Visual navigation easy-to-control intelligent weeder

Abstract

The present invention relates to an agricultural machine, namely a visual navigation easy-to-control intelligent weeding machine, comprising a visual navigation system and a traction frame, the rear of the traction frame is slidably connected to a sliding beam through a hydraulic device, and the sliding beam is connected to multiple weeding units. The characteristics are: there are two types of weeding units, one is an inter-row weeding unit, and the other is an inter-plant weeding unit. The inter-row weeding unit is installed on a fixed beam, and the fixed beam is rigidly connected to the traction frame. The inter-plant weeding unit is installed on the sliding beam, and the sliding beam can slide left and right under the drive of the hydraulic device. Beneficial effect: only the inter-plant weeding unit moves left and right, reducing the operating load and inertial impact. In addition, during operation, the wheel cutter of the support wheel in front of the machine and the knife wheel behind the machine cut into the ground, which enhances the stability of the machine, further offsets the inertia of the swinging parts, reduces energy consumption, improves the weeding effect, greatly reduces the difficulty and cost of intelligent control, and creates conditions for the popularization and application of intelligent weeding machines.

Description

Technical Field

[0001] The invention relates to an agricultural machine, namely a visual navigation easy-to-control intelligent weeding machine. Background Art

[0002] During the growth of crops such as corn, soil loosening and weeding are required. In the past, soil loosening and weeding were done by manpower, which was labor-intensive and inefficient. In recent years, a variety of weeding machines have been launched on the market to reduce the burden of manual labor. However, existing weeding machines can only remove weeds between crop rows, but not between crop plants. In order to completely remove weeds between plants, herbicides are currently used. However, herbicides do not have the effect of loosening the soil, and have certain side effects, which leads to a decrease in the grade of agricultural products. According to foreign data, the use of intelligent weeding machines can solve the problem of weeding between plants. However, advanced intelligent weeding machines are mostly based on technologies such as lasers and electric heating, which are expensive and difficult to popularize. Among them, intelligent weeding machines based on mechanical devices such as hydraulics have a lower cost and are more suitable for my country's national conditions. To this end, we introduced a relatively low-cost visual navigation intelligent weeding machine and improved it, see "A kind of intelligent weeding equipment and weeding method based on visual navigation (CN 117694087 A)". This visual navigation intelligent weeding machine is towed by a tractor, and multiple weeding units are mounted behind the tractor through a towing frame and other components. Each weeding unit is equipped with a plowshare for weeding between rows and a knife disc for weeding between rows. Cameras are installed at the left and right ends of the front of the weeding machine. During operation, the camera obtains the image in front and inputs it into the controller to determine the position of the grass seedlings in front. Then, the hydraulic device drives the weeding unit to swing left and right along the determined motion trajectory to remove the weeds between rows and plants. However, through experiments, it was found that this machine has the following problems: There are two main types of existing models. One is that the plowshare for weeding between rows and the knife disc for weeding between plants are installed on one unit. When working, the plowshare for weeding between rows and the knife disc for weeding between plants must be translated left and right together. The other is that although the plowshare for weeding between rows and the knife disc for weeding between plants are installed on two units, they are installed on the same sliding beam and still have to be translated left and right together. Since there are many trailer weeding machines and they are very heavy, moving back and forth horizontally will generate a lot of inertia, making it difficult to control and placing high demands on the control equipment. In particular, the plowshare and ground wheels of the inter-row weeding machine swing left and right with the whole machine, which is not only unnecessary, but also increases energy consumption, and increases the difficulty and cost of control. There are currently only two ways to solve the above problem: one is to slow down the speed, but this will greatly reduce the operating efficiency and further increase the operating cost. The second is to use more advanced intelligent control equipment, but it is expensive and difficult to popularize. Summary of the invention

[0003] The purpose of the present invention is to provide a visual navigation and easy-to-control intelligent weeding machine which can move stably under the drive of a tractor, can reduce or eliminate the swing inertia of the weeding unit, improve the movement accuracy of the weeding parts, and has a simple structure, low cost, energy saving and consumption reduction, and is easy to popularize and apply.

[0004] The above-mentioned purpose is achieved by the following technical solutions: providing a visual navigation easy-to-control intelligent weeding machine, including a visual navigation system and a traction frame, the visual navigation system including a camera, a voyage counter, an azimuth sensor, a driving state sensor and a controller, the output end of the camera is connected to the input end of the image processing module in the controller, the voyage counter, the azimuth sensor and the driving state sensor are all electrically connected to the controller, the controller outputs a control signal to the driver, the traction frame is provided with a hanging mechanism that can be engaged with a traction machine, the rear of the traction frame is slidably connected to a sliding beam through a hydraulic device, and the sliding beam hangs multiple weeding units, characterized in that: there are two types of weeding units, one is an inter-row weeding unit with a plowshare installed, and the other is an inter-row weeding unit with a knife disc installed, wherein the inter-row weeding unit is installed on a fixed beam, the fixed beam is rigidly connected to the traction frame, and the inter-row weeding unit is installed on a sliding beam, and the sliding beam can slide left and right under the drive of the hydraulic device.

[0005] The hydraulic device includes a translation cylinder and a single cylinder. The translation cylinder is transversely installed on the rear side of the traction frame and drives the sliding beam to move left and right under the control of the controller. The single cylinder is installed in front of the weeding unit between rows to control the rise and fall of the weeding unit and the pressure on the ground.

[0006] The sliding beam has multiple sections, which are respectively installed on the fixed beam and slidably cooperate with the fixed beam. Each section of the sliding beam is hung with a set of inter-row weeding units, and each section of the sliding beam slides left and right along the fixed beam driven by the translation cylinder.

[0007] A supporting wheel is respectively arranged on the left and right sides of the traction frame, and a protruding annular wheel cutter is arranged on the outer periphery of the supporting wheel.

[0008] The supporting wheel is spherical.

[0009] The traction frame is connected to a fixed beam installed in the heading direction through a longitudinal beam extending backward.

[0010] There are one or more fixed beams distributed front to back, and at least two or more disc-shaped cutter wheels are arranged under the rear fixed beam.

[0011] Under the fixed beam, two vertically installed limit clamps are provided relative to each inter-row weeding unit, and the tail end of the inter-row weeding unit extends between the two limit clamps.

[0012] The limit clamping plate is provided with a length-adjusting screw rod, and the inner end of the length-adjusting screw rod is provided with a buffer pad made of elastic material.

[0013] An inter-row plow weeding unit is installed below the fixed beam connected to the middle or rear end of the longitudinal beam.

[0014] The visual navigation system includes identifying weeds in the total area actually covered by the target weeder through a visual recognition algorithm; controlling the inter-row weeding units to remove various weeds in the area, and generating a corresponding removal completion signal after the removal is completed; transmitting the removal completion signal to the visual navigation system so that the visual recognition system performs a second recognition on the sub-area to be weeded and generates a corresponding removal completion confirmation signal; in response to the removal completion confirmation signal, controlling the target weeding device to move to the next sub-area to be weeded, and returning to the step of identifying the positions of all weeds in one of the sub-areas to be weeded in the total area actually covered by the target weeder through a visual recognition algorithm, until the weeds in the total area to be weeded are removed.

[0015] The steps of identifying weeds using the visual recognition algorithm are as follows: obtaining a plurality of first videos, wherein the first videos are videos of weeds in the field, and the first videos carry weed species information; obtaining a plurality of first picture sets based on the plurality of first videos; the first picture sets include a plurality of first pictures, and the first pictures carry weed species information of the first videos; searching for an optimal deep residual network structure to obtain a first deep residual network structure; training the deep residual network based on the plurality of first picture sets and the weed species information to obtain a rice field weed recognition model, wherein the deep residual network has the first deep residual network structure; obtaining a second picture; and inputting the second picture into the rice field weed recognition model to obtain a first recognition result; The deep residual network structure includes an input part, an intermediate part and an output part; the intermediate part is composed of 4 residual blocks, each residual block is composed of multiple convolution layers and bypass connections, each convolution layer is connected in sequence by BatchNorm batch normalization, ReLU activation function and Conv2D convolution, and the output part is in a multi-output form, each output is for a weed category, and the existence of multiple weeds can be identified and judged at the same time; the structure of the intermediate part is determined by finding the best deep residual network structure from the ENAS search space, and the content of the ENAS search space includes: the number of convolution layers contained in each residual block, the number range is 15; the step size of the first convolution layer in each residual block is 2, and the step size of the remaining convolution layers is 1; the number of channels of the convolution layer in each residual block, the number of channels of the first layer ranges from 16, 32, 64, 128, 256, 512, if the number of layers in the current residual block exceeds 2, the number of channels of the last convolution layer is 4 times that of the first convolution layer; the number of channels of the remaining layers is the same as that of the first layer; the convolution kernel size of each convolution layer is selected from one of the three sizes of 1x1, 3x3 and 5x5; the size of the ENAS search space is (5*6*3)^4, and the rules of the above search space are encoded as the search space of AutoKears. AutoKeras selects the one with the highest recognition rate by automatically training and comparing the deep residual network structure; during the search process, the loss function is calculated in such a way that each output corresponds to a cross entropy error, and the average value of the cross entropy errors of all outputs is calculated as the loss of the entire network; the obtaining of multiple first picture sets according to multiple first videos includes: obtaining a third picture set according to each of the first videos, the third picture set being a picture set composed of all picture frames of the first video according to a time series; constructing the first picture set;Obtain a first picture from the third picture set, detect whether the first picture is blurred, and detect whether the first picture is similar to existing pictures in the first picture set; if the first picture is not blurred and is not similar to any existing pictures in the first picture set, use the first picture as the first target picture, and obtain a picture with a first preset interval frame number as the next first picture; if the first picture is blurred or the first picture is similar to any existing picture in the first picture set, obtain a picture with a second preset interval frame number as the next first picture; store the first target picture in the first picture set; detecting whether the first picture is blurred includes: calculating a gradient matrix of the first picture, calculating the variance of all elements of the gradient matrix, and if the variance is less than a predetermined variance threshold, determining that the first picture is blurred; if the variance is greater than or equal to a predetermined variance threshold, determining that the first picture is not blurred; detecting whether the first picture is similar to existing pictures in the first picture set includes: compressing the size of the first picture and all pictures in the first picture set to a first size using bilinear interpolation; converting all compressed pictures into grayscale pictures using the super green method; Calculate the mean of all pixels of the grayscale image, perform binarization on the grayscale image, record pixels greater than the mean as 1, and record pixels less than the mean as 0, to obtain a matrix containing only 0 or 1; calculate the absolute value of the difference between the matrix corresponding to the first image and the matrix corresponding to each image in the first image set, to obtain a difference matrix containing only 0 and 1; calculate the sum of all elements of the difference matrix, when the sum of all elements of the difference matrix is ​​less than a preset difference threshold, it is determined that the first image and the corresponding image in the first image set are similar, when the sum of all elements of the difference matrix is ​​greater than or equal to the preset difference threshold, it is determined that the first image and the corresponding image in the first image set are not similar. ;

[0016] The beneficial effects of the present invention are: the weeding unit between rows does not move left and right, only the weeding unit between plants moves left and right, reducing the operating load and inertial impact, and achieving the purpose of clearing weeds between rows and plants. In addition, during operation, the wheel cutter of the support wheel in front of the machine and the knife wheel behind the machine cut into the ground, limiting the deviation of the machine, enhancing the stability of the machine, further offsetting the inertia of the swinging part, reducing energy consumption, improving the flexibility and accuracy of the swinging part, greatly reducing the difficulty and cost of intelligent control, and creating conditions for the popularization and application of intelligent weeders. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a front view of a first embodiment;

[0018] Figure 2 is a top view of the first embodiment;

[0019] Figure 3It is a right side view of the first embodiment;

[0020] Figure 4 is a control system structure block diagram of the first embodiment;

[0021] Figure 5 It is a right side view of the inter-row weeding unit of the first embodiment;

[0022] Figure 6 It is a right side view of the inter-row weeding unit of the first embodiment;

[0023] Figure 7 is a structural block diagram of the first embodiment;

[0024] Figure 8 is a structural block diagram of a second embodiment;

[0025] Fig. 9 is a partial top view of the second embodiment;

[0026] Fig.10 is a front view of the assembly of the traction frame and the support wheel of the third embodiment;

[0027] Fig.11 is an enlarged perspective view of a component support wheel of the third embodiment;

[0028] Fig.12 is a schematic diagram of the use state of the third embodiment;

[0029] Fig.13 is an assembly front view of a spherical support wheel component of a fourth embodiment;

[0030] Fig.14 is a schematic diagram of the use state of the fourth embodiment;

[0031] Fig.15 is a schematic top view of the machine structure of the fifth embodiment;

[0032] Fig.16 is a schematic top view of another machine mechanism of the fifth embodiment;

[0033] Fig.17 is a perspective view of a tool frame of a sixth embodiment;

[0034] Fig.18 is a schematic diagram of a working state of a component limiting clamp of the sixth embodiment;

[0035] Fig.19 is a schematic diagram of another working state of the component limiting clamp of the sixth embodiment;

[0036] Fig. 20 is a flow chart of the intelligent weed identification method of the seventh embodiment;

[0037] Fig.21 is a flow chart of obtaining a first picture set by the intelligent weed identification method according to the seventh embodiment;

[0038] Fig. 22 Schematic diagram of the structure of the deep residual network of the seventh embodiment.

[0039] As can be seen in the figure: traction frame 1, support wheel 2, sliding beam 3, camera 4, fixed beam 5, inter-row weeding unit 6, inter-plant weeding unit 7, longitudinal beam 8, ground wheel 9, plow beam 10, plowshare 11, single cylinder 12, cutter disc 13, translation cylinder 14, wheel knife 15, scraper brush 16, ridge platform 17, ridge ditch 18, crops 19, knife wheel 20, limit splint 21, lengthening screw 22, buffer pad 23, hanging line 24. DETAILED DESCRIPTION

[0040] The general concept of the present invention is to reduce the number of swinging parts, reduce the swinging inertia, and reduce the difficulty of intelligent control, so as to achieve the purpose of removing weeds between rows and plants. The following are seven embodiments based on this concept:

[0041] First embodiment: Figure 1 , Figure 2 , Figure 3 A visual navigation and easy-to-control intelligent weeder is introduced, comprising a visual navigation system and a traction frame 1 that can be coupled to a tractor.

[0042] Combination Figure 4 It can be seen that the visual navigation system includes a camera, a range counter, an orientation sensor, a drive state sensor and a controller. The output end of the camera is connected to the input end of the image processing module in the controller. The range counter, the orientation sensor and the drive state sensor are all electrically connected to the controller. The camera is used to collect video information of weeds in the field, the range counter is used to calculate the range of the weeding equipment, the orientation sensor is used to monitor the speed and direction of the weeding equipment, and the drive state sensor is used to collect the action parameters of the translation cylinder and the single cylinder. The controller outputs a control signal to the driver, and the driver drives the cutter motor to control the rotation of the cutter to perform weeding operations between rows.

[0043] The lower part of the left and right sides of the traction frame 1 is provided with a support wheel 2. The box behind the traction frame is provided with sensors of the intelligent control system and hydraulic devices and other components. Figure 7It can be seen that a slideway is provided behind the traction frame, and another slider slides on the slideway, and two translation cylinders 14 are oppositely installed on both sides of the slider. Of course, a double-acting cylinder can also be used, which is installed on one side of the slider. The translation cylinder, oil pipe and oil pump are not drawn in detail because they belong to the prior art. The oil pump drives the piston rod of the translation cylinder to perform telescopic action under the control of the solenoid valve and in accordance with the instructions of the intelligent control system, pushing the slider to translate left and right. The slider drives the sliding beam 3 to move left and right. Multiple weeding units are installed on the sliding beam 3. The improvement is that: there are two types of weeding units, one is the inter-row weeding unit 6 with the plowshare 11 installed, and the other is the inter-row weeding unit 7 with the knife disc 13 installed. Among them, the front end of the inter-row weeding unit is installed on the fixed beam 5, and the fixed beam is rigidly connected to the traction frame. The front end of the inter-row weeding unit is installed on the sliding beam 3 and translates left and right with the sliding beam.

[0044] The sliding beam in the figure has a complex structure, so multiple weeding units are directly installed on a sub-beam. Since the sub-beam and the sliding beam are rigidly connected as a whole, they are equivalent to a part of the sliding beam. Therefore, the weeding units connected to the sub-beam are equivalent to being installed on the sliding beam.

[0045] The sliding beam in the figure can also be extended to both sides to attach more weeding units. When the working width is large, the sliding beam can be divided into three sections, each of which is hinged with a rotating shaft. A hydraulic device can be provided to control the rise and fall of the sliding beams at both ends and the attached weeding units to adjust the width of the machine to meet the needs of farmlands or roads of different widths.

[0046] As can be seen in the figure, a camera 4 is installed at each end of the sliding beam through a high bracket. During operation, the camera obtains the image in front and transmits it to the intelligent control system, which determines the cutting trajectory of the weeding component after processing, and then drives the working component through the hydraulic device to complete the weeding action.

[0047] Combination Figure 5 , Figure 7It can be seen that the inter-row weeding unit 6 has a plow beam 10 extending forward and backward, and several plowshares 11 of different shapes are installed under the plow beam. Among them, there are plowshares that can dig furrows and plowshares that can shovel grass on both sides. The rear plowshare can build up soil for the roots of crops and repair the shape of the ridge. The front end of the plow beam is hinged to the fixed beam, and the plow beam is allowed to float up and down around the fixed beam to a certain extent through the hinge, so that the machine has good profiling performance. But the plow beam is not allowed to swing left and right. In order to better support the weight of the inter-row weeding unit, a ground wheel 9 is installed at the front and lower part of the plow beam. Obviously, the inter-row weeding unit is installed on the fixed beam and can only move forward with the machine, but cannot move left and right. The ground wheel and plowshare installed on the inter-row weeding unit do not need to move left and right, which will inevitably save a lot of power. The ground pressure of the inter-row weeding unit can be controlled by the deadweight of the machine, or it can be controlled by installing a single oil cylinder 12. For crops with straight rows such as corn, this method can completely remove the weeds between the rows.

[0048] Combination Figure 6 , Figure 7 It can be seen that the inter-row weeding unit 7 also has a plow beam 10, the front end of which is hinged on the sliding beam 3 and can float up and down relative to the sliding beam, but cannot swing left and right relative to the sliding beam. Since the plow beam is opposite to the crop plants in the vertical direction during operation, it is not suitable to install a ground wheel. The pressure of the plow beam on the ground can be controlled by the single oil cylinder 12 of the hydraulic device. At least one set of knife discs 13 suitable for inter-row weeding are installed on both sides below the plow beam 10. This knife disc is in the shape of a disc, with a micro motor in the middle, and multiple finger-shaped convex teeth are evenly distributed around, and the convex teeth are wrapped with soft materials. The knife disc and the plow beam are connected by elastic curved slats, so it has good profiling performance. During operation, the entire knife disc is at an angle of 25-40° with the ground, and can be extended into the inter-row, digging up the weeds and throwing them out. Since the root system of the crop plants is deeper than the weeds at this time, it will not cause damage even if it touches the plants. In particular, after adopting the above measures, weeding between rows and between plants is separated, the power required for weeding between plants is small and easy to control, the movement of the blade is more flexible and accurate, and the weeds between plants can be better removed. The damage of the blade to crop plants is also greatly reduced, reaching the leading level of similar models.

[0049] Second embodiment: A new structure is proposed based on the first embodiment. Figure 8 , Fig. 9 As shown, the sliding beam 3 has multiple sections, which are respectively mounted on the fixed beam 5 and slide with the fixed beam. Each section of the sliding beam is equipped with a set of weeding units 7 between rows, and each section of the sliding beam has a set of translation cylinders 14.

[0050] During operation, each translation cylinder 14 pushes each section of the sliding beam 3 to slide on the fixed beam 5 according to the instruction of the intelligent control system, thereby driving each inter-row weeding unit to complete the action of clearing the weeds between the rows by swinging left and right.

[0051] Obviously, this method can process images of multiple rows of crops separately, and each weeding unit between plants can perform weeding actions separately, so the weeding accuracy is higher and the damage to the plants is less. It is a more optimized technical solution.

[0052] The third embodiment is improved on the basis of the above embodiment. In order to improve efficiency, most of these implements have a larger working width, and generally have to tow more than a dozen weeding units to handle 10-20 rows of weeding at a time. The swinging of so many weeding units together has a great impact on the stability of the whole machine. Therefore, the tractor for towing operation is required to have sufficient tonnage and stability. However, the larger the tractor, the higher its cost and the more difficult it is to use. It can be seen that in order to reduce costs, it is necessary to improve the stability of the weeding implement without increasing the specifications of the tractor. Fig.10 A structure for improving the stability of a machine is introduced. As shown in the figure, a support wheel 2 is provided on each side of a traction frame 1, and a protruding annular wheel cutter 15 is provided on the periphery of the support wheel 2.

[0053] Combination Fig.11 It can be seen that the wheel cutter 15 is a hard circular ring and can be made of wear-resistant metal plate. In order to enhance the performance of entering the soil, the thickness of the wheel cutter cannot be too thick, generally ≤20mm, and the outer edge has a sharp edge formed by a bevel. Since the support wheel has the function of limiting the depth, the part where the outer periphery of the wheel disc contacts the ground, that is, the so-called "wheel surface", is very wide. Its width is ≥80mm. In fields with soft soil, in order to prevent sinking, the width of the wheel surface can reach more than 600mm. The wheel cutter is installed in the middle of the wheel surface, and the sharp edge protrudes outward. The difference between the outer diameter of the wheel cutter and the outer diameter of the wheel disc is ≥50mm. Depending on soil conditions, this value can be increased to more than 150mm.

[0054] When working, if Fig.12 As shown in the figure, the wheel cutter on the support wheel is inserted into the soil layer, and the squeezing of the soil sections on both sides limits the left and right deviation of the machine, offsetting the swing inertia of the sliding beam and the weeding unit between the plants. This support wheel can neither hinder the movement of the machine nor improve the stability of the traction frame, greatly reducing the difficulty of intelligent control. The running track of the plowshare and the cutter disc is more accurate, the weed removal rate in the field is greatly improved, and the damage rate of the plants is greatly reduced, producing unexpected results.

[0055] Fourth embodiment: Based on the above embodiment, another structure of the support wheel is proposed. Fig.13As shown, the support wheel 2 here is spherical. What is listed in the figure is a kind of ellipsoid, and the outer midline of the sphere is equipped with a circular ring-shaped wheel cutter 15. This support wheel is particularly suitable for the upper ridge planting method.

[0056] Combination Fig.14 It can be seen that the upper ridge is the crops 19 growing on the ridge platform 17, and the lower ridge ditch 18 is between the rows of crops. This method is often used in depressions with high groundwater levels. In arid farmland with high terrain and low groundwater levels, seeds are mostly placed in the ridge ditch during sowing. When the crops grow to a certain height, the original ridge platform is broken, and soil is piled up at the roots of the crops, and ridges will also form between the rows of crops. When weeding, the crops are on the ridge platform, and the support wheels have to walk in the ridge ditch. The support wheel is made into a spherical shape, especially an ellipsoidal shape, which coincides with the shape of the ridge ditch when walking, and has better stability. The wheel knife on the spherical support wheel is inserted into the soil layer on the center line of the ridge ditch when walking, which is more stable. The difficulty of intelligent control is lower, the composition of the intelligent system will be simpler, and the cost will be lower, which is more conducive to the popularization and application of intelligent weeders.

[0057] also, Fig.13 As can be seen in the figure, scraping brushes 16 are installed on both sides of the wheel cutter. The scraping brushes can be wire brushes as shown in the figure, which are installed on the outer side of the supporting wheel mounting arm through a brush holder. During operation, the scraping brushes can remove the dirt and debris stuck on the wheel cutter to maintain the soil penetration performance of the wheel cutter. At the same time, the scraping brushes can also be made of high-hardness blades. While removing debris, the sharp edge of the wheel cutter is sharpened to maintain the shearing performance of the blade.

[0058] The fifth embodiment: based on the above embodiments, the stability of the machine is further improved. Fig.15 As shown, the traction frame 1 is connected to the laterally installed fixed beam 5 through the longitudinal beam 8 extending backward. There can be multiple fixed beams, distributed front and back. At least two disc-shaped cutter wheels 20 are installed under the rear fixed beam. This cutter wheel is different from the ground wheel. It is a thin sheet-like circular plate, which has a strong cutting performance and will not flatten the surface of the ground that has just been mowed. The two cutter wheels are preferably installed at both ends of the fixed beam. When the working width is large, the cutter wheel can also be installed in the middle. When working, the cutter wheel is inserted into the soil layer between the rows of crops by gravity and rotates forward with the implement. In this way, the soil can be further loosened and it can cooperate with the wheel cutters of the two front support wheels, which is equivalent to providing a positioning and anti-deviation mechanism at the four corners of the implement, and the positioning effect is better.

[0059] Furthermore, the fixed beam in the figure can also be equipped with an inter-row weeding unit. Fig.15 As shown, the inter-row weeding unit 6 is installed under the fixed beam 5 in the middle of the machine. The weeds between the rows of crops are removed by the inter-row plowshare 11. The inter-row weeding unit is still installed on the sliding beam 3 and slides left and right with the sliding beam.

[0060] Further. Combine Fig.16 It can be seen that the inter-row plowshare 11 is directly installed below the fixed beam 5 through a section of column. In this way, the plow beam of the inter-row weeding unit can be omitted, and the structure is more simplified.

[0061] The sixth embodiment is improved on the basis of the above embodiment. Fig.17 It can be seen that under the fixed beam 5 behind the machine, two vertically installed limit clamps 21 are provided for each inter-row weeding unit 7, and the tail end of the inter-row weeding unit 7 extends between the two limit clamps. When the inter-row weeding unit swings to a specified limit, it will be blocked by the limit clamp and stop moving. Its movement inertia will also be absorbed by the limit clamp, so that some functions of the intelligent control system can be omitted, simplifying the complexity of the system and reducing the cost.

[0062] Further, combined with Fig.18 It can be seen that the two limit clamps 21 under the fixed beam are respectively equipped with length adjustment screws 22, and the inner ends of the length adjustment screws are equipped with buffer pads 23 made of elastic material. By turning the length adjustment screws, the distance between the two buffer pads can be adjusted, and the amplitude of the left and right swing of the weeding unit 7 between rows can be determined according to the operation needs. In addition, the buffer pad absorbs the inertia more gently and stably, and is conducive to the return of the weeding unit, and the difficulty of intelligent control is further reduced.

[0063] Further, if Fig.19 As shown, a length-adjustable soft suspension wire 24 is provided below the fixed beam 5 relative to each inter-row weeding unit 7. The suspension wire may have a spring. By adjusting the length of the suspension wire as needed, a suitable pulling force can be applied to the inter-row weeding unit to keep the blade disc below at an appropriate tillage depth.

[0064] This suspension wire can also be installed under the fixed beam in the middle of the machine. Since the plow beam of the inter-row weeding unit is placed above the crops, the ground wheel cannot be installed under the plow beam, and the weight of the inter-row weeding unit lacks strong support, which increases the load on the sliding beam. With this suspension wire, the original ground wheel can be replaced to give appropriate traction to the inter-row weeding unit, further reducing the load on the related sliding beam and optimizing the machine's operating performance.

[0065] The seventh embodiment: Based on the above embodiments, a method of intelligent control is provided. Fig. 20 , Fig.21 , Fig. 22 It can be seen that the intelligent weeding method includes:

[0066] Identify the weeds in the total area actually covered by the target weeder through a visual recognition algorithm;

[0067] Controlling the inter-row weeding units and the inter-plant weeding units to remove various weeds in the area, and generating a corresponding removal completion signal after the removal is completed;

[0068] Transmitting the clearing completion signal to the visual navigation system so that the visual recognition system can perform a second recognition on the sub-area to be weeded and generate a corresponding clearing completion confirmation signal;

[0069] In response to the clearing completion confirmation signal, the target weeding device is controlled to move to the next sub-area to be weeded, and the step of identifying the positions of all weeds in one of the sub-areas to be weeded in the total area actually covered by the target weeder is returned to, until the weeds in the total area to be weeded are cleared.

[0070] The intelligent weeding method based on visual recognition and motion control can be implemented as a computer program, which is tangibly contained in a computer-readable storage medium, and part or all of the computer program can be loaded and / or installed on an electronic device via a ROM and / or a communication unit. When the computer program is loaded into the RAM and executed by the processor, one or more steps of the intelligent weeding method based on visual recognition and motion control described above can be performed.

[0071] Alternatively, in other embodiments, the processor may be configured in any other appropriate manner to execute an intelligent weeding method based on visual recognition and motion control.

[0072] It includes: identifying the coordinate positions of all weeds in one of the sub-areas to be weeded in the total area actually covered by the target weeder by using a visual recognition algorithm;

[0073] The sub-area to be weeded is the area covered by the target weeder at one time; the manipulator is controlled to remove the weeds corresponding to the coordinate positions of all weeds, and a corresponding removal completion signal is generated after the removal is completed;

[0074] The cleaning completion signal is transmitted to the visual recognition system so that the visual recognition system can perform a secondary recognition on the sub-area to be weeded and generate a corresponding cleaning completion confirmation signal;

[0075] In response to the clearing completion confirmation signal, the target weeder is controlled to move to the next sub-area to be weeded, and the step of identifying the coordinate positions of all weeds in one of the sub-areas to be weeded in the total area actually covered by the target weeder is returned to, until the weeds in the total area to be weeded are cleared.

[0076] Further, combined with Fig. 20 , Fig.21 , Fig. 22 It can be seen that the steps of visual recognition algorithm to identify weeds are:

[0077] Acquire multiple first videos, where the first videos are videos of weeds in the field and carry weed species information;

[0078] Acquire multiple first picture sets according to the multiple first videos; the first picture sets include multiple first pictures, and the first pictures carry the weed species information of the first videos;

[0079] Searching for the best deep residual network structure to obtain a first deep residual network structure;

[0080] Training the deep residual network according to the first picture sets and the weed species information to obtain a weed recognition model, wherein the deep residual network has the first deep residual network structure;

[0081] Get the second picture;

[0082] Inputting the second image into a weed recognition model to obtain a first recognition result;

[0083] The deep residual network structure includes an input part, an intermediate part and an output part;

[0084] The middle part is composed of 4 residual blocks, each residual block is composed of multiple convolution layers and bypass connections, each convolution layer is composed of BatchNorm batch normalization, ReLU activation function and Conv2D convolution connected in sequence, and the output part is in a multi-output form, each output is for a weed category, and the presence of multiple weeds can be identified and judged at the same time.

[0085] The structure of the middle part is determined by searching for the best deep residual network structure from the ENAS search space, and the content of the ENAS search space includes:

[0086] The number of convolutional layers contained in each residual block ranges from 1 to 5;

[0087] The stride of the first convolution layer in each residual block is 2, and the stride of the remaining convolution layers is 1;

[0088] The number of channels of the convolution layer in each residual block is selected from 16, 32, 64, 128, 256, and 512. If the number of layers in the current residual block exceeds 2, the number of channels of the last convolution layer is 4 times that of the first convolution layer; the number of channels of the remaining layers is the same as that of the first layer;

[0089] The convolution kernel size of each convolution layer is selected from 1x1, 3x3 and 5x5.

[0090] The size of the ENAS search space is (5*6*3)^4. The rules of the above search space are encoded into the search space of AutoKeras. AutoKeras automatically trains and compares deep residual network structures to select the one with the highest recognition rate.

[0091] During the search process, the loss function is calculated as follows: each output corresponds to a cross entropy error, and the average of the cross entropy errors of all outputs is calculated as the loss of the entire network;

[0092] The acquiring a plurality of first picture sets according to the plurality of first videos comprises:

[0093] Acquire a third picture set according to each of the first videos, where the third picture set is a picture set composed of all picture frames of the first video according to a time sequence;

[0094] constructing the first picture set;

[0095] Obtain a first picture from the third picture set, detect whether the first picture is blurred, and detect whether the first picture is similar to existing pictures in the first picture set; if the first picture is not blurred and is not similar to any existing pictures in the first picture set, use the first picture as the first target picture, and obtain a picture with a first preset interval frame number as the next first picture; if the first picture is blurred or the first picture is similar to any existing picture in the first picture set, obtain a picture with a second preset interval frame number as the next first picture;

[0096] storing the first target picture in the first picture set;

[0097] The detecting whether the first image is blurred comprises: calculating a gradient matrix of the first image, calculating the variance of all elements of the gradient matrix, and if the variance is less than a predetermined variance threshold, determining that the first image is blurred; if the variance is greater than or equal to the predetermined variance threshold, determining that the first image is not blurred;

[0098] Detecting whether the first picture is similar to existing pictures in the first picture set includes:

[0099] Compressing the first image and all images in the first image set to a first size using bilinear interpolation; converting all compressed images into grayscale images using a super green method;

[0100] Calculate the mean of all pixels of the grayscale image, perform binarization on the grayscale image, record pixels greater than the mean as 1, and record pixels less than the mean as 0, to obtain a matrix containing only 0 or 1; calculate the absolute value of the difference between the matrix corresponding to the first image and the matrix corresponding to each image in the first image set, to obtain a difference matrix containing only 0 and 1;

[0101] The sum of all elements of the difference matrix is ​​calculated. When the sum of all elements of the difference matrix is ​​less than a preset difference threshold, it is judged that the first image and the corresponding image in the first image set are similar. When the sum of all elements of the difference matrix is ​​greater than or equal to the preset difference threshold, it is judged that the first image and the corresponding image in the first image set are not similar.

[0102] During the movement of the weeder, it is necessary to monitor the field in front of the weeder in real time. The camera in the visual navigation mechanism is used to collect weed videos in front of the weeder, identify the type of weeds in the field, and use the artificial intelligence recognition model of the deep residual network to identify the input image to obtain the recognition result. The weeds can be identified intelligently and efficiently, so that the control unit can adjust the moving weeder in real time. The hydraulic system controls the inter-row single cylinder and the inter-plant single cylinder to accurately control the rise and fall of each inter-row weeding unit and inter-plant weeding unit, thereby improving the maneuverability of the weeder.

Claims

1. A visual navigation and easy-to-control intelligent weeding machine, comprising a visual navigation system and a traction frame (1), wherein the visual navigation system comprises a camera, a voyage counter, an orientation sensor, a driving state sensor and a controller, wherein the output end of the camera is connected to the input end of an image processing module in the controller, the voyage counter, the orientation sensor and the driving state sensor are all electrically connected to the controller, and the controller outputs a control signal to the driver, wherein the traction frame is provided with a hanging mechanism that can be engaged with a traction machine, wherein the rear of the traction frame is slidably connected to a sliding beam (3) through a hydraulic device, and the sliding beam hangs a plurality of weeding units, and wherein: There are two types of weeding units, one is an inter-row weeding unit (6) equipped with a plowshare (11), and the other is an inter-row weeding unit (7) equipped with a knife disc (13), wherein the inter-row weeding unit (6) is installed on a fixed beam (5), which is rigidly connected to a traction frame, and the inter-row weeding unit (7) is installed on a sliding beam, which can slide left and right under the drive of a hydraulic device; The traction frame (1) is provided with a support wheel (2) on each of its left and right sides, and a protruding annular wheel cutter (15) is provided on the outer periphery of the support wheel; The traction frame (1) is connected to a transversely mounted fixed beam (5) via a longitudinal beam (8) extending backwards; The fixed beams (5) are provided in one or more sections and are distributed front to back, and at least two or more disc-shaped cutter wheels (20) are provided under the rear fixed beam (5); Two vertically installed limit clamps (21) are provided below the fixed beam (5) relative to each inter-row weeding unit (7), and the tail end of the inter-row weeding unit (7) extends between the two limit clamps (21); The limit clamping plate (21) is provided with a length-adjusting screw rod (22), and the inner end of the length-adjusting screw rod (22) is provided with a buffer pad (23) made of elastic material.

2. The visual navigation easy-to-control intelligent weeder according to claim 1, characterized in that: The hydraulic device comprises a translation cylinder (14) and a single cylinder (12); the translation cylinder is transversely mounted on the rear side of the traction frame (1) and drives the sliding beam (3) to move left and right under the control of a controller; the single cylinder (12) is mounted on the front of the weeding unit between rows and controls the rise and fall of the weeding unit and the pressure on the ground.

3. The visual navigation easy-to-control intelligent weeder according to claim 1 is characterized in that: The sliding beam (3) has multiple sections, which are respectively mounted on the fixed beam (5) and slidably cooperate with the fixed beam (5). Each section of the sliding beam (3) is connected to a set of inter-row weeding units (7). Each section of the sliding beam slides left and right along the fixed beam (5) under the drive of the translation oil cylinder (14).

4. The visual navigation easy-to-control intelligent weeder according to claim 1, characterized in that: The supporting wheel (2) is spherical.

5. The visual navigation easy-to-control intelligent weeder according to claim 1 is characterized by: An inter-row weeding unit (6) is installed below the fixed beam (5) connected to the middle or rear end of the longitudinal beam (8).

6. The visual navigation easy-to-control intelligent weeder according to claim 1, characterized in that: The method comprises a visual navigation system, which identifies weeds in a total area actually covered by a target weeder through a visual recognition algorithm; controls an inter-row weeding unit to remove various weeds in the area, and generates a corresponding removal completion signal after the removal is completed; transmits the removal completion signal to the visual navigation system, so that the visual recognition system performs a second recognition on the sub-area to be weeded, and generates a corresponding removal completion confirmation signal; in response to the removal completion confirmation signal, controls the target weeding device to move to the next sub-area to be weeded, and returns to the step of identifying the positions of all weeds in one of the sub-areas to be weeded in the total area actually covered by the target weeder through a visual recognition algorithm, until the weeds in the total area to be weeded are removed.

7. The visual navigation easy-to-control intelligent weeder according to claim 6, characterized in that: The steps of identifying weeds by the visual recognition algorithm are: obtaining a plurality of first videos, wherein the first video is a video of weeds in the field, and the first video carries weed species information; obtaining a plurality of first picture sets according to the plurality of first videos; the first picture set includes a plurality of first pictures, and the first pictures carry weed species information of the first video; searching for the best deep residual network structure to obtain a first deep residual network structure; training a deep residual network according to the plurality of first picture sets and the weed species information to obtain a rice field weed recognition model, wherein the deep residual network has a first deep residual network structure; Get the second picture; Inputting the second image into the paddy field weed recognition model to obtain a first recognition result; The deep residual network structure includes an input part, an intermediate part and an output part; The middle part consists of 4 residual blocks, each of which consists of multiple convolutional layers and bypass connections. Each convolutional layer is composed of BatchNorm batch normalization, ReLU activation function and Conv2D convolution connected in sequence. The output part is in multi-output form, each output is for a weed category, and the presence of multiple weeds can be identified at one time. The structure of the middle part is determined by finding the best deep residual network structure from the ENAS search space. The content of the ENAS search space includes: The number of convolutional layers contained in each residual block, the number range is 15; The stride of the first convolutional layer in each residual block is 2, and the stride of the remaining convolutional layers is 1; The number of channels of the convolutional layer in each residual block, the number of channels of the first layer is selected from 16, 32, 64, 128, 256, 512, if the layer in the current residual block If the number of convolution layers exceeds 2, the number of channels of the last convolution layer is 4 times that of the first convolution layer; the number of channels of the remaining layers is the same as that of the first layer; the convolution kernel size of each convolution layer is selected from 1x1, 3x3 and 5x5; the size of the ENAS search space is (5*6*3)^4, and the rules of the above search space are encoded into the search space of AutoKears. AutoKeras automatically trains and compares the deep residual network structure to select the one with the highest recognition rate; In the search process, the loss function is calculated in such a way that each output corresponds to a cross entropy error, and the average value of the cross entropy errors of all outputs is calculated as the loss of the entire network; the obtaining of multiple first picture sets according to multiple first videos includes: obtaining a third picture set according to each first video, the third picture set being a picture set composed of all picture frames of the first video according to a time series; constructing a first picture set; obtaining a first picture from the third picture set, detecting whether the first picture is blurred, and detecting whether the first picture is similar to existing pictures in the first picture set; if the first picture is not blurred and is not similar to any existing picture in the first picture set, taking the first picture as the first target picture, and obtaining a picture with a first preset interval frame number as the next first picture; if the first picture is blurred or the first picture is similar to any existing picture in the first picture set, obtaining a picture with a second preset interval frame number as the next first picture; storing the first target picture in the first picture set; detecting whether the first picture is blurred includes: calculating a gradient matrix of the first picture, calculating the variance of all elements of the gradient matrix, and if the variance is less than a predetermined variance threshold, judging that the first picture is blurred;If the variance is greater than or equal to a predetermined variance threshold, the first image is judged to be unblurred; Detecting whether the first image is similar to existing images in the first image set includes: Compressing the size of the first image and all images in the first image set to the first size using bilinear interpolation; Converting all compressed images into grayscale images using the super green method; Calculating the mean of all pixels in the grayscale image, binarizing the grayscale image, recording pixels greater than the mean as 1, and recording pixels less than the mean as 0, to obtain a matrix containing only 0 or 1; Calculating the absolute value of the difference between the matrix corresponding to the first image and the matrix corresponding to each image in the first image set, to obtain a difference matrix containing only 0 and 1; Calculating the sum of all elements of the difference matrix, when the sum of all elements of the difference matrix is ​​less than a preset difference threshold, it is judged that the first image and the corresponding image in the first image set are similar, and when the sum of all elements of the difference matrix is ​​greater than or equal to the preset difference threshold, it is judged that the first image and the corresponding image in the first image set are not similar. ;

Images