3D Vision-Based Cargo Recognition Method and System

Abstract

This invention relates to the field of image data processing, and particularly to a 3D vision-based cargo recognition method and system. The method includes: acquiring a grayscale image and a depth image of the cargo surface in a logistics scenario; calculating the noise probability of each pixel in the grayscale image, and calculating the neighborhood range and number of reliable neighbors of each pixel in the grayscale image; calculating the denoised grayscale value of the pixel at the center of the neighborhood range based on the grayscale values ​​of all pixels within the neighborhood range and the number of reliable neighbors; calculating the noise probability of the depth image, and calculating the neighborhood range and number of reliable neighbors of each pixel in the depth image; calculating the denoised depth value of the pixel at the center of the neighborhood range based on the depth values ​​of all pixels within the neighborhood range and the number of reliable neighbors; inputting the data information corresponding to the denoised grayscale image and depth image into a neural network for cargo recognition, and outputting the cargo recognition result. This recognition method and system improves the accuracy of cargo recognition.

Description

Technical Field

[0001] This invention relates to the field of image data processing. In particular, it relates to a method and system for cargo recognition based on 3D vision. Background Technology

[0002] In logistics scenarios, cargo identification is a crucial step. Intelligent identification technology, as one of the foundations for the automation and intelligent development of logistics, is applied across all key stages of product production, logistics, and delivery. In the warehousing and logistics sector, with the continuous improvement of automation levels, the importance of intelligent identification technology is becoming increasingly prominent, and its market demand is growing. Current technologies primarily utilize RGB-D cameras for cargo identification in logistics.

[0003] However, existing technologies for point cloud recognition of goods on metal and glass surfaces suffer from low recognition accuracy and poor noise processing due to the unique characteristics of these surfaces. Furthermore, current processing methods struggle to remove grayscale and depth noise in stages, leading to inaccurate feature extraction and impacting the accuracy and efficiency of goods recognition. This causes inconvenience for goods management and sorting in logistics scenarios. Summary of the Invention

[0004] To address the problem that reflections from metal and glass surfaces interfere with the collected point cloud data during point cloud recognition of goods on metal and glass surfaces, resulting in poor noise processing and low goods recognition accuracy, this invention provides solutions in the following aspects.

[0005] In the first aspect, the 3D vision-based cargo recognition method includes: acquiring a grayscale image and a depth image of the cargo surface in a logistics scenario; calculating the curvature features of each pixel in the grayscale image; calculating the noise probability of each pixel in the grayscale image based on the curvature features; calculating the neighborhood range and the number of reliable neighbors of each pixel in the grayscale image based on the noise probability; acquiring the grayscale values ​​of all pixels within the neighborhood range; selecting multiple pixels with the smallest difference in grayscale values ​​at the center position within the neighborhood range based on the number of reliable neighbors; and calculating the average of the grayscale values ​​of the selected pixels as the denoised value of the pixel at the center position. The process involves: obtaining grayscale values; acquiring density features from the depth map; calculating the noise probability of pixels on the depth map based on the density features; calculating the neighborhood range and number of reliable neighbors of each pixel on the depth map based on the noise probability; extracting the depth values ​​of all pixels within the neighborhood range; selecting multiple pixels with the smallest difference in depth value from the pixel at the center position within the neighborhood range based on the number of reliable neighbors; calculating the average depth value of the selected pixels as the denoised depth value of the pixel at the center position; inputting the data information corresponding to the denoised grayscale map and depth map into a neural network for cargo recognition; and outputting the cargo recognition result.

[0006] By employing dual denoising extrema, the false detection rate of goods is significantly reduced. The noise probability is dynamically determined by the curvature feature entropy value, adaptively adjusting the neighborhood range and the number of reliable neighbors. Pixels with similar gray values ​​are selected for averaging based on the number of reliable neighbors, effectively suppressing gray-scale noise caused by uneven lighting, complex textures, or reflections, while preserving edge details. The depth noise probability is evaluated using density features, and depth outliers are filtered in a targeted manner, avoiding the damage to the geometric structure of objects caused by traditional fixed window filtering. The noise probability directly drives the neighborhood window size and the number of reliable neighbors, exhibiting higher adaptability and scene generalization ability. This achieves high-precision, low-latency goods recognition in complex logistics scenarios, effectively improving the accuracy of goods recognition.

[0007] Preferably, the method for calculating the noise probability of pixels on the grayscale image is as follows: calculate the Hessian matrix of each pixel within the window using the second derivative of Gaussian, and calculate the primary and secondary curvature directions of each pixel. Use the primary curvature direction as the curvature vector direction of the pixel, divide the curvature vector direction into multiple continuous intervals, quantify the uniformity of pixel distribution in each interval by normalizing the entropy value, and use it as the noise probability of the pixel at the center of the window. Traverse all windows to obtain the noise probability of each pixel on the grayscale image.

[0008] By adopting the above technical solution, the noise probability is estimated by "main curvature direction + normalized entropy". It can adaptively classify pixels with disordered orientation as high noise and pixels with consistent orientation as low noise. This can remove random noise while preserving the edge and texture details of the goods to the maximum extent, thereby improving the accuracy of subsequent recognition.

[0009] Preferably, the method for calculating the neighborhood range of a pixel in the grayscale image is as follows: a base window is preset, the noise probability is incremented by 1 and then the result is rounded by exponential mapping, and the base window is superimposed as the neighborhood range.

[0010] By adopting the above technical solution, the neighborhood of high-noise pixels is dynamically enlarged and the neighborhood of low-noise pixels is reduced through "noise probability → exponential mapping → superimposed basic window". This can not only collect reliable pixels in areas with severe noise to suppress artifacts, but also maintain the smallest neighborhood in clean areas to preserve edge details, thus achieving the optimal balance between denoising accuracy and computational efficiency.

[0011] Preferably, the method for calculating the number of reliable neighbors of a pixel in the grayscale image is as follows: preset a base number of neighbors, add 1 to the noise probability of the pixel, multiply it by the base number of neighbors and take the integer part, and use it as the number of reliable neighbors of the pixel.

[0012] By adopting the above technical solution, the number of reliable neighboring points is dynamically determined by multiplying the noise probability by 1 and the number of basic neighboring points and taking the integer part. High noise pixels automatically obtain more neighboring points to participate in smoothing, while low noise pixels only use a small number of neighboring points to avoid excessive blurring. This suppresses noise while accurately protecting the edges and details of the goods and improving the accuracy of subsequent recognition.

[0013] Preferably, the method for calculating the noise probability of pixels on the depth map is as follows: a local region range is preset, the difference between the depth values ​​of each pixel and other pixels within the local region range is calculated, the number of pixels whose depth difference is less than a preset threshold is counted and divided by the total number of pixels within the local region range to obtain the local density value, the local density value is normalized to obtain the density feature, which is used as the noise probability of the pixel at the center of the local region range, and the noise probability of each pixel on the depth map is obtained by traversing all local regions ranges.

[0014] Preferably, the method for calculating the neighborhood range of a pixel on the depth map is as follows: a base window is preset, the noise probability is incremented by 1 and then the result is rounded by exponential mapping, and the base window is superimposed as the neighborhood range.

[0015] By adopting the above technical solution, the high-noise area neighborhood is adaptively enlarged and the low-noise area neighborhood is reduced through "noise probability → exponential mapping → superimposed basic window". This not only fully searches for reliable pixels to complete the data in deep holes or flying point concentrations, but also avoids over-filtering in smooth areas. Thus, the real-time performance and reliability of depth denoising are improved while maintaining the geometric accuracy of the object.

[0016] Preferably, the method for calculating the number of reliable neighbors of a pixel on the depth map is as follows: preset a base number of neighbors, add 1 to the noise probability of the pixel, multiply it by the base number of neighbors and take the integer part, and use it as the number of reliable neighbors of the pixel.

[0017] By adopting the above technical solution, the number of basic neighbor points is linearly amplified by multiplying the "noise probability + 1" by the number of basic neighbor points and taking the integer part. This allows depth anomaly areas to automatically obtain more reliable neighbor points to participate in smoothing, while real surfaces only need a small number of neighbor points to avoid over-smoothing. This effectively maintains the object's edges and details while suppressing depth noise, thereby improving the accuracy and stability of subsequent 3D recognition and positioning.

[0018] Secondly, a cargo recognition system based on 3D vision includes: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the cargo recognition method based on 3D vision described in any one of the claims is implemented.

[0019] The present invention has the following effects:

[0020] 1. By employing dual denoising extrema, the false detection rate of goods is significantly reduced. The noise probability is dynamically determined by the curvature feature entropy value, adaptively adjusting the neighborhood range and the number of trusted neighbors. Pixels with similar gray values ​​are selected for averaging based on the number of trusted neighbors, effectively suppressing gray-scale noise caused by uneven lighting, complex textures, or reflections, while preserving edge details. The depth noise probability is evaluated using density features, and depth outliers are filtered in a targeted manner, avoiding the destruction of object geometry caused by traditional fixed window filtering. The noise probability directly drives the neighborhood window size and the number of trusted neighbors, exhibiting higher adaptability and scene generalization ability. High-precision, low-latency goods recognition is achieved in complex logistics scenarios, effectively improving the accuracy of goods recognition.

[0021] 2. By dynamically expanding the neighborhood of high-noise pixels and shrinking the neighborhood of low-noise pixels through "noise probability → exponential mapping → superimposed base window", the reliability of noise reduction can be improved.

[0022] 3. The number of reliable neighboring points is dynamically determined by "multiplying the noise probability by 1 and the base number of neighboring points and taking the integer part". High noise pixels automatically obtain more neighboring points to participate in smoothing, while low noise pixels only use a few neighboring points to avoid excessive blurring. While suppressing noise, the edges and details of objects are accurately preserved, improving the recognition accuracy of subsequent goods. Attached Figure Description

[0023] Figure 1 This is a flowchart of steps S1-S4 in the 3D vision-based cargo recognition method of the present invention.

[0024] Figure 2 This is a structural block diagram of the cargo recognition system based on 3D vision according to the present invention. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0026] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] Reference Figure 1 The 3D vision-based cargo recognition method includes steps S1-S4, as follows:

[0028] S1: Obtain the grayscale image and depth map of the surface of goods in the logistics scenario.

[0029] Raw data of goods in a logistics scenario is acquired using an RGB-D camera. This raw data includes 2D color image RGB (Red, Green, Blue) information and depth image D information. The 2D color image RGB information includes the two-dimensional spatial information of pixels. The coordinates, where D represents the z-coordinate of a pixel in 3D space, are used to generate initial point cloud data. Radius filtering is then applied to the initial point cloud data to remove obvious outliers, resulting in well-formed point cloud data.

[0030] The grayscale image is obtained by converting the RGB information of a two-dimensional color image to grayscale. The grayscale values ​​in the grayscale image are then replaced with the D information of the depth image to obtain the depth image.

[0031] S2: Calculate the curvature features of each pixel in the grayscale image, calculate the noise probability of each pixel in the grayscale image based on the curvature features, calculate the neighborhood range and number of reliable neighbors of each pixel in the grayscale image based on the noise probability, obtain the grayscale values ​​of all pixels in the neighborhood range, and select multiple pixels with the smallest difference in grayscale values ​​at the center position of the neighborhood range based on the number of reliable neighbors. Calculate the average value of the grayscale values ​​of the selected pixels as the denoised grayscale value of the pixel at the center position.

[0032] The method for calculating noise probability is as follows:

[0033] The Hessian matrix of each pixel within the window is calculated using the second derivative of Gaussian, and the primary and secondary curvature directions of each pixel are calculated. The primary curvature direction is selected as the curvature vector direction of the pixel, and the curvature vector direction is divided into multiple continuous intervals. The uniformity of the pixel distribution in each interval is quantified by the normalized entropy value, which is used as the noise probability of the pixel at the center of the window. The noise probability of each pixel on the grayscale image is obtained by traversing all windows.

[0034] Specifically, adopt The convolution kernel performs Gaussian filtering on the grayscale image to suppress the interference of high-frequency noise on the second derivative. Based on each pixel in the filtered grayscale image... The Hessian matrix is ​​constructed by calculating the second-order partial derivatives of the pixel in the x-direction, y-direction, and mixed second-order partial derivatives using the Gaussian second-order derivative template. ;in, This represents the Hessian matrix composed of pixels. , pixels Second-order partial derivatives in the x and y directions, It is a mixed second-order partial derivative.

[0035] Based on the Hessian matrix, the primary and secondary curvature directions of pixels in the x and y directions are calculated using characteristic polynomials. The primary curvature direction is taken as the curvature vector direction of the pixel, and the curvature vector direction is divided into multiple continuous intervals. In this embodiment, the continuous intervals include six intervals: 0°-30°, 30°-60°, 60°-90°, 90°-120°, 120°-150°, and 150°-180°. The number of pixels in each interval is counted, and a histogram is constructed. The curvature vector direction of noise points is randomly distributed, and the histogram shows a uniform distribution without obvious peaks. The curvature vector direction of real structures is consistent, and the histogram shows a concentrated distribution, with significant peaks appearing in 1-2 intervals.

[0036] The uniformity of histogram distribution is quantified by entropy value, and the curvature feature entropy value is defined. for:

[0037] ;

[0038] in, The curvature feature entropy value, For the histogram of the th Pixel count in each interval The total number of intervals, Total number of pixels. Curvature feature entropy value. The larger the value, the more uniform the histogram distribution, and the more likely it is to be noise.

[0039] Normalized curvature eigenentiation value: ;

[0040] in, The normalized curvature feature entropy value, The curvature feature entropy value, = The maximum entropy value when the distribution is uniform. The closer it is to 1, the higher the probability that it is noise.

[0041] The pixels on the grayscale image are selected sequentially according to a certain window size. Specifically, in this embodiment, a 5×5 pixel window can be selected, and the noise probability of the pixels within each window range on the grayscale image is calculated as the noise probability of the pixel at the center of the window. The noise probability of each pixel on the grayscale image is calculated in this way.

[0042] The specific method for calculating the neighborhood range of a pixel on a grayscale image is as follows:

[0043] A base window is preset, and the noise probability is incremented by 1, then exponentially mapped and rounded. This base window is then superimposed to form the neighborhood range. In this embodiment, the preset base window size is 3×3 pixels, used for areas with extremely low noise probability to reduce damage to true features. In other embodiments, different base window sizes can be selected based on experience.

[0044] For each pixel The formula for calculating the size of the neighborhood of a pixel in the grayscale image is as follows: ;

[0045] in, This represents the size of the neighborhood of a pixel in the grayscale image. This represents the base window size of a pixel in a grayscale image. This represents the noise probability of a pixel in a grayscale image. This indicates that the natural exponent is being performed. This indicates that the calculated value is rounded down to the nearest integer.

[0046] This indicates the addition of noise probability to the pixel. And perform natural exponential calculation to add the noise probability of each pixel. Processing can The range of values ​​is from translate to This ensures that there is still sufficient increment when the noise probability is at its lowest, maintaining the algorithm's adaptive range and preventing the neighborhood size from remaining almost unchanged due to excessively small values. It means to The calculated value is rounded to the nearest integer.

[0047] The specific method for calculating the number of reliable neighbors of a pixel in a grayscale image is as follows:

[0048] A preset base neighbor number is used. The noise probability of a pixel is incremented by 1, multiplied by the base neighbor number, and the result is the integer part of the product. This integer part is the number of reliable neighbors for that pixel. In this embodiment, the preset base neighbor number is three, which is used for regions with extremely low noise probability to reduce damage to real features. In other embodiments, different numbers of base neighbors can be selected based on experience.

[0049] For each pixel, calculate its number of reliable neighbors in the grayscale image. The specific calculation formula is as follows:

[0050] ;

[0051] in, This represents the number of reliable neighbors of a pixel in the grayscale image. This represents the number of basic neighbors of a pixel in the grayscale image. This represents the noise probability of a pixel in a grayscale image. Represents the number of basic neighbors. This indicates rounding the calculated value to the nearest integer. It means to The calculated value is rounded to the nearest integer.

[0052] Obtain the grayscale values ​​of all pixels within the neighborhood range, compare the differences between each pixel and the grayscale value of the pixel at the center of the neighborhood range, select the pixels with the smallest differences from the pixel at the center of the neighborhood range based on the number of trusted neighbors, and ensure that the number of selected pixels is consistent with the number of trusted neighbors. Calculate the average grayscale value of each selected pixel as the denoised grayscale value of the pixel at the center of the neighborhood, and output the denoised grayscale image.

[0053] S3: Obtain the density features of the depth map, calculate the noise probability of pixels on the depth map based on the density features, calculate the neighborhood range and number of reliable neighbors of pixels on the depth map based on the noise probability, extract the depth values ​​of all pixels within the neighborhood range, and select multiple pixels with the smallest difference in depth value from the pixel at the center position within the neighborhood range based on the number of reliable neighbors. Calculate the average of the depth values ​​of the selected pixels as the denoised depth value of the pixel at the center position.

[0054] The method for calculating the noise probability of pixels on the depth map is as follows:

[0055] A local region is preset. The difference between the depth values ​​of each pixel within the local region and other pixels in the depth map is calculated. The number of pixels whose depth difference is less than a preset threshold is counted and divided by the total number of pixels in the local region to obtain the local density value. The local density value is normalized to obtain the density feature, which is used as the noise probability of the pixel at the center of the local region. All local regions are traversed to obtain the noise probability of each pixel in the depth map.

[0056] Specifically, with each pixel Centered on a fixed-size local area, in this embodiment the size of the selected local area can be 5×5 pixels. In other embodiments, the size of the selected local area can be adjusted according to the data density.

[0057] Calculate the depth value of each pixel within this local region. The number of pixels whose depth values ​​differ from other pixels is less than a preset threshold. In this embodiment, the preset threshold for a pixel is 10mm. In other embodiments, the preset threshold can be set according to the depth variation range.

[0058] Number of pixels Divide by the total number of pixels in the region To obtain local density values For local density values Normalization is performed to obtain the normalized density features: ;

[0059] in, This represents the normalized density feature. Represents the local density value. This represents the mean density characteristic within the region. Standard deviation, with normalized density features This represents the noise probability of a pixel at the center of a local region.

[0060] Calculate the noise probability of pixels within each local region on the depth map, and use this probability as the noise probability of the pixel at the center of that local region. Calculate the noise probability of each pixel on the depth map in this manner.

[0061] The method for calculating the neighborhood range of a pixel on the depth map is as follows:

[0062] A base window is preset, and the noise probability is incremented by 1, then exponentially mapped and rounded. This base window is then superimposed to form the neighborhood range. In this embodiment, the preset base window size is 3×3 pixels, used for areas with extremely low noise probability to reduce damage to true features. In other embodiments, different base window sizes can be selected based on experience.

[0063] For each pixel The size of its neighborhood is calculated using the following formula: ;

[0064] in, This represents the size of the neighborhood of a pixel on the depth map. This represents the base window size of a pixel on the depth map. This represents the noise probability of a pixel in the depth map. This indicates that the natural exponent is being performed. This indicates that the calculated value is rounded down to the nearest integer.

[0065] Specifically, This indicates the addition of noise probability to the pixel. And perform natural exponential calculation, by adding the noise probability of the pixel. Processing can reduce noise probability The range of values ​​is from translate to This ensures that there is still sufficient increment when the noise probability is at its lowest, maintaining the algorithm's adaptive range and preventing the neighborhood size from remaining almost unchanged due to excessively small values. It means to The calculated value is rounded to the nearest integer.

[0066] The method for calculating the number of reliable neighbors of a pixel on a depth map is as follows:

[0067] A preset base neighbor count is used. The noise probability of a pixel is incremented by 1, multiplied by the base neighbor count, and the result is the integer part of the product. This integer represents the number of reliable neighbors for that pixel. In this embodiment, the preset base neighbor count is 3 pixels, used for regions with extremely low noise probability to reduce damage to true features. In other embodiments, different numbers of base neighbors can be selected based on experience.

[0068] For each pixel, the number of its reliable neighbors is calculated using the following formula:

[0069] ;

[0070] in, This represents the number of reliable neighbors of a pixel in the depth map. This represents the number of basic neighbors of a pixel in the depth map. This represents the noise probability of a pixel in the depth map. This indicates that the calculated value is rounded down to the nearest integer. It means to The calculated value is rounded to the nearest integer.

[0071] Obtain the depth values ​​of all pixels within the neighborhood range, compare the differences between each pixel and the pixel at the center of the neighborhood range, select the pixels with the smallest differences from the pixel at the center of the neighborhood range based on the number of trusted neighbors, and ensure that the number of selected pixels is consistent with the number of trusted neighbors. Calculate the average depth value of each selected pixel as the denoised depth value of the pixel at the center of the neighborhood, and output the denoised depth map.

[0072] S4: Input the data information corresponding to the denoised grayscale image and depth image into the neural network to identify the goods and output the goods identification result.

[0073] The 3D vision-based cargo identification system includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the 3D vision-based cargo identification method according to the first aspect of the present invention.

[0074] Reference Figure 2 The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0075] The 3D vision-based cargo recognition method and system provided by this invention denoises the grayscale image and depth image of the cargo surface, and dynamically adjusts the neighborhood range and the number of reliable neighboring points according to the noise probability of each pixel during the denoising process, thereby improving the denoising effect and ensuring the accuracy of cargo recognition.

[0076] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for cargo identification based on 3D vision, characterized in that, The method comprises the following steps: obtain a grayscale image and a depth image of a cargo surface in a logistics scene; calculate the curvature feature of each pixel point on the grayscale image, calculate the noise probability of the pixel point on the grayscale image based on the curvature feature, calculate the neighborhood range of the pixel point on the grayscale image and the number of trusted neighboring points based on the noise probability, obtain the grayscale values of all pixel points in the neighborhood range, and select multiple pixel points in the neighborhood range with the smallest difference in grayscale value based on the number of trusted neighboring points, and calculate the average value of the grayscale values of the selected pixel points as the denoised grayscale value of the pixel point at the center position; obtain the density feature of the depth image, calculate the noise probability of the pixel point on the depth image based on the density feature, calculate the neighborhood range of the pixel point on the depth image and the number of trusted neighboring points based on the noise probability, extract the depth values of all pixel points in the neighborhood range, and select multiple pixel points with the smallest difference in depth value based on the number of trusted neighboring points, and calculate the average value of the depth values of the selected pixel points as the denoised depth value of the pixel point at the center position; input the data information corresponding to the denoised grayscale image and depth image into a neural network for cargo identification, and output the cargo identification result; The way to calculate the noise probability of the pixel point on the grayscale image is: calculate the Hessian matrix of each pixel in the window through the second derivative of Gaussian, and calculate the principal bending direction and the secondary bending direction of each pixel point, take the principal bending direction as the curvature vector direction of the pixel point, divide the curvature vector direction into multiple continuous intervals, quantify the uniformity of the pixel point in each interval through normalized entropy value, take the noise probability of the pixel point at the center position of the window, and traverse all windows to obtain the noise probability of each pixel point on the grayscale image; The way to calculate the noise probability of the pixel point on the depth image is: preset a local region range, calculate the difference between each pixel point and other pixel points in the local region range of the depth image, count the number of pixel points with a depth difference less than a preset threshold and divide the total number of pixel points in the local region range to obtain a local density value, normalize the local density value to obtain a density feature as the noise probability of the pixel point at the center position of the local region range, and traverse all local region ranges to obtain the noise probability of each pixel point on the depth image. 2.The 3D vision-based cargo identification method of claim 1, wherein, The way to calculate the neighborhood range of the pixel point on the grayscale image is: preset a basic window, take the noise probability plus 1 and perform exponential mapping and rounding, and superimpose the basic window as the neighborhood range. 3.The 3D vision-based cargo identification method of claim 1, wherein, The way to calculate the number of trusted neighboring points of the pixel point on the grayscale image is: preset a basic number of neighboring points, take the noise probability of the pixel point plus 1 and multiply it by the basic number of neighboring points to obtain the number of trusted neighboring points of the pixel point. 4.The 3D vision-based cargo identification method of claim 1, wherein, The way to calculate the neighborhood range of the pixel point on the depth image is: preset a basic window, take the noise probability plus 1 and perform exponential mapping and rounding, and superimpose the basic window as the neighborhood range. 5.The 3D vision-based cargo identification method of claim 1, wherein, The way to calculate the number of trusted neighboring points of the pixel point on the depth image is: preset a basic number of neighboring points, take the noise probability of the pixel point plus 1 and multiply it by the basic number of neighboring points to obtain the number of trusted neighboring points of the pixel point.

6. A goods identification system based on 3D vision, characterized in that, The method comprises the following steps: A processor and a memory, the memory storing computer program instructions which, when executed by the processor, implement the 3D vision based cargo identification method according to any one of claims 1-5.

Images