AGV dynamic obstacle avoidance control system based on multi-sensor fusion

Abstract

The invention relates to the technical field of sensor fusion, in particular to an AGV (Automatic Guided Vehicle) dynamic obstacle avoidance control system based on multi-sensor fusion, which comprises a visual path planning module used for determining an obstacle boundary position according to an obstacle image detected by a visual sensor through an obstacle recognition model based on an improved YOLOv10 neural network, generating a visual obstacle avoidance map according to the obstacle boundary position; the electromagnetic tracing module is used for generating an electromagnetic obstacle position according to the electromagnetic wire track detected by the electromagnetic sensor; the infrared obstacle avoidance module is used for generating an infrared obstacle position according to the obstacle distance detected by the infrared sensor; and the dynamic obstacle avoidance control module is used for substituting the visual obstacle avoidance map, the electromagnetic obstacle position and the infrared obstacle position into an improved local path planning model based on a dynamic window method to generate a linear velocity and an angular velocity for AGV dynamic obstacle avoidance control. According to the invention, intelligent dynamic obstacle avoidance control of the AGV is realized.

Description

Technical Field

[0001] The present invention relates to the field of sensor fusion technology, and in particular to an AGV dynamic obstacle avoidance control system based on multi-sensor fusion. Background Art

[0002] Path planning is one of the core technologies of AGV intelligent vehicles, which includes two levels: global path planning and local path planning. Global path planning refers to determining the best route in the entire driving environment, while local path planning targets specific obstacles and real-time dynamic information, and decides how to adjust the driving trajectory under the current circumstances. Common path planning methods include A algorithm, artificial potential field method, dynamic window method, etc. During the search process, the A algorithm only considers the static distance from the starting point to the end point and cannot adapt to real-time changes in the environment. When there is a complex environment, the potential field function of the artificial potential field method becomes more difficult, which will lead to suboptimal path planning. The dynamic window approach (DWA) algorithm can better cope with dynamic environmental conditions.

[0003] Therefore, the current dynamic obstacle avoidance process of AGV mostly adopts the dynamic window approach (DWA). The dynamic window approach is a path planning algorithm. Its core is to achieve safe and efficient navigation by dynamically evaluating feasible trajectories in the speed space.

[0004] However, the path planning trajectory generated by the existing dynamic window method may fall into the local optimal path in narrow channels or dense obstacles. This is because the decision-making ability of the evaluation function of the trajectory scoring mechanism is not intelligent enough.

[0005] Therefore, how to realize the intelligent dynamic obstacle avoidance control of AGV based on the dynamic window method of multi-sensor fusion is a technical problem that needs to be solved. Summary of the Invention

[0006] To this end, the present invention provides an AGV dynamic obstacle avoidance control system based on multi-sensor fusion. The accurate obstacle boundary position is determined by an obstacle recognition model based on an improved YOLOv10 neural network, thereby realizing the intelligent perception of the AGV. The obstacle boundary position is combined with the detection data of electromagnetic sensors and infrared sensors to improve the decision-making ability of the trajectory scoring mechanism of the dynamic window method, thereby realizing the intelligent dynamic obstacle avoidance control of the AGV.

[0007] To achieve the above objectives, the present invention proposes an AGV dynamic obstacle avoidance control system based on multi-sensor fusion. The AGV is equipped with a visual sensor, an electromagnetic sensor, and an infrared sensor for dynamic obstacle avoidance. The AGV dynamic obstacle avoidance control system includes:

[0008] A visual path planning module, configured to use the obstacle image detected by the visual sensor to determine the obstacle boundary position through an obstacle recognition model based on an improved YOLOv10 neural network, and generate a visual obstacle avoidance map based on the obstacle boundary position;

[0009] An electromagnetic tracking module, configured to generate the position of an electromagnetic obstacle based on the electromagnetic line trajectory detected by the electromagnetic sensor;

[0010] an infrared obstacle avoidance module, configured to generate an infrared obstacle position according to the obstacle distance detected by the infrared sensor;

[0011] The dynamic obstacle avoidance control module is used to substitute the visual obstacle avoidance map, electromagnetic obstacle positions and infrared obstacle positions into an improved local path planning model based on the dynamic window method to generate linear velocity and angular velocity for AGV dynamic obstacle avoidance control.

[0012] Furthermore, the backbone network of the obstacle recognition model includes a first-level composite layer to a fourth-level composite layer, and each composite layer includes a composite unit and a C3 unit, wherein the output of the C3 unit of the fourth-level composite layer is connected to an attention mechanism, and the attention mechanism is used to capture obstacle boundary information.

[0013] Furthermore, the attention mechanism includes a residual connection layer, a pooling operation layer, a feature processing layer, a split activation layer and a weighted layer;

[0014] The residual connection layer is used to filter the features of the attention input image to generate a first feature map;

[0015] The pooling operation layer is used to perform average pooling operations in three directions on the first feature map to generate three second feature maps;

[0016] The feature processing layer is used to perform feature splicing, convolution operation, batch normalization and activation function operation on the three second feature maps to generate a third feature map;

[0017] The split activation layer is used to perform channel splitting on the third feature map and perform convolution activation operations on each of the channels to generate a fourth feature map;

[0018] The weighted layer is used to perform weighted multiplication of the fourth feature map and the attention input image in the channel dimension to generate an attention output image.

[0019] Furthermore, the composite unit includes a convolution layer, a batch normalization layer and a SiLU activation function layer, wherein the convolution sizes of the convolution layers from the first-level composite layer to the fourth-level composite layer are 1×1, 1×1, 3×3, and 9×9, respectively.

[0020] Furthermore, the number of channels of the bottleneck of the main branch of the C3 unit from the first-level composite layer to the fourth-level composite layer is 64, 128, 256 and 512 respectively.

[0021] Furthermore, the obstacle recognition model is provided with an improved loss function, which is constructed based on the intersection-over-union ratio minus the variance of the distance loss term, the angle loss term and the shape loss term.

[0022] In the above solution, the AGV system can achieve safe, efficient and smooth obstacle avoidance navigation in complex and dynamic industrial scenarios.

[0023] Furthermore, the dynamic obstacle avoidance control module includes a normalization unit, a dynamic obstacle avoidance function generation unit, a weight generation unit and an evaluation function generation unit;

[0024] The normalization unit is used to calculate a normalization coefficient based on the obstacle avoidance starting coordinates, obstacle avoidance target coordinates and obstacle avoidance map area determined by the visual obstacle avoidance map;

[0025] The dynamic obstacle avoidance function generating unit is configured to generate the obstacle avoidance correction function according to the obstacle avoidance target coordinates, the normalization coefficient, and the current obstacle avoidance coordinates determined by the visual obstacle avoidance map;

[0026] The weight generating unit is used to generate a direction weight and a speed sub-function according to the electromagnetic obstacle position and the infrared obstacle position respectively;

[0027] The evaluation function generating unit is used to construct the improved local path planning model according to the obstacle avoidance correction function, the speed weight and the orientation weight, and the improved local path planning model generates the linear velocity and angular velocity for performing dynamic obstacle avoidance control of the AGV.

[0028] Furthermore, the improved local path planning model is constructed based on a dynamic window method, which includes a distance sub-function, and the dynamic obstacle avoidance control module further includes a distance sub-function generation unit;

[0029] The distance sub-function generating unit is used to determine the distance sub-function according to the visual obstacle avoidance map and the infrared obstacle position.

[0030] Further, the weight generation unit includes a direction weight generation subunit;

[0031] The weight generating subunit is configured to calculate a direction weight based on a ratio of a derivative of the distance subfunction to an obstacle width, wherein the obstacle width is determined according to a visual obstacle avoidance map.

[0032] Further, the weight generating unit includes a speed weight generating subunit;

[0033] The speed weight generating subunit is used to calculate the speed weight according to the obstacle weight extreme value determined by the infrared obstacle position and the number of obstacles determined by the visual obstacle avoidance map.

[0034] Furthermore, the improved local path planning model is constructed based on a dynamic window method, which includes a distance sub-function, and the dynamic obstacle avoidance control module further includes a distance sub-function generation unit;

[0035] The distance sub-function generating unit is used to determine the distance sub-function according to the visual obstacle avoidance map and the infrared obstacle position.

[0036] In the above solution, through multi-sensor data fusion, dynamic path planning and refined weight distribution, the AGV's obstacle avoidance ability in dynamic and complex environments, the real-time and safety of path planning are significantly improved, while taking into account computing efficiency. It is suitable for the scenario requirements of AGV autonomous navigation and dynamic obstacle avoidance.

[0037] Compared with the prior art, the present invention has the following advantages:

[0038] 1. By accurately determining the obstacle boundary location using an improved YOLOv10 neural network obstacle recognition model, the AGV's intelligent perception is realized. By combining the obstacle boundary location with the detection data of electromagnetic sensors and infrared sensors, the decision-making ability of the dynamic window method's trajectory scoring mechanism is improved, realizing intelligent dynamic obstacle avoidance control for the AGV.

[0039] 2. The AGV system can achieve safe, efficient and smooth obstacle avoidance navigation in complex and dynamic industrial scenarios.

[0040] 3. Through multi-sensor data fusion, dynamic path planning and refined weight distribution, the AGV's obstacle avoidance capability in dynamic and complex environments, the real-time performance and safety of path planning are significantly improved, while taking into account computing efficiency. It is suitable for scenarios where AGVs require autonomous navigation and dynamic obstacle avoidance. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Schematic diagram of the process of an AGV dynamic obstacle avoidance control system based on multi-sensor fusion according to an embodiment of the present invention;

[0042] Figure 2 Schematic diagram of the structure of the obstacle recognition model of the AGV dynamic obstacle avoidance control system based on multi-sensor fusion according to an embodiment of the present invention;

[0043] Figure 3 Schematic diagram of the calculation flow of the improved local path planning model of the obstacle recognition model of the AGV dynamic obstacle avoidance control system based on multi-sensor fusion according to an embodiment of the present invention;

[0044] Figure 4 The present invention provides a flowchart of an improved local path planning model for an obstacle recognition model of an AGV dynamic obstacle avoidance control system based on multi-sensor fusion according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are merely used to explain the present invention and are not intended to limit the present invention.

[0046] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0047] It should be noted that, in the description of the present invention, terms such as "up", "down", "left", "right", "inside", and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation on the present invention.

[0048] Furthermore, it should be noted that, in the description of the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0049] like Figures 1 to 4 As shown, the present invention provides an AGV dynamic obstacle avoidance control system based on multi-sensor fusion. The obstacle boundary position is accurately determined by an obstacle recognition model based on an improved YOLOv10 neural network, thereby realizing the intelligent perception of the AGV. The obstacle boundary position is combined with the detection data of the electromagnetic sensor and the infrared sensor to improve the decision-making ability of the trajectory scoring mechanism of the dynamic window method, thereby realizing the intelligent dynamic obstacle avoidance control of the AGV.

[0050] like Figures 1 to 4 As shown, this embodiment proposes an AGV dynamic obstacle avoidance control system based on multi-sensor fusion. The AGV is equipped with a visual sensor, an electromagnetic sensor, and an infrared sensor for dynamic obstacle avoidance. The AGV dynamic obstacle avoidance control system includes:

[0051] A visual path planning module, configured to use the obstacle image detected by the visual sensor to determine the obstacle boundary position through an obstacle recognition model based on an improved YOLOv10 neural network, and generate a visual obstacle avoidance map based on the obstacle boundary position;

[0052] An electromagnetic tracking module, configured to generate the position of an electromagnetic obstacle based on the electromagnetic line trajectory detected by the electromagnetic sensor;

[0053] an infrared obstacle avoidance module, configured to generate an infrared obstacle position according to the obstacle distance detected by the infrared sensor;

[0054] The dynamic obstacle avoidance control module is used to substitute the visual obstacle avoidance map, electromagnetic obstacle positions and infrared obstacle positions into an improved local path planning model based on the dynamic window method to generate linear velocity and angular velocity for AGV dynamic obstacle avoidance control.

[0055] Specifically, visual sensors capture images of the environment using cameras or 3D vision devices, then combine image processing algorithms to detect obstacles and model the environment. They then construct a real-time 2D / 3D map of the surrounding environment, identify obstacle shapes, and calculate obstacle distances using binocular vision or Time-of-Flight cameras, thus addressing the blind spots of electromagnetic tracking. Their advantages include a large amount of information, adaptability to scenarios without fixed paths, and lower cost than laser navigation systems. Their limitations lie in their dependence on lighting conditions, which can lead to misjudgments due to strong light or shadows. Infrared sensors are used for close-range safety protection. They emit infrared light and detect reflected signals to determine the presence and distance of obstacles, excelling in short-range obstacle avoidance and emergency braking. Their advantages lie in their low cost and fast response time, as well as their sensitivity to non-reflective objects, making them suitable for complex industrial environments. Their limitations lie in their limited detection range and susceptibility to interference from ambient temperature or smoke. Electromagnetic sensors guide AGVs along fixed paths by detecting electromagnetic wires or magnetic spikes buried in the ground. Their advantages lie in their high stability and suitability for structured environments, such as fixed production lines. However, their limitations lie in their lack of flexibility and inability to handle dynamic obstacles. They also require pre-deployed infrastructure, resulting in high costs. Therefore, the coordinated application of the three is required to further improve the obstacle avoidance robustness and environmental adaptability of AGV and realize the intelligent dynamic obstacle avoidance control of AGV.

[0056] Furthermore, the backbone network of the obstacle recognition model includes a first-level composite layer to a fourth-level composite layer, and each composite layer includes a composite unit and a C3 unit, wherein the output of the C3 unit of the fourth-level composite layer is connected to an attention mechanism, and the attention mechanism is used to capture obstacle boundary information.

[0057] Furthermore, the attention mechanism includes a residual connection layer, a pooling operation layer, a feature processing layer, a split activation layer and a weighted layer;

[0058] The residual connection layer is used to filter the features of the attention input image to generate a first feature map;

[0059] The pooling operation layer is used to perform average pooling operations in three directions on the first feature map to generate three second feature maps;

[0060] The feature processing layer is used to perform feature splicing, convolution operation, batch normalization and activation function operation on the three second feature maps to generate a third feature map;

[0061] The split activation layer is used to perform channel splitting on the third feature map and perform convolution activation operations on each of the channels to generate a fourth feature map;

[0062] The weighted layer is used to perform weighted multiplication of the fourth feature map and the attention input image in the channel dimension to generate an attention output image.

[0063] Specifically, the newly added attention mechanism in the improved YOLOv10 obstacle recognition model significantly enhances the ability to capture boundary information through multi-directional feature fusion and dynamic weight adjustment. The residual connection layer can retain key features and enhance gradient flow. By adding the original input features with the features after attention processing, it avoids feature degradation in the deep network, prevents gradient vanishing, improves training stability, retains the original details of the obstacle contour, and avoids edge blurring caused by multiple convolutions.

[0064] The multi-directional pooling operation layer can capture spatial context information, perform average pooling along the three directions of height, width, and diagonal, generate multi-view feature maps, and extract the directional features of the horizontal / vertical edges of the obstacle boundaries, enhance the adaptability to irregular shapes, expand the receptive field, fuse local and global information, and reduce the missed detection of small obstacles.

[0065] The feature processing layer can fuse multi-dimensional information. After splicing multi-directional features, it performs nonlinear fusion through convolution, batch normalization (BN), and SiLU activation function, thereby generating highly discriminative features, enhancing the contrast between boundaries and backgrounds, accelerating convergence through batch normalization, and alleviating overfitting.

[0066] Splitting the activation layer can refine channel features. By splitting the channels and performing independent convolutional activation, multiple sets of fine-grained features are generated. This allows different channels to focus on differentiated color and edge features, improving the ability to analyze complex obstacles. This avoids feature coupling between channels and enhances model flexibility.

[0067] The weighted layer enables dynamic feature enhancement by multiplying the refined feature map with the original input according to channel weights, highlighting important features. This enables adaptive feature selection, suppressing noise and non-critical areas. It also allows for precise location of obstacle boundaries in complex scenes with occlusion and changing lighting.

[0068] Furthermore, the composite unit includes a convolution layer, a batch normalization layer and a SiLU activation function layer, wherein the convolution sizes of the convolution layers from the first-level composite layer to the fourth-level composite layer are 1×1, 1×1, 3×3, and 9×9, respectively.

[0069] Specifically, the stepped design of the convolution kernel size in the four-level composite layer significantly improves the model's perception of obstacles of different sizes through multi-scale feature fusion and hierarchical receptive field expansion. The receptive field of the 1×1 convolution in the first-level composite layer is local pixel association, capturing underlying features such as texture and edges. The receptive field of the 1×1 convolution in the second-level composite layer is channel interaction optimization, compressing redundant channels and strengthening the transmission of key features. The receptive field of the 3×3 convolution in the third-level composite layer is medium-area coverage, enabling the recognition of the outlines of conventional obstacles. The receptive field of the 9×9 convolution in the fourth-level composite layer is global perception, enabling the recognition of large obstacles and distant targets.

[0070] Therefore, the shallow first and second composite layers use 1×1 convolutions to focus on channel dimensionality adjustment, reducing the dimensionality to 64 channels to avoid premature introduction of spatial noise. The deeper third and fourth composite layers gradually expand the spatial perception range through 3×3 and 9×9 convolutions, achieving coverage of multi-scale obstacles. The 1×1 convolutions achieve dual functions: lightweighting and feature reorganization. The lightweighting is reflected in parameter compression: the first two 1×1 convolutions compress the number of channels from the input dimension of 256 to an intermediate dimension of 64, reducing the number of parameters in the subsequent 3×3 / 9×9 convolutions by approximately 75%. The 9×9 large kernel convolution's global modeling capability covers over 80% of the original image, enabling the identification of the overall structure of large obstacles and local-global correlations in occlusion scenarios. This synergistic effect is achieved with the attention mechanism. The attention mechanism following the fourth composite layer forms a coarse localization and fine correction process.

[0071] Furthermore, the number of channels of the bottleneck of the main branch of the C3 unit from the first-level composite layer to the fourth-level composite layer is 64, 128, 256 and 512 respectively.

[0072] Therefore, efficient detection of multi-scale obstacles is achieved through progressive feature capacity expansion and hierarchical semantic abstraction.

[0073] Furthermore, the obstacle recognition model is provided with an improved loss function, which is constructed based on the intersection-over-union ratio minus the variance of the distance loss term, the angle loss term, and the shape loss term, specifically:

[0074]

[0075] Where, L iou To improve IOU, IOU is the intersection-over-union ratio, Δ is the distance penalty term, Ω is the shape penalty term, and ∧ is the angle penalty term.

[0076] In the above solution, the AGV system can achieve safe, efficient and smooth obstacle avoidance navigation in complex and dynamic industrial scenarios.

[0077] It should be noted that the local path planning model in this embodiment uses the speed model in the kinematic model of the mobile nursing service robot as the calculation model. According to the conditions of the mobile nursing service robot itself and the limitations of the environment, there is a range of feasible speeds (speed group) (v, w). Then, sampling is performed within this speed range, and the motion path of each speed is calculated through the kinematic model. Finally, the evaluation function is used to score the simulated path, and the optimal path and speed group are selected to control the AGV movement.

[0078] Specifically, the kinematic model of the two-wheel differentially driven AGV is:

[0079]

[0080] In the formula, the next moment is recorded as t+1, the current moment is recorded as t, and θ is the angle between the center line and the x-axis.

[0081] Specifically, the kinematic model of the four-wheeled AGV is:

[0082]

[0083] Where, point P is the center of mass of the wheeled AGV, d is the distance from point O to point P, b is half the distance between the two rear-drive wheels, 2r is the diameter of the rotating wheel, and θ is the angle between the driving direction of the wheeled AGV and the horizontal direction in the rectangular coordinate system. Respectively represent the deflection angles of the driving wheels on both sides.

[0084] Specifically, the improved evaluation function in this embodiment is:

[0085]

[0086] Where ε is the normalization coefficient, is the orientation weight, α is the setting coefficient, R is the horizontal projection size of the largest obstacle in the detection window, head(v,w) is the orientation function, β is the obstacle weight, dist(v,w) is the obstacle function, δ is the velocity weight, vel(v,w) is the velocity function, and (v,w) is the velocity group.

[0087] Preferably, in the formula, the coefficient α is set to 1. When the detection window of the AGV sensor is small, the coefficient α can be set to 0.2 to ensure that the proportion of the orientation weight is not too large.

[0088] It should be noted that the conventional evaluation function is:

[0089] G(v,w)=ε(αhead(v,w)+δdist(v,w)+βvel(v,w))

[0090] Among them, the orientation weight α, obstacle weight γ and speed weight β are all set values, which leads to the poor environmental adaptability of the algorithm. Specifically, since it cannot be adjusted according to the density of obstacles, especially the obstacle weight must be set to a large value to ensure that it does not encounter obstacles, the robot cannot directly cross the dense obstacle area.

[0091] It can be understood that the main function of the heading score function head(v,w) is to continuously guide the AGV to move towards the target point during its movement. Its value can be calculated by the following formula:

[0092] head(v,w)=180-θ

[0093] Here, θ represents the angular difference between the end of the path simulated using the kinematic model and the target point. Because changes in the AGV speed group directly affect the AGV posture state, the orientation score function also changes with changes in the speed group and decreases as the AGV speed value increases.

[0094] The obstacle scoring function dist(v,w) represents the distance from the path simulated using the kinematic model to the nearest obstacle. A smaller distance value indicates a greater likelihood of the AGV colliding with an obstacle. Conversely, a larger value indicates a path further away from obstacles using the current speed profile, making the AGV safer. The speed scoring function vel(v,w) represents the AGV's movement speed.

[0095] Furthermore, the dynamic obstacle avoidance control module includes a normalization unit, a dynamic obstacle avoidance function generation unit, a weight generation unit and an evaluation function generation unit;

[0096] The normalization unit is used to calculate a normalization coefficient based on the obstacle avoidance starting coordinates, obstacle avoidance target coordinates and obstacle avoidance map area determined by the visual obstacle avoidance map;

[0097] The dynamic obstacle avoidance function generating unit is configured to generate the obstacle avoidance correction function according to the obstacle avoidance target coordinates, the normalization coefficient, and the current obstacle avoidance coordinates determined by the visual obstacle avoidance map;

[0098] The weight generating unit is used to generate a direction weight and a speed sub-function according to the electromagnetic obstacle position and the infrared obstacle position respectively;

[0099] The evaluation function generating unit is used to construct the improved local path planning model according to the obstacle avoidance correction function, the speed weight and the orientation weight, and the improved local path planning model generates the linear velocity and angular velocity for performing dynamic obstacle avoidance control of the AGV.

[0100] Specifically, the normalization unit and the dynamic obstacle avoidance function generation unit are:

[0101]

[0102] number(v,ω)=x e +y e -μln(|x i -1|×|y i -1|)

[0103] Where M is the area of the obstacle avoidance map, (x f ,y f ) is the starting coordinate of obstacle avoidance, (x e ,y e ) is the coordinate of the obstacle avoidance target, (v,w) is the velocity group including linear velocity and angular velocity, (x i ,y i ) is the current obstacle avoidance coordinate, number(v,w) is the obstacle number quantum function, and μ is the normalization coefficient.

[0104] Furthermore, the improved local path planning model is constructed based on a dynamic window method, which includes a distance sub-function, and the dynamic obstacle avoidance control module further includes a distance sub-function generation unit;

[0105] The distance sub-function generating unit is used to determine the distance sub-function according to the visual obstacle avoidance map and the infrared obstacle position.

[0106] The distance sub-function is:

[0107] dist(v,w)=σ(dist1(v,w)+etadist2(v,w))

[0108] Where dist(v,w) represents the distance function, σ and η represent two setting coefficients, which are determined according to the classification of the infrared obstacle position. dist1(v,w) and dist2(v,w) represent the distance function determined by the visual obstacle avoidance map and the distance function determined by the infrared obstacle position.

[0109] Further, the weight generation unit includes a direction weight generation subunit;

[0110] The weight generating subunit is configured to calculate a direction weight based on a ratio of a derivative of the distance subfunction to an obstacle width, wherein the obstacle width is determined according to a visual obstacle avoidance map.

[0111] The orientation weight is Among them, α is the direction weight setting value, Dist(v,w) represents the distance sub-function, and W represents the obstacle width.

[0112] Further, the weight generating unit includes a speed weight generating subunit;

[0113] The speed weight generating subunit is used to calculate the speed weight according to the obstacle weight extreme value determined by the infrared obstacle position and the number of obstacles determined by the visual obstacle avoidance map.

[0114] Specifically, the speed weight is:

[0115]

[0116] Where, β is the speed weight, β min The minimum value of the speed weight determined by the infrared obstacle position, β max is the maximum value of the speed weight determined by the infrared obstacle position, and N is the number of obstacles determined by the visual obstacle avoidance map.

[0117] The improved evaluation function of the improved local path planning model is:

[0118]

[0119] Where G(v,w) is the improved evaluation function, the heading weight setting value α, the speed weight β, the distance weight setting value δ, and the obstacle number weight τ are all fixed values, Dist(v,w) represents the distance sub-function, W represents the obstacle width, number(v,w) is the obstacle number sub-function, β is the speed weight, head(v,w) represents the heading scoring function, dist(v,w) represents the distance sub-function, vel(v,w) is the speed sub-function, and ε is the normalization coefficient.

[0120] In the above solution, through multi-sensor data fusion, dynamic path planning and refined weight distribution, the AGV's obstacle avoidance ability in dynamic and complex environments, the real-time and safety of path planning are significantly improved, while taking into account computing efficiency. It is suitable for the scenario requirements of AGV autonomous navigation and dynamic obstacle avoidance.

[0121] In this embodiment, the intelligent perception of the AGV is achieved by determining the accurate obstacle boundary position through an obstacle recognition model based on an improved YOLOv10 neural network. By combining the obstacle boundary position with the detection data of electromagnetic sensors and infrared sensors, the decision-making ability of the trajectory scoring mechanism of the dynamic window method is improved, and the intelligent dynamic obstacle avoidance control of the AGV is realized. The AGV system is able to achieve safe, efficient, and smooth obstacle avoidance navigation in complex and dynamic industrial scenarios. Through multi-sensor data fusion, dynamic path planning, and refined weight allocation, the obstacle avoidance capability of the AGV in dynamic and complex environments, the real-time and safety of path planning are significantly improved, while taking into account computational efficiency, and is suitable for the scenario requirements of AGV autonomous navigation and dynamic obstacle avoidance.

[0122] Thus far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

[0123] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that the present invention is susceptible to various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. An AGV dynamic obstacle avoidance control system based on multi-sensor fusion, characterized in that: The AGV is equipped with visual sensors, electromagnetic sensors, and infrared sensors for dynamic obstacle avoidance. The AGV dynamic obstacle avoidance control system includes: A visual path planning module, configured to use the obstacle image detected by the visual sensor to determine the obstacle boundary position through an obstacle recognition model based on an improved YOLOv10 neural network, and generate a visual obstacle avoidance map based on the obstacle boundary position; An electromagnetic tracking module, configured to generate the position of an electromagnetic obstacle based on the electromagnetic line trajectory detected by the electromagnetic sensor; an infrared obstacle avoidance module, configured to generate an infrared obstacle position according to the obstacle distance detected by the infrared sensor; The dynamic obstacle avoidance control module is used to substitute the visual obstacle avoidance map, electromagnetic obstacle positions and infrared obstacle positions into an improved local path planning model based on the dynamic window method to generate linear velocity and angular velocity for AGV dynamic obstacle avoidance control.

2. The AGV dynamic obstacle avoidance control system based on multi-sensor fusion according to claim 1 is characterized in that: The backbone network of the obstacle recognition model includes the first to fourth composite layers, each composite layer includes a composite unit and a C3 unit, wherein the output of the C3 unit of the fourth composite layer is connected to an attention mechanism, which is used to capture obstacle boundary information.

3. The AGV dynamic obstacle avoidance control system based on multi-sensor fusion according to claim 2 is characterized in that: The attention mechanism includes a residual connection layer, a pooling operation layer, a feature processing layer, a split activation layer and a weighted layer; The residual connection layer is used to filter the features of the attention input image to generate a first feature map; The pooling operation layer is used to perform average pooling operations in three directions on the first feature map to generate three second feature maps; The feature processing layer is used to perform feature splicing, convolution operation, batch normalization and activation function operation on the three second feature maps to generate a third feature map; The split activation layer is used to perform channel splitting on the third feature map and perform convolution activation operations on each of the channels to generate a fourth feature map; The weighted layer is used to perform weighted multiplication on the fourth feature map and the attention input image in the channel dimension to generate an attention output image.

4. The AGV dynamic obstacle avoidance control system based on multi-sensor fusion according to claim 2 is characterized in that: The composite unit includes a convolution layer, a batch normalization layer and a SiLU activation function layer, wherein the convolution sizes of the convolution layers from the first-level composite layer to the fourth-level composite layer are 1×1, 1×1, 3×3, and 9×9, respectively.

5. The AGV dynamic obstacle avoidance control system based on multi-sensor fusion according to claim 2 is characterized in that: The composite unit includes a convolution layer, a batch normalization layer and a SiLU activation function layer, wherein the convolution sizes of the convolution layers from the first-level composite layer to the fourth-level composite layer are 1×1, 1×1, 3×3, and 9×9, respectively.

6. The AGV dynamic obstacle avoidance control system based on multi-sensor fusion according to claim 1 is characterized in that: The obstacle recognition model is provided with an improved loss function, which is constructed based on the intersection-over-union ratio minus the variance of the distance loss term, the angle loss term, and the shape loss term.

7. The AGV dynamic obstacle avoidance control system based on multi-sensor fusion according to any one of claims 1 to 6, characterized in that: The dynamic obstacle avoidance control module includes a normalization unit, a dynamic obstacle avoidance function generation unit, a weight generation unit and an evaluation function generation unit; The normalization unit is used to calculate a normalization coefficient based on the obstacle avoidance starting coordinates, obstacle avoidance target coordinates and obstacle avoidance map area determined by the visual obstacle avoidance map; The dynamic obstacle avoidance function generating unit is configured to generate the obstacle avoidance correction function according to the obstacle avoidance target coordinates, the normalization coefficient, and the current obstacle avoidance coordinates determined by the visual obstacle avoidance map; The weight generating unit is used to generate a direction weight and a speed sub-function according to the electromagnetic obstacle position and the infrared obstacle position respectively; The evaluation function generating unit is used to construct the improved local path planning model according to the obstacle avoidance correction function, the speed weight and the orientation weight, and the improved local path planning model generates the linear velocity and angular velocity for performing dynamic obstacle avoidance control of the AGV.

8. The AGV dynamic obstacle avoidance control system based on multi-sensor fusion according to claim 7 is characterized in that: The improved local path planning model is constructed based on a dynamic window method, which includes a distance sub-function, and the dynamic obstacle avoidance control module further includes a distance sub-function generation unit; The distance sub-function generating unit is used to determine the distance sub-function according to the visual obstacle avoidance map and the infrared obstacle position.

9. The AGV dynamic obstacle avoidance control system based on multi-sensor fusion according to claim 7, characterized in that: The weight generation unit includes a direction weight generation subunit; The weight generating subunit is configured to calculate a direction weight based on a ratio of a derivative of the distance subfunction to an obstacle width, wherein the obstacle width is determined according to a visual obstacle avoidance map.

10. The AGV dynamic obstacle avoidance control system based on multi-sensor fusion according to claim 7, characterized in that: The improved local path planning model is constructed based on a dynamic window method, which includes a distance sub-function, and the dynamic obstacle avoidance control module further includes a distance sub-function generation unit; The distance sub-function generating unit is used to determine the distance sub-function according to the visual obstacle avoidance map and the infrared obstacle position.

Images