Intelligent picking hand based on improved YOLOv5 model

Abstract

The invention relates to an intelligent picking hand based on an improved YOLOv5 model, which comprises a base and a driving controller, the base and the driving controller are connected through a connecting rod mechanism, a picking turntable is arranged on the driving controller, a tray is arranged on one side, far away from the driving controller, of the picking turntable, and a plurality of mechanical claws are arranged in the circumferential direction of the tray. Each mechanical claw is slidably mounted on a sliding rail of the picking turntable through a moving block assembly on the lower side of the mechanical claw, the mechanical claws are rotatably connected with the moving block assembly, all the mechanical claws and the tray jointly form a picking space, and a camera in which an improved YOLOv5 model is arranged is mounted on the driving controller. During picking, the driving controller recognizes and positions fruits needing to be picked through the camera, the mechanical claw is controlled to approach the fruits, then the picking space is enlarged to cover the fruits, then the picking space is reduced to clamp the fruits, and finally the picking rotary disc is controlled to rotate to complete picking. The fruit picking device can be matched with picking of fruits of various sizes, and the picking efficiency is higher.

Description

Technical Field

[0001] The present invention relates to the field of fruit picking, and in particular to an intelligent picking hand based on an improved YOLOv5 model. Background Art

[0002] At present, with the development of robotics and intelligent control technology, mechanical picking hands are more and more widely used in agriculture. However, the current intelligent picking mechanical claws have a single application scenario. Usually, the picking manipulator can only harvest one type of fruit and has a slow response speed. To this end, the present invention proposes an intelligent picking hand based on an improved YOLOv5 model. Summary of the invention

[0003] Aiming at the problems that the mechanical pickers in the prior art can only harvest a single type of fruit, have a limited scope of use, and have slow recognition speed of the target detection algorithm and low picking efficiency, the present invention proposes an intelligent picker based on an improved YOLOv5 model, which can pick fruits of various sizes while improving work efficiency. The specific technical solution is as follows:

[0004] An intelligent picking hand based on an improved YOLOv5 model comprises a base and a driving controller. The base and the driving controller are connected by a connecting rod mechanism. A picking turntable is arranged at the other end of the driving controller. A tray is arranged on a side of the picking turntable away from the driving controller. A plurality of mechanical claws are arranged in a circumferential direction of the tray. Each mechanical claw is slidably mounted on a slide rail of the picking turntable through a moving block assembly at its lower side. The mechanical claw is rotatably connected to the moving block assembly. All the mechanical claws and the tray form a picking space together. A camera is installed on the driving controller. An improved YOLOv5 model is deployed in the camera. The improved YOLOv5 model uses an improved Shuff leNetV2 is used as the backbone network, GhostConv is used as the convolution in the neck network, the fused bidirectional feature pyramid is used as the connection method of the neck network, and the EDCA attention mechanism is introduced; the drive controller is communicated with the camera, the connecting rod mechanism, each mechanical claw and the moving block component respectively. When picking, the drive controller identifies and locates the fruit to be picked through the camera, and then controls the connecting rod mechanism to rotate close to the fruit, and then controls the mechanical claw to rotate on the moving block component and the moving block component to slide on the picking turntable to expand the picking space to cover the fruit, and then reduce the picking space to clamp the fruit, and finally controls the picking turntable to rotate to complete the picking.

[0005] Furthermore, the connecting rod mechanism includes a telescopic rod one fixedly connected to the base, a telescopic rod two rotatably connected to the telescopic rod one, a telescopic rod three rotatably connected to the telescopic rod two, and a telescopic rod four rotatably connected to the telescopic rod three. The telescopic rod four is rotatably connected to the drive controller, and the rotation and extension of each telescopic rod and the rotation of the drive controller relative to the telescopic rod four are controlled by the drive controller.

[0006] Furthermore, the moving block assembly includes a mechanical claw moving block located on the upper side and rotatably connected to the mechanical claw, and a track moving block located on the lower side and slidably connected to the slide rail, and the mechanical claw moving block and the track moving block are rotatably connected; an angle sensor is provided at the rotational connection between the mechanical claw and the mechanical claw moving block; the slide rail is provided with a track baffle at the edge of the picking turntable, and a distance sensor for measuring the distance between each mechanical claw and its corresponding track baffle is installed in the slide rail.

[0007] Furthermore, a spring is provided between the picking turntable and the tray, a rubber pad is provided on the mechanical claw, and a pressure sensor is provided on the side of the rubber pad that contacts the fruit.

[0008] Further, the steps to improve the YOLOv5 model construction are as follows:

[0009] S1. Constructing a data set: Use a photographing device to photograph fruits under natural conditions to construct an original data set;

[0010] S2. Data preprocessing: perform data augmentation on the original dataset, then label the fruits in the pictures, and finally randomly divide the labeled dataset into training set, validation set, and test set according to the proportion;

[0011] S3. Build an improved YOLOv5 model: Use the improved ShuffleNetV2 to replace the original YOLOv5 backbone network, use GhostConv to replace the conventional convolution in the neck network, use the fused bidirectional feature pyramid to replace the connection method of the neck network, and introduce the EDCA attention mechanism;

[0012] S4. Use the training set and validation set to train and validate the improved YOLOv5 model, save the optimal weights during the training process, and then use the test set for testing.

[0013] Furthermore, in the improved ShuffleNetV2, a residual connection was introduced in the basic feature extraction module of the original model, the maximum pooling layer was used in the residual connection of its downsampling module to complete the downsampling, the number of channels was adjusted through 1×1 convolution, and a dual-channel feature extraction module was introduced to replace the 3*3 deep convolution in the basic feature extraction module.

[0014] Furthermore, in S1, the sample image is adjusted so that the image size is 640×640×3.

[0015] Furthermore, in S2, the labeled dataset is randomly divided into training set, validation set, and test set in a ratio of 8:1:1.

[0016] Furthermore, in S2, the data set is divided, and the initial anchor boxes of the labeled data set are reset using k-means clustering and genetic learning algorithms.

[0017] Furthermore, the EDCA attention mechanism uses average pooling and standard deviation pooling to process the input features, compressing them into 1×1×C feature maps, where C is the number of channels in the feature map, and generates weights for each channel through a one-dimensional convolution of size K. The elements of the feature maps obtained by the two paths are added together, and the weights of each channel are obtained through the Sigmoid function. Finally, the weight ω is multiplied by the original input feature map to obtain the output feature map.

[0018] The beneficial effects of the present invention are:

[0019] 1. A sliding mechanical claw is provided, so that the picking space formed can be large or small, and the picking and grading of fruits of various sizes and shapes can be realized.

[0020] 2. Use the improved ShuffleNetV2 model to replace the backbone network of YOLOv5 to reduce model parameters and improve the detection speed of the model; use GhostConv to replace the conventional convolution in the neck network to further reduce model parameters, and introduce the self-designed EDCA attention mechanism in the C3 module of the neck structure of the YOLOV5 model to improve the feature expression ability of the model, thereby enhancing the model's ability to detect targets; use the BiFPN structure to replace the PANet layer in YOLOv5, thereby reducing the loss of the model between different structural layers and effectively improving the model's ability to extract multi-scale features. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0022] Figure 1 is a schematic diagram of the structure of the intelligent picking hand of the present invention;

[0023] Figure 2 is a schematic diagram of the mechanical claw moving block of the present invention;

[0024] Figure 3 is a schematic diagram of the mechanical claw of the present invention;

[0025] Figure 4 It is a network structure diagram of the improved YOLOv5 model described in the present invention;

[0026] Figure 5 It is the network structure diagram of the improved ShuffleNetV2 of the present invention;

[0027] Figure 6 It is a network structure diagram of the fused bidirectional feature pyramid of the present invention;

[0028] Figure 7 It is a structural diagram of the EDCA attention mechanism of the present invention;

[0029] Figure 8 It is a working schematic diagram of the intelligent picking hand of the present invention picking oval fruits.

[0030] In the figure: 1. base; 2. telescopic rod 1; 3. telescopic rod 2; 4. telescopic rod 3; 5. telescopic rod 4; 6. drive controller; 7. camera; 8. picking turntable; 9. moving track A; 10. moving track B; 11. track moving block; 12. mechanical claw moving block; 12.1. rotating shaft; 12.2. rotating ring; 13. spring; 14. tray; 15. track baffle; 16. rubber pad; 17. mechanical claw; 17.1. shaft hole. DETAILED DESCRIPTION

[0031] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and the like are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first", "second", and the like may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0032] Taking apple picking as an example, the present invention provides the following specific implementation scheme:

[0033] like Figure 1-3As shown, the present invention provides an intelligent picking hand based on the improved YOLOv5 model, including a base 1 and a drive controller 6. The base 1 and the drive controller 6 are connected by a connecting rod mechanism. A picking turntable 8 is provided at the other end of the drive controller 6. The picking turntable 8 can rotate according to the drive controller 6. A tray 14 is provided on the side of the picking turntable 8 away from the drive controller 6. A plurality of mechanical claws 17 are provided in the circumferential direction of the tray 14. Each mechanical claw 17 is slidably mounted on the slide rail of the picking turntable 8 through a moving block assembly on the lower side thereof. The mechanical claw 17 is rotatably connected to the moving block assembly. All the mechanical claws 17 and the tray 14 form a picking space together. A camera 7 is installed on the drive controller 6, and an improved YOLOv5 model is deployed in the camera 7. The improved YOLOv5 model uses an improved ShuffleNetV2 As the backbone network, GhostConv is used to replace the conventional convolution in the neck network, and the fused bidirectional feature pyramid is used as the connection method of the neck network, and the EDCA attention mechanism is introduced; the drive controller is respectively connected to the camera, the connecting rod mechanism, and each mechanical claw. When picking, the drive controller 6 identifies and locates the fruit to be picked through the camera 7, and then controls the connecting rod mechanism to rotate close to the fruit, and then controls the mechanical claw 17 to rotate on the moving block assembly and the sliding of the moving block assembly on the picking turntable 8 to expand the picking space to cover the fruit and then move in the opposite direction to narrow the picking space to clamp the fruit, and finally controls the picking turntable 8 to rotate to complete the picking. In order to adapt to the picking of fruits of different sizes and shapes, the mechanical claw 17 can rotate 360 ​​degrees on the mechanical claw moving block 12 to facilitate the picking of more types of fruits.

[0034] Preferably, the linkage mechanism is a four-bar linkage mechanism, which includes a telescopic rod 2 fixedly connected to the base 1, a telescopic rod 2 3 rotatably connected to the telescopic rod 2 , a telescopic rod 3 rotatably connected to the telescopic rod 2 3 , and a telescopic rod 4 5 rotatably connected to the telescopic rod 3 4 , the telescopic rod 4 5 is rotatably connected to the drive controller 6 , and the rotation and extension of each telescopic rod and the rotation of the drive controller 6 relative to the telescopic rod 4 5 are controlled by the drive controller 6.

[0035] Preferably, the moving block assembly includes a mechanical claw moving block 12 located on the upper side and rotatably connected to the mechanical claw 17 and a track moving block 11 located on the lower side and slidably connected to the slide rail. The mechanical claw moving block 12 is rotatably connected to the track moving block 11 through a rotating ring 12.2 located at the lower part thereof. Correspondingly, a groove for mounting the rotating ring 12.2 is provided on the track moving block. The rotation of the mechanical claw moving block 12 is controlled by a driving controller. The inner concave surface of the mechanical claw 17 faces the tray 14. The mechanical claw 17 is provided with an axial hole 17.1. The upper part of the mechanical claw moving block 12 is provided with a rotating shaft 12.1. The mechanical claw 17 and the mechanical claw moving block 12 are rotatably connected through the rotating shaft 12.1 and the axial hole 17.1. An angle sensor is provided at the rotating connection between the mechanical claw 17 and the mechanical claw moving block 12, that is, the rotating shaft 12.1; the slide rail is provided with a track baffle 15 at the edge of the picking turntable 8, and a distance sensor for measuring the distance between each mechanical claw 17 and its corresponding track baffle 15 is installed in the slide rail.

[0036] Preferably, the camera 7 is a binocular depth camera, a spring 13 is provided between the picking turntable 8 and the tray 14, and a rubber pad 16 is also provided on the mechanical claw 17 to prevent the fruit from being damaged by pressure; the tray 14 is made of rubber; a pressure sensor is provided on the side of the rubber pad 16 that contacts the fruit to determine whether it is in contact with the fruit and the amount of force applied to the fruit.

[0037] In this embodiment, the slide rail on the side of the picking turntable 8 away from the drive controller 6 includes a moving track A9 and a moving track B10 that intersect each other perpendicularly. Four mechanical claws 17 are provided, one pair of mechanical claws 17 is installed on the moving track A9, and the other pair of mechanical claws 17 is installed on the moving track B10.

[0038] Preferably, taking Apple as an example, the steps for building an improved YOLOv5 model are as follows:

[0039] S1. Construct an apple dataset: Use shooting equipment such as mobile phones and cameras to shoot fruits under natural conditions to construct the original dataset. Natural conditions include cloudy days, rainy days, front light and back light. To ensure the recognition speed of the model, adjust the image so that the input image size is 640×640×3.

[0040] S2, data preprocessing: mosaic data enhancement and brightness, contrast, Gaussian blur and other enhancement methods are used to enhance the original data set, and Labelimg software is used to mark the fruits in the image. During the marking process, the fruits can be marked in multiple levels according to their colors. Finally, the initial anchor frame of the marked data set is reset using k-means clustering + genetic learning algorithm, and then the data set is randomly divided into training set, verification set and test set according to the ratio of 8:1:1. In this embodiment, the number of anchor frames, input image size, maximum threshold of the aspect ratio of the annotation frame and the number of iterations in the k-means clustering + genetic learning algorithm are set to 9, 640, 4 and 1000 respectively, and the sizes most suitable for apple data set recognition are obtained as [12, 15, 17, 23, 27, 35], [46, 55, 92, 97, 199, 146], [237, 244, 325, 255, 330, 331]. (When picking other fruits, you can follow this process to preprocess the data and initial anchor frames)

[0041] S3, such as Figure 4 As shown in the figure, an improved YOLOv5 model is built: the original YOLOv5 backbone network is replaced with the improved ShuffleNetV2 to reduce the number of model parameters and improve the recognition speed of the model; the conventional convolution in the neck network is replaced with GhostConv to further reduce the model parameters and facilitate model deployment; the connection mode of the neck network is replaced with the fused bidirectional feature pyramid (BiPFN), and the ability to detect objects of different scales is enhanced by bidirectionally fusing low-level spatial features and high-level semantic features. The network structure diagram of BiPFN is shown in the figure. Figure 6 As shown in the figure, the EDCA attention mechanism is introduced to improve the recognition ability of the target. In the improved YOLOv5 structure diagram, the GBS feature extraction module is composed of GhostConv convolution layer + batch normalization layer (BatchNormalization) + activation function stack. The C3 module is composed of conventional convolution, batch normalization layer and activation function. In the improved YOLOv5 structure diagram, the backbone network is downsampled five times, and then after passing through the neck network, three detection layers of 160*160, 80*80, and 40*40 are generated in the head network for multi-scale detection of the target.

[0042] Among them, Figure 5As shown in the figure, the improved ShuffleNetV2 introduces residual connection in the basic feature extraction module of the original model to solve the gradient disappearance problem, so that the model can learn more complex feature representations, and uses the maximum pooling layer in the residual connection of its downsampling module to complete downsampling, and adjusts the number of channels through 1×1 convolution to ensure the consistency of the output channel number; introduces a dual-channel feature extraction module to replace the 3*3 deep convolution in the basic feature extraction module of the original model to improve the model's ability to extract lesions of different scales. To facilitate the use of the model in YOLOv5, the basic feature extraction module of the improved ShuffleNetV2 model is named SFB1, and the downsampling feature extraction module is named SFB2.

[0043] like Figure 7 As shown in the figure, X is used as the input feature in the EDCA attention mechanism, and the size of the feature map is H×W×C, where H represents the height of the feature map, W represents the width of the feature map, and C represents the number of channels of the feature map. The EDCA attention mechanism uses average pooling (AvgPool) and standard deviation pooling (StdPool) to process the input, compressing it into 1×1×C feature maps respectively, and generates weights for each channel through a one-dimensional convolution of size K. The calculation of the mean and standard deviation is shown in formulas (1) and (2):

[0044]

[0045]

[0046] In formula (1) and (2), A c and S c Represents the mean and standard deviation of each element in the channel.

[0047] The convolution kernel size K can be adaptively determined by performing nonlinear mapping on the channel dimension. The adaptation function is defined as shown in formula (3):

[0048]

[0049] In formula (3), C represents the input feature channel dimension, |x| odd Represents the odd number closest to x. γ and b are used to change the ratio between the number of channels C and the size of the convolution kernel. According to the empirical values ​​in the literature, they are set to γ ​​= 2 and b = 1. Then, the elements of the feature maps obtained by the two paths are added together, and the weight ω of each channel is obtained through the Sigmoid function. Finally, the weight ω is multiplied by the original input feature map to obtain the output feature map. The calculation of the weight ω is shown in formula (4):

[0050] ω=σ(C1D k (y1)+C1D k(y2)) (4)

[0051] In formula (4), σ represents the Sigmoid activation function, C1D represents the one-dimensional convolution, K represents the one-dimensional convolution kernel size, y1 represents the feature map output by the average pooling path, and y2 represents the feature map output by the standard deviation pooling path.

[0052] S4. The number of images input into the improved YOLOv5 model in each batch is 32; the loop iteration is 300 rounds; the learning rate is 0.0024. The improved YOLOv5 model is trained and verified using the training set and the validation set, the optimal weights in the training process are saved, the model is tested using the test set, and the model is deployed in the camera after passing the test.

[0053] In this embodiment, picking is a process as follows: in the process of fruit picking, the camera 7 first obtains the distance between the target to be picked and the target, locks the target, and transmits it to the drive controller 6. The drive controller 6 controls the rotation and extension of the telescopic rod 1 2, the telescopic rod 2 3, the telescopic rod 3 4 and the telescopic rod 4 5 to realize the movement and extension of the mechanical claw 17, thereby realizing the approach of the mechanical claw 17 to the fruit, and then controls the mechanical claw 17 to rotate and open, and the track moving block 11 slides in the direction away from the tray 14. After the tray 14 reaches the fruit, the mechanical claw 17 rotates and contracts, and the track moving block 11 slides toward the tray 14 until the pressure sensor on the mechanical claw 17 senses the pressure, the mechanical claw 17 stops moving, and the picking turntable 8 rotates to complete the picking.

[0054] The circuit controller required in the process of fruit picking is placed on the drive controller 6, which controls the shooting of the camera 7, the rotation of the picking turntable 8, the rotation and extension of each telescopic rod in the connecting rod mechanism, the rotation of the mechanical claw 17 relative to the mechanical claw moving block 12 and the sliding of the track moving block 11.

[0055] When picking round fruits, four mechanical claws 17 are located in four directions, and the round fruits can be evenly stressed during the picking process, making it easy to pick. The rubber pads 16 on the mechanical claws 17 can prevent damage to the fruits during the picking process. The pressure sensor on the rubber pads 16 can transmit the pressure of the mechanical claws 17 during the grabbing process to the mobile terminal in real time and record it at all times, avoiding damage to the fruits caused by excessive pressure. The sliding of each mechanical claw 17 on the slide rail is controlled separately by the drive controller 6, and the slightly irregular shapes of round fruits can still be picked well. The distance sensors arranged on the moving track A9 and the moving track B10 can measure the distance moved by the mechanical claw 17 through the relative distance between the track moving block 11 and the track baffle 15, wherein the mechanical claw 17 and the rotating shaft 12.1 and the shaft hole 17.1 in the mechanical claw moving block 12 are connected together to realize the rotation of the mechanical claw 17, and the angle sensor arranged at the rotating shaft 12.1 can sense the inclination angle of the mechanical claw 17. The size of the fruit can be measured by the distance sensor and the angle sensor, so as to realize the size grading of the fruit during the picking process, and the color grading is realized by the camera 7. After the fruit is picked, one of the mechanical claws is placed horizontally with the ground, and the weight of the fruit can be calculated by the pressure sensors on the other three mechanical claws to realize the grading of the weight of the fruit, and finally the picking and grading of the fruit can be realized.

[0056] like Figure 8 As shown, when picking oval fruits, through the rotation connection between the rotating ring 12.2 on the mechanical claw moving block 12 and the track moving block 11, four manipulators are parallel in pairs of non-diagonal positions, which facilitates better picking of oval fruits.

[0057] The present application sets a slidable mechanical claw 17, so that the formed picking space can be large or small, which can match the picking and grading of fruits of various sizes and shapes. The improved ShuffleNetV2 model is used to replace the backbone network of YOLOv5, reduce model parameters, and improve the detection speed of the model; the self-designed EDCA attention mechanism is introduced in the C3 module of the neck structure of the YOLOV5 model to improve the feature expression ability of the model, thereby enhancing the model's detection ability of the target; the BiFPN structure is used to replace the PANet layer in YOLOv5, thereby reducing the loss of the model between different structural layers, and effectively improving the model's ability to extract multi-scale features.

[0058] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected.

Claims

1. An intelligent picking hand based on an improved YOLOv5 model, characterized in that: The invention comprises a base (1) and a drive controller (6), wherein the base (1) and the drive controller (6) are connected via a connecting rod mechanism, a picking turntable (8) is provided at the other end of the drive controller (6), a tray (14) is provided on the side of the picking turntable (8) away from the drive controller (6), a plurality of mechanical claws (17) are provided in the circumferential direction of the tray (14), each mechanical claw (17) is slidably mounted on a slide rail of the picking turntable (8) via a moving block assembly at its lower side, the mechanical claw (17) is rotatably connected to the moving block assembly, and all the mechanical claws (17) and the tray (14) together form a picking space, a camera (7) is installed on the drive controller (6), and an improved YOLOv5 model is deployed in the camera (7); the improved YOLOv5 model The model uses the improved ShuffleNetV2 as the backbone network, uses GhostConv as the convolution in the neck network, uses the fused bidirectional feature pyramid as the connection method of the neck network, and introduces the EDCA attention mechanism; the driving controller (6) is respectively connected to the camera (7), the connecting rod mechanism, the mechanical claw (17) and the moving block component. When picking, the driving controller (6) identifies and locates the fruit to be picked through the camera (7), and then controls the connecting rod mechanism to rotate and approach the fruit, and then controls the mechanical claw (7) to rotate on the moving block component and the moving block component to slide on the picking turntable (8) to expand the picking space to cover the fruit, and then reduce the picking space to clamp the fruit, and finally controls the picking turntable (8) to rotate to complete the picking.

2. According to claim 1, the intelligent picking hand based on the improved YOLOv5 model is characterized in that: The connecting rod mechanism comprises a telescopic rod 1 (2) fixedly connected to a base (1), a telescopic rod 2 (3) rotatably connected to the telescopic rod 1 (2), a telescopic rod 3 (4) rotatably connected to the telescopic rod 2 (3), and a telescopic rod 4 (5) rotatably connected to the telescopic rod 3 (4); the telescopic rod 4 (5) is rotatably connected to a driving controller (6); the rotation and extension of each telescopic rod and the rotation of the driving controller (6) relative to the telescopic rod 4 (5) are controlled by the driving controller (6).

3. The intelligent picking hand based on the improved YOLOv5 model according to claim 1, characterized in that: The moving block assembly comprises a mechanical claw moving block (12) located on the upper side and rotatably connected to the mechanical claw (17) and a track moving block (11) located on the lower side and slidably connected to the slide rail, wherein the mechanical claw moving block (12) and the track moving block (11) are rotatably connected; an angle sensor is provided at the rotatable connection between the mechanical claw (17) and the mechanical claw moving block (12); a track baffle (15) is provided at the edge of the picking turntable (8), and a distance sensor for measuring the distance between each mechanical claw (17) and its corresponding track baffle (15) is installed in the slide rail.

4. The intelligent picking hand based on the improved YOLOv5 model according to claim 1, characterized in that: A spring (13) is also provided between the picking turntable (8) and the tray (14), a rubber pad (16) is also provided on the mechanical claw (17), and a pressure sensor is provided on the side of the rubber pad (16) that contacts the fruit.

5. An intelligent picking hand based on an improved YOLOv5 model according to any one of claims 1 to 4, characterized in that: The steps to improve the YOLOv5 model construction are as follows: S1. Constructing a data set: Use a photographing device to photograph fruits under natural conditions and construct an original data set; S2. Data preprocessing: perform data augmentation on the original dataset, then label the fruits in the images, and finally randomly divide the labeled dataset into training set, validation set, and test set according to the proportion; S3. Build an improved YOLOv5 model: Use the improved ShuffleNetV2 to replace the original YOLOv5 backbone network, use GhostConv to replace the conventional convolution in the neck network, use the fused bidirectional feature pyramid to replace the connection method of the neck network, and introduce the EDCA attention mechanism; S4. Use the training set and validation set to train and validate the improved YOLOv5 model, save the optimal weights during the training process, and finally use the test set for testing.

6. The intelligent picking hand based on the improved YOLOv5 model according to claim 5, characterized in that: In the improved ShuffleNetV2, a residual connection is introduced in the basic feature extraction module of the ShuffleNetV2 model, and the maximum pooling layer is used in the residual connection of its downsampling module to complete the downsampling. The number of channels is adjusted by 1×1 convolution, and a dual-channel feature extraction module is introduced to replace the 3*3 depth convolution in the basic feature extraction module.

7. The intelligent picking hand based on the improved YOLOv5 model according to claim 5, characterized in that: In S1, the sample image is adjusted to a size of 640×640×3.

8. The intelligent picking hand based on the improved YOLOv5 model according to claim 5, characterized in that: In S2, the labeled dataset is randomly divided into training set, validation set and test set in the ratio of 8:1:

1.

9. The intelligent picking hand based on the improved YOLOv5 model according to claim 5, characterized in that: In S2, the data set is divided, and the initial anchor boxes of the labeled data set are reset using k-means clustering and genetic learning algorithms.

10. The intelligent picking hand based on the improved YOLOv5 model according to claim 5, characterized in that: The EDCA attention mechanism uses average pooling and standard deviation pooling to process the input features, compressing them into 1×1×C feature maps, where C is the number of channels in the feature map, and generates weights for each channel through one-dimensional convolution. The elements of the feature maps obtained by the two paths are added together, and the weights of each channel are obtained through the Sigmoid function. Finally, the weights are multiplied by the original input feature map to obtain the output feature map.

Images