A recognition and positioning method based on garbage inspection robot

Abstract

This invention discloses a recognition and positioning method based on a garbage inspection robot. This method combines YOLOv5-ghost with an improved ORB-SLAM2 to identify and locate garbage. This enhances the robot's ability to extract feature points, providing a more accurate and stable foundation for the robot's positioning and navigation within an environment. This enables the robot to more effectively identify and collect garbage targets, further improving its efficiency and performance in garbage inspections. Furthermore, by improving the recognition and positioning algorithms, the invention adds an attention mechanism and a lightweight Ghost module to the YOLOv5 module, enhancing the algorithm's ability to identify and locate target objects. This reduces parameters and computational complexity, lowering the model's storage and computational overhead, and making deep neural networks adaptable to devices with varying computing power.

Description

Technical Field

[0001] The present invention belongs to the field of visual SLAM technology, and in particular relates to an identification and positioning method based on a garbage inspection robot. Background Art

[0002] SLAM (Simultaneous Localization and Mapping) is a recognized cutting-edge positioning technology, primarily used to address the positioning and mapping challenges faced by robots in unknown environments. With the rise of robotics, autonomous driving, AR / VR, and other fields, SLAM technology has experienced a surge in development. SLAM is key to autonomous mobility for robots and other intelligent entities, helping to address the positioning and mapping challenges faced by robots in unknown environments. With the increasing importance of SLAM technology and the continued expansion of its application market, an increasing number of companies are investing in its research and development.

[0003] SLAM technology currently falls into two categories, based on the type of sensor used: laser SLAM based on lidar (LiDAR), and visual SLAM based on vision sensors. LiDAR is relatively expensive and consumes high power during operation. Furthermore, atmospheric conditions and stray light can affect its performance. The visual SLAM technology used in this article has a much wider range of application scenarios and can be used in both indoor and outdoor environments.

[0004] ORB-SLAM2 is a powerful visual SLAM system that can perform positioning and map construction under conditions with high real-time requirements, and is suitable for many real-time application scenarios. By using sparse feature points, ORB-SLAM2 can achieve high-precision camera positioning, which is suitable for application scenarios that require high-precision positioning. However, it also has some disadvantages that require a lot of computing resources, especially when real-time requirements are high, which may require high-performance hardware support. Like many visual SLAM systems, ORB-SLAM2 is sensitive to environmental changes (such as lighting changes, dynamic objects), which may lead to instability in positioning and mapping.

[0005] With the rapid development of deep learning, image recognition technology has been fully empowered in various fields and has achieved remarkable progress. Image recognition technology can adjust and process images in complex video streams. Based on artificial intelligence technology, it can achieve comprehensive analysis of image information, rapidly process and recognize graphics with high accuracy. This technology is currently widely used in fields such as intelligent transportation and industry, significantly promoting the development of information technology in the industry and playing a significant role in promoting industrial upgrading and transformation in my country. Summary of the Invention

[0006] The purpose of the present invention is to provide an identification and positioning method based on a garbage inspection robot.

[0007] The present invention provides a method for identifying and locating garbage inspection robots, comprising the following steps:

[0008] Step 1: The garbage inspection robot collects images and inputs them into the YOLOv5 module for recognition. If there is garbage in the image, the location of the garbage is selected and the confidence score is obtained.

[0009] Step 2: If the confidence of the acquired image is greater than the confidence threshold, mask the image to obtain the image of the region of interest; otherwise, use the image obtained in step 1 directly as the target image and execute step 6;

[0010] Step 3: Perform histogram equalization and Gaussian filtering on the image of the region of interest;

[0011] Step 4: Set the excessive motion threshold a. If the total translation w between two adjacent frames of the ROI image is greater than the excessive motion threshold a, the scale factor is reduced by the ORB feature extractor in the ORB-SLAM2 algorithm.

[0012] The total translation amount w between the two adjacent frames of the region of interest image is expressed as:

[0013]

[0014] Wherein, J is the translation coefficient; K is the dimensional change coefficient; is the average translation vector between two adjacent frames of ROI images; Δd is the size change value of the ROI image.

[0015] Step 5: Fuse the image of the region of interest obtained in step 3 with the image outside the mask to obtain a new image as the target image and execute step 6.

[0016] Step 6: Use the ORB-SLAM2 algorithm to locate the target garbage in the image.

[0017] Preferably, in step 4, the scale factor is reduced by adding a setScaleFactor function in the ORB-SLAM2 algorithm.

[0018] Preferably, in step 4, the method for obtaining the excessive movement threshold a is: the garbage inspection robot runs the ORB-SLAM2 algorithm, and the camera provided on the mobile garbage inspection robot is calculated to obtain the excessive movement threshold a by calculating the total translation w when the camera loses the tracking thread.

[0019] Preferably, in the step 1, the YOLOv5 module includes an input end, a backbone module, a neck module and a head module; the backbone module is used for feature extraction; the neck module is used for feature fusion; and the head module is used for detection output.

[0020] Preferably, in the step 1, the YOLOv5 module is trained before using the YOLOv5 module to recognize the collected image. The training process is: obtaining a data set containing junk images and inputting it into the YOLOv5 module for training, measuring the YOLOv5 module obtained after each training by the loss function, and selecting the optimal YOLOv5 module.

[0021] Preferably, the backbone module uses the C0 module to slice the input feature map, uses the SPPF module to pool the output feature map, and adds the SimAM attention mechanism in the 1-C3 layers of the backbone module.

[0022] Preferably, the Neck module includes an FPN structure and a PAN structure, and a lightweight module Ghost is added to the convolution module in the PAN structure. The expression of the lightweight module Ghost is:

[0023] Y=X*f+b

[0024] Among them, Y is the output feature map; X is the input feature map; f is the convolution filter; b is the bias term.

[0025] Preferably, the loss function adopts one of CIoU Loss and EIoU Loss.

[0026] As an advantage, in the step 1, the YOLOv5 module updates the confidence of the candidate box through the flexible non-maximum suppression algorithm to obtain the best target bounding box, and the updated candidate box confidence S' i The expression is:

[0027]

[0028] Among them, S i is the initial confidence of the i-th candidate box; M is the candidate box with the maximum confidence; b i is the i-th candidate box; IoU is the intersection over union ratio; σ is the attenuation coefficient, which is 0.5; i = 1, 2, ... m; m is the number of candidate boxes.

[0029] Preferably, in step 2, the confidence threshold is 0.8.

[0030] As an example, in the step 3, the gray level S of the region of interest image after histogram equalization is used. k The expression is:

[0031]

[0032] Where L is the maximum grayscale level of the image in the region of interest; k = 0, 1, 2, ... L-1; p(r j ) is the jth gray level r j Probability of occurrence.

[0033] The jth gray level r j The probability of occurrence p(r j ) is:

[0034]

[0035] Where N is the total number of pixels in the image of the region of interest; n j The gray level r in the region of interest image j The number of pixels.

[0036] Preferably, in step 3, the region of interest image G(x, y) processed by the Gaussian filter is:

[0037]

[0038] Where x and y are the horizontal and vertical coordinates of the image of the region of interest respectively; σ x and σ y are the standard deviations of the Gaussian function in the x and y directions, respectively.

[0039] Preferably, in step 4, the expression of the size change value Δd of the ROI image is:

[0040]

[0041] Among them, d2 is the size of the ROI image in the current frame; d1 is the size of the ROI image in the previous frame; (x 1n ,y 1n ) is the pixel coordinate of the nth vertex of the ROI image in the previous frame; (x 2n ,y 2n ) is the pixel coordinate of the nth vertex of the ROI image of the current frame; the value of n is 1 or 2, corresponding to the upper left vertex and lower right vertex of the ROI image, respectively.

[0042] The average translation vector of the region of interest images in two adjacent frames The expression is:

[0043]

[0044] Among them, (Δx n ,Δy n ) is the translation vector of the image of the region of interest, and its expression is:

[0045]

[0046] Preferably, in step 4, the value of the translation coefficient J is 0.78, and the value of the dimensional change coefficient K is 0.22.

[0047] The present invention has the following beneficial effects:

[0048] 1. This paper combines YOLOv5-ghost with the improved ORB-SLAM2 to identify and locate garbage in the target image, enhancing the garbage inspection robot's extraction of feature points, providing a more accurate and stable basis for the garbage inspection robot's positioning and navigation in the environment, enabling the garbage inspection robot to more effectively identify and collect garbage targets in the environment, further improving its garbage inspection efficiency and performance.

[0049] 2. By improving the recognition and positioning algorithms, the present invention enhances the algorithm's ability to identify and locate target objects while reducing parameters and computational complexity, lowering the model's storage and computational overhead, and enabling deep neural networks to be adapted to devices of various computing capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is the main flow chart of the present invention.

[0051] Figure 2 Improve the YOLOv5 flowchart for this invention.

[0052] Figure 3 This is a schematic diagram of the improved YOLOv5-ghost model of the present invention.

[0053] Figure 4 This is a precision-recall curve of the YOLOv5-ghost model trained by the present invention.

[0054] Figure 5 Schematic diagram of images in the training set of the present invention.

[0055] Figure 6 Schematic diagrams of the ORB-SLAM2 algorithm extracting feature points on a target object before and after the improvement of the present invention; wherein, (a) is the schematic diagram before improvement; (b) is the schematic diagram after improvement.

[0056] Figure 7Schematic diagram of the number of target point clouds extracted by the ORB-SLAM2 algorithm before and after the improvement of the present invention; (a) is the schematic diagram before the improvement; (b) is the schematic diagram after the improvement. DETAILED DESCRIPTION

[0057] The present invention will be further described below with reference to the accompanying drawings.

[0058] like Figure 1 As shown, a method for identifying and locating a garbage inspection robot includes the following steps:

[0059] Step 1: Use the YOLOv5-ghost module to detect images

[0060] 1-1. Obtain the open-source VOC dataset. Most images in the VOC dataset contain a large number of instances of varying types. Divide the obtained VOC dataset into a training set and a test set. Maintain a consistent data distribution between the training and test sets to avoid negative impacts on model training due to imbalanced instance class ratios.

[0061] 1-2. Build the YOLOv5-ghost module

[0062] like Figure 2 and 3 As shown in the figure, the YOLOv5-ghost module includes an input end, a backbone module, a neck module, and a head module; the backbone module is used for feature extraction; the neck module is used for feature fusion; and the head module is used for detection output.

[0063] 1-2-1. Improve the backbone module

[0064] In the backbone module, the C0 module in YOLOv5s replaces the Focus module to perform image slicing. This slicing operation reduces computational complexity and increases model speed, thereby improving object detection performance. The Focus module typically includes several convolutional layers and pooling operations, requiring a certain number of parameters. Compared to the Focus module, the C0 module uses a standard convolutional layer, reducing the number of model parameters and lowering the model's storage and computational overhead.

[0065] To improve the accuracy of the backbone module and simplify the model without compromising overall performance, an attention mechanism, SimAM, focusing on 3D weights, was added to layers 1-C3 of the backbone module. The attention mechanism mimics perception—i.e., how human vision perceives and focuses attention in the brain. It can focus on required information, quickly analyze complex scenes, avoid complex calculations, and simplify the model. Compared to attention mechanisms such as SE and ECA, the SimAM attention mechanism is highly stable and requires no additional parameters, resulting in better performance. SimAM also quickly ignores unimportant information, enabling efficient detection, thereby reducing model computation and increasing detection speed.

[0066] Before the feature map is input into the C2 layer of the neck module, SPPF (Spatial Pyramid Pooling-Fast) is used to pool the feature map. Similar to SPP (Spatial Pyramid Pooling), SPPF can convert feature maps of varying sizes to a uniform size without requiring pre-processing such as cropping and scaling. SPPF is also faster than SPP.

[0067] 1-2-2. Improve the neck module

[0068] The Neck module contains a large amount of feature fusion information. To reduce the computational errors caused by this information, a lightweight module, Ghost, is deployed in the Neck module to reduce model parameters and better adapt to devices with different computing capabilities. The neck module includes the FPN (Feature Pyramid Network) structure and the PAN (Path Aggregation Network) structure. Through ablation experiments, it was determined that the lightweight module Ghost is deployed within the convolution module of the PAN structure. The lightweight module Ghost is used to improve the performance and efficiency of the model, reducing parameters and computational complexity while increasing the information displayed by the model. The expression of the lightweight module Ghost is:

[0069] Y=X*f+b

[0070] Among them, Y is the output feature map; X is the input feature map; f is the convolution filter; b is the bias term; * is the convolution operation.

[0071] 1-3. Select non-maximum suppression algorithm

[0072] NMS (Non-Maximum Suppression) is an important part of the object detection process, used to search for local maxima. During the target detection process, a large number of candidate boxes will be generated at the same location. NMS can be used to find the best target bounding box and eliminate the remaining bounding boxes. However, in the case of dense boxes, using NMS will miss the target, so Soft-NMS (Flexible Non-Maximum Suppression Algorithm) is used to update the confidence of the candidate box to find the best target bounding box. The updated candidate box confidence S' i The expression is:

[0073]

[0074] Among them, S i is the initial confidence of the i-th candidate box; M is the candidate box with the maximum confidence; b i is the i-th candidate box; IoU is the intersection over union ratio; σ is the attenuation coefficient, which is 0.5; i = 1, 2, ... m; m is the number of candidate boxes.

[0075] 1-4 Determine the loss function.

[0076] In order to consider factors such as the center point offset and aspect ratio deviation of the predicted box and the target box, and to make the loss function more comprehensive, CIoU Loss is used as the loss function of the YOLOv5-ghost module, which can consider the distance and shape difference between the predicted box and the real box. However, CIoU Loss may have a certain degree of excessive penalty in some special cases, resulting in unstable training or difficulty in convergence.

[0077] In some embodiments, EIoU Loss is used as the loss function of the YOLOv5-ghost module. EIoU has a fast convergence speed, but has a high computational complexity and requires more computing resources and time for training.

[0078] 1-5. Use the training set to train the improved YOLOv5-ghost module, optimize the parameters of the YOLOv5-ghost module, and obtain the optimal YOLOv5-ghost module. During the training process, the precision-recall curve is as follows: Figure 3 As shown, the images in the training set are Figure 4 shown.

[0079] 1-6. Use the test set to test the optimal YOLOv5-ghost module and the baseline version. The images in the training set are as follows Figure 5The test results are shown in Table 1. The baseline version achieved mAP@0.5 of 79.47%, mAP@0.5:0.95 of 55.28%, and an average FPS of only around 60 frames per second. Compared to the baseline version, YOLOv5-ghost reduced the number of parameters by 1,403,280, the computational overhead by 2.5 GB, and the average FPS by around 80 frames per second, demonstrating significant improvement.

[0080] Table 1 Comparative experimental results of improved YOLOv5s

[0081]

[0082] 1-7. The garbage inspection robot collects images and inputs them into the optimized YOLOv5-ghost module for detection. If garbage is present in the image, the image box is used to select the location of the garbage as the region of interest and obtain the confidence score.

[0083] In step 2, if the confidence score of the acquired image exceeds the confidence threshold, the image is masked. Otherwise, the image obtained in step 1 is used directly as the target image and step 6 is executed. Masking involves partially blocking an image to control the area of ​​image processing. The ROI mask is multiplied by the original image to obtain the ROI image. The image values ​​within the ROI remain unchanged, while the image values ​​outside the ROI are set to 0.

[0084] In this embodiment, the confidence threshold is 0.8.

[0085] Step 3: Process the image using histogram equalization and Gaussian filter

[0086] Use histogram equalization to enhance the contrast of the masked image. In the masked image, the kth gray level r k The probability of occurrence p(r k ) is:

[0087]

[0088] Where N is the total number of pixels in the image; n k The gray level r in the image k The number of pixels; k = 0, 1, 2, ... L-1; L is the maximum grayscale level of the image.

[0089] The gray level S of the image after histogram equalization k The expression is:

[0090]

[0091] Among them, S kis the new gray level after the current gray level is mapped by the cumulative distribution function; is the cumulative probability.

[0092] Use Gaussian filter to process the image of the region of interest. The Gaussian filter will smooth the details and textures in the image, help reduce unnecessary details and noise in the image, make the image clearer, and facilitate ORB-SLAM2 to accurately extract feature points and match, thereby improving the stability and robustness of ORB-SLAM2 and enhancing its focus on the target object. The image G(x,y) after two-dimensional Gaussian filter processing is:

[0093]

[0094] Among them, x and y are the horizontal and vertical coordinates of the image respectively; σ x and σ y are the standard deviations of the Gaussian function in the x and y directions, respectively.

[0095] Step 4: Speed ​​Smoothing Model Detection

[0096] Smooth images can be better processed by the tracking thread, but when the image moves too fast, the image will have a scale mismatch problem, making the ORB-SLAM2 algorithm unable to correctly match feature points or unable to track enough feature points for positioning, resulting in ORB-SLAM2 tracking loss. By setting the speed smoothing model to detect image speed changes, dynamically adjust the feature point extraction density, and reduce the probability of tracking thread loss. The speed smoothing model is set as follows:

[0097] Assume that the lower left point of the ROI image is the origin, establish a plane rectangular coordinate system with the horizontal right direction as the x-axis and the vertical upward direction as the y-axis, and the translation vector (Δx n ,Δy n ) is:

[0098]

[0099] Among them, (x 1n ,y 1n ) is the pixel coordinate of the nth vertex of the ROI image in the previous frame; (x 2n ,y 2n ) is the pixel coordinate of the nth vertex of the ROI image in the current frame; the value of n is 1 or 2, which are the upper left vertex and lower right vertex of the image respectively.

[0100] The average translation vector of the region of interest images in two adjacent frames The expression is:

[0101]

[0102] The expression of the size change value Δd between two adjacent frames of the ROI image is:

[0103]

[0104] Wherein, d2 is the size of the ROI image in the current frame; d1 is the size of the ROI image in the previous frame.

[0105] The expression of the total translation w between two adjacent frames of the region of interest image is:

[0106]

[0107] Among them, J is the translation coefficient; K is the dimensional change coefficient.

[0108] In this embodiment, the value of the translation coefficient J is 0.78, and the value of the dimensional change coefficient K is 0.22.

[0109] Set the excess motion threshold a. Run ORB-SLAM2 on the camera and move it at different speeds. The total translation w when the camera loses the tracking thread is calculated to obtain the excess motion threshold a. When the total translation w exceeds the excess motion threshold a, it indicates that the camera is moving too fast. The scale factor can be adjusted by adding the setScaleFactor function to the ORB-SLAM2 object. A smaller scale factor results in more pyramid layers, a higher density of feature points to extract, and more complex calculations. Adjusting the scale factor can improve ORB-SLAM2's robustness to fast-moving scenes.

[0110] In this embodiment, the value of the excessive movement threshold a is 52.

[0111] Step 5: Fusion processing

[0112] The processed ROI image is fused with the image outside the mask to obtain a new image as the target image, which is then input into the ORB-SLAM2 algorithm to locate the garbage.

[0113] Step 6: Use ORB-SLAM2 algorithm for positioning.

[0114] The threads of the ORB-SLAM2 algorithm include tracking, local optimization, and global loop detection and optimization; the tracking thread is responsible for real-time camera pose tracking, feature point matching, and real-time processing of target images and selection of key frames. Figure 6 and Figure 7 These are the feature points extracted by the initial ORB-SLAM2 algorithm and the improved ORB-SLAM2 algorithm. As can be seen from the figure, compared with the initial ORB-SLAM2 algorithm, the improved ORB-SLAM2 algorithm can extract more feature points for the target object. The comparative experimental results are shown in Table 2.

[0115] Table 2 Comparative experimental results of improved feature extraction

[0116]

[0117] The local optimization thread is responsible for creating and maintaining the local map, processing new feature points to update the local map, and performing local bundle adjustment (BA) and verifying the keyframes after obtaining the processed keyframes. During the local map optimization process, adding feature points can provide more reprojection constraints, thereby more accurately estimating the camera motion.

[0118] The global loop detection and optimization thread is responsible for detecting and processing loop closures, that is, detecting situations where the camera returns to a previously passed location. This can correct the consistency of the map and reduce cumulative errors. During the loop detection process, the system compares the similarities between the current frame and certain keyframes in the map. Adding feature points can make the map richer and more detailed, providing more comparable information, thereby making loop detection more reliable. Finally, a sparse point cloud map is output to help the garbage inspection robot better understand the surrounding environment, providing a more accurate and stable foundation for the garbage inspection robot to locate and navigate in the environment, enabling the garbage inspection robot to more effectively identify and collect garbage targets in the environment, further improving its garbage inspection efficiency and performance.

Claims

1. A method for identifying and locating garbage based on a garbage inspection robot, characterized by: The following steps are involved: Step 1: The garbage inspection robot collects images and inputs them into the YOLOv5 module for recognition. If there is garbage in the image, the location of the garbage is selected and the confidence score is obtained. The YOLOv5 module updates the confidence of the candidate box through the flexible non-maximum suppression algorithm to obtain the best target bounding box. The updated candidate box confidence The expression is: ； Among them, S i is the initial confidence of the i-th candidate box; M is the candidate box with the maximum confidence; b i is the i-th candidate box; IoU is the intersection over union ratio; σ is the attenuation coefficient, which is 0.5; i=1,2,...m; m is the number of candidate boxes; Step 2: If the confidence of the acquired image is greater than the confidence threshold, mask the image to obtain the image of the region of interest; otherwise, use the image obtained in step 1 directly as the target image and execute step 6; Step 3: Perform histogram equalization and Gaussian filtering on the image of the region of interest; Gray level S of the region of interest image after histogram equalization k The expression is: ； Where L is the maximum grayscale level of the image in the region of interest; k=0,1,2,...L-1; p(r j ) is the jth gray level r j Probability of occurrence; j-th gray level r j The probability of occurrence p(r j ) is: ； Where N is the total number of pixels in the image of the region of interest; n j The gray level r in the region of interest image j The number of pixels; Step 4: Set the excessive motion threshold a. If the total translation w between two adjacent frames of the ROI image is greater than the excessive motion threshold a, the scale factor is reduced by the ORB feature extractor in the ORB-SLAM2 algorithm. The total translation amount w between the two adjacent frames of the region of interest image is expressed as: ； Where J is the translation coefficient; K is the size change coefficient; (Δ ,Δ ) is the average translation vector between two adjacent frames of ROI images; Δd is the size change value of the ROI image; Step 5: Fuse the image of the region of interest obtained in step 3 with the image outside the mask to obtain a new image as the target image and execute step 6; Step 6: Use the ORB-SLAM2 algorithm to locate the target garbage in the image.

2. The identification and positioning method based on a garbage inspection robot according to claim 1, characterized in that: In step 4, the scale factor is reduced by adding the setScaleFactor function in the ORB-SLAM2 algorithm.

3. The identification and positioning method based on a garbage inspection robot according to claim 1, characterized in that: In step 4, the method for obtaining the excessive movement threshold a is as follows: the garbage inspection robot runs the ORB-SLAM2 algorithm, moves the camera on the garbage inspection robot, and obtains the excessive movement threshold a by calculating the total translation w when the camera loses the tracking thread.

4. The identification and positioning method based on a garbage inspection robot according to claim 1, characterized in that: In the step 1, the YOLOv5 module includes an input end, a backbone module, a neck module and a head module; the backbone module is used for feature extraction; the neck module is used for feature fusion; and the head module is used for detection output. Before using the YOLOv5 module to recognize the collected image, the YOLOv5 module is trained. The training process is as follows: a data set containing junk images is obtained and input into the YOLOv5 module for training. The YOLOv5 module obtained after each training is measured by a loss function, and the optimal YOLOv5 module is selected.

5. The identification and positioning method based on a garbage inspection robot according to claim 4, characterized in that: The backbone module uses the C0 module to slice the input feature map, uses the SPPF module to pool the output feature map, and adds the SimAM attention mechanism in the 1-C3 layers of the backbone module; the Neck module includes the FPN structure and the PAN structure, and adds the lightweight module Ghost to the convolution module in the PAN structure. The expression of the lightweight module Ghost is: ； Among them, Y is the output feature map; X is the input feature map; f is the convolution filter; b is the bias term; The loss function adopts one of CIoU Loss and EIoU Loss.

6. The identification and positioning method based on a garbage inspection robot according to claim 1, characterized in that: In step 3, the region of interest image G(x, y) after Gaussian filter processing is: ； Where x and y are the horizontal and vertical coordinates of the image of the region of interest respectively; σ x and σ y are the standard deviations of the Gaussian function in the x and y directions, respectively.

7. The identification and positioning method based on a garbage inspection robot according to claim 1, characterized in that: In the step 4, the expression of the size change value Δd of the ROI image is: ； Among them, (x 1n , y 1n ) is the pixel coordinate of the nth vertex of the ROI image in the previous frame; (x 2n , y 2n ) is the pixel coordinate of the nth vertex of the ROI image in the current frame; the value of n is 1 or 2, corresponding to the upper left vertex and lower right vertex of the ROI image, respectively; The average translation vector (Δ ,Δ ) is: ； Among them, (Δx n , Δy n ) is the translation vector of the image of the region of interest, and its expression is: 。 8. The identification and positioning method based on a garbage inspection robot according to claim 1, characterized in that: In the step 4, the value of the translation coefficient J is 0.78, and the value of the dimensional change coefficient K is 0.22.

Images