Full-space AMR forklift obstacle avoidance method and AMR forklift

Abstract

The invention relates to a full-space AMR forklift obstacle avoidance method and an AMR forklift. The method comprises the steps that depth information of a measured object is obtained through a sensor array; analyzing the depth information by using a deep learning framework so as to identify the category of the tested object; according to the position coordinates of each sensor in the sensor array and the pixel point coordinates of the depth information, a pinhole camera model and inverse projection transformation are utilized, and multi-view distance data are obtained through fusion; for the multi-view distance data, an improved iterative nearest point algorithm model is adopted for analysis, a shape constraint condition is introduced, and size information of the measured object is obtained through iterative solution; and outputting an obstacle avoidance result based on the category of the measured object, the position of the measured object, the size information of the measured object and the real-time environment dynamic state, and generating an obstacle avoidance action sequence in combination with the dynamic constraint of AMR. Therefore, through the highly integrated sensor array, the advanced deep learning algorithm and the complex path planning strategy, accurate and efficient obstacle avoidance operation is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of safety obstacle avoidance, and in particular to a full-space AMR forklift obstacle avoidance method and an AMR forklift. Background Art

[0002] There are generally some technical deficiencies in the existing AGV (Automated Guided Vehicle) and AMR (Autonomous mobile robots) full-vehicle safety obstacle avoidance systems. Although traditional safety obstacle avoidance methods, such as the use of safety laser radars, anti-collision strips, and contact switches, can provide safety protection to a certain extent, these methods are often single or lack tight integration, resulting in safety blind spots, and thus unable to fully ensure the driving safety of forklifts. Especially during the driving process of forklifts, traditional obstacle avoidance systems often have difficulty effectively identifying dynamic obstacles on the path or obstacles in blind spots, which increases the risk of accidents when the forklift is driving along the fork tine direction.

[0003] Moreover, most traditional obstacle avoidance measures can only react passively, but cannot achieve active prevention. When faced with obstacles of special geometric shapes, such as low-height inclined objects, these traditional methods may fail to identify them and cause safety accidents. At the same time, since the safety measures are relatively simple, once a measure fails, it may lead to the collapse of the entire safety system.

[0004] In addition, the existing obstacle avoidance system is also insufficient in terms of cost-effectiveness and intelligence. Using traditional technology to achieve more complete safety functions often leads to a significant increase in costs, and traditional systems are usually unable to dynamically determine the size and position of obstacles in real time, which means that AGVs and AMRs can only adopt a strategy of stopping and waiting when encountering obstacles, and cannot choose the optimal driving path. These problems limit the flexibility and efficiency of AGVs and AMRs in complex environments.

[0005] Therefore, the existing AGV and AMR safety obstacle avoidance systems have significant deficiencies in all-round protection, dynamic obstacle recognition, active obstacle avoidance, special obstacle handling, system cost-effectiveness, and real-time decision-making. These deficiencies seriously affect the safety and efficiency of AGV and AMR in practical applications. Summary of the invention

[0006] 1. Technical issues to be resolved

[0007] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a full-space AMR forklift obstacle avoidance method and an AMR forklift, which solve the technical problems that the existing AGV and AMR safety obstacle avoidance systems have significant deficiencies in all-round protection, dynamic obstacle recognition, active obstacle avoidance, special obstacle handling, system cost-effectiveness and real-time decision-making.

[0008] (II) Technical solution

[0009] In order to achieve the above object, the main technical solutions adopted by the present invention include:

[0010] In a first aspect, an embodiment of the present invention provides a full-space adaptive AMR forklift obstacle avoidance method, comprising:

[0011] Acquire the depth information of the object at both ends or evenly distributed along the length direction synchronously or asynchronously through the sensor array;

[0012] Use deep learning framework to analyze the depth information to identify the category of the object being tested;

[0013] According to the position coordinates of each sensor in the sensor array and the pixel coordinates of the depth information, the actual distance value between each sensor in the sensor array and each point on the surface of the object to be measured is calculated using the pinhole camera model and inverse projection transformation, and multi-view distance data is obtained by fusion;

[0014] For multi-view distance data, an improved iterative closest point algorithm model with shape prior knowledge is used for analysis, shape constraints are introduced, and the size information of the object being measured is iteratively solved.

[0015] Based on the category, location, size and real-time environment dynamics of the object being measured, the obstacle avoidance result is output, and combined with the dynamic constraints of the AMR, an obstacle avoidance action sequence is generated.

[0016] Optionally, synchronously or asynchronously acquiring depth information uniformly distributed at both ends of the object or along the length direction through the sensor array includes:

[0017] The sensor array is started to collect depth information at both ends or along the length direction of the object under test in a synchronous or asynchronous manner;

[0018] Record the depth information and corresponding timestamp collected by each sensor and store them as a depth data set containing spatial coordinates and timestamps;

[0019] De-noising, registration and normalization are performed on the collected depth information to obtain pre-processed depth information;

[0020] The preprocessed depth information is organized according to timestamps and spatial coordinates to generate a depth information dataset.

[0021] Optionally, the deep information is analyzed using a deep learning framework to identify the category of the object under test including:

[0022] Determine the deep learning framework and train the deep learning framework based on the depth information;

[0023] The preprocessed depth information is input into the trained deep learning framework for analysis. Through multi-layer nonlinear transformation, high-level features in the depth information are automatically extracted and feature representation vectors are output.

[0024] Map the feature representation vector into a predefined category space and compare the similarity with the prototypes of each category in the category space or the samples in the preset training set;

[0025] Based on the similarity results, the probability or score of the object being tested belonging to each category is calculated, and the category of the object being tested is determined in combination with the preset threshold or rule.

[0026] Optionally, according to the position coordinates of each sensor in the sensor array and the pixel coordinates of the depth information, the actual distance value between each sensor in the sensor array and each point on the surface of the object to be measured is calculated using a pinhole camera model and an inverse projection transformation, and the multi-view distance data obtained by fusion includes:

[0027] Get the position coordinates (x i ,y i ,z i ), where i represents the index of the sensor;

[0028] Get the pixel coordinates (u, v) of the depth information corresponding to each sensor in the sensor array;

[0029] For each sensor, the pinhole camera model considering radial distortion is used to convert the pixel coordinates (u, v) of the depth information into the three-dimensional point coordinates (X c ',Y c ',Z c ');

[0030] The inverse projection transformation formula is used to transform the three-dimensional point coordinates (X c ',Y c ',Z c ') is converted to the global coordinate system to obtain the global coordinate (X g ,Y g ,Z g );

[0031] In the global coordinate system, according to the position coordinates of the sensor (x i ,y i ,zi ) and the coordinates of the surface point of the object after the inverse projection transformation (X g ,Y g ,Z g ), and use the Euclidean distance formula to calculate the actual distance value d between each sensor and each point on the surface of the object being measured i ;

[0032] The actual distance value d calculated by all sensors i Fusion is performed to form a multi-view distance data set D = {d 1 ,d 2 ,…,d n}, where n is the number of sensors;

[0033] in,

[0034] The conversion formula of the pinhole camera model considering radial distortion is:

[0035]

[0036] In the formula, k 1 and k 2 are the first-order and second-order radial distortion coefficients, c x and c y is the optical center coordinate, f x and f y is the focal length of the camera, r is the distance from the pixel to the optical center, d uv is the depth value, u and v are the pixel coordinates;

[0037] The inverse projection transformation formula is:

[0038]

[0039] In the formula, s x 、s y 、s z They are the scaling factors in the X, Y, and Z axes, respectively. ij is the element of the rotation matrix R used to describe the rotation relationship between the global coordinate system and the camera coordinate system. i and j represent the row and column respectively. x ,t y ,t z They are the translation vectors in the X, Y, and Z axes respectively.

[0040] Optionally, for multi-view distance data, an improved iterative closest point algorithm model with shape prior knowledge is used for analysis, shape constraints are introduced, and the size information of the object being measured is iteratively solved, including:

[0041] Using the acquired shape prior knowledge, the initial shape of the object to be measured is estimated to obtain an initial geometric model;

[0042] Establish shape constraints based on prior knowledge of the object being measured;

[0043] Using an improved iterative closest point algorithm model, the closest distance between corresponding point sets is calculated based on multi-view distance data and the current initial geometric model, and the initial geometric model is updated at least once to obtain a new geometric model.

[0044] Applying shape constraints to a new geometric model, adjusting the model to satisfy the shape constraints, obtaining an adjusted geometric model, and determining the difference between the geometric models before and after the adjustment. If the difference is less than a set threshold, the adjusted geometric model is used as the final geometric model.

[0045] The dimensional information of the object under test, including length, width, height and volume, is extracted from the final geometric model.

[0046] Optionally, based on the category of the object to be measured, the position of the object to be measured, the size information of the object to be measured, and the real-time environmental dynamics, the obstacle avoidance result is output, and combined with the dynamic constraints of the AMR, the obstacle avoidance action sequence is generated, including:

[0047] Construct a three-dimensional environment model including the object to be measured and environmental elements according to the type of the object to be measured, the location of the object to be measured, the size information of the object to be measured and the real-time environmental dynamics;

[0048] Define nodes in the three-dimensional environment model, each node represents a location point or a small area in the model;

[0049] Set status attributes for each node, including whether it is reachable, whether it has been visited, and whether it belongs to the open list or the closed list;

[0050] Initialize the starting point and the end point, add the current position of the object to be measured to the open list, use the A algorithm to search for the optimal path from the starting point to the end point in the model, and output the preliminary obstacle avoidance path;

[0051] Introduce reinforcement learning mechanism to dynamically adjust the initial obstacle avoidance path and output the optimized obstacle avoidance path;

[0052] Generate obstacle avoidance action sequences based on the optimized obstacle avoidance path and the dynamic constraints of the AMR:

[0053] in,

[0054] The obstacle avoidance action sequence is:

[0055] θ k =θ k-1 +Δθ×sign(P t -P c )

[0056]

[0057]

[0058] In the formula, θ k is the steering angle at time step k, θ k-1 is the steering angle of the previous time step, Δθ is the adjustment step of the steering angle, P t , P c are the target position and the current position respectively, and sign is the sign function used to calculate the value of the target position P according to the target position P. t and the current position P c The difference between the two determines the steering direction, v k is the driving speed at time step k, v max is the maximum driving speed, D t is the distance threshold, h k is the fork tine lifting height at time step k, h t is the target fork height, Δh is the adjustment step of the fork height, H t is the height threshold, h c is the current fork tine height, H o is the height of the obstacle.

[0059] Optionally, after outputting the obstacle avoidance result based on the category of the object to be measured, the position of the object to be measured, the size information of the object to be measured and the real-time environment dynamics, and generating the obstacle avoidance action sequence in combination with the dynamic constraints of the AMR, it also includes:

[0060] Convert the optimized obstacle avoidance action sequence into control instructions and send them to the AMR's actuator to drive the AMR to perform obstacle avoidance actions;

[0061] During the obstacle avoidance process, the obstacle avoidance action is monitored in real time, and closed-loop control is performed on the obstacle avoidance action based on the acquired feedback data and dynamic updates of the environment;

[0062] When it is detected that the obstacle cannot be avoided safely or the AMR fails, emergency braking or avoidance mechanism is initiated.

[0063] Optionally, the sensor array includes: a plurality of sensor modules arranged in a surrounding manner on the AMR, each sensor being configured with a specific horizontal field of view angle and a vertical field of view angle to form an interlaced and fully covered conical sensing space;

[0064] Among them, the representation range of the cone perception space is:

[0065]

[0066] Where n is the number of sensors, θ h1 ,θ h2are the horizontal and vertical field of view angles of the i-th sensor module, respectively. is the effective detection distance of the i-th sensor module.

[0067] In a second aspect, an embodiment of the present invention provides a full-space monitoring AMR forklift, including:

[0068] AMR forklift;

[0069] a sensor array; and,

[0070] The central controller is used to execute the method as described above.

[0071] Optionally, the sensor array includes a plurality of sensors arranged in a surrounding manner on the AMR, each sensor being configured with a specific horizontal field of view angle and a vertical field of view angle to form an interlaced and fully covered conical sensing space;

[0072] Among them, the representation range of the cone perception space is:

[0073]

[0074] Where n is the number of sensors, θ h1 ,θ h2 are the horizontal and vertical field of view angles of the i-th sensor module, respectively. is the effective detection distance of the i-th sensor module.

[0075] (III) Beneficial effects

[0076] The beneficial effects of the present invention are:

[0077] First, the present invention uses a sensor array to synchronously or asynchronously acquire depth information evenly distributed at both ends of the object or along the length direction. This step greatly enhances the forklift's ability to perceive the surrounding environment, ensures the comprehensiveness and accuracy of the data, and lays a solid foundation for subsequent in-depth analysis.

[0078] Next, the present invention uses a deep learning framework to analyze the acquired depth information to identify the category of the object being measured. This enables the AMR forklift to more intelligently understand and identify objects in the surrounding environment, thereby more accurately determining the obstacle avoidance strategy. This step not only improves the accuracy of obstacle avoidance, but also greatly shortens the forklift's reaction time.

[0079] After obtaining the depth information and the type of the object being measured, the pinhole camera model and inverse projection transformation are further used to calculate the actual distance value between each sensor in the sensor array and each point on the surface of the object being measured, and multi-view distance data is obtained through fusion. The innovation of this step is that it realizes the precise conversion from two-dimensional depth information to three-dimensional space distance, providing the forklift with a more three-dimensional and comprehensive environmental perception.

[0080] Furthermore, for multi-view distance data, the present invention uses an improved iterative closest point algorithm model with shape prior knowledge for analysis. By introducing shape constraints, the size information of the object to be measured is iteratively solved. Through this step, the size information of the object to be measured can be obtained more accurately, providing more accurate data support for obstacle avoidance decisions. At the same time, the improved algorithm model also improves the efficiency and accuracy of the calculation, and can respond to obstacle avoidance needs more quickly.

[0081] Finally, based on the category, position, size information of the object being measured and the real-time environmental dynamics, the method outputs the obstacle avoidance result and generates an obstacle avoidance action sequence in combination with the dynamic constraints of the AMR. The implementation of this step can quickly formulate the optimal obstacle avoidance strategy and generate the corresponding obstacle avoidance action sequence based on the real-time environmental dynamics and the actual situation of the object being measured. This not only improves the accuracy and safety of obstacle avoidance, but also enables the AMR forklift to respond more flexibly to the needs of various complex scenarios.

[0082] Therefore, the present invention achieves accurate and efficient obstacle avoidance operations through a highly integrated sensor array, advanced deep learning algorithms, and complex path planning strategies. It can not only acquire and analyze environmental depth information in real time, but also dynamically adjust obstacle avoidance strategies to adapt to different scenarios and obstacle types. The present invention solves the significant deficiencies of existing obstacle avoidance systems in terms of all-round protection, dynamic obstacle recognition, active obstacle avoidance, special obstacle handling, and system cost-effectiveness, and also improves the safety and operating efficiency of AMR forklifts in complex environments, ensuring the safe and efficient operation of AMR forklifts in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 A schematic flow chart of a full-space AMR forklift obstacle avoidance method provided by an embodiment of the present invention;

[0084] Figure 2 A schematic diagram of a specific flow chart of step S1 of a full-space AMR forklift obstacle avoidance method provided by an embodiment of the present invention;

[0085] Figure 3 A schematic diagram of a specific flow chart of step S2 of a full-space AMR forklift obstacle avoidance method provided by an embodiment of the present invention;

[0086] Figure 4 A schematic diagram of a specific flow chart of step S3 of a full-space AMR forklift obstacle avoidance method provided by an embodiment of the present invention;

[0087] Figure 5 A schematic diagram of a specific flow chart of step S4 of a full-space AMR forklift obstacle avoidance method provided by an embodiment of the present invention;

[0088] Figure 6 A schematic diagram of a specific flow chart of step S5 of a full-space AMR forklift obstacle avoidance method provided by an embodiment of the present invention;

[0089] Figure 7 A schematic diagram of the front and side structure of a full-space monitoring AMR forklift provided by an embodiment of the present invention;

[0090] Figure 8 A multi-view partial enlarged view of a sensor array for full-space monitoring of an AMR forklift provided by an embodiment of the present invention;

[0091] Fig. 9 A schematic diagram of the field of view of a 3D sensor for full-space monitoring of an AMR forklift provided by an embodiment of the present invention;

[0092] [Description of Reference Numerals]

[0093] 101: sensor array;

[0094] 102: Identify artifacts;

[0095] 103: Identify the carrier of the workpiece;

[0096] 104: AMR machine;

[0097] 105: 3D sensor. DETAILED DESCRIPTION

[0098] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation modes in conjunction with the accompanying drawings.

[0099] like Figure 1 As shown, an embodiment of the present invention proposes a full-space adaptive AMR forklift obstacle avoidance method, including: synchronously or asynchronously acquiring depth information uniformly distributed at both ends of the object to be measured or along the length direction through a sensor array; using a deep learning framework to analyze the depth information to identify the category of the object to be measured; according to the position coordinates of each sensor in the sensor array and the pixel coordinates of the depth information, using a pinhole camera model and an inverse projection transformation, the actual distance value between each sensor in the sensor array and each point on the surface of the object to be measured is calculated, and multi-view distance data is obtained by fusion; for the multi-view distance data, an improved iterative closest point algorithm model with shape prior knowledge is used for analysis, shape constraints are introduced, and the size information of the object to be measured is iteratively solved; based on the category of the object to be measured, the position of the object to be measured, the size information of the object to be measured, and the real-time environmental dynamics, the obstacle avoidance result is output, and in combination with the dynamic constraints of the AMR, an obstacle avoidance action sequence is generated.

[0100] First, the present invention uses a sensor array to synchronously or asynchronously acquire depth information evenly distributed at both ends of the object or along the length direction. This step greatly enhances the forklift's ability to perceive the surrounding environment, ensures the comprehensiveness and accuracy of the data, and lays a solid foundation for subsequent in-depth analysis.

[0101] Next, the present invention uses a deep learning framework to analyze the acquired depth information to identify the category of the object being measured. This enables the AMR forklift to more intelligently understand and identify objects in the surrounding environment, thereby more accurately determining the obstacle avoidance strategy. This step not only improves the accuracy of obstacle avoidance, but also greatly shortens the forklift's reaction time.

[0102] After obtaining the depth information and the type of the object being measured, the pinhole camera model and inverse projection transformation are further used to calculate the actual distance value between each sensor in the sensor array and each point on the surface of the object being measured, and multi-view distance data is obtained through fusion. The innovation of this step is that it realizes the precise conversion from two-dimensional depth information to three-dimensional space distance, providing the forklift with a more three-dimensional and comprehensive environmental perception.

[0103] Furthermore, for multi-view distance data, the present invention uses an improved iterative closest point algorithm model with shape prior knowledge for analysis. By introducing shape constraints, the size information of the object to be measured is iteratively solved. Through this step, the size information of the object to be measured can be obtained more accurately, providing more accurate data support for obstacle avoidance decisions. At the same time, the improved algorithm model also improves the efficiency and accuracy of the calculation, and can respond to obstacle avoidance needs more quickly.

[0104] Finally, based on the category, position, size information of the object being measured and the real-time environmental dynamics, the method outputs the obstacle avoidance result and generates an obstacle avoidance action sequence in combination with the dynamic constraints of the AMR. The implementation of this step can quickly formulate the optimal obstacle avoidance strategy and generate the corresponding obstacle avoidance action sequence based on the real-time environmental dynamics and the actual situation of the object being measured. This not only improves the accuracy and safety of obstacle avoidance, but also enables the AMR forklift to respond more flexibly to the needs of various complex scenarios.

[0105] Therefore, the present invention achieves accurate and efficient obstacle avoidance operations through a highly integrated sensor array, advanced deep learning algorithms, and complex path planning strategies. It can not only acquire and analyze environmental depth information in real time, but also dynamically adjust obstacle avoidance strategies to adapt to different scenarios and obstacle types. The present invention solves the significant deficiencies of existing obstacle avoidance systems in terms of all-round protection, dynamic obstacle recognition, active obstacle avoidance, special obstacle handling, and system cost-effectiveness, and also improves the safety and operating efficiency of AMR forklifts in complex environments, ensuring the safe and efficient operation of AMR forklifts in complex environments.

[0106] In order to better understand the above technical solution, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a clearer and more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0107] Specifically, an embodiment of the present invention provides a full-space adaptive AMR forklift obstacle avoidance method, comprising:

[0108] S1. Acquire the depth information of the object at both ends or evenly distributed along the length direction synchronously or asynchronously through the sensor array.

[0109] Furthermore, if Figure 2 As shown, step S1 includes:

[0110] S11, start the sensor array to collect depth information at both ends or along the length direction of the object under test in a synchronous or asynchronous manner.

[0111] S12. Record the depth information collected by each sensor and the corresponding timestamp, and store them as a depth data set D(x, y, z, t) including spatial coordinates (x, y, z) and timestamp t.

[0112] It is important to understand that when using a sensor array, choose synchronous or asynchronous mode to collect depth information according to actual needs. In synchronous mode, ensure that all sensors start and end collection at the same time to maintain time consistency. In asynchronous mode, allow sensors to collect data independently according to their own timestamps, and then process it through timestamp alignment technology.

[0113] S13, performing denoising, registration and normalization processing on the collected depth information to obtain pre-processed depth information.

[0114] S14. Organize the preprocessed depth information according to timestamps and spatial coordinates to generate a depth information data set.

[0115] S2. Use the deep learning framework to analyze the depth information to identify the category of the object being tested.

[0116] Furthermore, if Figure 3 As shown, step S2 includes:

[0117] S21. Determine a deep learning framework, and train the deep learning framework based on the depth information.

[0118] Choose a suitable deep learning framework based on the actual application scenario and data characteristics. For example, you can choose a convolutional neural network (CNN), a recurrent neural network (RNN) or its variants, such as a long short-term memory network (LSTM).

[0119] S22. The preprocessed depth information is input into a trained deep learning framework for analysis. Through multi-layer nonlinear transformation, high-level features in the depth information are automatically extracted, and a feature representation vector is output.

[0120] S23, mapping the feature representation vector to a predefined category space, and comparing the similarity with the prototypes of each category in the category space or the samples in a preset training set.

[0121] S24. Calculate the probability or score of the object under test belonging to each category based on the similarity result, and determine the category of the object under test in combination with a preset threshold or rule.

[0122] S3. According to the position coordinates of each sensor in the sensor array and the pixel coordinates of the depth information, the actual distance value between each sensor in the sensor array and each point on the surface of the object to be measured is calculated using the pinhole camera model and inverse projection transformation, and multi-view distance data is obtained through fusion.

[0123] Furthermore, if Figure 4 As shown, step S3 includes:

[0124] S31, obtain the position coordinates (x i ,y i ,z i ), where i represents the index of the sensor.

[0125] S32, obtaining pixel coordinates (u, v) of depth information corresponding to each sensor in the sensor array.

[0126] S33, for each sensor, using a pinhole camera model that takes into account radial distortion, the pixel point coordinates (u, v) of the depth information are converted into three-dimensional point coordinates (X, V) in the camera coordinate system after radial distortion correction. c ',Y c ',Z c ').

[0127] Among them, the conversion formula of the pinhole camera model is:

[0128]

[0129] In the formula, k 1 and k 2 are the first-order and second-order radial distortion coefficients, c x and c y is the optical center coordinate, fx and f y is the focal length of the camera, r is the distance from the pixel to the optical center, d uv is the depth value, u and v are the pixel coordinates.

[0130] S34, using the inverse projection transformation formula, the three-dimensional point coordinates (X c ',Y c ',Z c ') is converted to the global coordinate system to obtain the global coordinate (X g ,Y g ,Z g ).

[0131] Among them, the inverse projection transformation formula is:

[0132]

[0133] In the formula, s x 、s y 、s z They are the scaling factors in the X, Y, and Z axes, which are used to adjust the scale of the point in the global coordinate system. ij is the element of the rotation matrix R, i and j represent the row and column respectively. The rotation matrix is ​​used to describe the rotation relationship between the camera coordinate system and the global coordinate system. x ,t y ,t z They are the translation vectors in the X, Y, and Z axes, which are used to describe the translation relationship between the camera coordinate system and the global coordinate system.

[0134] The general form of the rotation matrix R is:

[0135]

[0136] S35. In the global coordinate system, according to the position coordinates of the sensor (x i ,y i ,z i ) and the coordinates of the surface point of the object after the inverse projection transformation (X g ,Y g ,Z g ), and use the Euclidean distance formula to calculate the actual distance value d between each sensor and each point on the surface of the object being measured i .

[0137] Among them, the Euclidean distance formula is:

[0138]

[0139] Where ε is the correction factor, dm is the distance value measured by the sensor, dn is the nominal distance value (such as the midpoint of the nominal measurement range of the sensor), and X obj , Y obj , Z obj It is the coordinate of the surface point of the object being measured in the global coordinate system.

[0140] S36, calculate the actual distance value d obtained by all sensors i Fusion is performed to form a multi-view distance data set D = {d 1 ,d 2 ,…,d n}, where n is the number of sensors.

[0141] S4. For multi-view distance data, an improved iterative closest point algorithm model with shape prior knowledge is used for analysis, shape constraints are introduced, and the size information of the object being measured is iteratively solved.

[0142] Furthermore, if Figure 5 As shown, step S4 includes:

[0143] S41. Using the acquired shape prior knowledge, the initial shape of the object to be measured is estimated to obtain an initial geometric model. Let the initial geometric model be M, which can be a parameterized model based on prior knowledge (such as an ellipse, a cylinder, etc.) or a statistical model (such as a shape obtained by principal component analysis).

[0144] S42. Establish shape constraints based on prior knowledge of the object being measured.

[0145] Specifically, there are the following shape constraints:

[0146] (1) Size constraints: Constrain certain sizes of the model (such as length, width, and height) based on prior knowledge, such as Lmin≤L(M)≤Lmax, where L(M) is the size measure of the model M, and Lmin and Lmax are the minimum and maximum allowed values ​​of the size, respectively.

[0147] (2) Volume or area constraint: The volume or surface area of ​​the model can be constrained within a certain range. For example: Vt-δV≤V(M)≤Vt+δV, where V(M) is the volume of the model M, Vt is the target volume, and δV is the allowed volume deviation.

[0148] (3) Feature point constraint: There are specific feature points (such as corner points and vertices) for the object being measured, and the positions or relative relationships of these feature points can be maintained during the iteration process. For example, for a feature point set F, for any feature point f∈F, its position should satisfy: ‖T(f)-ft‖<∈, where ft is the target position of the feature point and ∈ is the allowable value of the position deviation.

[0149] (4) The model surface is required to be smooth to a certain extent, that is, the normal vector change between adjacent points should not be too large. For example, for any adjacent points m and n on the model, their normal vectors nm and nn should satisfy: ‖nm-nn‖<θ, where θ is the maximum allowable angle of the normal vector change.

[0150] S43. Using an improved iterative closest point algorithm model, by calculating the closest distance between corresponding point sets based on multi-view distance data and the current initial geometric model, the initial geometric model is updated at least once to obtain a new geometric model.

[0151] In a specific embodiment, based on the multi-view distance dataset D, the following iterative steps are performed until a convergence condition is met:

[0152] First, start the initial step: for each point in the initial geometric model M, find the nearest point in the multi-view distance dataset D to form a corresponding point pair set P = {(m, d)|m∈M, d∈D}.

[0153] Secondly, the least square method or other optimization methods are used to calculate the rigid body transformation matrix T (including rotation and translation) that minimizes the distance between corresponding point pairs.

[0154] Next, the transformation matrix T is applied to update the model M to obtain a new model M', the shape constraint C(M') is applied to the new model M', and the model is adjusted to satisfy the shape constraint to obtain an updated model Mnew.

[0155] Finally, the difference between the model before and after the update is calculated. If the difference is less than the preset threshold, the algorithm is considered to have converged, otherwise it returns to the initial step to continue iterating.

[0156] S44, applying shape constraints to the new geometric model, adjusting the model to satisfy the shape constraints, obtaining an adjusted geometric model, and determining the difference between the geometric models before and after the adjustment. If the difference is less than a set threshold, the adjusted geometric model is used as the final geometric model.

[0157] S45. Extracting the dimension information including length, width, height and volume of the object to be measured from the final geometric model.

[0158] S5. Based on the category, position, size and real-time environment dynamics of the object being measured, the obstacle avoidance result is output, and combined with the dynamic constraints of the AMR, an obstacle avoidance action sequence is generated.

[0159] Furthermore, if Figure 6 As shown, step S5 includes:

[0160] S51, constructing a three-dimensional environment model including the object to be measured and environmental elements according to the type of the object to be measured, the position of the object to be measured, the size information of the object to be measured and the real-time environmental dynamics.

[0161] S52. Define nodes in the three-dimensional environment model, each node representing a location point or a small area in the model.

[0162] S53. Set status attributes for each node, including whether it is reachable, whether it has been visited, and whether it belongs to an open list or a closed list.

[0163] S54, initialize the starting point and the end point, add the current position of the object to be measured to the open list, use the A algorithm to search for the optimal path from the starting point to the end point in the model, and output the preliminary obstacle avoidance path.

[0164] In a specific embodiment, first, a constructed three-dimensional environment model is loaded, which includes an accurate three-dimensional representation of obstacles, objects to be measured, ground and other environmental elements, and nodes are defined in the three-dimensional environment model, each node representing a location point or a small area in the model. The status of each node is set, including whether it is reachable, whether it has been visited, whether it is a target node, etc. The starting node and the target node are initialized, the starting node corresponds to the current position of the object to be measured, and the target node corresponds to the expected position.

[0165] Next, for each node, the heuristic value to the target node is calculated, which is used to estimate the cost or distance from the current node to the target node. The selection of the heuristic function should be based on the actual requirements of the obstacle avoidance path, such as considering factors such as path length, time cost, and safety.

[0166] Add the starting node to the open list and set its cost to 0. Loop through the following steps until the target node is found or the open list is empty: a. Select the node with the smallest cost from the open list as the current node. b. Check the neighboring nodes of the current node, and for each neighboring node: If the neighboring node has not been visited and is reachable, calculate the total cost from the starting node to the neighboring node (including the actual cost and the heuristic value). If the total cost of the neighboring node is less than its existing cost in the open list (if any), update its cost and parent node, and add or re-add it to the open list. c. Mark the current node as visited, remove it from the open list, and add it to the closed list.

[0167] When the target node is found, start from the target node and trace back along the parent node to the starting node to build an obstacle avoidance path. Output the preliminary obstacle avoidance path, including the node sequence, direction, distance and other information on the path. If the target node is not found and the open list is empty, it means that a valid obstacle avoidance path cannot be generated, and the corresponding prompt information should be output.

[0168] S55. Introduce a reinforcement learning mechanism to dynamically adjust the initial obstacle avoidance path and output the optimized obstacle avoidance path;

[0169] In another specific embodiment, a suitable reinforcement learning algorithm (such as Q-learning, DQN, PPO, etc.) is selected, and the state space, action space and reward function are designed according to the requirements of the obstacle avoidance task.

[0170] The preliminary obstacle avoidance path generated by the A* algorithm is used as the input of the reinforcement learning model. The path is represented in the form of a sequence of nodes, each of which contains location information and possible action options.

[0171] For each node in the path, the reinforcement learning model selects an action based on its current state (including position, obstacle distribution, historical actions, etc.). The action space should contain all possible obstacle avoidance actions, such as forward, backward, turning, acceleration, deceleration, etc., and the selection of actions should take into account the kinematic constraints of the object being measured.

[0172] The selected action is simulated in the 3D environment model and the results are observed. The results include the new position, collision, whether the target is reached, etc. The reward is calculated based on the simulation results. The reward function should reflect the obstacle avoidance effect, such as positive rewards for successful obstacle avoidance and negative rewards for collision or deviation from the path.

[0173] Then, the current state, the selected action, the reward, and the next state are stored in the experience replay buffer. A batch of experiences are randomly drawn from the experience replay buffer to train the reinforcement learning model. The model parameters are updated by methods such as gradient descent to maximize the cumulative reward.

[0174] During the simulation execution and model update process, the learning model of the obstacle avoidance strategy is continuously reinforced. For each node in the preliminary obstacle avoidance path, the action selection is dynamically adjusted according to the output of the reinforcement learning model to generate a better obstacle avoidance path. The steps of simulation execution, reward calculation, experience storage and model update are repeated until the preset number of iterations or convergence conditions are reached.

[0175] Finally, an optimized obstacle avoidance path is generated based on the final output of the reinforcement learning model.

[0176] S56. Based on the optimized obstacle avoidance path and the dynamic constraints of the AMR, an obstacle avoidance action sequence including steering angle, driving speed, fork tine lifting height, etc. is generated.

[0177] Specifically, the obstacle avoidance action sequence is:

[0178] θ k =θ k-1 +Δθ×sign(P t -P c )

[0179]

[0180]

[0181] In the formula, θ k is the steering angle at time step k, θ k-1 is the steering angle of the previous time step, Δθ is the adjustment step of the steering angle, P t , P c are the target position and the current position respectively, and sign is the sign function used to calculate the value of the target position P according to the target position P. t and the current position P c The difference between the two determines the steering direction, v k is the driving speed at time step k, v max is the maximum driving speed, D t is the distance threshold, h k is the fork tine lifting height at time step k, h t is the target fork height, Δh is the adjustment step of the fork height, H t is the height threshold, h c is the current fork tine height, H o is the height of the obstacle.

[0182] In addition, the present invention also includes: converting the optimized obstacle avoidance action sequence into a control instruction and sending it to the actuator of the AMR to drive the AMR to perform the obstacle avoidance action; during the obstacle avoidance process, monitoring the execution process of the obstacle avoidance action in real time, and performing closed-loop control of the obstacle avoidance action based on the acquired feedback data and environmental dynamic updates; when it is detected that the obstacle cannot be avoided safely or the AMR fails, initiating an emergency braking or avoidance mechanism.

[0183] It should be understood that the sensor array used in the present invention includes: multiple sensor modules arranged in a surrounding manner on the AMR, each sensor is configured with a specific horizontal field of view angle and a vertical field of view angle, forming an interlaced and fully covered cone perception space. Among them, the representation range of the cone perception space is:

[0184] Among them, the representation range of the cone perception space is:

[0185]

[0186] Where n is the number of sensors, θ h1 ,θ h2 are the horizontal and vertical field of view angles of the i-th sensor module, respectively, and d i 2 is the effective detection distance of the i-th sensor module.

[0187] In addition, an embodiment of the present invention provides a full-space monitoring AMR forklift, comprising: an AMR forklift; a sensor array; and a central controller for executing the above method. The sensor array comprises a plurality of sensor modules arranged in a surrounding manner on the AMR, each sensor being configured with a specific horizontal field of view angle and a vertical field of view angle, forming an interlaced and fully covered conical sensing space.

[0188] The design and layout of the sensor array used in the present invention fully consider the limitations of traditional AMR obstacle avoidance systems, especially those that only choose to arrange sensors diagonally on the chassis, which cannot achieve full-space obstacle avoidance and have blind spots. In order to solve these problems, the present invention proposes an innovative solution, which is described as follows:

[0189] Figure 7 For full-space monitoring of AMR forklifts. 101 is a sensor array, 102 is an identification workpiece, 103 is a carrier for identifying the workpiece, and 104 is an AMR machine. Figure 8 It is a partial enlarged view of the sensor array 101, which is provided with multiple 3D sensors 105, including TOF cameras, structured light cameras, etc. The 3D sensors 105 are distributed in the upper, lower, left and right positions of the sensor array 101. These modules are not simply distributed in a certain local area of ​​the AMR, but are arranged around the AMR in all directions, thereby ensuring a comprehensive perception of the surrounding environment. This layout method fundamentally breaks the limitations of traditional obstacle avoidance systems and lays a solid foundation for achieving full-space obstacle avoidance.

[0190] It is important to emphasize that each sensor module is configured with a specific horizontal and vertical field of view. Figure 8 The placement of the 3D sensor 105 is not looking straight ahead, but has a certain downward tilt angle in all four directions. This adjustment is to achieve zero blind spot. The downward tilt angle needs to be determined according to the size of the camera's air defense range. The conical space extending from the front end of the 3D sensor 105 is the air defense range of the 3D sensor 105. Fig. 9 is the air defense range of the 3D sensor 105. More importantly, the sensing space formed by these sensor modules is a staggered and fully covered cone. This means that the sensing ranges of different sensor modules are staggered with each other, forming a full-range sensing network without blind spots. No matter where the obstacle is located, it can be captured by this sensing network in time, thereby ensuring comprehensive obstacle avoidance capabilities.

[0191] The sensor array 101 has a processor with specific computing power to implement some more complex algorithm requirements. The sensor array 101 can be replicated and applied in scene requirements of 4, 5, or n 3D sensors. If it is to be realized whether there are obstacles in the protection space, it is necessary to provide early warning and alarm information to the AMR machine 104. The 3D sensor 105 can cover a wide field of view and capture the depth information of the entire scene. In this way, no matter where the obstacle is located, it can be detected in time to ensure comprehensive obstacle avoidance capabilities. It has good adaptability to lighting conditions and is not easily affected by changes in ambient light. Whether it is bright daytime or dim night, they can work stably. In addition, the camera is less dependent on surface characteristics such as color and texture, and can work effectively on various surface types.

[0192] In summary, the embodiments of the present invention provide a full-space AMR forklift obstacle avoidance method and an AMR forklift. The solution of the present invention achieves accurate and efficient obstacle avoidance operations through a highly integrated sensor array, advanced deep learning algorithms, and complex path planning strategies. It can not only acquire and analyze environmental depth information in real time, but also dynamically adjust obstacle avoidance strategies to adapt to different scenarios and obstacle types, thereby significantly improving the safety and operating efficiency of AMR forklifts in complex environments. In particular, by abandoning traditional sensors such as safety laser radars and using 3D sensors to achieve 360° high-altitude obstacle avoidance without blind spots, the performance of the obstacle avoidance system is further enhanced. 3D sensors rely on their accurate depth perception, efficient recognition capabilities in complex environments, and important applications in multiple fields such as robot navigation, autonomous driving, and industrial automation, making the AMR forklift obstacle avoidance method of the present invention more advantageous and practical.

[0193] Since the system / device described in the above embodiments of the present invention is a system / device used to implement the method of the above embodiments of the present invention, a person skilled in the art can understand the specific structure and deformation of the system / device based on the method described in the above embodiments of the present invention, and thus will not be described in detail here. All systems / devices used in the method of the above embodiments of the present invention belong to the scope of protection of the present invention.

[0194] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0195] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present invention. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions.

[0196] It should be noted that in the claims, any reference numerals placed between brackets shall not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention may be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In the claims enumerating several means, several of these means may be embodied by the same hardware. The use of the words first, second, third, etc., is for convenience of expression only and does not indicate any order. These words may be understood as part of the component name.

[0197] In addition, it should be noted that, in the description of this specification, the description of the terms "one embodiment", "some embodiments", "embodiment", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.

[0198] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments after knowing the basic creative concept. Therefore, the claims should be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present invention.

[0199] Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention should also include these modifications and variations.

Claims

1. A full-space adaptive AMR forklift obstacle avoidance method, characterized in that: include: Acquire the depth information of the object at both ends or evenly distributed along the length direction synchronously or asynchronously through the sensor array; Use deep learning framework to analyze the depth information to identify the category of the object being tested; According to the position coordinates of each sensor in the sensor array and the pixel coordinates of the depth information, the actual distance value between each sensor in the sensor array and each point on the surface of the object to be measured is calculated using the pinhole camera model and inverse projection transformation, and multi-view distance data is obtained by fusion; For multi-view distance data, an improved iterative closest point algorithm model with shape prior knowledge is used for analysis, shape constraints are introduced, and the size information of the object being measured is iteratively solved. Based on the category, location, size and real-time environment dynamics of the object being measured, the obstacle avoidance result is output, and combined with the dynamic constraints of the AMR, an obstacle avoidance action sequence is generated.

2. The full-space adaptive AMR forklift obstacle avoidance method according to claim 1, characterized in that: Acquiring depth information uniformly distributed at both ends or along the length of the object through a sensor array synchronously or asynchronously includes: The sensor array is started to collect depth information at both ends or along the length direction of the object under test in a synchronous or asynchronous manner; Record the depth information and corresponding timestamp collected by each sensor and store them as a depth data set containing spatial coordinates and timestamps; De-noising, registration and normalization are performed on the collected depth information to obtain pre-processed depth information; The preprocessed depth information is organized according to timestamps and spatial coordinates to generate a depth information dataset.

3. The full-space adaptive AMR forklift obstacle avoidance method according to claim 1, characterized in that: The deep learning framework is used to analyze the depth information to identify the categories of the objects being tested, including: Determine the deep learning framework and train the deep learning framework based on the depth information; The preprocessed depth information is input into the trained deep learning framework for analysis. Through multi-layer nonlinear transformation, high-level features in the depth information are automatically extracted and feature representation vectors are output. Map the feature representation vector into a predefined category space and compare the similarity with the prototypes of each category in the category space or the samples in the preset training set; Based on the similarity results, the probability or score of the object being tested belonging to each category is calculated, and the category of the object being tested is determined in combination with the preset threshold or rule.

4. The full-space adaptive AMR forklift obstacle avoidance method according to claim 1, characterized in that: According to the position coordinates of each sensor in the sensor array and the pixel coordinates of the depth information, the actual distance value between each sensor in the sensor array and each point on the surface of the object to be measured is calculated using the pinhole camera model and inverse projection transformation. The multi-view distance data obtained by fusion includes: Get the position coordinates (x i ,y i ,z i ), where i represents the index of the sensor; Get the pixel coordinates (u, v) of the depth information corresponding to each sensor in the sensor array; For each sensor, the pinhole camera model considering radial distortion is used to convert the pixel coordinates (u, v) of the depth information into the three-dimensional point coordinates (X c ',Y c ',Z c '); The inverse projection transformation formula is used to transform the three-dimensional point coordinates (X c ',Y c ',Z c ') is converted to the global coordinate system to obtain the global coordinate (X g ,Y g ,Z g ); In the global coordinate system, according to the position coordinates of the sensor (x i ,y i ,z i ) and the coordinates of the surface point of the object after the inverse projection transformation (X g ,Y g ,Z g ), and use the Euclidean distance formula to calculate the actual distance value d between each sensor and each point on the surface of the object being measured i ; The actual distance value d calculated by all sensors i Fusion is performed to form a multi-view distance data set D = {d1, d2, ..., d n }, where n is the number of sensors; in, The conversion formula of the pinhole camera model considering radial distortion is: Where k1 and k2 are the first-order and second-order radial distortion coefficients, respectively, and c x and c y is the optical center coordinate, f x and f y is the focal length of the camera, r is the distance from the pixel to the optical center, d uv is the depth value, u and v are the pixel coordinates; The inverse projection transformation formula is: In the formula, s x 、s y 、s z They are the scaling factors in the X, Y, and Z axes, respectively. ij is the element of the rotation matrix R used to describe the rotation relationship between the global coordinate system and the camera coordinate system. i and j represent the row and column respectively. x ,t y ,t z They are the translation vectors in the X, Y, and Z axes respectively.

5. The full-space adaptive AMR forklift obstacle avoidance method according to claim 1, characterized in that: For multi-view distance data, an improved iterative closest point algorithm model with shape prior knowledge is used for analysis. Shape constraints are introduced and the size information of the object being measured is iteratively solved, including: Using the acquired shape prior knowledge, the initial shape of the object to be measured is estimated to obtain an initial geometric model; Establish shape constraints based on prior knowledge of the object being measured; Using an improved iterative closest point algorithm model, the closest distance between corresponding point sets is calculated based on multi-view distance data and the current initial geometric model, and the initial geometric model is updated at least once to obtain a new geometric model; Applying shape constraints to a new geometric model, adjusting the model to satisfy the shape constraints, obtaining an adjusted geometric model, and determining the difference between the geometric models before and after the adjustment. If the difference is less than a set threshold, the adjusted geometric model is used as the final geometric model. The dimensional information of the object under test, including length, width, height and volume, is extracted from the final geometric model.

6. The full-space adaptive AMR forklift obstacle avoidance method according to claim 1, characterized in that: Based on the type, location, size and real-time environment of the object being measured, the obstacle avoidance result is output, and combined with the dynamic constraints of the AMR, the obstacle avoidance action sequence is generated, including: Construct a three-dimensional environment model including the object to be measured and environmental elements according to the type of the object to be measured, the location of the object to be measured, the size information of the object to be measured and the real-time environmental dynamics; Define nodes in the three-dimensional environment model, each node represents a location point or a small area in the model; Set status attributes for each node, including whether it is reachable, whether it has been visited, and whether it belongs to the open list or the closed list; Initialize the starting point and the end point, add the current position of the object to be measured to the open list, use the A algorithm to search for the optimal path from the starting point to the end point in the model, and output the preliminary obstacle avoidance path; Introduce reinforcement learning mechanism to dynamically adjust the initial obstacle avoidance path and output the optimized obstacle avoidance path; Generate obstacle avoidance action sequences based on the optimized obstacle avoidance path and the dynamic constraints of the AMR: in, The obstacle avoidance action sequence is: i k =θ k-1 +Δθ×sign(P t -P c ) In the formula, θ k is the steering angle at time step k, θ k-1 is the steering angle of the previous time step, Δθ is the adjustment step of the steering angle, P t , P c are the target position and the current position respectively, and sign is the sign function used to calculate the value of the target position P according to the target position P. t and the current position P c The difference between the two determines the steering direction, v k is the driving speed at time step k, v max is the maximum driving speed, D t is the distance threshold, h k is the fork tine lifting height at time step k, h t is the target fork height, Δh is the adjustment step of the fork height, H t is the height threshold, h c is the current fork tine height, H o is the height of the obstacle.

7. The full-space adaptive AMR forklift obstacle avoidance method according to claim 1, characterized in that: Based on the type, location, size and real-time environment of the object being measured, the obstacle avoidance result is output, and after the obstacle avoidance action sequence is generated in combination with the dynamic constraints of the AMR, the following is also included: Convert the optimized obstacle avoidance action sequence into control instructions and send them to the AMR's actuator to drive the AMR to perform obstacle avoidance actions; During the obstacle avoidance process, the obstacle avoidance action is monitored in real time, and closed-loop control is performed on the obstacle avoidance action based on the acquired feedback data and dynamic updates of the environment; When it is detected that the obstacle cannot be avoided safely or the AMR fails, emergency braking or avoidance mechanism is initiated.

8. The full-space adaptive AMR forklift obstacle avoidance method according to any one of claims 1 to 7, characterized in that: The sensor array includes: multiple sensor modules arranged in a surrounding pattern on the AMR, each sensor is configured with a specific horizontal field of view angle and a vertical field of view angle, forming an interlaced and fully covered cone-shaped sensing space; Among them, the representation range of the cone perception space is: Where n is the number of sensors, θ h1 ,θ h2 are the horizontal and vertical field of view angles of the i-th sensor module, respectively. is the effective detection distance of the i-th sensor module.

9. A full-space monitoring AMR forklift, characterized in that: include: AMR forklift; Sensor arrays; as well as, A central controller, configured to execute the method according to any one of claims 1 to 8.

10. The full-space monitoring AMR forklift according to claim 9, characterized in that: The sensor array includes multiple sensors arranged in a circle on the AMR, each sensor is configured with a specific horizontal field of view angle and vertical field of view angle, forming an interlaced and fully covered cone perception space; Among them, the representation range of the cone perception space is: Where n is the number of sensors, θ h1 ,θ h2 are the horizontal and vertical field of view angles of the i-th sensor module, respectively. is the effective detection distance of the i-th sensor module.

Images