Front-end picture self-adaptive obstacle avoidance analysis method based on AGV (Automatic Guided Vehicle)

Abstract

The invention discloses a front-end picture self-adaptive obstacle avoidance analysis method based on an AGV transport vehicle, relates to the technical field of computer vision, realizes accurate understanding and path planning of a complex driving environment by fusing multi-source sensing data of a camera and a millimeter wave radar, and realizes accurate obstacle avoidance by synchronously acquiring gray images and pixel-level distance information. A reliable ground area is extracted based on gray features and morphological processing, a dynamic drivable area is constructed in combination with an image geometric structure, a strong edge detection and linear clustering technology is utilized, left and right vanishing points are automatically identified, a geometric corridor is constructed, and the robustness of environment modeling is improved; secondly, through distance mapping and clustering analysis, obstacles are identified, an influence circle is generated, and a safety corridor is dynamically constructed; the optimal path is planned in real time in the safe corridor, it is ensured that the AGV runs safely and efficiently in a narrow or dynamic environment, and the method is suitable for an intelligent transportation system in complex scenes such as storage and manufacturing.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, specifically, it relates to an adaptive obstacle avoidance analysis method for the front-end image of an AGV transport vehicle. Background Technology

[0002] Automated Guided Vehicles (AGVs) have become a key infrastructure in modern factories, warehousing and logistics, and flexible manufacturing systems. In complex and dynamic environments, how to achieve efficient and safe autonomous navigation has become a core issue in promoting the intelligent upgrading of industries.

[0003] In existing technologies, obstacle avoidance methods for AGV transport vehicles often rely on a single sensor system, such as using only cameras for visual analysis or millimeter-wave radar for distance detection. This leads to incomplete environmental perception data and is prone to misjudgments due to changes in lighting, sensor noise, or object occlusion, failing to achieve high-precision obstacle recognition and ground area segmentation. Furthermore, existing methods often employ simple threshold segmentation or fixed rules for obstacle detection, lacking in-depth analysis of the geometric structure in the image. This results in a coarse definition of the drivable area, often ignoring dynamic factors in the actual environment, such as changes in ground texture or temporary obstacles. Consequently, AGVs are prone to getting stuck in blind spots or colliding in complex scenarios. In terms of path planning, existing technologies are often based on static maps or preset paths, without dynamically generating paths using real-time sensor data. This makes them unsuitable for narrow or unstructured environments, and their low computational efficiency causes response delays, affecting the AGV's real-time obstacle avoidance capabilities and increasing the probability of incorrect obstacle avoidance or path deviation.

[0004] To address the aforementioned issues, this invention proposes an adaptive obstacle avoidance analysis method based on the front-end view of an AGV transport vehicle. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an adaptive obstacle avoidance analysis method based on the front-end image of an AGV transport vehicle, which solves the problems of inaccurate obstacle avoidance, non-optimized path planning, and poor environmental adaptability in existing technologies.

[0006] The objective of this invention can be achieved through the following technical solutions: An adaptive obstacle avoidance analysis method based on the front-end image of an AGV transport vehicle, the method comprising: Step 1: Simultaneously acquire grayscale images of the AGV during its driving process and the straight-line distance between the AGV and each grayscale pixel in the grayscale image based on the camera and millimeter-wave radar. Step 2: Based on the gray values ​​of each gray pixel, identify suspected ground areas and non-ground areas, extract all suspected ground areas for secondary refinement and segmentation, and determine the binarized ground areas. Then, extract strong edge pixels from the grayscale image, determine all straight lines in the grayscale image based on the strong edge pixels, classify the straight lines based on their slope, and construct left-leaning straight line groups and right-leaning straight line groups. Intersection point analysis is performed on the left-leaning straight line group and the right-leaning straight line group respectively to determine the left intersection point set and the right intersection point set. Then, cluster analysis is performed on the two intersection point sets respectively, and the two largest cluster centers after the two intersection point sets are taken as candidate points for the left vanishing point and the right vanishing point. Step 3: Extract the bottom center point of the grayscale image and form a geometric corridor with the left and right vanishing point candidate points. Combine the binarized ground area with the geometric corridor to calculate the fusion area and determine the drivable area. Step 4: Map the straight-line distance of the corresponding grayscale pixels onto the plane of the drivable area, form a set of obstacle reference points based on the value of the straight-line distance, and generate an obstacle influence circle within the drivable area; Within the drivable area, starting from the decision-making starting line, and combining the boundary of the drivable area and the obstacle influence circle, the left and right boundary lines of the safety corridor are determined, and an effective safety corridor is generated. Based on all decision starting points on the decision starting line, the optimal travel path of the AGV is determined within the effective safety corridor.

[0007] As a further aspect of the present invention, the specific method for simultaneously acquiring grayscale images of the AGV during its operation and the straight-line distance between the AGV and each grayscale pixel in the grayscale image based on the camera and millimeter-wave radar in step one is as follows: The AGV-based control system sends synchronous acquisition signals to the camera and millimeter-wave radar equipped on the AGV. The camera takes the AGV as the origin of the camera coordinates and captures the first frame of RGB image within F meters of the front end of the AGV, and converts it into a grayscale image G_gray, where F is the preset monitoring distance; The millimeter-wave radar collects point cloud data within F meters of the AGV front end, with the AGV as the spatial coordinate origin. The spatial coordinate system is then unified with the camera coordinate system. In the unified coordinate system, the straight-line distance between all grayscale pixels in the grayscale image G_gray and the AGV is determined. Arrange all grayscale pixels in the grayscale image G_gray in order from left to right and from top to bottom to obtain the grayscale pixel sequence g1, g2, ..., gj, where j is the total number of grayscale pixels; Extract the straight-line distance between each grayscale pixel and the AGV, and record the straight-line distance sequence d1, d2, ..., dj according to the order of the grayscale pixel sequence.

[0008] As a further aspect of the present invention, the specific method for identifying suspected ground areas and non-ground areas based on the grayscale values ​​of each grayscale pixel in step two is as follows: Traverse the sequence of grayscale pixels g1, g2, ..., gj, and extract the grayscale value of each grayscale pixel to obtain the sequence of grayscale pixel values ​​v1, v2, ..., vj; Calculate the pixel value difference rate between j grayscale pixel values ​​and the preset ground grayscale reference value, extract grayscale pixels with a pixel value difference rate less than α%, and determine all connected regions as suspected ground regions, where α% is a preset percentage; All areas composed of grayscale pixels, except for suspected ground areas, are considered non-ground areas.

[0009] As a further aspect of the present invention, the specific method for determining the binarized ground region in step two is as follows: Set all grayscale pixels in the identified suspected ground areas as foreground pixels, and set grayscale pixels in non-ground areas as background pixels. Extract all foreground pixels to generate a binary ground region mask image G_mask; A morphological closing operation, involving dilation followed by erosion, is performed on the binarized ground region mask image G_mask to smooth the boundaries; Next, the binarized ground region mask image G_mask is labeled with connected components, the pixel area of ​​each connected region is calculated, and connected regions with a pixel area less than the preset pixel area threshold T_area are identified as noise and removed from the foreground pixels. In the remaining connected regions, select the connected region with the largest pixel area and verify whether it is connected to the bottom boundary of the grayscale image G_gray. If connected, then this connected region is determined to be the main ground region; If not connected, the connected region with the largest area connected to the bottom boundary of the grayscale image G_gray is determined as the main ground region; Perform convex hull calculation on the main ground region, fill the internal holes, and generate a binary ground region R_ground.

[0010] As a further aspect of the present invention, the specific method for determining the left and right vanishing point candidate points in the grayscale image G_gray in step two is as follows: Extract strong edge pixels from the grayscale image G_gray, perform probabilistic Hough transform on all strong edge pixels, detect all straight lines, and randomly arrange them into a set of straight lines L={L1,L2,...,Lm}, where m is the total number of straight lines; Calculate the slopes of m straight lines to obtain the slope set K = {K1, K2, ..., Km}, where the slope Kn corresponds to the line Ln, and n is the counting index, with a value range from 1 to m; Obtain the slope threshold K_th preset by the operator; Lines with a slope less than -K_th are assigned to the left-leaning line group Z_left, lines with a slope greater than K_th are assigned to the right-leaning line group Z_right, and the rest are discarded. For the left-leaning line group Z_left, calculate the intersection points between all pairs of lines in the group and construct the left intersection point set P_left. Similarly, construct the right intersection point set P_right based on the right-leaning line group Z_right. Remove all intersection points that are outside the grayscale image G_gray from the left intersection point set P_left and the right intersection point set P_right, respectively; Next, mean-shift clustering analysis is performed on the left intersection point set P_left with preset bandwidth parameters, and the cluster center with the highest density is extracted as the candidate point VP_left for the left vanishing point. Similarly, the candidate point VP_right for the right vanishing point is determined based on the right intersection point set P_right.

[0011] As a further aspect of the present invention, the specific method for determining the drivable area in step three is as follows: Obtain the bottom center point O(W / 2,H) of the grayscale image G_gray, where W is the width of the grayscale image G_gray, H is the height of the grayscale image G_gray, and the bottom center point O and the origin of the unified coordinate system share the same horizontal axis scale. Connect the left vanishing point candidate point VP_left, the right vanishing point candidate point VP_right, and the bottom center point O to form a triangular region Δ(VP_left,VP_right,O), and define the triangular region as the geometric corridor R_geo; Perform a logical AND operation between the geometric corridor R_geo and the binary ground region R_ground to obtain the fused region R_fused. Where R_fused(x,y)=R_geo(x,y)∧R_ground(x,y), R_fused(x,y)=1 if and only if the pixel (x,y) belongs to both the geometric corridor R_geo and the binary ground region R_ground, and the pixel (x,y) represents the grayscale pixel of any coordinate in the grayscale image G_gray in the unified coordinate system; The identified fused region R_fused is marked as the drivable region R_dvb.

[0012] As a further aspect of the present invention, the specific method for generating the obstacle influence circle within the drivable area in step four is as follows: Based on the straight-line distance sequence d1, d2, ..., dj and the gray-level pixel sequence g1, g2, ..., gj, determine the straight-line distance between all gray-level pixels in the drivable area R_dvb and the AGV; Obtain the operator's preset minimum avoidance distance, and combine the preset semantic segmentation model and the minimum avoidance distance to identify obstacles in all grayscale pixels within the drivable area R_dvb; All grayscale pixels that are classified as obstacles and whose straight-line distance from the AGV is less than the minimum avoidance distance are marked as obstacle reference points and included in the obstacle reference point set P_obs; DBSCAN clustering based on Euclidean distance is performed on the obstacle reference point set P_obs, and the obstacle reference points that are clustered into one class are fitted into an obstacle cluster. For each cluster of obstacles, fit it with the smallest circle and mark the fitted area as the obstacle influence circle; If any two obstacle influence circles intersect within the drivable region R_dvb, then the two obstacle influence circles are fitted with the smallest circle to form a single obstacle influence circle.

[0013] As a further aspect of the present invention, the specific method for generating the safe corridor in step four is as follows: Obtain the maximum width of the AGV body and the horizontal line of the position of the front of the vehicle in the grayscale image G_gray, and denote it as the decision start line L_start. The length of the decision start line L_start is equal to the maximum width of the AGV body. Starting from the decision-making starting line L_start, extend upwards parallel to the upper boundary L_high of the drivable region R_dvb; Initialize the left boundary line L_left and the right boundary line L_right as two vertical lines, located at the left and right endpoints of the decision start line L_start, respectively; Divide the region between the decision start line L_start and the upper boundary L_high into u horizontal layers, where u is the total number of column pixels between the decision start line L_start and the upper boundary L_high. For each horizontal layer, scan from left to right and determine the first pixel that does not belong to any obstacle influence circle and is located within the drivable area R_dvb, as the left boundary candidate point P_left_hx; Then scan from right to left to determine the first pixel that does not belong to any obstacle's influence circle and is located within the drivable area R_dvb, as the right boundary candidate point P_right_hx; Connecting all the candidate points for the left boundary of the horizontal layer forms the final left boundary of the safety corridor; Connecting all the candidate points for the right boundary of the horizontal layers forms the final right boundary of the safety corridor; The final left boundary, the final right boundary, the decision start line L_start, and the upper boundary L_high form a safe corridor; Determine the distance between the left and right boundary candidate points of all horizontal layers in the safety corridor and compare it with the length of the decision starting line L_start. If the distance between the left and right boundary candidate points of any horizontal layer is less than the length of the decision starting line L_start, the safety corridor is deemed invalid and an inaccessible alarm is issued. The safety corridor is updated in real time until it is deemed valid and recorded as a valid safety corridor.

[0014] As a further aspect of the present invention, the specific method for determining the optimal travel path of the AGV within the effective safety corridor in step four is as follows: Obtain the midpoint of the distance between the left and right boundary candidate points of all horizontal layers in the effective safety corridor, and connect all midpoints from bottom to top to obtain the optimal driving path for the AGV.

[0015] The beneficial effects of this invention are: This invention achieves efficient adaptive obstacle avoidance for AGV transport vehicles in complex environments by fusing data from cameras and millimeter-wave radar. Its core advantages lie in improving the accuracy and reliability of obstacle avoidance through multi-sensor collaboration and intelligent image processing. It utilizes grayscale images and distance information to identify ground areas and vanishing points, constructing geometric corridors and drivable areas to effectively distinguish between obstacles and safe zones. Furthermore, by mapping distance data to generate obstacle influence circles and safety corridors, it ensures the real-time performance and safety of path planning, thereby achieving adaptation to various terrain changes, reducing environmental misjudgments, and realizing efficient and safe autonomous AGV navigation in dynamic environments. This invention enables the AGV control system to send synchronous acquisition signals, achieving collaborative operation between a camera and a millimeter-wave radar. Based on grayscale images acquired by the camera and combined with point cloud data provided by the millimeter-wave radar, visual and distance information are fused through coordinate system unification, enhancing the AGV's comprehensive perception of the environment ahead. This multi-sensor data integration not only simplifies subsequent data processing but also improves the accuracy of obstacle detection and path planning. Furthermore, by sequentially arranging pixels and straight-line distances, it provides structured input for real-time navigation algorithms. This invention improves the accuracy and robustness of ground region recognition through pixel analysis and morphological processing. Its advantages lie in its efficient screening of suspected ground regions based on grayscale differences, followed by the effective elimination of noise interference and accurate localization of the main ground region through closing operations, area screening, and bottom connectivity verification, thereby enhancing the segmentation stability in complex scenes. Secondly, it utilizes slope screening and cluster analysis to extract candidate vanishing point positions from edge lines, providing a reliable spatial reference benchmark for subsequent visual perception tasks. It also features automated processing capabilities and anti-interference performance, thus ensuring the continuity of ground detection results. This invention improves the accuracy of drivable area detection by combining geometric constraints and ground segmentation information. It constructs triangular geometric corridors using the bottom center point and left and right vanishing point candidate points of grayscale images to effectively capture the perspective structure of the road. Then, by performing a logical AND operation with the binarized ground area, it ensures that only areas that simultaneously satisfy geometric and ground features are marked as drivable areas. This fusion strategy effectively reduces false detections caused by shadows, obstacles, or noise, improves adaptability in complex environments, and is computationally simple and efficient, providing a reliable environmental perception foundation for AGV autonomous driving. This invention improves the intelligence of AGV navigation by integrating obstacle influence circle generation, safety corridor construction, and optimal path planning. Its advantages lie in the accurate identification of obstacles using grayscale pixel sequences and straight-line distance data, and the effective simplification of complex environment modeling through DBSCAN clustering to fit the obstacle influence circle, enhancing adaptability to dynamic obstacles. Specifically, the generation of the safety corridor uses horizontal layer scanning to determine boundaries in real time, ensuring the matching of AGV passage space with vehicle width, and combining this with an alarm mechanism to prevent potential collision risks. Finally, by connecting the midpoints of the safety corridor to generate the optimal driving path, path safety is guaranteed, improving the AGV's autonomous navigation capability and operating efficiency in complex environments. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart illustrating the method described in this invention; Figure 2 This is a flowchart illustrating the method described in Embodiment 2 of the present invention; Figure 3 This is a flowchart illustrating the method described in Embodiment 3 of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 An adaptive obstacle avoidance analysis method based on the front-end image of an AGV transport vehicle, such as Figure 1 As shown, this method includes the following: This method is an adaptive obstacle avoidance analysis method based on the front-end image of AGV transport vehicle. Its core lies in integrating visual perception and radar ranging, and through multi-level region segmentation and specific geometric algorithm processing, it plans a safe path for AGV to travel in real time in a dynamic environment. First, grayscale images of the AGV during its operation and the straight-line distance between the AGV and each grayscale pixel in the grayscale image are simultaneously acquired using a camera and millimeter-wave radar. Specifically: The AGV's control system synchronously triggers the camera and millimeter-wave radar to perform data acquisition operations. The camera acquires RGB images from the front of the AGV and converts them into grayscale images, while the millimeter-wave radar acquires point cloud data from the front of the AGV. Because grayscale images and point cloud data have inconsistent coordinate systems, a coordinate system unification operation is required. This can be achieved using existing coordinate system fitting technology. After coordinate system unification, each pixel in the grayscale image is assigned a straight-line distance value between itself and the AGV. The straight-line distance value is determined based on the distance between the point cloud data and the AGV, forming a mapping sequence between pixels and distances. It should be explained that the purpose of the synchronous triggering operation is to ensure the consistency of visual information and spatial distance information in time, and to overcome the data misalignment problem caused by AGV movement. Secondly, by mapping the 2D planar data of vision and the 3D spatial data of radar to the same reference system through coordinate system one, the complementary and fusion of information is realized. This step is the basis for subsequent operations.

[0020] Next, based on the gray values ​​of each gray pixel, suspected ground regions and non-ground regions are identified, and all suspected ground regions are extracted for secondary refinement and segmentation to determine the binarized ground regions. Then, extract strong edge pixels from the grayscale image, determine all straight lines in the grayscale image based on the strong edge pixels, classify the straight lines based on their slope, and construct left-leaning straight line groups and right-leaning straight line groups. Intersection point analysis is performed on the left-leaning and right-leaning straight line groups respectively to determine the left intersection point set and the right intersection point set. Then, cluster analysis is performed on the two intersection point sets respectively, and the two largest cluster centers after the two intersection point sets are selected as candidate points for the left vanishing point and the right vanishing point. Specifically: Suspected ground regions are initially segmented from grayscale images based on grayscale similarity. Then, through morphological processing, connected component analysis, and connectivity verification with the bottom of the image, reliable binarized ground regions are obtained.

[0021] Meanwhile, edge detection and Hough transform (both edge detection and Hough transform can be implemented based on existing technologies, so this solution will not elaborate further) are used to extract straight lines in the image. The straight lines are classified into left-leaning straight line groups and right-leaning straight line groups according to the slope threshold. The intersection clusters of straight lines in each group are calculated, and the most significant left and right vanishing point candidate points are found through clustering. It should be explained that the bottom of the image must be a passable road surface, because the grayscale image data is taken by the AGV while it is on the road surface. If the bottom of the image is a non-passable road surface, then no further operations are needed, and it can be directly determined that the AGV cannot pass.

[0022] When AGVs are used in warehousing scenarios, operators will treat the ground, generally using green asphalt paint, to ensure a uniform warehousing environment and avoid errors in AGV automatic navigation due to pixel differences in the ground. The warehousing scenario of the AGV described in this solution meets the above conditions.

[0023] Suspected ground areas are initially segmented using grayscale thresholding, then morphological operations are used to eliminate noise and small holes. Finally, connected component analysis and bottom connectivity verification are used to ensure that the identified ground areas are continuous areas that the AGV can actually reach, rather than suspended objects with similar textures.

[0024] The method for determining the candidate points of the left and right vanishing points is based on the geometric principle of perspective projection. In a structured road or corridor environment, parallel lane lines or boundary lines will intersect at a point in the image, which is the vanishing point. By classifying lines by slope, the boundary lines on the left and right sides can be separated. Because these straight lines may not strictly intersect at a single point due to noise and occlusion, the intersection points are clustered by mean shift to robustly find the common intersection area of ​​these straight lines. The cluster center is then used as the best estimate of the vanishing point to define the main visual direction of the scene where the AGV is currently located.

[0025] Then, the bottom center point of the grayscale image is extracted, and the candidate points of the left and right vanishing points are used to construct a geometric corridor. The binarized ground region and the geometric corridor are combined to calculate the fusion region and determine the drivable area. Specifically: By connecting the left and right vanishing point candidate points with the bottom center point of the image, a triangular geometric corridor is formed; then, this geometric corridor is logically ANDed with the binarized ground region determined in the above steps to obtain the intersection region, which is defined as the drivable region. It should be explained that the geometric corridor defines the fan-shaped space that can theoretically extend forward, allowed by the perspective structure of the scene where the AGV is currently located, while the binarized ground area defines the physical, flat semantic area that the AGV can run over. Therefore, by combining the two through logic and operation, it means that a pixel must simultaneously meet the two conditions of being in a reasonable direction of visual perspective and belonging to the actual ground in order to be determined as a drivable area. This operation can effectively eliminate two abnormal situations: first, the edge area that is on the ground but deviates from the main line of sight; second, the misjudged area that is within the perspective direction but is not the ground (such as walls or obstacles), thereby improving the accuracy of drivable area identification.

[0026] Finally, the straight-line distance of the corresponding grayscale pixels is mapped onto the plane of the drivable area, a set of obstacle reference points is formed based on the value of the straight-line distance, and an obstacle influence circle is generated within the drivable area; Within the drivable area, starting from the decision-making starting line, and combining the boundary of the drivable area and the obstacle influence circle, the left and right boundary lines of the safety corridor are determined, and an effective safety corridor is generated. Based on all decision-making starting points on the decision-making starting line, the optimal travel path of the AGV is determined within the effective safety corridor. Specifically: Within the drivable area, obstacle reference points below the safety threshold are identified based on the straight-line distance of pixels, and then fitted into obstacle influence circles after clustering. Starting from the decision starting line at the AGV's front position, the system scans upwards layer by layer, avoiding obstacle influence circles on each layer, and finding the leftmost and rightmost safety boundary points. These points are connected to form the left and right boundaries of the safety corridor. Finally, the center points of each layer of the safety corridor are connected to form the optimal driving path.

[0027] This step is the core of real-time obstacle avoidance and path planning. Among them, the calibration of the obstacle influence circle abstracts discrete obstacle points into circular areas with a safe radius. Through this simple, conservative but efficient modeling method, a safe buffer area is provided for the AGV. The safety corridor is essentially a dynamic boundary planning process. By determining the boundary layer by layer from the front of the vehicle upwards, which is equivalent to determining the boundary from near to far, not only is a safe area drawn on the static map, but the spatial continuity of the corridor is also guaranteed, ensuring that there is enough width for the AGV to pass from the start to the end. Finally, by verifying whether the width of the safety corridor is greater than the width of the vehicle, the feasibility of the path is ensured. Finally, the centerline principle is adopted, which connects the center points of each layer of the safety corridor to determine an optimal driving path. This optimal driving path mathematically maximizes the minimum distance from the AGV to the obstacles or boundaries on both sides, that is, it keeps the AGV as far away from danger as possible in any position, thus providing the highest safety redundancy.

[0028] Example 2 This embodiment, based on Embodiment 1, further discloses a method for acquiring grayscale images of an AGV during its operation using a camera and millimeter-wave radar, and determining a binarized ground region by identifying suspected ground regions and non-ground regions. Figure 2 As shown, it specifically includes the following: The ultimate goal of this embodiment is to use the data collected by the camera and millimeter-wave radar to perform a comprehensive analysis, and finally determine the ground area and non-ground area for subsequent operation processing.

[0029] First, based on the content described in Example 1, both the camera and the millimeter-wave radar are installed at the front end of the AGV. In actual use, the AGV's control system simultaneously sends synchronous acquisition signals to the camera and millimeter-wave radar equipped at the front end of the AGV to ensure that the camera and millimeter-wave radar acquire data at the same time, thereby reducing errors caused by time deviation.

[0030] The camera uses the AGV as the camera coordinate origin, acquires RGB images of the front end of the AGV, and quickly processes the RGB images to convert them into grayscale images G_gray. The camera's monitoring distance is limited to within F meters of the front end of the AGV as preset by the operator, in order to focus on the relevant area of ​​the front end of the AGV. The RGB image is the first frame of the RGB image. Next, the millimeter-wave radar collects point cloud data within F meters of the AGV's front end, with the AGV as the spatial coordinate origin. The point cloud data is collected at the same time as the first frame of the RGB image.

[0031] It is important to note that because the spatial coordinate origin (spatial coordinate system) associated with the millimeter-wave radar and the camera coordinate origin (camera coordinate system) associated with the camera are misaligned, it is necessary to align the two coordinate systems to obtain a single coordinate system, denoted as the unified coordinate system. This step can be achieved using existing translation matrices. After aligning the two coordinate systems, the point cloud data corresponds to the grayscale pixels in the grayscale image G_gray, and the straight-line distance between all grayscale pixels in the grayscale image G_gray and the AGV can be determined in the unified coordinate system.

[0032] Next, extract all grayscale pixels from the grayscale image G_gray, and sort them according to their order from left to right and from top to bottom in the grayscale image G_gray to obtain the sorted grayscale pixel sequence g1, g2, ..., gj, where j represents the total number of grayscale pixels in the grayscale image G_gray.

[0033] Then, the straight-line distances between all grayscale pixels in the grayscale pixel sequence g1, g2, ..., gj and the AGV are determined using point cloud data, and sorted according to the order of the grayscale pixel sequence g1, g2, ..., gj to obtain the straight-line distance sequence d1, d2, ..., dj.

[0034] Based on the determined grayscale pixel sequence g1, g2, ..., gj, extract the grayscale value of each grayscale pixel, resulting in a total of j grayscale pixel values. Sort these values ​​according to the order of the grayscale pixel sequence g1, g2, ..., gj to obtain the grayscale pixel value sequence v1, v2, ..., vj.

[0035] Next, the operator's preset ground grayscale reference value is obtained. The ground grayscale reference value is the grayscale value of the green asphalt coating described in Example 1. Then, the grayscale pixel value sequence v1, v2, ..., vj is compared with the ground grayscale reference value in turn. All grayscale pixels with a pixel value difference rate of less than α% between their grayscale pixel values ​​and the ground grayscale reference value are extracted. Connected component processing is performed on all extracted grayscale pixels to obtain all connected regions, all of which are marked as suspected ground areas. Here, α% is a percentage preset by the operator. α% is an adjustable threshold that allows for some changes in ground grayscale due to long-term use or changes in lighting.

[0036] Except for grayscale pixels identified as suspected ground areas, all other grayscale pixels are considered as grayscale pixels in non-ground areas, and these grayscale pixels constitute the non-ground areas.

[0037] Then, the determined grayscale image G_gray is copied to obtain a copy of the grayscale image G_gray. The following operations are performed on this copy, which is referred to as the grayscale image G_gray copy in subsequent operations. All suspected ground regions and non-ground regions identified above are marked in the grayscale image G_gray copy. Grayscale pixels in suspected ground regions are set as foreground pixels, and grayscale pixels in non-ground regions are set as background pixels.

[0038] Then, extract all foreground pixels to generate a binary ground region mask image G_mask. Perform a dilation-erosion closing operation on the binary ground region mask image G_mask. The dilation operation connects neighboring regions and fills small holes, while the erosion operation restores the region size but maintains connectivity, thereby smoothing the boundaries and reducing noise. Next, connected component labeling is performed on the binarized ground region mask image G_mask after the closing operation, the pixel area of ​​each connected region is calculated, and small connected regions with pixel areas smaller than the operator's preset threshold T_area are removed. These small connected regions are regarded as noise and are removed from the foreground pixels.

[0039] Then, from the remaining connected regions, select the region that is connected to the bottom boundary of the grayscale image G_gray copy. Since the ground usually extends from the bottom of the AGV, if there are multiple connected regions that are connected to the bottom boundary of the grayscale image G_gray copy, select the region with the largest connected area as the main ground region to ensure that the main drivable area is selected.

[0040] Next, convex hull calculation is performed on the determined main ground area to find the region enclosed by the smallest convex polygon. Then, the internal holes of the region enclosed by the smallest convex polygon are filled to make the region continuous and complete, and to avoid the holes in the ground affecting navigation. Finally, a binary ground area is generated based on the region enclosed by the filled smallest convex polygon, and is marked as R_ground.

[0041] It needs to be explained that selecting the largest area connected to the bottom ensures that the main ground where the AGV can travel is identified, which is consistent with the actual scenario. Since the ground usually touches the bottom of the AGV, when the camera takes pictures, the bottom of the captured image is regarded as the ground where the AGV is currently located. If the largest area connected to the bottom is not selected, a floating area, such as the ceiling or wall, may be selected as the ground, which will lead to serious navigation errors.

[0042] Example 3 This embodiment, based on embodiment 2, further discloses a method for determining drivable areas by combining a grayscale image G_gray with a binarized ground region R_ground, such as... Figure 3 As shown, it specifically includes the following: The main purpose of this embodiment is to define a macroscopic geometric corridor that conforms to physical laws by using the perspective principle and the property that parallel lines such as lane lines intersect at the vanishing point in the image. Then, a drivable area is constructed by combining the binarized ground area R_ground determined in Embodiment 2.

[0043] A grayscale image G_gray is obtained. Strong edge pixels can be directly extracted from the grayscale image G_gray using existing edge detection operators. Then, by performing a probabilistic Hough transform on all strong edge pixels, the endpoints of all line segments composed of strong edge pixels are directly output. In this way, all line segments can be determined. Then, the line segments are extended infinitely to obtain the straight lines associated with all line segments. The determined straight lines are randomly arranged and combined into a set of straight lines L, represented as: L={L1,L2,...,Lm}, where m represents the total number of determined straight lines.

[0044] Then, calculate the slopes of the m lines in the grayscale image G_gray to obtain the slope set K corresponding to the arrangement order of the line set L={L1,L2,...,Lm}, denoted as: K={K1,K2,...,Km}, where the slope Kn corresponds to the line Ln, and n is the counting index, with a value range from 1 to m; Then obtain the slope threshold K_th preset by the operator based on the actual data; Then, all lines with slopes less than -K_th in the slope set K={K1,K2,...,Km} are grouped into the left-leaning line group Z_left; All lines with a slope greater than K_th are grouped into the right-leaning line group Z_right, and the remaining lines with slopes are discarded. It should be noted that the principle of slope comparison is that the slope of the left-leaning straight lines in the left-leaning straight line group Z_left is less than -K_th, corresponding to the right lane lines in the real world. They extend from the bottom of the grayscale image G_gray to the upper right. The slope of the right-leaning straight line group Z_right is greater than K_th, corresponding to the left lane lines in the real world. They extend from the bottom of the image to the upper left.

[0045] Straight lines with slopes between -K_th and K_th are considered to be nearly horizontal and may be vehicles, road markings, or other obstructions. Since they are irrelevant to the vanishing point, they are discarded. This is achieved by filtering through a preset slope threshold K_th to prevent near-horizontal noise lines from being included in subsequent calculations.

[0046] Next, extract the left-leaning line group Z_left, calculate the intersection points of all pairs of lines within the left-leaning line group Z_left, and finally obtain the set of left intersection points P_left associated with the left-leaning line group Z_left. Repeat the above steps to obtain the set of right intersection points P_right associated with the right-leaning line group Z_right.

[0047] Next, obtain the bandwidth parameter preset by the operator for the mean-shift clustering analysis method, perform mean-shift clustering analysis on the left intersection point set P_left, and extract the cluster center with the highest density after the mean-shift clustering analysis of the left intersection point set P_left, and mark it as the left vanishing point candidate point VP_left. Similarly, the right vanishing point candidate point VP_right associated with the right intersection point set P_right can be determined. The bandwidth parameter determines the window size for finding the density, so the operator needs to adjust it according to the resolution of the grayscale image and the intersection point distribution.

[0048] Next, extract the bottom center point of the grayscale image G_gray and mark it as O. The bottom center point O is represented as O(W / 2,H), where W is the width of the grayscale image G_gray and the length from left to right of the grayscale image G_gray is W. H is the height of the grayscale image G_gray. The height from the top to the bottom of the grayscale image G_gray is H. The bottom center point O is on the same horizontal axis scale as the origin of the unified coordinate system to ensure that it is consistent with the movement direction of the AGV.

[0049] Next, in the grayscale image G_gray, the determined left vanishing point candidate point VP_left, right vanishing point candidate point VP_right, and bottom center point O are connected in sequence to form a triangular region Δ(VP_left,VP_right,O), and the triangular region is defined as the geometric corridor R_geo. The triangular region Δ is a strong geometric constraint that defines a fan-shaped area extending from the distant view to the front of the vehicle in the foreground, determined by the left and right vanishing points. This area is almost necessarily the core area of ​​the drivable road, unless there is a sharp turn.

[0050] Then, extract the binarized ground region R_ground determined in Example 2, and perform a logical AND operation between the geometric corridor R_geo and the binarized ground region R_ground to obtain a fused region R_fused; Among them, the geometric corridor R_geo may have inaccurate vanishing point detection when the road is curved or the lane lines are unclear, causing the triangular region to deviate from the real road. The binarized ground region R_ground may misrepresent some non-road areas that look like the ground (such as shadowed areas) or may be interfered with by obstacles in the lane. By performing a logical AND operation between the geometric corridor R_geo and the binarized ground region R_ground, the geometric corridor R_geo and the binarized ground region R_ground are mutually corrected; Correcting the binary ground region R_ground with the geometric corridor R_geo can eliminate misclassifications in the binary ground region R_ground that resemble the ground but are located outside the triangular region Δ; By correcting the geometric corridor R_geo using the binarized ground region R_ground, non-ground regions within the triangular region Δ can be eliminated; By using logic and operations, the probability of misjudging non-drivable areas as drivable areas can be reduced, thereby improving the safety of autonomous driving.

[0051] The logical AND operation is specifically represented as follows: R_fused(x,y)=R_geo(x,y)∧R_ground(x,y) R_fused(x,y)=1 if and only if the pixel (x,y) belongs to both the geometric corridor R_geo and the binary ground region R_ground. The pixel (x,y) represents the grayscale pixel of any coordinate in the unified coordinate system of the grayscale image G_gray. Ultimately, the fused region R_fused is a binary image in which grayscale pixels with a grayscale value of 1 constitute a high-confidence drivable region that conforms to road perspective geometry and is recognized as ground by the visual model. This region is labeled as the drivable region R_dvb.

[0052] Example 4 This embodiment, based on embodiment 3, further discloses a method for determining the optimal driving path of an AGV at its current position based on the drivable area R_dvb, specifically including the following: First, extract the grayscale pixel sequence g1, g2, ..., gj associated with the grayscale image G_gray. Based on the straight-line distance sequence d1, d2, ..., dj associated with the grayscale pixel sequence g1, g2, ..., gj, and combining the corresponding grayscale pixels in the grayscale pixel sequence g1, g2, ..., gj within the drivable area R_dvb, determine the straight-line distance between all grayscale pixels within the drivable area R_dvb and the AGV.

[0053] Next, obtain the minimum avoidance distance preset by the operator for the AGV. The minimum avoidance distance refers to the distance when any obstacle is detected that is less than or equal to the straight-line distance from the AGV, and then the avoidance operation is started.

[0054] Then, using the minimum avoidance distance, the current position of the AGV, and a preset semantic segmentation model, obstacles are identified in all grayscale pixels within the drivable area R_dvb. All grayscale pixels classified as obstacles, as well as grayscale pixels whose straight-line distance from the AGV is less than the minimum avoidance distance, are marked as obstacle reference points and included in the obstacle reference point set P_obs.

[0055] It should be explained that the obstacle categories in the semantic segmentation model are added by the operator based on obstacles that have appeared in the actual environment, such as pedestrians, vehicles, and traffic cones.

[0056] Next, DBSCAN clustering based on Euclidean distance is performed on the obstacle reference point set P_obs, where the Euclidean distance is adjusted according to the clustering progress required by the operator; Finally, the obstacle reference points clustered into one class in the obstacle reference point set P_obs are fitted into an obstacle cluster; For each obstacle cluster, fit it with the smallest circle (construct the smallest circumcircle), and mark the fitted smallest circle region as the obstacle influence circle. In this way, in subsequent calculation operations, to determine whether a point or a path intersects with the circle, it is only necessary to calculate the distance from the point to the center of the circle, which is simpler and faster than finding the intersection with complex polygons.

[0057] If any two obstacle influence circles intersect within the drivable area R_dvb, then the two obstacle influence circles are fitted with the smallest circle as a single obstacle influence circle.

[0058] Obtain the widest width of the AGV body and determine the horizontal line of the position of the AGV's front in the grayscale image G_gray. Record this horizontal line as the decision start line L_start, where the length of the decision start line L_start is equal to the widest width of the AGV body. Then, starting from the decision start line L_start, extend parallel to the upper boundary of the grayscale image G_gray until it reaches the upper boundary of the drivable region R_dvb, marked as L_high; Next, a left boundary line L_left and a right boundary line L_right are initialized and constructed at the left and right endpoints of the decision start line L_start, respectively. The left boundary line L_left and the right boundary line L_right are both vertical lines that pass through the left and right endpoints of the decision start line L_start and are perpendicular to the upper and lower boundaries of the grayscale image G_gray.

[0059] Then, the region between the decision start line L_start and the upper boundary L_high (that is, the closed region composed of the left boundary line L_left, the right boundary line L_right, the decision start line L_start, and the upper boundary L_high) is divided into u horizontal layers, where u is the total number of column pixels between the decision start line L_start and the upper boundary L_high, that is, each row of pixels is a horizontal layer.

[0060] For the determined u horizontal layers, each horizontal layer is scanned from left to right in turn, and the first pixel in the corresponding horizontal layer that does not belong to any obstacle influence circle and is located in the drivable area R_dvb is identified and marked as the left boundary candidate point P_left_hx; The first pixel that does not belong to any obstacle influence circle and is located within the drivable area R_dvb is identified by scanning the horizontal layer from right to left and marked as the right boundary candidate point P_right_hx. Similarly, after determining the left boundary candidate point P_left_hx and right boundary candidate point P_right_hx associated with each of the u horizontal layers, if there is no pixel in any horizontal layer that does not belong to any obstacle influence circle and is located within the drivable area R_dvb, it means that all pixels in this horizontal layer are obstacles, and it is directly determined that the AGV cannot pass.

[0061] By performing bidirectional scanning in all horizontal layers, the left and right limits of the safety passage in each horizontal layer can be determined efficiently. Then, the left boundary candidate points of all horizontal layers are connected from top to bottom to form the final left boundary of the safety corridor. Similarly, the right boundary candidate points of all horizontal layers are connected from top to bottom to form the final right boundary of the safety corridor. At this point, a safe corridor is formed based on the determined final left boundary, final right boundary decision start line L_start, and upper boundary L_high. Then, the distance between the left and right boundary candidate points of each horizontal layer in the safe corridor is extracted, and the distance between the left and right boundary candidate points of each horizontal layer is compared with the length of the decision start line L_start, which is compared with the widest part of the AGV body. If, among the u horizontal layers, the distance between the left and right boundary candidate points of any horizontal layer is less than the length of the decision starting line L_start, then the safe corridor is deemed invalid, and an inaccessible alarm is issued. The safe corridor is updated in real time until it is deemed valid, and is then recorded as a valid safe corridor.

[0062] Next, within the effective safety corridor, determine the optimal travel path for the AGV, extract the midpoint of the distance between the candidate points on the left and right boundaries of all horizontal layers, and connect all the midpoints in order from bottom to top to form a path from the current position of the AGV to the distance. This path is geometrically as far away from the obstacles on both sides as possible, and is denoted as the optimal travel path of the AGV.

[0063] All data in the formulas described above are numerical calculations performed after removing their dimensions. Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0064] The above description is merely an example and illustration of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0065] It should be stated that all user data collected in this application was collected with the user's consent and authorization. Furthermore, the uses of user data are legal and compliant, and the use and processing of user data comply with the relevant laws, regulations, and standards of the relevant regions.

Claims

1. An adaptive obstacle avoidance analysis method based on the front-end image of an AGV transport vehicle, characterized in that, The method includes: Step 1: Simultaneously acquire grayscale images of the AGV during its driving process and the straight-line distance between the AGV and each grayscale pixel in the grayscale image based on the camera and millimeter-wave radar. Step 2: Based on the gray values ​​of each gray pixel, identify suspected ground areas and non-ground areas, extract all suspected ground areas for secondary refinement and segmentation, and determine the binarized ground areas. Then, extract strong edge pixels from the grayscale image, determine all straight lines in the grayscale image based on the strong edge pixels, classify the straight lines based on their slope, and construct left-leaning straight line groups and right-leaning straight line groups. Intersection point analysis is performed on the left-leaning straight line group and the right-leaning straight line group respectively to determine the left intersection point set and the right intersection point set. Then, cluster analysis is performed on the two intersection point sets respectively, and the two largest cluster centers after the two intersection point sets are taken as candidate points for the left vanishing point and the right vanishing point. Step 3: Extract the bottom center point of the grayscale image and form a geometric corridor with the left and right vanishing point candidate points. Combine the binarized ground area with the geometric corridor to calculate the fusion area and determine the drivable area. Step 4: Map the straight-line distance of the corresponding grayscale pixels onto the plane of the drivable area, form a set of obstacle reference points based on the value of the straight-line distance, and generate an obstacle influence circle within the drivable area; Within the drivable area, starting from the decision-making starting line, and combining the boundary of the drivable area and the obstacle influence circle, the left and right boundary lines of the safety corridor are determined, and an effective safety corridor is generated. Based on all decision starting points on the decision starting line, the optimal travel path of the AGV is determined within the effective safety corridor.

2. The method according to claim 1, characterized in that, In step one, the specific method for simultaneously acquiring grayscale images of the AGV during its driving process and the straight-line distance between the AGV and each grayscale pixel in the grayscale image based on the camera and millimeter-wave radar is as follows: The AGV-based control system sends synchronous acquisition signals to the camera and millimeter-wave radar equipped on the AGV. The camera takes the AGV as the origin of the camera coordinates and captures the first frame of RGB image within F meters of the front end of the AGV, and converts it into a grayscale image G_gray, where F is the preset monitoring distance; The millimeter-wave radar collects point cloud data within F meters of the AGV front end, with the AGV as the spatial coordinate origin. The spatial coordinate system is then unified with the camera coordinate system. In the unified coordinate system, the straight-line distance between all grayscale pixels in the grayscale image G_gray and the AGV is determined. Arrange all grayscale pixels in the grayscale image G_gray in order from left to right and from top to bottom to obtain the grayscale pixel sequence g1, g2, ..., gj, where j is the total number of grayscale pixels; Extract the straight-line distance between each grayscale pixel and the AGV, and record the straight-line distance sequence d1, d2, ..., dj according to the order of the grayscale pixel sequence.

3. The method according to claim 2, characterized in that, In step two, the specific method for identifying suspected ground areas and non-ground areas based on the gray values ​​of each gray pixel is as follows: Traverse the sequence of grayscale pixels g1, g2, ..., gj, and extract the grayscale value of each grayscale pixel to obtain the sequence of grayscale pixel values ​​v1, v2, ..., vj; Calculate the pixel value difference rate between j grayscale pixel values ​​and the preset ground grayscale reference value, extract grayscale pixels with a pixel value difference rate less than α%, and determine all connected regions as suspected ground regions, where α% is a preset percentage; All areas composed of grayscale pixels, except for suspected ground areas, are considered non-ground areas.

4. The method according to claim 3, characterized in that, In step two, the specific method for determining the binarized ground region is as follows: Set all grayscale pixels in the identified suspected ground areas as foreground pixels, and set grayscale pixels in non-ground areas as background pixels. Extract all foreground pixels to generate a binary ground region mask image G_mask; A morphological closing operation, involving dilation followed by erosion, is performed on the binarized ground region mask image G_mask to smooth the boundaries; Next, the binarized ground region mask image G_mask is labeled with connected components, the pixel area of ​​each connected region is calculated, and connected regions with a pixel area less than the preset pixel area threshold T_area are identified as noise and removed from the foreground pixels. In the remaining connected regions, select the connected region with the largest pixel area and verify whether it is connected to the bottom boundary of the grayscale image G_gray. If connected, then this connected region is determined to be the main ground region; If not connected, the connected region with the largest area connected to the bottom boundary of the grayscale image G_gray is determined as the main ground region; Perform convex hull calculation on the main ground region, fill the internal holes, and generate a binary ground region R_ground.

5. The method according to claim 4, characterized in that, In step two, the specific method for determining the candidate points for the left and right vanishing points in the grayscale image G_gray is as follows: Extract strong edge pixels from the grayscale image G_gray, perform probabilistic Hough transform on all strong edge pixels, detect all straight lines, and randomly arrange them into a set of straight lines L={L1,L2,...,Lm}, where m is the total number of straight lines; Calculate the slopes of m straight lines to obtain the slope set K = {K1, K2, ..., Km}, where the slope Kn corresponds to the line Ln, and n is the counting index, with a value range from 1 to m; Obtain the slope threshold K_th preset by the operator; Lines with a slope less than -K_th are assigned to the left-leaning line group Z_left, lines with a slope greater than K_th are assigned to the right-leaning line group Z_right, and the rest are discarded. For the left-leaning line group Z_left, calculate the intersection points between all pairs of lines in the group and construct the left intersection point set P_left. Similarly, construct the right intersection point set P_right based on the right-leaning line group Z_right. Remove all intersection points that are outside the grayscale image G_gray from the left intersection point set P_left and the right intersection point set P_right, respectively; Next, mean-shift clustering analysis is performed on the left intersection point set P_left with preset bandwidth parameters, and the cluster center with the highest density is extracted as the candidate point VP_left for the left vanishing point. Similarly, the candidate point VP_right for the right vanishing point is determined based on the right intersection point set P_right.

6. The method according to claim 5, characterized in that, In step three, the specific method for determining the drivable area is as follows: Obtain the bottom center point O(W / 2,H) of the grayscale image G_gray, where W is the width of the grayscale image G_gray, H is the height of the grayscale image G_gray, and the bottom center point O and the origin of the unified coordinate system share the same horizontal axis scale. Connect the left vanishing point candidate point VP_left, the right vanishing point candidate point VP_right, and the bottom center point O to form a triangular region Δ(VP_left,VP_right,O), and define the triangular region as the geometric corridor R_geo; Perform a logical AND operation between the geometric corridor R_geo and the binary ground region R_ground to obtain the fused region R_fused. Where R_fused(x,y)=R_geo(x,y)∧R_ground(x,y), R_fused(x,y)=1 if and only if the pixel (x,y) belongs to both the geometric corridor R_geo and the binary ground region R_ground, and the pixel (x,y) represents the grayscale pixel of any coordinate in the grayscale image G_gray in the unified coordinate system; The identified fused region R_fused is marked as the drivable region R_dvb.

7. The method according to claim 6, characterized in that, In step four, the specific method for generating the obstacle influence circle within the drivable area is as follows: Based on the straight-line distance sequence d1, d2, ..., dj and the gray-level pixel sequence g1, g2, ..., gj, determine the straight-line distance between all gray-level pixels in the drivable area R_dvb and the AGV; Obtain the operator's preset minimum avoidance distance, and combine the preset semantic segmentation model and the minimum avoidance distance to identify obstacles in all grayscale pixels within the drivable area R_dvb; All grayscale pixels that are classified as obstacles and whose straight-line distance from the AGV is less than the minimum avoidance distance are marked as obstacle reference points and included in the obstacle reference point set P_obs; DBSCAN clustering based on Euclidean distance is performed on the obstacle reference point set P_obs, and the obstacle reference points that are clustered into one class are fitted into an obstacle cluster. For each cluster of obstacles, fit it with the smallest circle and mark the fitted area as the obstacle influence circle; If any two obstacle influence circles intersect within the drivable region R_dvb, then the two obstacle influence circles are fitted with the smallest circle to form a single obstacle influence circle.

8. The method according to claim 7, characterized in that, In step four, the specific method for generating the safe corridor is as follows: Obtain the maximum width of the AGV body and the horizontal line of the position of the front of the vehicle in the grayscale image G_gray, and denote it as the decision start line L_start. The length of the decision start line L_start is equal to the maximum width of the AGV body. Starting from the decision-making starting line L_start, extend upwards parallel to the upper boundary L_high of the drivable region R_dvb; Initialize the left boundary line L_left and the right boundary line L_right as two vertical lines, located at the left and right endpoints of the decision start line L_start, respectively; Divide the region between the decision start line L_start and the upper boundary L_high into u horizontal layers, where u is the total number of column pixels between the decision start line L_start and the upper boundary L_high. For each horizontal layer, scan from left to right and determine the first pixel that does not belong to any obstacle influence circle and is located within the drivable area R_dvb, as the left boundary candidate point P_left_hx; Then scan from right to left to determine the first pixel that does not belong to any obstacle's influence circle and is located within the drivable area R_dvb, as the right boundary candidate point P_right_hx; Connecting all the candidate points for the left boundary of the horizontal layer forms the final left boundary of the safety corridor; Connecting all the candidate points for the right boundary of the horizontal layers forms the final right boundary of the safety corridor; The final left boundary, the final right boundary, the decision start line L_start, and the upper boundary L_high form a safe corridor; Determine the distance between the left and right boundary candidate points of all horizontal layers in the safety corridor and compare it with the length of the decision starting line L_start. If the distance between the left and right boundary candidate points of any horizontal layer is less than the length of the decision starting line L_start, the safety corridor is deemed invalid and an inaccessible alarm is issued. The safety corridor is updated in real time until it is deemed valid and recorded as a valid safety corridor.

9. The method according to claim 8, characterized in that, In step four, the specific method for determining the optimal travel path of the AGV within the effective safety corridor is as follows: Obtain the midpoint of the distance between the left and right boundary candidate points of all horizontal layers in the effective safety corridor, and connect all midpoints from bottom to top to obtain the optimal driving path for the AGV.

Images