Pallet identification method based on 3D point cloud for intelligent transfer robot

Abstract

The invention provides a pallet identification method based on 3D point cloud for an intelligent transfer robot. The pallet identification method comprises the following steps: S1, intercepting a target area from the 3D point cloud acquired by the intelligent transfer robot; s2, filtering noisy points through statistical filtering to obtain filtered point clouds; s3, identifying the ground from the filtered point clouds, removing the ground, and segmenting pallet point clouds according to the Euclidean distance; s4, projecting the pallet point cloud into the front view to obtain a 2D image; s5, constructing a front view template of the pallet and searching similar contours in the 2D image to obtain a pallet area; s6, the pallet 3D point cloud and the middle supporting leg 3D point cloud are reversely deduced; s7, extracting a front point cloud close to the front surface of the pallet; s8, obtaining a pallet placement angle according to the front point cloud, and obtaining the center position of the front surface of the pallet according to the 3D point cloud of the middle supporting leg; and S9, outputting, and completing pallet identification. The method is suitable for various 3D sensors, is high in universality and expansibility, does not depend on training and reasoning of a neural network, can be rapidly deployed, and is low in requirement for calculation performance.

Description

Technical Field

[0001] The present invention relates to the technical field of target recognition of transport robots, and in particular to a pallet recognition method based on 3D point cloud for an intelligent transport robot. Background Art

[0002] In the field of intelligent warehousing and logistics, intelligent handling robots (AGVs) with autonomous perception, positioning, navigation and other functions are gradually replacing manual handling labor. One of the foundations for the stable operation of intelligent handling robots (AGVs) is the accurate identification and positioning of the handling targets. How to correctly identify cargo carriers, especially the commonly used pallets, is a hot issue in the field of intelligent logistics and a research direction for continuous improvement.

[0003] A more popular pallet recognition method in the existing technology is pallet recognition based on deep learning, which has high environmental adaptability, but the underlying neural network requires a lot of training. Changes in input data from different sensors may also cause the neural network to need to be retrained. In addition, the requirements for processor computing performance during operation are high, and the scope of application is somewhat limited. Summary of the invention

[0004] The purpose of the present invention is to solve one or more problems in the prior art and to provide a pallet recognition method based on 3D point cloud for an intelligent handling robot.

[0005] To achieve the above object, the technical solution adopted by the present invention is a pallet recognition method based on 3D point cloud for an intelligent handling robot, characterized in that it includes the following steps: S1. intercepting a target area from the 3D point cloud obtained by the intelligent handling robot, where the target area is an area containing a pallet to be identified; S2. filtering out noise points in the target area by statistical filtering to obtain a filtered point cloud; S3. Based on the prior features of the ground point cloud, the ground is identified and removed from the filtered point cloud, and then the pallet point cloud is segmented according to the Euclidean distance; S4. Project the pallet point cloud into the front view to obtain a 2D image; S5. Construct a front view template of the pallet based on the pallet contour parameters input externally, search for contours similar to the front view template in the 2D image, and extract the area whose similarity to the front view template exceeds a specified threshold as the pallet area in the 2D image; S6. According to the correspondence between the 2D image and the 3D point cloud, the 3D point cloud of the pallet is inferred from the pallet area of ​​the 2D image, and the 3D point cloud of the middle support leg is inferred from the 2D image; S7. performing plane fitting on the 3D point cloud of the pallet, and extracting the front point cloud close to the front surface of the pallet; S8. Obtain the placement angle of the pallet according to the front point cloud, and obtain the center position of the front surface of the pallet according to the 3D point cloud of the middle support leg; S9. Output the placement angle of the pallet and the center position of the front surface of the pallet to complete the pallet identification.

[0006] Preferably, the step of intercepting the target area in step S1 includes: S11. With the sensor on the intelligent handling robot that obtains the 3D point cloud as the center of the coordinate system, set the minimum and maximum value ranges of the three directions of x, y, and z respectively to ensure that the pallet to be identified is within the range, where the positive direction of x points forward, the positive direction of y points to the left, and the positive direction of z points upward, satisfying the right-hand coordinate system rule; S12. Given the external parameters of the relative position between the pallet to be identified and the sensor, in the 3D point cloud, data points within the specified range of x, y, and z are intercepted as the target area.

[0007] Further preferably, the step of statistical filtering in step S2 includes: S21. Calculate the average distance between each point in the target area and the nearest K points, and adjust the K value according to the filtering effect; S22. Sort all points from small to large according to the average distance calculated in S21; S23. retain a set proportion of data points that are at the front of the sort, and remove the remaining points to obtain a filtered point cloud.

[0008] Further preferably, the step of segmenting the pallet point cloud in step S3 includes: S31. Use the RANSAC algorithm to fit a virtual plane in the filtered point cloud, where the angle between the virtual plane and the horizontal plane does not exceed a set angle and contains the most data points; S32. Eliminate the virtual plane in S31 and the data points whose distance from the virtual plane does not exceed the set value from the filtered point cloud; S33. Cluster the point cloud obtained in S32 according to the Euclidean distance, divide the point cloud into several relatively independent point cloud clusters, select the point cloud cluster close to the center of the point cloud field of view and the number of data points in the cluster exceeds the set threshold for segmentation, and obtain the pallet point cloud.

[0009] Further preferably, the step of converting the pallet point cloud into a 2D image in step S4 includes: S41. Traverse the stack point cloud in step S33, calculate the extreme values ​​of the y coordinates and z coordinates of all data points: y_min, y_max, z_min, z_max, and construct a 2D initial image with the y coordinates and z coordinates of all data points. The width of the 2D initial image is (y_max-y_min) and the height is (z_max-z_min). S42. Converting each data point in the pallet point cloud in step S33 into a white pixel in the 2D initial image; S43. Perform a closing operation on the 2D initial image in step S42 to fill the sparse white pixels and reduce the internal hollowing to obtain a 2D image.

[0010] Further preferably, the searching step in step S5 includes: S51 constructs a pallet front view template A based on external input pallet contour parameters, the pallet front view template A parameters include pallet support leg width, pallet support leg height, pallet overall width, pallet overall height; S52. Constructing a widened template B according to the pallet front view template A in step S51, the width of the widened template B is 80% of the pallet front view template A, the pallet front view template A and the widened template B together constitute the front view template of the pallet; S53. Traverse all rows and columns in the 2D image obtained in step S43, and search for the area with the highest matching degree with the front view template of the pallet and exceeding the matching degree threshold, as the pallet area in the 2D image; S54. According to the pallet support leg width and pallet support leg height parameters given in step S51, extract the middle support leg area in the pallet area in step S53.

[0011] Further preferably, the step of extracting the middle supporting leg area in step S54 includes: Taking the vertical center line of the pallet area in step S53 as a reference, expand 1 / 2 of the pallet support leg width to the left and right, and expand 1 pallet support leg height upward from the bottom of the pallet area in step S53 to obtain the middle support leg area.

[0012] Further preferably, the step of inferring the 3D point cloud from the 2D image in step S6 includes: S61. Convert each data point in the stack point cloud of step S33 into a pixel point of the 2D initial image in step S42. If the pixel point is located in the stack area in step S53, extract the data point corresponding to the pixel point to form a stack 3D point cloud; S62. Convert each data point in the stack point cloud in step S33 into a pixel point of the 2D initial image in step S42. If the pixel point is located in the middle supporting leg area in step S54, extract the data point corresponding to the pixel point to form the middle supporting leg 3D point cloud.

[0013] Further preferably, the step of extracting the front point cloud in step S7 includes: S71. Using the RANSAC algorithm to perform plane fitting on the 3D point cloud in S61, a plane corresponding to the front surface of the pallet is obtained; S72. Extract points whose distance from the plane in step S71 is no more than 10 mm to obtain a front point cloud close to the front surface of the pallet.

[0014] Further preferably, the step of obtaining the placement angle and the center position in step S8 includes: S81. Perform plane fitting on the previous point cloud in step S72, and the corresponding plane angle is the placement angle of the pallet; S82. Take the average of the 3D point cloud of the middle supporting leg obtained in step S62 to obtain the center position of the front surface of the pallet.

[0015] Due to the application of the above technical solution, the present invention has the following advantages compared with the prior art: The 3D point cloud-based pallet recognition method for the intelligent handling robot provided by the present invention can be applied to a variety of 3D sensors and has strong scalability. Regardless of whether it is a standard pallet or a non-standard pallet, it only needs to obtain the pallet contour parameters in advance to perform template construction and 2D image matching. In addition, compared with the deep learning method, the present invention does not rely on the training and reasoning of the neural network, can be deployed quickly, and has lower requirements on computing performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of the present invention.

[0017] Figure 2 It is a top view schematic diagram of intercepting the target area in the present invention.

[0018] Figure 3 It is a schematic diagram of the 2D image (including goods) in the present invention.

[0019] Figure 4 It is a schematic diagram of the front view template in the present invention.

[0020] Figure 5 It is a schematic diagram of the output result of the present invention.

[0021] Among them: 10. Target area; 20. Pallet. DETAILED DESCRIPTION

[0022] The technical solution protected by the present invention is described in detail below with reference to the accompanying drawings.

[0023] like Figures 1 to 5 As shown, the pallet recognition method based on 3D point cloud for the intelligent handling robot provided by the present invention comprises the following steps: S1. intercepting a target area from the 3D point cloud obtained by the intelligent handling robot, where the target area is an area containing a pallet to be identified; Specifically, the step of intercepting the target area in step S1 includes: S11. Take the sensor on the intelligent handling robot that obtains the 3D point cloud as the center of the coordinate system (usually also the origin of the sensor), set the minimum and maximum value ranges of the three directions of x, y, and z respectively, and ensure that the pallet to be identified is within the range, where the positive direction of x points forward, the positive direction of y points to the left, and the positive direction of z points upward, satisfying the right-hand coordinate system rule; S12. Given the external parameters of the relative position between the pallet to be identified and the sensor, in the 3D point cloud, data points within the specified range of x, y, and z are intercepted as the target area.

[0024] It should be noted that the parameters x_min, x_max, y_min, y_max, z_min, and z_max of the target area vary depending on the specific application scenario, and these parameters need to be input externally by the operator. When intercepting, all data points in the 3D point cloud can be traversed, and only those x, y, and z coordinate values ​​that are within the interception range are retained. After traversal, the target area is obtained, and subsequent operations are all within this area.

[0025] S2. filtering out noise points in the target area by statistical filtering to obtain a filtered point cloud; Specifically, the statistical filtering step in step S2 includes: S21. Calculate the average distance between each point in the target area and the nearest K points, and adjust the K value according to the filtering effect; S22. Sort all points from small to large according to the average distance calculated in S21; S23. retain a set proportion of data points that are at the front of the sort, and remove the remaining points to obtain a filtered point cloud.

[0026] It should be noted that no matter it comes from 3D lidar or depth camera, there will always be noise points in 3D point cloud, especially near the edge of the object, the noise points will be more obvious. From the performance point of view, the noise points are generally obviously far away from the surrounding points, often appearing as single isolated points. These noise points do not help to identify and locate the target, so they need to be filtered out.

[0027] The filtering method of the present invention is based on distance statistics. Specifically, the average distance between each point in the 3D point cloud and the nearest K (for example, K=30) points is first calculated. Since the noise points are far away from the surrounding points, the calculated average distance will be significantly greater than that of the normal points. All points are sorted according to the size of the average distance, and only the data points with the highest ranking, such as the first 10% of the points, are retained. The remaining points are eliminated. After statistical filtering, only the data points that are relatively close to the surrounding points are retained in the point cloud. These data points can more accurately reflect the actual position of the target.

[0028] S3. Based on the prior features of the ground point cloud, the ground is identified and removed from the filtered point cloud, and then the pallet point cloud is segmented according to the Euclidean distance; Specifically, the step of segmenting the pallet point cloud in step S3 includes: S31. Use the RANSAC algorithm to fit a virtual plane in the filtered point cloud, where the angle between the virtual plane and the horizontal plane does not exceed a set angle and contains the most data points; S32. Eliminate the virtual plane in S31 and the data points whose distance from the virtual plane does not exceed a set value from the filtered point cloud; S33. Cluster the point cloud obtained in S32 according to the Euclidean distance, divide the point cloud into several relatively independent point cloud clusters, select the point cloud cluster close to the center of the point cloud field of view and the number of data points in the cluster exceeds the set threshold for segmentation, and obtain the pallet point cloud.

[0029] Specifically, based on the prior knowledge of the basic level of the ground, the RANSAC algorithm is used to fit a plane in the filtered point cloud. The angle between the plane and the horizontal plane is required to not exceed a specific angle (for example, 30°) and to contain enough data points. In the specific implementation, the existing point cloud processing software, such as PCL, Open3D, etc., can be used to conveniently perform plane fitting. Taking into account the accuracy error of the data points, data points within a certain distance from the fitting plane should be considered to be included in the fitting plane (for example, the distance is set to ±20mm). Through this fitting, the point cloud of the ground area can be extracted and removed from the filtered point cloud.

[0030] In addition, considering that there is usually a certain gap between the pallet and the surrounding objects, the point cloud can be clustered according to the Euclidean distance based on this prior knowledge. After clustering, the overall point cloud is divided into several relatively independent point cloud clusters. The distance between the points in the cluster is relatively close, and the distance between the points between clusters is relatively far. From several point cloud clusters, the point cloud cluster where the pallet is located is selected according to the following principles: (1) The number of data points contained in the point cloud cluster should exceed the set threshold; (2) The point cloud cluster should be close to the center of the point cloud field of view. This is because when performing pallet recognition, the intelligent handling robot has generally moved to the front of the pallet, which may have a slight deviation. The purpose of pallet recognition is to accurately locate the goods and then plan a corrective delivery route. According to the above principles, the point cloud corresponding to the pallet and the goods it carries can be segmented.

[0031] S4. Project the pallet point cloud into the front view to obtain a 2D image; Specifically, the steps from the pallet point cloud to the 2D image in step S4 include: S41. Traverse the stack point cloud in step S33, calculate the extreme values ​​of the y coordinates and z coordinates of all data points: y_min, y_max, z_min, z_max, and construct a 2D initial image with the y coordinates and z coordinates of all data points. The width of the 2D initial image is (y_max-y_min) and the height is (z_max-z_min). S42. Converting each data point in the pallet point cloud in step S33 into a white pixel in the 2D initial image; S43. Perform a closing operation on the 2D initial image in step S42 to fill the sparse white pixels, reduce or eliminate the internal hollowing, and obtain a 2D image.

[0032] The purpose of step S4 is to speed up the search of the pallet area. This step requires projecting the pallet point cloud space into a 2D image. When projecting, the width and height of the 2D image must first be determined to accommodate the projection of the entire pallet point cloud.

[0033] Specifically, by traversing the pallet point cloud (including goods) in step S33, the extreme values ​​of the y coordinates and z coordinates of all data points therein are calculated: y_min, y_max, z_min, z_max, and a 2D initial image is constructed, the width of the image is (y_max-y_min), and the height is (z_max-z_min).

[0034] Then, each data point in the stack point cloud in step S33 is converted into a pixel with a value of 255 in the above 2D initial image, that is, a white pixel. Since the density of the data points in the stack point cloud may not be sufficient to fill the white pixel area in the 2D initial image after projection, this will affect the calculation of the matching value during template matching. Therefore, it is necessary to perform a closing operation on the 2D initial image to fill the sparse white pixels and reduce or eliminate the internal hollowing, as shown in the following figure. Figure 3 The 2D image shown in the figure can be processed by existing image processing software such as opencv during specific implementation to form the 2D image and perform closing operations.

[0035] S5. Construct a front view template of the pallet based on the pallet contour parameters input externally, search for contours similar to the front view template in the 2D image, and extract the area whose similarity to the front view template exceeds a specified threshold as the pallet area in the 2D image; Specifically, the search step in step S5 includes: S51 constructs a pallet front view template A based on external input pallet contour parameters, the pallet front view template A parameters include pallet support leg width, pallet support leg height, pallet overall width, pallet overall height; S52. Constructing a widened template B according to the pallet front view template A in step S51, the width of the widened template B is 80% of the pallet front view template A, the pallet front view template A and the widened template B together constitute the front view template of the pallet; S53. Traverse all rows and columns in the 2D image obtained in step S43, and search for the area with the highest matching degree with the front view template of the pallet and exceeding the matching degree threshold, as the pallet area in the 2D image; S54. According to the pallet support leg width and pallet support leg height parameters given in step S51, the middle support leg area is extracted in the pallet area in step S53, the step comprising: Taking the vertical center line of the pallet area in step S53 as a reference, expand 1 / 2 of the pallet support leg width to the left and right, and expand 1 pallet support leg height upward from the bottom of the pallet area in step S53 to obtain the middle support leg area.

[0036] It should be noted that step S5 first constructs a front view template of the pallet according to the pallet contour parameters input externally. These parameters include the pallet support leg width a, the pallet support leg height b, the pallet overall width c, the pallet height d, etc. Figure 4 As shown; considering the problem of the pallet placement angle, the pallet front view area in the 2D image obtained in step S43 may be narrower than the actual size. Therefore, in actual operation, the front view template of the pallet includes the pallet front view template A (full-size template) and the widened template B. The pallet front view template A is constructed according to the pallet contour parameters input above, and the width of the widened template B is 80% of the pallet front view template A. When traversing step S53, it is necessary to search for the area with the highest matching degree with the pallet front view template A and the widened template B and exceeding the matching degree threshold, as the pallet area in the 2D image; then using the vertical center line of the pallet area as a reference, expand 1 / 2 of the pallet support leg width to the left and right, and expand 1 pallet support leg height from the bottom of the pallet area upward to obtain the middle support leg area.

[0037] S6. According to the correspondence between the 2D image and the 3D point cloud, the 3D point cloud of the pallet is inferred from the pallet area of ​​the 2D image, and the 3D point cloud of the middle support leg is inferred from the 2D image; Specifically, the step of inferring the 3D point cloud from the 2D image in step S6 includes: S61. Convert each data point in the stack point cloud of step S33 into a pixel point of the 2D initial image in step S42. If the pixel point is located in the stack area in step S53, extract the data point corresponding to the pixel point to form a stack 3D point cloud; S62. Convert each data point in the stack point cloud in step S33 into a pixel point of the 2D initial image in step S42. If the pixel point is located in the middle supporting leg area in step S54, extract the data point corresponding to the pixel point to form the middle supporting leg 3D point cloud.

[0038] It should be noted that, since step S43 fills the sparse white pixels, when directly inferring the stack 3D point cloud from the data points contained in the stack area in the 2D image, some data points will not have x-coordinates and the inferencing cannot be completed. In order to solve this problem, the present invention converts each data point in the stack point cloud of step S33 into a pixel point of the 2D initial image in step S42 again, and compares the positions of these pixel points with the stack area of ​​step S53 and the middle support leg area of ​​step S54. The data points corresponding to all the pixel points in the stack area in step S53 are extracted to form the stack 3D point cloud, and the data points corresponding to all the pixel points in the middle support leg area in step S54 are extracted to form the middle support leg 3D point cloud, thereby completing the inferencing from 2D image to 3D point cloud.

[0039] S7. performing plane fitting on the 3D point cloud of the pallet, and extracting the front point cloud close to the front surface of the pallet; Specifically, the step of extracting the front point cloud in step S7 includes: S71. Using the RANSAC algorithm to perform plane fitting on the 3D point cloud in S61, a plane corresponding to the front surface of the pallet is obtained; S72. Extract points whose distance from the plane in step S71 is no more than 10 mm to obtain a front point cloud close to the front surface of the pallet.

[0040] It should be noted that when fitting the plane of the 3D point cloud in S61, it can be achieved with the help of point cloud processing software such as PCL and Open3D. It is more convenient to fit the plane from the point cloud, including points with a distance of no more than 10mm, and the front point cloud close to the front surface of the pallet can be obtained more accurately.

[0041] S8. Obtain the placement angle of the pallet according to the front point cloud, and obtain the center position of the front surface of the pallet according to the 3D point cloud of the middle support leg; Specifically, the step of obtaining the placement angle and the center position in step S8 includes: S81. Perform plane fitting on the previous point cloud in step S72, and the corresponding plane angle is the placement angle of the pallet; S82. Take the average of the 3D point cloud of the middle supporting leg obtained in step S62 to obtain the center position of the front surface of the pallet.

[0042] S9. Output pallet placement angle and pallet front surface center position, such as Figure 5, complete the pallet identification.

[0043] The 3D point cloud-based pallet recognition method for the intelligent handling robot provided by the present invention can be applied to a variety of 3D sensors, has high sensor universality, and strong scalability. Regardless of whether it is a standard pallet or a non-standard pallet, it is only necessary to obtain the pallet contour parameters in advance to perform template construction and 2D image matching. In addition, compared with the deep learning method, the present invention does not rely on the training and reasoning of neural networks, can be deployed quickly, has lower requirements on computing performance, and avoids dependence on high-performance computing platforms.

[0044] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with this technology to understand the contents of the present invention and implement them accordingly. They cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A pallet recognition method based on 3D point cloud for an intelligent handling robot, characterized in that: The following steps are involved: S1. intercepting a target area from the 3D point cloud obtained by the intelligent handling robot, where the target area is an area containing a pallet to be identified; S2. filtering out noise points in the target area by statistical filtering to obtain a filtered point cloud; S3. Based on the prior features of the ground point cloud, the ground is identified and removed from the filtered point cloud, and then the pallet point cloud is segmented according to the Euclidean distance; S4. Project the pallet point cloud into the front view to obtain a 2D image; S5. Construct a front view template of the pallet based on the pallet contour parameters input externally, search for contours similar to the front view template in the 2D image, and extract the area whose similarity to the front view template exceeds a specified threshold as the pallet area in the 2D image; S6. According to the correspondence between the 2D image and the 3D point cloud, the 3D point cloud of the pallet is inferred from the pallet area of ​​the 2D image, and the 3D point cloud of the middle support leg is inferred from the 2D image; S7. performing plane fitting on the 3D point cloud of the pallet, and extracting the front point cloud close to the front surface of the pallet; S8. Obtain the placement angle of the pallet according to the front point cloud, and obtain the center position of the front surface of the pallet according to the 3D point cloud of the middle support leg; S9. Output the placement angle of the pallet and the center position of the front surface of the pallet to complete the pallet identification.

2. The method for identifying pallets based on 3D point clouds for an intelligent handling robot according to claim 1, characterized in that: The step of intercepting the target area in step S1 includes: S11. With the sensor on the intelligent handling robot that obtains the 3D point cloud as the center of the coordinate system, set the minimum and maximum value ranges of the three directions of x, y, and z respectively to ensure that the pallet to be identified is within the range, where the positive direction of x points forward, the positive direction of y points to the left, and the positive direction of z points upward, satisfying the right-hand coordinate system rule; S12. Given the external parameters of the relative position between the pallet to be identified and the sensor, in the 3D point cloud, data points within the specified range of x, y, and z are intercepted as the target area.

3. The method for identifying pallets based on 3D point clouds for an intelligent handling robot according to claim 2, characterized in that: The statistical filtering step in step S2 includes: S21. Calculate the average distance between each point in the target area and the nearest K points, and adjust the K value according to the filtering effect; S22. Sort all points from small to large according to the average distance calculated in S21; S23. retain a set proportion of data points that are at the front of the sort, and remove the remaining points to obtain a filtered point cloud.

4. The method for identifying pallets based on 3D point clouds for an intelligent handling robot according to claim 3, characterized in that: The step of segmenting the pallet point cloud in step S3 includes: S31. Use the RANSAC algorithm to fit a virtual plane in the filtered point cloud, where the angle between the virtual plane and the horizontal plane does not exceed a set angle and contains the most data points; S32. Eliminate the virtual plane in S31 and the data points whose distance from the virtual plane does not exceed a set value from the filtered point cloud; S33. Cluster the point cloud obtained in S32 according to the Euclidean distance, divide the point cloud into several relatively independent point cloud clusters, select the point cloud cluster close to the center of the point cloud field of view and the number of data points in the cluster exceeds the set threshold for segmentation, and obtain the pallet point cloud.

5. The method for identifying pallets based on 3D point clouds for an intelligent handling robot according to claim 4, characterized in that: The steps from the pallet point cloud to the 2D image in step S4 include: S41. Traverse the stack point cloud in step S33, calculate the extreme values ​​of the y coordinates and z coordinates of all data points: y_min, y_max, z_min, z_max, and construct a 2D initial image with the y coordinates and z coordinates of all data points. The width of the 2D initial image is (y_max-y_min) and the height is (z_max-z_min). S42. Converting each data point in the pallet point cloud in step S33 into a white pixel in the 2D initial image; S43. Perform a closing operation on the 2D initial image in step S42 to fill the sparse white pixels and reduce the internal hollowing to obtain a 2D image.

6. The method for identifying pallets based on 3D point clouds for an intelligent handling robot according to claim 5, characterized in that: The searching step in step S5 includes: S51 constructs a pallet front view template A based on external input pallet contour parameters, the pallet front view template A parameters include pallet support leg width, pallet support leg height, pallet overall width, pallet overall height; S52. Constructing a widened template B according to the pallet front view template A in step S51, the width of the widened template B is 80% of the pallet front view template A, the pallet front view template A and the widened template B together constitute the front view template of the pallet; S53. Traverse all rows and columns in the 2D image obtained in step S43, and search for the area with the highest matching degree with the front view template of the pallet and exceeding the matching degree threshold, as the pallet area in the 2D image; S54. According to the pallet support leg width and pallet support leg height parameters given in step S51, extract the middle support leg area in the pallet area in step S53.

7. The method for identifying pallets based on 3D point clouds for an intelligent handling robot according to claim 6, characterized in that: The step of extracting the middle supporting leg area in step S54 includes: Taking the vertical center line of the pallet area in step S53 as a reference, expand 1 / 2 of the pallet support leg width to the left and right, and expand 1 pallet support leg height upward from the bottom of the pallet area in step S53 to obtain the middle support leg area.

8. The method for identifying pallets based on 3D point clouds for an intelligent handling robot according to claim 6, characterized in that: The step of inferring the 3D point cloud from the 2D image in step S6 includes: S61. Convert each data point in the stack point cloud of step S33 into a pixel point of the 2D initial image in step S42. If the pixel point is located in the stack area in step S53, extract the data point corresponding to the pixel point to form a stack 3D point cloud; S62. Convert each data point in the stack point cloud in step S33 into a pixel point of the 2D initial image in step S42. If the pixel point is located in the middle supporting leg area in step S54, extract the data point corresponding to the pixel point to form the middle supporting leg 3D point cloud.

9. The method for identifying pallets based on 3D point clouds for an intelligent handling robot according to claim 8, characterized in that: The step of extracting the front point cloud in step S7 includes: S71. Using the RANSAC algorithm to perform plane fitting on the 3D point cloud in S61, a plane corresponding to the front surface of the pallet is obtained; S72. Extract points whose distance from the plane in step S71 is no more than 10 mm to obtain a front point cloud close to the front surface of the pallet.

10. The method for identifying pallets based on 3D point clouds for an intelligent handling robot according to claim 9, characterized in that: The steps of obtaining the placement angle and the center position in step S8 include: S81. Perform plane fitting on the previous point cloud in step S72, and the corresponding plane angle is the placement angle of the pallet; S82. Take the average of the 3D point cloud of the middle supporting leg obtained in step S62 to obtain the center position of the front surface of the pallet.

Images