynamic obstacles to obtain trajectory prediction information;

[0128] S206, analyzing the driving risk of the target storage robot along the predetermined driving trajectory based on the environmental perception information and the motion trajectory information, and obtaining the risk diffusion range of the dynamic obstacle and the risk diffusion range of the obstacle;

[0129] S208, superimposing the risk diffusion range of the dynamic obstacle with the risk diffusion range of the obstacle to generate a risk heat map, and analyzing the driving risk of the target storage robot using the driving monitoring information to obtain real-time driving trend characteristics to obtain driving risk analysis information;

[0130] S210: Perform a driving safety analysis based on the driving risk analysis information. If the safety analysis information indicates that there is a safety risk in the current driving trend, perform a driving trajectory adjustment analysis to generate a driving strategy for driving assistance.

[0131] Furthermore, the safety analysis method for a complex driving scenario of a warehouse robot provided by the present invention further includes the following steps:

[0132] Based on data retrieval, warehouse robot guidance instances in different scenarios are obtained, and feature extraction is performed on the obtained warehouse robot guidance instances to obtain guidance scene features and guidance strategy features of each guidance instance;

[0133] Associating the guidance scene features with the guidance strategy features, presetting a membership function, calculating the membership between each guidance scene feature and the corresponding guidance strategy feature, and generating a guidance instance data set;

[0134] The Apriority algorithm is introduced to mine fuzzy rules through the guidance instance data set, the triggering rules corresponding to different guidance strategies are analyzed, and the minimum membership degree of each rule is defined as the matching degree, thereby obtaining several guidance strategy triggering rules;

[0135] The Euclidean distance between each guidance strategy triggering rule is calculated using the Euclidean clustering algorithm. The calculated Euclidean distance is compared with a preset Euclidean distance threshold. The guidance strategy triggering rules are classified according to the judgment results to generate several category subsets.

[0136] Each category subset is associated with the corresponding guidance scene feature to generate an association set, and the guidance scene features of each association subset are analyzed by principal component analysis. The trigger rule category feature profile of each category subset is constructed by principal component analysis;

[0137] Based on the trigger rule feature portrait, guidance strategy trigger rules and guidance strategies, a complex scenario guidance knowledge graph is constructed with the guidance trigger rule feature-guidance strategy trigger rule-guidance strategy as the meta-path;

[0138] When the target warehouse robot is driving in a complex scene, the real-time environmental perception data is input into the complex scene guidance knowledge graph for guidance strategy analysis, and the navigation mode of the warehouse robot is switched according to the obtained guidance strategy.

[0139] It should be noted that when warehouse robots navigate complex scenarios, safety requirements often necessitate one or more environmental monitoring methods, such as magnetic strip guidance, laser monitoring technology, and machine vision monitoring technology. For example, in complex, dynamic environments, such as workshops where humans and machines work together or areas where goods are temporarily stored, fixed-path guidance may not be able to adapt to changing environments. In these situations, switching to more flexible guidance methods, such as laser or visual navigation, ensures driving safety. Switching to magnetic strip guidance in efficient transport areas along fixed routes reduces computing resource consumption. The essence of switching guidance methods is to match dynamic technology to a dynamic environment. Through scenario-adaptive guidance strategies, warehouse robots can strike a balance between efficiency, cost, and safety, providing core support for the continuous optimization of smart factories and warehouses.

[0140] Figure 3 A safety analysis system 3 for complex driving scenarios of a warehouse robot is provided in one embodiment of the present invention. The system includes: a memory 31 and a processor 32. The memory 31 contains a safety analysis method program for complex driving scenarios of the warehouse robot. When the safety analysis method program for complex driving scenarios of the warehouse robot is executed by the processor 32, the following steps are implemented:

[0141] Acquiring driving monitoring information based on a driving monitoring module set on the target warehouse robot, and sensing the driving environment of the target warehouse robot according to the driving monitoring information to obtain environmental perception information;

[0142] Performing dynamic obstacle detection based on the environmental perception information to detect dynamic obstacles in the real-time driving environment of the target warehouse robot, and predicting the motion trajectory of the detected dynamic obstacles to obtain trajectory prediction information;

[0143] Performing scenario risk modeling based on the environmental perception information and the motion trajectory information, analyzing the driving risk of the target warehouse robot following the predetermined driving trajectory, and obtaining driving risk analysis information;

[0144] When a driving risk is detected in the planned driving trajectory, the driving trajectory is adjusted by combining environmental perception information and action trajectory prediction information, and a driving strategy is generated for driving assistance.

[0145] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0146] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0147] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0148] Those skilled in the art will appreciate that all or part of the steps of the above-mentioned method embodiments may be implemented by hardware associated with program instructions, and the aforementioned program may be stored in a computer-readable storage medium. When the program is executed, the program executes the steps of the above-mentioned method embodiments. The aforementioned storage medium includes various media that can store program codes, such as mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0149] Alternatively, if the integrated units described above are implemented as software modules and sold or used as standalone products, they can also be stored on a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present invention, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product, stored on a storage medium, includes instructions for enabling a computer device (such as a personal computer, server, or network device) to execute all or part of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as removable storage devices, ROM, RAM, magnetic disks, or optical disks.

[0150] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A safety analysis method for complex driving scenarios of warehouse robots, characterized by: include: Acquiring driving monitoring information based on a driving monitoring module set on the target warehouse robot, and sensing the driving environment of the target warehouse robot according to the driving monitoring information to obtain environmental perception information; Dynamic obstacle detection is performed based on the environmental perception information to detect dynamic obstacles in the real-time driving environment of the target warehouse robot, and motion trajectory prediction is performed on the detected dynamic obstacles to obtain trajectory prediction information, specifically including: Use the extended Kalman filter to estimate the target's motion state, calculate the state transfer matrix of the locked target through first-order approximation, and generate the predicted motion state at the next moment based on the state transfer matrix; Obtaining the observed motion state of the locked target based on the cosine metric algorithm, extracting the initial position of the locked target through the dynamic obstacle detection information, and generating observation trajectory information in combination with the observed motion state of the locked target; Importing the observed trajectory information as training data into a Bayesian inference network for training to obtain a transient posterior distribution, and inferring the next motion state of the locked target in the Bayesian inference network through the transient posterior distribution to generate an inferred motion state at the next moment; Extract the next moment inferred motion state with the highest inference probability value, calculate the deviation between the next moment predicted motion state and the next moment predicted motion state to obtain a deviation value, and compare the calculated deviation value with a preset threshold value; If the deviation value is greater than a preset threshold, the next moment predicted motion state is corrected by weighted fusion based on the next moment inferred motion state, and trajectory prediction information is generated by the corrected next moment predicted motion state; Performing scenario risk modeling based on the environmental perception information and trajectory prediction information, analyzing the driving risk of the target warehouse robot following the predetermined driving trajectory, and obtaining driving risk analysis information; When a driving risk is detected in the predetermined driving trajectory, the driving trajectory is adjusted by combining environmental perception information and trajectory prediction information, and a driving strategy is generated for driving assistance.

2. A safety analysis method for complex driving scenarios of a warehouse robot according to claim 1, characterized in that: The driving monitoring module set on the target warehouse robot obtains driving monitoring information, and perceives the driving environment of the target warehouse robot according to the driving monitoring information to obtain environmental perception information, specifically including: A driving monitoring module is set on the target storage robot. The driving monitoring module is composed of a laser monitoring array and a visual monitoring array. The driving environment is monitored by the set driving monitoring module to obtain driving monitoring information, including laser driving monitoring information and visual driving monitoring information; Extracting laser driving monitoring information based on the driving monitoring information, performing feature extraction on the laser driving monitoring information to obtain a laser environment detection feature signal of the target warehouse robot in the driving area; Extracting the local maximum and local minimum of the laser environment detection characteristic signal, and performing signal interpolation processing based on the local maximum and local minimum to construct the upper envelope and lower envelope of the laser environment detection characteristic signal; Preset an envelope threshold interval, calculate the average value of the upper envelope and the lower envelope to obtain an envelope mean, and if the envelope mean is within the envelope threshold interval, remove the local signal trend corresponding to the envelope mean from the laser environment detection characteristic signal; Repeat the envelope operation judgment step to obtain the remaining laser environment detection feature signals of the laser environment detection feature signals, perform voxel processing based on the remaining laser environment detection feature signals, and finally obtain the driving environment point cloud features of the target warehouse robot in the driving area.

3. The safety analysis method for complex driving scenarios of a warehouse robot according to claim 1 is characterized in that: The method further includes: obtaining driving monitoring information based on a driving monitoring module set on the target warehouse robot, sensing the driving environment of the target warehouse robot according to the driving monitoring information, and obtaining environmental perception information; Establishing an environmental perception model based on BEVFormer, using the driving monitoring information to obtain visual driving monitoring information, and importing the visual driving monitoring information into the environmental perception model for analysis; Extract multi-view image features from the input visual driving monitoring information through a set feature extraction layer, generate a learnable query vector on a preset BEV grid, and project the multi-view image features into the BEV space using a deformable attention mechanism; Obtain the point cloud features of the target warehouse robot's driving environment, obtain visual driving monitoring image features through the BEV space, perform feature alignment processing with the driving environment point cloud features, and fuse the aligned features to obtain multimodal driving monitoring features; An environmental semantic analysis is performed based on the multimodal driving monitoring features to analyze the semantics of the real-time driving environment of the target warehouse robot, and the passable areas and obstacles are calibrated to obtain environmental perception information.

4. The safety analysis method for complex driving scenarios of a warehouse robot according to claim 1 is characterized in that: Before using the extended Kalman filter to estimate the motion state of the target, the following steps are also included: Obtaining environmental perception information, building a driving obstacle analysis model based on the YOLOv5 framework, inputting the environmental perception information into the driving obstacle analysis model to perform dynamic obstacle analysis and obtain dynamic obstacle detection information; Target locking is performed based on the dynamic obstacle detection information, and a target detection frame is generated. Position features and category features of each dynamic obstacle are extracted, and a unique data label defining each locked target is generated.

5. The safety analysis method for complex driving scenarios of a warehouse robot according to claim 1 is characterized in that: The scenario risk modeling is performed based on the environmental perception information and trajectory prediction information, and the driving risk of the target warehouse robot according to the predetermined driving trajectory is analyzed to obtain driving risk analysis information, specifically including: Acquiring driving monitoring information and real-time transportation information of a target warehouse robot, calculating dynamic obstacle state characteristics based on the driving monitoring information, and extracting static obstacle state characteristics, extracting real-time transportation characteristics of the target warehouse robot from the real-time transportation information, including transportation type, transportation speed, and transportation status, and combining the dynamic obstacle state characteristics, the static obstacle state characteristics, and the real-time transportation characteristics to generate first characteristic information; Obtaining environmental perception information and trajectory prediction information, introducing a Monte Carlo algorithm to perform risk simulation based on the first feature information and the trajectory prediction information, initializing parameters using the first feature information, and obtaining a trajectory prediction distribution for each dynamic obstacle based on the trajectory prediction information; The risk diffusion range of each predicted trajectory is calculated based on the predicted trajectory distribution of each dynamic obstacle, and the risk diffusion range of all predicted trajectories is obtained by taking the weighted average of the risk diffusion ranges of all predicted trajectories. The Gaussian diffusion model is used to describe the risk impact range of static obstacles. The risk diffusion range of dynamic obstacles is superimposed on the risk diffusion range of static obstacles to obtain the real-time driving environment risk field of the target warehouse robot. A risk heat map is generated through the real-time driving environment risk field, and the real-time driving trend characteristics of the target storage robot are obtained based on the driving monitoring information. The driving risk is analyzed in combination with the risk heat map to obtain driving risk analysis information.

6. The safety analysis method for complex driving scenarios of a warehouse robot according to claim 1 is characterized in that: When a driving risk is detected in the predetermined driving trajectory, the driving trajectory is adjusted by combining the environmental perception information and the trajectory prediction information, and a driving strategy is generated to provide driving assistance, specifically including: Acquiring driving risk analysis information, extracting driving risk characteristics of the target warehouse robot according to the current driving trend through the driving risk analysis information, performing a driving safety analysis based on the extracted driving risk characteristics, and obtaining safety analysis information; If the safety analysis information indicates that the target warehouse robot has a safety risk according to the current driving trend, the driving trajectory is adjusted based on the environmental perception information and trajectory prediction information; The target warehouse robot's driving environment is gridded, and the cuckoo search algorithm is used to adjust the driving trajectory. Constraints and objective functions are generated based on environmental perception information and trajectory prediction information, and the adjusted path is initialized to generate the initial path population. A penalty term is set based on the path curvature, the fitness value of each path in the initial path population is calculated, paths with fitness values ​​greater than a preset threshold are selected as elite paths, a path replacement operation is performed using the elite paths, and an iterative analysis is performed to obtain an initial optimal path set; The tabu search algorithm is introduced for improvement. The population is initialized based on the initial optimal path set and a tabu table is constructed. A neighborhood solution is generated based on the neighborhood movement rule. The optimal solution is selected based on the generated neighborhood solutions, and the tabu table is updated. Perform iterative search and update the taboo table to obtain the optimized individuals after the search to update the population. When the number of iterations is reached or the stopping criteria are met, the final population is obtained. The optimal solution is output through the final population to obtain the adjusted path information. Based on the adjusted path information, a driving strategy is generated to provide driving assistance.

7. A safety analysis system for complex driving scenarios of warehouse robots, characterized by: The system includes: a memory and a processor, wherein the memory contains a safety analysis method program applied to complex driving scenarios of a warehouse robot, and when the safety analysis method program applied to complex driving scenarios of a warehouse robot is executed by the processor, the steps of the safety analysis method applied to complex driving scenarios of a warehouse robot as described in claims 1-6 are implemented.

Images