AGV automatic driving control method and controller based on positioning perception

Abstract

The application belongs to the technical field of AGV, and provides an AGV automatic driving control method and controller based on positioning perception. The method comprises the following steps: an AGV detects the surrounding environment by using a positioning perception device arranged thereon to obtain detection data, and sends the detection data to a positioning perception calculation platform; the positioning perception calculation platform identifies static obstacles and dynamic obstacles from the detection data, and determines the positions of all the obstacles; for the identified dynamic obstacles, motion trend data is predicted, and the probability of change of the motion trend is evaluated; a central decision calculation platform plans an optimal driving path for the AGV by using a swarm intelligence algorithm; and a controller of the AGV executes the optimal driving path. The application can significantly improve the driving safety, path smoothness and passing efficiency of the AGV in a complex dynamic industrial environment.

Description

Technical Field

[0001] This invention relates to the field of AGV technology, and more specifically, to an AGV automatic driving control method and controller based on positioning perception. Background Technology

[0002] In industrial environments such as spinning mills, AGVs can be used for the automated handling of yarn packages and spindles, thereby improving the production efficiency of spinning mills. To enhance the intelligent obstacle avoidance capabilities of AGVs, researchers have introduced positioning and sensing devices such as LiDAR and visual sensors in recent years, enabling AGVs to possess a certain degree of environmental detection capabilities. Based on sensor data, swarm intelligence algorithms are used to plan driving trajectories based on obstacle avoidance targets, thereby controlling the AGVs to autonomously avoid obstacles.

[0003] Spinning factories contain both static obstacles such as spinning equipment, shelves, columns, and walls, and dynamic obstacles such as moving workers, handcarts, and other AGVs. The appearance of these dynamic obstacles is random, and their movement trajectories are uncertain. Most existing methods can only identify the current position of the obstacle, lacking effective prediction of its future movement trends (such as speed and direction), and even more so, the ability to assess the probability of a change in its movement intention. This leads to AGV obstacle avoidance behavior often being either too conservative or too aggressive, failing to achieve smooth and efficient human-machine / vehicle collaborative operations.

[0004] In summary, when applied to dynamic and complex indoor industrial environments such as spinning mills, existing AGV driving control still faces the problem of weak intelligent prediction and response capabilities to dynamic obstacles. Therefore, an AGV path planning method capable of intelligently predicting and assessing the risks of dynamic obstacles is needed to ensure the driving safety and optimal path of AGVs in complex and dynamic environments such as spinning mills. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an AGV automatic driving control method and controller based on positioning perception.

[0006] This invention provides an AGV (Automated Guided Vehicle) autonomous driving control method based on positioning perception, comprising the following steps: the AGV uses its onboard positioning perception device to detect the surrounding environment to obtain detection data, and sends the detection data to a positioning perception computing platform; the positioning perception computing platform identifies static and dynamic obstacles from the detection data and determines the location of all obstacles; and, for the identified dynamic obstacles, predicts their motion trend data and assesses the probability of their motion trend changing; a central decision computing platform receives and, based on the location of each obstacle, the motion trend data of each dynamic obstacle, and the probability of their motion trend changing as sent by the positioning perception computing platform, uses a swarm intelligence algorithm to plan the optimal driving path for the AGV; wherein, the swarm intelligence algorithm embeds an attention mechanism; and the AGV's controller receives and executes the optimal driving path.

[0007] As an example, for an identified dynamic obstacle, predicting its motion trend data and assessing the probability of its motion trend changing includes: extracting a multi-dimensional feature vector based on the type identifier of the dynamic obstacle, its current motion trend data, and its relative positional relationship with other obstacles within a preset range. This feature vector includes: prior features determined based on the type identifier, representing the inherent randomness of the corresponding type of dynamic obstacle's behavior; ontological dynamic features determined based on the current motion trend data, representing the dynamic obstacle's own motion stability; and environmental constraint features determined based on its relative positional relationship with surrounding obstacles, representing the interaction with the environment. The multi-dimensional feature vector is then input into a pre-trained probability assessment model, which outputs the probability value of the dynamic obstacle's motion trend changing.

[0008] As an example, the type identifier is used to distinguish dynamic obstacles as personnel, manual transport vehicles, or other AGVs; the prior features include typical rate of change of acceleration and typical turning frequency corresponding to the type identifier, obtained based on historical statistical data; the ontological dynamic features include the current instantaneous acceleration and angular velocity of the direction of motion; and the environmental constraint features include the distance to the nearest obstacle and an estimated width of the passage ahead.

[0009] As an example, based on the location of each obstacle, the motion trend data of each dynamic obstacle, and the probability of changes in their motion trends sent by the positioning and perception computing platform, a swarm intelligence algorithm is used to plan the optimal driving path for the AGV. This includes: constructing a path search space with the current position of the AGV as the starting point and the target position as the ending point; calculating a dynamic attention weight for each obstacle based on its location, the motion trend data of each dynamic obstacle, and the probability of changes in their motion trends; wherein the attention weight of a dynamic obstacle is positively correlated with its probability of changes in motion trends; constructing a comprehensive cost function, including a path length cost and an obstacle risk cost weighted by the attention weights; and optimizing the comprehensive cost function in the path search space using a swarm intelligence algorithm, selecting the path with the minimum comprehensive cost as the optimal driving path.

[0010] As an example, based on the location of each obstacle, the motion trend data of each dynamic obstacle, and the probability of changes in their motion trends, a dynamic attention weight is calculated for each obstacle. This includes: constructing a change probability matrix based on the probability of changes in the motion trends of all dynamic obstacles; extracting several key features from the change probability matrix, including at least the global average change probability and the degree of deviation between the change probability of a specific dynamic obstacle and the global average change probability; and calculating a dynamic attention weight for each obstacle based on the location of each obstacle, the motion trend data of each dynamic obstacle, and each of the key features.

[0011] The present invention also provides a controller, including a receiving unit, a processing unit, and an output unit; the output unit is used to send the detection data obtained by the AGV using the positioning and sensing device equipped on it to the positioning and sensing computing platform; the receiving unit is used to receive the optimal driving path planned by the aforementioned central decision computing platform; the processing unit is used to parse the optimal driving path into specific driving instructions and control the AGV to drive along the optimal driving path.

[0012] As an example, the processing unit is also used to control the positioning sensing device equipped on the AGV to perform variable-period detection.

[0013] As an example, the variable periodicity detection refers to dynamically determining the detection period length for the next time period based on historical detection data at the current time.

[0014] As an example, the processing unit has an embedded storage unit that stores computer program code that can be executed by the processing unit.

[0015] As an example, the controller is deployed on an AGV operating in a spinning factory environment.

[0016] This invention constructs a comprehensive environmental situation map by accurately identifying and locating static and dynamic obstacles in a spinning mill environment, and predicting and evaluating the movement trends and probabilities of dynamic obstacles. Based on this, the central decision-making platform employs a swarm intelligence algorithm with an embedded attention mechanism for path planning, enabling the AGV to have forward-looking decision-making and dynamic risk assessment capabilities. This implementation effectively overcomes the problems in existing technologies where AGVs, unable to predict the behavior of dynamic obstacles, exhibit conservative or reckless obstacle avoidance behavior and short-sighted path planning, significantly improving the AGV's driving safety, path smoothness, and passage efficiency in complex and dynamic industrial environments. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an AGV automatic driving control method based on positioning perception disclosed in an embodiment of the present invention.

[0018] Figure 2 This is the overall architecture diagram of the AGV cluster scheduling system disclosed in the present invention.

[0019] Figure 3 This is a schematic diagram of the structure of a controller disclosed in an embodiment of the present invention.

[0020] Figure 4 This is another structural schematic diagram of a controller disclosed in an embodiment of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0024] like Figure 1 As shown, this embodiment of the invention discloses an AGV automatic driving control method based on positioning perception, including the following steps: 100, the AGV uses the positioning perception device equipped on it to detect the surrounding environment to obtain detection data, and sends the detection data to the positioning perception computing platform.

[0025] In this step, such as Figure 2 As shown, the solution of the present invention is applied to the AGV cluster scheduling system of a spinning mill. The system consists of AGVs deployed on the spinning mill site, a positioning and perception computing platform as an edge computing node, and a central decision computing platform as the central brain.

[0026] In this step, the AGV, acting as a mobile sensing terminal, utilizes its onboard positioning and sensing devices, such as one or more combinations of LiDAR, depth cameras, and millimeter-wave radar, to continuously scan and detect the local environment around its travel path. The raw data obtained from the detection is point cloud data or image data containing information such as distance and reflection intensity. The AGV then uploads this unprocessed raw detection data to the positioning and sensing computing platform in real time via its built-in wireless communication module.

[0027] 200, the positioning and perception computing platform identifies static and dynamic obstacles from the detection data, determines the location of all obstacles; and, for the identified dynamic obstacles, predicts their motion trend data and assesses the probability of changes in their motion trend.

[0028] In this step, the positioning and perception computing platform uses target recognition algorithms such as point cloud segmentation and computer vision to identify and distinguish obstacles in the surrounding environment from the raw detection data. Specifically, stationary objects such as spinning equipment, shelves, and walls are identified as static obstacles, while mobile objects such as workers, handcarts, and other AGVs are identified as dynamic obstacles.

[0029] Building upon the aforementioned identification, the positioning and perception computing platform also precisely determines the location of all obstacles. This location includes not only the obstacle's position coordinates in the global coordinate system but also its direction, distance, and angle relative to the AGV itself. For example, it determines that an obstacle is located at a 30-degree angle to the left front of the AGV, at a distance of 5 meters. It should be noted that the positioning and perception device equipped on the AGV should also include a locator, which can obtain its real-time location in the high-precision map of the spinning factory—that is, its coordinates in the factory's global coordinate system—through methods such as wireless node positioning (based on triangulation principles) or GPS positioning. Then, based on the detected location information of each obstacle, it indirectly calculates the position coordinates of each obstacle in the factory's global coordinate system.

[0030] Secondly, for identified dynamic obstacles, the positioning and perception computing platform not only records their instantaneous location but also predicts their motion. By analyzing the historical trajectory sequence of the dynamic obstacle (for example, using time-series prediction models such as Kalman filtering or long short-term memory networks), it predicts their future motion trend data, which includes at least their speed and direction of motion.

[0031] Furthermore, the positioning and perception computing platform also assesses the probability of changes in its movement trends. This assessment can be based on factors such as the type of dynamic obstacles (workers have higher randomness compared to AGVs) and real-time behavior patterns (such as whether they decelerate or turn), which will be explained in detail later and will not be repeated here.

[0032] 300, the central decision computing platform receives and, based on the location of each obstacle, the movement trend data of each dynamic obstacle and the probability of changes in their movement trends sent by the positioning and perception computing platform, plans the optimal driving path for the AGV using a swarm intelligence algorithm; wherein, the swarm intelligence algorithm embeds an attention mechanism.

[0033] In this step, the positioning and perception computing platform obtains environmental situation information after deep processing. This information includes the location of all obstacles, the movement trend data of dynamic obstacles, and the probability of changes in their movement trends. This environmental situation information is then sent to the central decision-making computing platform. Based on this comprehensive information, the central decision-making computing platform uses swarm intelligence algorithms (such as ant colony optimization or particle swarm optimization) to plan the globally optimal driving path for the AGV.

[0034] It is understood that the significant difference from existing technologies lies in the fact that this invention embeds an attention mechanism into the swarm intelligence algorithm. This mechanism enables the path planning algorithm to dynamically focus on high-risk targets. This improved swarm intelligence algorithm comprehensively utilizes the location information of obstacles (such as whether they are on the AGV's path), the movement trend data of dynamic obstacles, and the probability of changes in their movement trends, to automatically calculate and assign different attention weights. For example, dynamic obstacles with a high risk of intersecting with the AGV path and uncertain movement intentions (high probability of movement trend changes) will be given higher weights, thus having a greater impact on the path cost function and guiding the planned path to proactively and in advance avoid such high-risk areas. By introducing an attention mechanism and intelligently weighting multi-dimensional information, including location, this embodiment evolves path planning into an intelligent optimization process under dynamic risk assessment, ensuring that the optimal driving path calculated achieves the best balance between length, travel time, and safety.

[0035] 400, the AGV controller receives and executes the optimal travel path.

[0036] In this step, the central decision-making computing platform sends the planned optimal driving path to the controller of the corresponding AGV via a wireless network. After receiving the optimal driving path, the controller parses it into specific driving commands (such as the target speed and angle of each drive wheel) and controls the AGV to travel accurately and smoothly along the planned path.

[0037] This invention constructs a comprehensive environmental situation map by accurately identifying and locating static and dynamic obstacles in a spinning mill environment, and predicting and evaluating the movement trends and probabilities of dynamic obstacles. Based on this, the central decision-making platform employs a swarm intelligence algorithm with an embedded attention mechanism for path planning, enabling the AGV to have forward-looking decision-making and dynamic risk assessment capabilities. This implementation effectively overcomes the problems in existing technologies where AGVs, unable to predict the behavior of dynamic obstacles, exhibit conservative or reckless obstacle avoidance behavior and short-sighted path planning, significantly improving the AGV's driving safety, path smoothness, and passage efficiency in complex and dynamic industrial environments.

[0038] As an example, the step of predicting the motion trend data of the identified dynamic obstacle and assessing the probability of its motion trend changing includes: 201, extracting a multi-dimensional feature vector based on the type identifier of the dynamic obstacle, the current motion trend data, and its relative positional relationship with other obstacles within a preset range, including: prior features determined based on the type identifier that characterize the inherent randomness of the corresponding type of dynamic obstacle's behavior; ontological dynamic features determined based on the current motion trend data that characterize the motion stability of the dynamic obstacle itself; and environmental constraint features determined based on its relative positional relationship with surrounding obstacles that characterize the interaction with the environment.

[0039] As an example, the type identifier is used to distinguish dynamic obstacles as personnel, manual transport vehicles, or other AGVs; the prior features include typical rate of change of acceleration and typical turning frequency corresponding to the type identifier, obtained based on historical statistical data; the ontological dynamic features include the current instantaneous acceleration and angular velocity of the direction of motion; and the environmental constraint features include the distance to the nearest obstacle and an estimated width of the passage ahead.

[0040] In this step, after completing the identification of dynamic obstacles and the prediction of basic motion trends, the positioning and perception computing platform first extracts multi-dimensional feature vectors to accurately assess the probability of changes in their motion trends. These multi-dimensional feature vectors integrate the following three key features that reflect the uncertainty of obstacle behavior: (1) Prior features: These features originate from the type identification of dynamic obstacles. The positioning and perception computing platform classifies the identified dynamic obstacles into different categories, such as personnel, manual transport vehicles, or other AGVs. Each category has different behavioral randomness due to its inherent motion characteristics. It is understandable that, based on a large amount of historical operational statistics, a typical acceleration change rate and typical turning frequency are preset for each type identification. For example, the typical acceleration change rate and turning frequency of the personnel type are significantly higher than those of other AGVs, representing the higher uncertainty and randomness of human walking compared to the programmed operation of AGVs.

[0041] (2) On-body dynamic characteristics: These characteristics are directly derived from the current motion trend data of the dynamic obstacle and are used to quantify its instantaneous motion stability. Specifically, they include the current instantaneous acceleration and the angular velocity of the motion direction. It is understandable that the probability of a change in the motion trend of an obstacle undergoing violent acceleration, deceleration, or rapid turning is naturally higher than that of an obstacle maintaining uniform linear motion. These characteristics reflect the dynamic uncertainty of the obstacle itself in real time.

[0042] (3) Environmental constraint characteristics: These characteristics are based on the relative positional relationship between the dynamic obstacle and other obstacles (including static and dynamic ones) within a preset range (such as its perception radius), and are used to characterize the constraints and interactive effects of the environment on its movement. Specifically, these include the distance to the nearest obstacle and the estimated width of the passage ahead. For example, when the passage ahead of a dynamic obstacle (such as a person) is narrow or close to other obstacles, the probability of it being forced to change its current movement trend (such as decelerating or turning) will increase significantly.

[0043] 202. Input the multidimensional feature vector into the pre-trained probability evaluation model and output the probability value of the change in the motion trend of the dynamic obstacle.

[0044] In this step, after extracting the multidimensional feature vector, the positioning and perception computing platform inputs this multidimensional feature vector into a pre-trained probability evaluation model. This probability evaluation model is a machine learning model (such as a gradient boosting decision tree or neural network) that has been trained on massive amounts of historical dynamic obstacle trajectory data and corresponding motion trend change labels (i.e., data recording whether the motion trend changes under a certain feature state). This probability evaluation model can learn the non-linear mapping relationship between complex multidimensional features and the probability of motion trend change. Through the model's calculation, a quantified probability value is finally output, which is the probability that the motion trend of the dynamic obstacle has changed.

[0045] For example, a gradient boosting decision tree model can be used to implement a probability evaluation model. This model integrates multiple weak decision tree models and trains them sequentially in a gradient boosting manner, ultimately obtaining a probability evaluation model suitable for handling the complex nonlinear relationship between the aforementioned multidimensional feature vectors and the probability of changes in motion trends.

[0046] The model structure of the probability evaluation model includes: base learners: the model is composed of several (e.g., 100 to 500) CART regression trees strung together as base learners.

[0047] Additive model: The final output of the model is a weighted sum of all the predictions from these decision trees. For the i-th dynamic obstacle sample, the final output is the probability score. It can be represented as: ;in: These are initial values ​​(such as the mean of the target variable in the training set). It is the learning rate, used to control the contribution of each tree to the final result and prevent overfitting; It is the total number of decision trees; It is the k-th decision tree for the samples The predicted value.

[0048] Probability score The probability value of the change in the motion trend of the dynamic obstacle is obtained by mapping the Sigmoid function to the interval [0,1]. .

[0049] Compared to existing technologies, this implementation method does not simply make a binary judgment on the behavior of dynamic obstacles, but comprehensively considers the innate behavior patterns, real-time motion status and environmental pressure of dynamic obstacles, making the probability of changes in the final motion trend more scientific and accurate, thereby improving the accuracy of the central decision-making platform in identifying high-risk targets and planning optimal paths.

[0050] As an example, based on the location of each obstacle, the movement trend data of each dynamic obstacle and the probability of changes in their movement trends sent by the positioning and perception computing platform, a swarm intelligence algorithm is used to plan the optimal driving path for the AGV, including: 301, constructing a path search space with the current position of the AGV as the starting point and the target position as the ending point.

[0051] In this step, the central decision-making computing platform constructs a defined path search space on the digital map (grid map or topology map) of the spinning mill. This path search space starts from the real-time current position of the AGV and ends at the target position of its current handling task (such as a designated spinning machine or warehouse shelf). It is understood that this path search space defines the set of all possible paths, and subsequent path planning should be limited to this path search space.

[0052] 302. Based on the location of each obstacle, the motion trend data of each dynamic obstacle and the probability of its motion trend changing, calculate a dynamic attention weight for each obstacle; wherein, the attention weight of a dynamic obstacle is positively correlated with the probability of its motion trend changing.

[0053] In this step, the central decision-making computing platform does not treat all obstacles equally, but instead calculates a unique, dynamic attention weight for each obstacle (including static and dynamic ones). ,in This indexes obstacles. The calculation rule for this weight is as follows: For static obstacles (such as equipment or shelves), the attention weight can be set to a basic constant value. The inherent risk is represented; for dynamic obstacles (such as workers, other AGVs), the attention weight is... The probability of its movement trend changing Positive correlation. That is, The higher, The larger it is. Specifically, its weight can be determined by the formula... Calculation, where It is a weighting coefficient greater than 0.

[0054] This mechanism allows the system to allocate more attentional resources to high-risk dynamic obstacles that are prone to sudden turns or speed changes, much like an experienced driver.

[0055] 303. Construct a comprehensive cost function, including path length cost and obstacle risk cost weighted by the attention weights.

[0056] In this step, a comprehensive cost function is constructed based on the aforementioned attention weights. This is used to quantitatively evaluate the merits of any candidate path. The comprehensive cost function is a weighted sum of multiple costs, specifically including: path length cost. Encourage the planning of shorter routes to improve operational efficiency.

[0057] Obstacle risk cost weighted by attention This cost comprehensively considers the risks of the path and all obstacles. Its calculation method is as follows: .in, It is a point on the path to an obstacle. The shortest distance, It is a variable with distance Functions that increase and then decrease sharply (e.g.) Among them, the attention weight of the obstacle itself. The higher it is, and the closer the path is to it ( The smaller the value, the greater the risk and cost, thus strongly deterring a path from approaching the obstacle.

[0058] Understandable, It can be and The weighted sum of the values ​​can, of course, be expressed in other forms, and no specific restrictions are imposed on it.

[0059] 304. In the path search space, a swarm intelligence algorithm is used to optimize and solve the comprehensive cost function, and the path with the minimum comprehensive cost is taken as the optimal driving path.

[0060] In this step, within the defined search space, a swarm intelligence algorithm (such as ant colony optimization or particle swarm optimization) is used to synthesize the cost function. The algorithm performs optimization to find the optimal solution. It simulates group behavior (such as pheromone positive feedback in ant colonies or social learning in bird flocks), generating and iteratively optimizing a large number of candidate paths. Its optimization objective is to find the path that maximizes the overall cost function. The path with the smallest overall cost is selected. Finally, this path is output as the optimal driving path.

[0061] In this implementation, a dynamic attention weight is calculated for each obstacle, which is positively correlated with the probability of its motion trend change. This weight is then embedded into the cost function of path planning, thereby improving path planning from simple geometric optimization to intelligent decision-making based on dynamic risk perception. This allows the swarm intelligence algorithm to simulate the risk perception process of a human driver, which can significantly improve the safety, path smoothness, and decision-making intelligence when driving in complex dynamic environments such as spinning mills. It effectively avoids the risk of sudden stops or collisions caused by sudden changes in the behavior of dynamic obstacles.

[0062] As an example, based on the location of each obstacle, the motion trend data of each dynamic obstacle and the probability of its motion trend changing, a dynamic attention weight is calculated for each obstacle, including: 3021, constructing a change probability matrix based on the probability of the motion trend changing of all dynamic obstacles, and extracting several key features from the change probability matrix, including at least the global average change probability and the degree of deviation between the change probability of a specific dynamic obstacle and the global average change probability.

[0063] The aforementioned implementation provides a basic method for calculating dynamic attention weights, namely, the weights are positively correlated with the probability of changes in the obstacle's own motion trend. However, in the complex scenarios of actual spinning mills, relying solely on individual probabilities may have limitations. For example, in a densely populated area with high overall dynamism (such as an intersecting passageway), a high probability of change for a single obstacle may be the norm; while in a normally quiet area, an obstacle with a moderate probability of change may appear unusually prominent, requiring heightened vigilance.

[0064] To allocate attention resources more intelligently and in a more context-sensitive manner, this embodiment assesses the risk of individual obstacles within the risk context of the overall current environment, thereby improving the perception of risk from individual risk to contextual risk. By analyzing the overall distribution of the probabilities of all dynamic obstacles, truly high-risk sources relative to the current environment can be identified more accurately, avoiding overreaction in highly dynamic environments or negligence in low-dynamic environments.

[0065] In this step, the central decision-making computing platform calculates the probability of changes in the motion trends of all N dynamic obstacles within the current AGV's sensing range. (in, Organized into a variable probability matrix , can be represented as This probability matrix represents the overall dynamic uncertainty of the local area where the AGV is located.

[0066] From the probability matrix Extract the key features describing its statistical distribution characteristics: (1) Global average probability of variation ( The calculation formula is: This value reflects the average level of uncertainty in the behavior of dynamic obstacles under the current environment. High This indicates that the AGV is in a highly unstable region.

[0067] (2) The degree of probability deviation of a specific obstacle ( ): Calculate the probability of change for each dynamic obstacle Compared with the global average The difference, i.e. This key feature quantifies the degree of anomaly of a single obstacle relative to the average environmental risk level. Positive This indicates that the obstacle is more unpredictable than most surrounding obstacles and is a local high-risk source that requires close attention.

[0068] 3022. Based on the location of each obstacle, the motion trend data of each dynamic obstacle, and each key feature, calculate a dynamic attention weight for each obstacle.

[0069] In this step, the central decision-making computing platform integrates the obstacle's location information (relative position and distance between the obstacle and the AGV), the obstacle's motion trend data (speed, direction, etc.), and key features extracted from the probability matrix (global average probability of change, probability deviation of specific obstacles) to calculate dynamic attention weights for each obstacle. .

[0070] Specifically, dynamic attention weights The computation function can be evolved into: .

[0071] For example, for dynamic obstacles, their weights It can be designed not only with Positive correlation, and also with Positive correlation. Therefore, obstacles that appear particularly uncertain relative to the average environmental level will be given higher attention weights, even if their absolute probability value is not the highest. Meanwhile, the function... You can consider it. The effect on the overall baseline weights can be expressed as follows:

[0072] For dynamic obstacles: ;in, It is the basic weight. These are weighting coefficients; This represents the probability of change in the motion trend of the dynamic obstacle itself. It is a sigmoid function; This represents the current distance between the obstacle and the AGV. This represents the current speed of the obstacle.

[0073] For static obstacles: Attention weights for static obstacles are primarily determined by their distance from the AGV and do not include terms related to dynamic probabilities.

[0074] like Figure 3 As shown, this embodiment of the invention also provides a controller 200, including a receiving unit 2001, an output unit 2002, and a processing unit 2003.

[0075] The output unit 2002 is used to send the detection data obtained by the AGV using its onboard positioning and sensing device to the positioning and sensing computing platform.

[0076] The receiving unit 2001 is used to receive the optimal driving route planned by the aforementioned central decision-making computing platform.

[0077] The processing unit 2003 is used to parse the optimal driving path into specific driving instructions and control the AGV to drive along the optimal driving path.

[0078] As an example, the processing unit 2003 is also used to control the positioning sensing device equipped on the AGV to perform variable periodic detection.

[0079] As an example, the variable periodicity detection refers to dynamically determining the detection period length for the next time period based on historical detection data at the current time.

[0080] In this embodiment, the processing unit 2003 can dynamically determine the detection cycle length for the next time period based on historical detection data from the positioning sensing device. For example, if it is determined that there are no dynamic obstacles in the surrounding environment ahead of the AGV, or that all dynamic obstacles are moving away from the AGV, based on historical detection data from a certain time period (or multiple sets of historical detection data from a certain time period), then there is no risk of collision. In this case, the detection cycle length for the next time period is extended, i.e., the detection frequency is reduced. Conversely, if it is determined that there are many dynamic obstacles in the surrounding environment ahead of the AGV, and that the dynamic obstacles are moving towards the AGV, then there is a high risk of collision. In this case, the detection cycle length for the next time period is shortened, i.e., the detection frequency is increased, so as to update the optimal driving path in a timely manner.

[0081] As an example, such as Figure 4 As shown, a storage unit 2004 is embedded in the processing unit 2003, and the storage unit 2004 stores computer program code that can be executed by the processing unit 2003.

[0082] In this embodiment, the processing unit 2003 can execute the computer program code stored in the storage unit 2004 to achieve the corresponding function, that is, to parse the optimal driving path into specific driving instructions and control the AGV to drive along the optimal driving path.

[0083] As an example, the controller is deployed on an AGV operating in a spinning factory environment.

[0084] Of course, the solution of the present invention can also be applied to other scenarios, such as logistics warehouses, other types of production workshops, etc., which will not be elaborated further.

[0085] Although the invention has been specifically shown and described with reference to preferred embodiments, those skilled in the art will understand that various modifications in form and detail may be made without departing from the spirit and scope of the invention. Accordingly, the disclosed invention should be considered merely illustrative and limited only by the scope specified in the appended claims.

Claims

1. A method for automatic driving control of AGVs based on positioning perception, characterized in that, The process includes the following steps: The AGV uses its onboard positioning and sensing device to detect the surrounding environment and obtain detection data, and sends the detection data to the positioning and sensing computing platform; The positioning and sensing computing platform identifies static and dynamic obstacles from the detection data and determines the location of all obstacles; Furthermore, for identified dynamic obstacles, predict their motion trend data and assess the probability of changes in their motion trend; The central decision-making computing platform receives and, based on the location of each obstacle, the motion trend data of each dynamic obstacle, and the probability of changes in their motion trends, sent by the positioning and perception computing platform, plans the optimal driving path for the AGV using a swarm intelligence algorithm. The swarm intelligence algorithm incorporates an attention mechanism. The AGV controller receives and executes the optimal driving path. The process of planning the optimal driving path for the AGV using the swarm intelligence algorithm, based on the location of each obstacle, the motion trend data of each dynamic obstacle, and the probability of changes in their motion trends, includes: constructing a path search space with the AGV's current position as the starting point and the target position as the ending point; and calculating a dynamic attention weight for each obstacle based on its location, the motion trend data of each dynamic obstacle, and the probability of changes in their motion trends. The attention weight of a dynamic obstacle is related to its probability of changes in its motion trend. Positive correlation; construct a comprehensive cost function, including path length cost and obstacle risk cost weighted by the attention weights; in the path search space, use a swarm intelligence algorithm to optimize and solve the comprehensive cost function, and take the path with the minimum comprehensive cost as the optimal driving path; based on the orientation of each obstacle, the motion trend data of each dynamic obstacle and the probability of its motion trend changing, calculate a dynamic attention weight for each obstacle, including: constructing a change probability matrix based on the probability of the motion trend changing of all dynamic obstacles, extracting several key features from the change probability matrix, including at least the global average change probability and the degree of deviation of the change probability of a specific dynamic obstacle from the global average change probability; based on the orientation of each obstacle, the motion trend data of each dynamic obstacle, and each of the key features, calculate a dynamic attention weight for each obstacle.

2. The AGV automatic driving control method based on positioning perception according to claim 1, characterized in that: For identified dynamic obstacles, predict their motion trend data and assess the probability of changes in their motion trend. This includes: extracting a multi-dimensional feature vector based on the type identifier of the dynamic obstacle, its current motion trend data, and its relative positional relationship with other obstacles within a preset range. This feature vector includes: prior features determined based on the type identifier, representing the inherent randomness of the corresponding type of dynamic obstacle's behavior; ontological dynamic features determined based on the current motion trend data, representing the stability of the dynamic obstacle's own motion; and environmental constraint features determined based on its relative positional relationship with surrounding obstacles, representing the interaction with the environment. The multi-dimensional feature vector is then input into a pre-trained probability assessment model, which outputs the probability value of changes in the motion trend of the dynamic obstacle.

3. The AGV automatic driving control method based on positioning perception according to claim 2, characterized in that: The type identifier is used to distinguish dynamic obstacles as personnel, manual transport vehicles, or other AGVs; the prior features include typical acceleration change rate and typical turning frequency corresponding to the type identifier, obtained based on historical statistical data; the body dynamic features include the current instantaneous acceleration and angular velocity of the direction of motion; the environmental constraint features include the distance to the nearest obstacle and the estimated width of the passage ahead.

4. A controller, characterized in that, The system includes a receiving unit, a processing unit, and an output unit. The output unit is used to send the detection data obtained by the AGV using its onboard positioning and sensing device to the positioning and sensing computing platform. The receiving unit is used to receive the optimal driving path planned by the central decision computing platform in any one of claims 1-3. The processing unit is used to parse the optimal driving path into specific driving instructions and control the AGV to drive along the optimal driving path.

5. A controller according to claim 4, characterized in that: The processing unit is also used to control the positioning and sensing device equipped on the AGV to perform variable-period detection.

6. A controller according to claim 5, characterized in that: The variable-period detection refers to dynamically determining the detection period length for the next time period based on historical detection data at the current time.

7. A controller according to claim 6, characterized in that: The processing unit has an embedded storage unit, which stores computer program code that can be executed by the processing unit.

8. A controller according to claim 7, characterized in that: The controller is deployed on the AGV operating in the spinning factory environment.

Images