A machine vision-based water spraying head water volume automatic adjusting system and method for a watering vehicle

Abstract

The application belongs to the technical field of intelligent watering cart spraying system, and discloses a watering cart nozzle water quantity automatic adjusting system and method based on machine vision, which comprises an image acquisition device, an image processing device and a watering control device. The watering cart nozzle water quantity automatic adjusting method is as follows: first, collecting road images to obtain the road width according to the images; then, putting the images into a neural network for training to obtain the target category, speed and distance; and finally, realizing the automatic adjustment of the watering cart nozzle water quantity by controlling the water pump pressure and the start-stop time of the water pump of the watering cart. The application can effectively realize water saving by detecting the road edge center point to determine the road distance and thereby controlling the watering quantity of the three-bend front-flush nozzle. The application can realize the speed and distance detection of pedestrians and non-motor vehicles by combining the neural network with the binocular camera distance measuring module, thereby adjusting and controlling the watering quantity of the three-bend front-flush nozzle and realizing the real-time avoidance of pedestrians and non-motor vehicles.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent watering truck spraying systems, in particular to a watering truck nozzle water quantity automatic adjustment system and method based on machine vision. BACKGROUND

[0002] As a kind of sanitation vehicle, watering truck is used for cleaning in urban roads, parks, construction sites and other places, and has the functions of watering and dust reduction, greening irrigation and ground washing, which can effectively improve environmental quality and maintain environmental hygiene, and is an important tool for urban management and environmental protection.

[0003] At present, the high-pressure water gun nozzle connection on the watering truck is mostly a three-bend front-charging nozzle. Compared with the traditional straight-jet nozzle, the three-bend design of the three-bend front-charging nozzle makes the water flow more gentle when sprayed, reducing the impact of water flow on the ground, which helps to protect the ground vegetation and the road surface. However, since the three-bend front-charging nozzle of the watering truck needs to spray the entire road surface, the flow rate is often set to a large fixed value and cannot be changed, which causes waste of water resources. In addition, due to the small nozzle diameter of the three-bend front-charging nozzle, the water pressure is higher than that of the duckbill nozzle under the same flow rate, which not only splashes the pedestrians on the roadside, but also affects the passing non-motor vehicles, and cannot achieve intelligent spraying effect. SUMMARY

[0004] The present application aims to provide a watering truck nozzle water quantity automatic adjustment system and method based on machine vision, which uses an image acquisition device to collect original road images, uses an image processing device to calculate road edge center points and target categories, speeds and distances, and uses a watering control device to control the time interval and distance of watering. The problems in the background art are solved.

[0005] In order to achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0006] A watering truck nozzle water quantity automatic adjustment system based on machine vision, comprising an image acquisition device, an image processing device and a watering control device.

[0007] The image acquisition device is used to collect original road images and transmit the original images to the image processing device, and convert the original images into depth maps.

[0008] The image processing device is used to process the received original images, calculate the road edge center points and target categories, speeds and distances, and adjust the watering conditions.

[0009] The watering control device is used to control the watering time interval and watering distance to achieve intelligent watering.

[0010] Further, the image acquisition device comprises a binocular camera with a camera support, the binocular camera comprises a left camera and a right camera, and the left camera and the right camera are connected at the same horizontal height on the vehicle head of the watering vehicle.

[0011] Further, the image acquisition device is a Raspberry Pi.

[0012] Further, the watering control device comprises a water pump and a three-bend forward-jet nozzle, the three-bend forward-jet nozzle is communicated with the water pump, the water pump is electrically connected with the Raspberry Pi, and the Raspberry Pi realizes the adjustment of the watering distance by controlling the pressure of the water pump and controls the watering time interval by controlling the start and stop of the water pump.

[0013] Further, in the process of collecting the original road image, the calibration method of the binocular camera is as follows: the Zhang Zhengyou calibration method is used to solve the intrinsic matrix A, the extrinsic matrix [R T] is obtained through the intrinsic matrix and the homography matrix H, and the radial distortion and tangential distortion are corrected by using the distortion correction model, so as to obtain the radial distortion parameters k1, k2 and k3 and the tangential distortion parameters p1 and p2 of the binocular camera, and the forward projection mapping between the world coordinate system and the image coordinate system is obtained.

[0014]

[0015] In the formula, s is a scale factor from the world coordinate system to the image coordinate system, is the image coordinate.

[0016] Further, the method for converting the original image into a depth map by the binocular camera is as follows: the left camera and the right camera shoot the same target in the same picture, find the matching point of each pixel point in the left camera image in the right camera image, then calculate the parallax by comparing the position difference of the corresponding pixel points in the left camera image and the right camera image, so as to obtain the depth information of the target.

[0017] Further, the method for calculating the road edge center point by the image processing device is as follows: the obtained extrinsic parameters and distortion parameters of the binocular camera are used to correct the distortion and the epipolar line of the image, the depth map of the road image is obtained by the binocular camera, and the road edge center point coordinates are obtained by the operations of binarization, corrosion, expansion and finding the largest color block image.

[0018] Further, the image processing device calculates the target class and distance by using a neural network, wherein the neural network comprises a backbone network, a fusion network and a head network; the backbone network adopts different sizes of convolution kernels to extract different scale feature layers from input data; the fusion network diversifies and fuses the different scale feature layers to generate new fusion feature maps, and inputs the fusion feature maps into the head network of target detection; the head network of target detection calculates the target class and distance based on the fusion feature maps by a decision unit of the head network, and predicts the coordinates of the target to generate a prediction box.

[0019] Further, the backbone network is a feature extraction network adopting a CSPNet structure, which divides a feature map into two parts and processes them in different paths to reduce the amount of calculation.

[0020] The fusion network adopts a bottom-up manner to up-sample a low-resolution feature map and fuse it with a high-resolution feature map by a feature pyramid structure, so as to enrich the semantic information and detailed information of the features to identify and locate the target.

[0021] The head network converts the feature map into which category each pixel belongs to by a plurality of fully connected layers, and generates a prediction box.

[0022] Further, the position coordinates of the target predicted by the head network include the coordinates of the bounding box, the class probability and the confidence.

[0023] Further, the method for calculating the target speed of the image processing device comprises the following steps: first, reading the video image frame by frame; then, obtaining the coordinates of the target box by ByteTrack tracking, calculating the height and distance of the target box, and calculating the target speed according to the frame rate and the distance.

[0024] Further, the water spraying control device controls the water spraying time interval as follows:

[0025]

[0026] wherein t is the time for controlling water spraying, specifically, reducing the amount of water spraying after t seconds, d l is the distance between the pedestrian or non-motor vehicle and the front end of the water spraying vehicle along the road direction, v m is the relative speed of the pedestrian or non-motor vehicle and the water spraying vehicle; t x is a correction coefficient, which includes the visual recognition time, the system judgment time and the water spraying braking time.

[0027] The method for automatically adjusting the water volume of the sprinkler of the watering cart based on the machine vision automatically adjusts the water volume of the sprinkler of the watering cart by first calibrating the binocular camera by using the Zhang Zhengyou calibration method and collecting road images, detecting the road edge center point of the collected images to obtain the road width, then putting the collected images into the neural network for training to obtain the target category, speed and distance, and finally processing the data by using the Raspberry Pi to control the water pump pressure and the start-stop time of the water pump to automatically adjust the water volume of the sprinkler of the watering cart.

[0028] The technical scheme has the beneficial effects that:

[0029] 1. The image acquisition device adopts a binocular camera, and the left and right cameras are located at the same horizontal height, which ensures the accuracy and consistency of image acquisition, and provides a high-quality data basis for subsequent image processing; the image acquisition device adopts a Raspberry Pi, which has powerful computing capability, good expandability and programmability, making the system more flexible and easy to maintain; by controlling the pressure and start-stop time of the water pump, the watering distance and watering time interval can be accurately adjusted to meet the watering needs under different road and weather conditions.

[0030] 2. The image acquisition device adopts a binocular camera, and the left and right cameras are located at the same horizontal height, which ensures the accuracy and consistency of image acquisition, and provides a high-quality data basis for subsequent image processing; the image acquisition device adopts a Raspberry Pi, which has powerful computing capability, good expandability and programmability, making the system more flexible and easy to maintain; by controlling the pressure and start-stop time of the water pump, the watering distance and watering time interval can be accurately adjusted to meet the watering needs under different road and weather conditions.

[0031] In summary, the present application can effectively realize water saving by detecting the road edge center point to determine the road distance and controlling the water volume of the three-bend front spray head. By combining the neural network with the binocular camera distance measurement module, the speed and distance of the target (pedestrian, non-motor vehicle) can be detected to adjust the water volume of the three-bend front spray head, realizing real-time pedestrian and non-motor vehicle avoidance. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 Flow chart of the water spraying vehicle nozzle water quantity automatic adjusting system based on machine vision of the present application;

[0033] Figure 2 Flow chart of the binocular camera ranging of the present application;

[0034] Figure 3 Flow chart of the road edge ranging of the present application;

[0035] Figure 4 Network structure schematic diagram of the present application;

[0036] Figure 5 Pedestrian and non-motor vehicle speed and distance detection result diagram of the present application;

[0037] Figure 6 Distance judgment schematic diagram of the present application. DETAILED DESCRIPTION

[0038] The present application will be further described in detail below in combination with the drawings and embodiments:

[0039] A water spraying vehicle nozzle water quantity automatic adjusting system based on machine vision, comprising an image acquisition device, an image processing device and a water spraying control device;

[0040] The image acquisition device comprises a binocular camera with a camera support, the binocular camera comprises a left eye camera and a right eye camera, and the left eye camera and the right eye camera are connected at the same horizontal height on the head of the water spraying vehicle, the binocular camera is used for collecting original road images, and the original images are transmitted into the image processing device, and the original images are converted into depth maps;

[0041] The image acquisition device is a Raspberry Pi, which is used for processing the received original images, calculating the road edge center point, target category, speed and distance, and adjusting the water spraying condition;

[0042] The water spraying control device comprises a water pump and a three-bend front flushing nozzle, the three-bend front flushing nozzle is communicated with the water pump, the water pump is electrically connected with the Raspberry Pi, and the Raspberry Pi adjusts the water spraying distance by controlling the pressure of the water pump to realize intelligent water spraying.

[0043] As Figure 1As shown, a method for automatically adjusting the water flow of a sprinkler head of a water sprinkler based on machine vision, first, the binocular camera shoots the road image and converts it into a depth map, extracts the road edge center point from the depth map, and obtains the distance from the road edge; Then use neural network to obtain pedestrian and non-motor vehicle information, and obtain its distance from the front end of the car, and then obtain the distance along the road direction by combining the distance from the road edge, estimate the speed through the distance on two frames of images, and then obtain the meeting time, and control the water pump pressure of the water sprinkler to control the water flow, so as to achieve the purpose of real-time detection and control.

[0044] As shown in Figure 2 , target detection based on neural network and binocular ranging, the specific method is as follows:

[0045] Binocular camera calibration is performed on the road image: Zhang Zhengyou calibration method is used to solve the internal parameter matrix A, the external parameter matrix [R T] is obtained through the internal parameter matrix and the homography matrix H, and the radial distortion and tangential distortion are corrected by using the distortion correction model, so as to obtain the radial distortion parameters k1, k2, k3 and the tangential distortion parameters p1, p 2， Then complete the forward projection mapping between the world coordinate system and the image coordinate system:

[0046]

[0047] In the formula, s is the scale factor from the world coordinate system to the image coordinate system, is the image coordinate.

[0048] The principle of binocular camera ranging is that the binocular camera is composed of left and right cameras, and the two cameras are placed within a certain distance, which is called baseline. Two cameras shoot the same scene to generate different images, the left eye image is L, and the right eye image is R.

[0049] Due to the different placement positions, there is a difference in the viewing angle of the two cameras, and the pixel positions corresponding to the same object on L and R also have a difference, which is called parallax, so as to obtain the object distance z:

[0050]

[0051] In the formula, b is the baseline distance between the two cameras, f is the focal length of the camera, x r and x l are the horizontal coordinates of the feature points on L and R, and the target detection steps are as follows:

[0052] Collect road pictures, make training set labels and train, get the weight file after training, and evaluate the weight file parameters, so as to get the best weight file, and import the trained weight file into the neural network for prediction.

[0053] As Figure 3 shown, the road edge ranging based on machine vision, the specific method is as follows:

[0054] First, the road image is binarized, using global threshold method, the pixel greater than the threshold is set to white, and the pixel less than the threshold is set to black, the formula is as follows:

[0055]

[0056] In the formula, f(x, y) is the pixel value after operation, g(x, y) is the original pixel value, and T is the threshold.

[0057] The binarized image is preprocessed, and the image is opened. The convolution kernel slides along the image. If all the pixel values of the original image corresponding to the convolution kernel are 1, the center element keeps the original pixel value, otherwise it is 0, that is, the image is eroded.

[0058] Then, the inflation operation is performed, that is, as long as one of the pixel values of the original image corresponding to the convolution kernel is 1, the pixel value of the center element is 1. The opening operation can eliminate noise and fill small holes in the image, making the depth map more smooth. After preprocessing, the largest color block in the image is found, so as to obtain the road edge information. The pixel values of the part except the road edge are all set to 0, and the road center point coordinates are calculated.

[0059] Finally, the distance between the water cart and the road edge center point is calculated:

[0060] The conversion formula between pixel and image is:

[0061]

[0062] In the formula, (u, v) is the pixel point coordinate, d x , d y respectively represent the length corresponding to a single pixel in x direction and y direction.

[0063] Therefore, the coordinates of the pixel point in the camera coordinate system are:

[0064]

[0065] In the formula, D(u, v) is the depth information of the pixel point (u, v), (f x , f y , u0, v0) is the camera parameter obtained by camera calibration.

[0066] Then, the coordinates in the camera coordinate system are converted into the world coordinate system:

[0067] PW = [R T] * P C

[0068] In the formula, [R T] is the camera parameter obtained by camera calibration, P W (X W , Y W , Z W ) is its coordinate in the world coordinate system, P C is the coordinate in the camera coordinate system.

[0069] The actual distance between the last obtained camera and the center point of the road edge is:

[0070]

[0071] If the camera height is h, the lateral distance is:

[0072]

[0073] As Figure 4 shown, the neural network includes a backbone network, a fusion network and a head network; its specific structure is:

[0074] The backbone network is a feature extraction network, which adopts the CSPNet structure, which can effectively reduce the amount of calculation by dividing the feature map into two parts and processing them in different paths;

[0075] The fusion network uses a bottom-up approach to upsample low-resolution feature maps and fuse them with high-resolution feature maps through a feature pyramid structure, enriching the semantic information and detailed information of the features, so that the subsequent detection head can more accurately identify and locate the target;

[0076] The head network converts the feature map to which category each pixel belongs through several fully connected layers and generates a prediction box.

[0077] As Figure 5 shown, the speed and distance detection results of pedestrians and non-motor vehicles are as follows:

[0078] First, read the frame rate of the video and read the video image frame by frame; get the coordinates of the target box (X1, y1), (x2, y2) through ByteTrack tracking, and calculate the height of the target box:

[0079] H = y2-y1

[0080] Estimate the distance of the target box by the height of the target box:

[0081]

[0082] In the formula, H rFor the actual height of the object, since the pedestrians and non-motor vehicles are detected, H r is set to a certain value, f is the focal length of the camera, and since the coordinate frame completely surrounds the target object, the height of the coordinate frame is slightly higher than the height of the object, so the modified target frame height H* is used.

[0083] Speed measurement is achieved by comparing the distance changes of the object between different frames. First, record the distance of the object in each frame of image, then calculate the moving speed of the object through the frame interval:

[0084]

[0085] In the formula, D2 is the distance of the target frame, D1 is the distance of the last frame, t is the time interval, which is calculated by the frame rate.

[0086] As Figure 6 shown, the distance judgment principle is as follows:

[0087]

[0088] In the formula, d l is the distance between the pedestrian or non-motor vehicle and the front end of the water truck along the road direction, D is the distance between the camera and the target, h is the height of the camera, and L is the horizontal distance between the camera and the road edge.

[0089] According to the obtained target distance and speed, the control water spraying time can be obtained as:

[0090]

[0091] In the formula, t is the control water spraying time, i.e. after t seconds, the water spraying amount is reduced, d l is the distance between the pedestrian or non-motor vehicle and the front end of the water truck along the road direction, v m is the relative speed of the pedestrian or non-motor vehicle and the water truck. t x is a correction coefficient, which includes visual recognition time, system judgment time and water spraying braking time.

[0092] The water spraying amount is controlled by adjusting the pressure and rotating speed of the water pump. The water pump uses mechanical force to pump water from low to high or long distance, and its core components are pump body and impeller. When the water pump works, the impeller rotates to generate centrifugal force, which sucks water from the inlet of the pump and accelerates it, and then discharges it through the outlet.

[0093] In which, the principle of controlling the water spraying amount by controlling the water pump pressure is:

[0094] The pressure of the water pump can be controlled by adjusting the outlet valve of the pump. When the outlet valve is opened less, the flow rate of water increases and the pressure increases, which will result in a farther and more concentrated water column, thus reducing the amount of sprayed water per unit area. Conversely, if the valve opening is increased, the pressure decreases and the flow rate of water slows down, and the amount of sprayed water will be more and the coverage area will be wider.

[0095] By controlling the water pump pressure to control the flow of the nozzle, the specific formula is as follows:

[0096]

[0097] In the formula, q is the flow of the nozzle, P is the working pressure of the nozzle, K is the flow coefficient of the nozzle, g is the acceleration of gravity, and h is the height of the nozzle from the ground.

[0098] As can be seen from the above, the water spraying vehicle nozzle water amount automatic adjusting system and method based on machine vision provided by the application can effectively realize water saving by detecting the road edge center point to determine the road distance and thus controlling the three-bend front spray nozzle water amount. The speed and distance of the target (pedestrian, non-motor vehicle) are detected by combining the neural network with the binocular camera distance measuring module, so as to control the three-bend front spray nozzle water amount, and thus realizing real-time pedestrian and non-motor vehicle avoidance.

[0099] The above-mentioned is only an embodiment of the application, and the well-known specific technical solutions or characteristics in the scheme are not described in detail. It should be noted that for those skilled in the art, without departing from the technical solutions of the application, some deformations and improvements can be made, which should also be considered as the protection scope of the application, and these will not affect the effect and practicality of the application. The protection scope of the present application should be subject to the content of its claims, and the specific implementation mode and the like in the specification can be used to explain the content of the claims.

Claims

1. A machine vision-based automatic water volume adjustment system for sprinkler truck nozzles, characterized in that, Includes image acquisition device, image processing device, and sprinkler control device; The image acquisition device is used to acquire original road images, transmit the original images to the image processing device, and convert the original images into depth maps; The image processing device is used to process the received raw image, calculate the center point of the road edge and the target category, speed and distance, and adjust the water spraying situation; The water spraying control device is used to control the water spraying time interval and water spraying distance to achieve intelligent water spraying; The image acquisition device includes a binocular camera with a camera bracket. The binocular camera includes a left eye camera and a right eye camera, and the left eye camera and the right eye camera are connected at the same horizontal height on the front of the sprinkler truck. The image processing device calculates the center point of the road edge as follows: It performs distortion correction and epipolar correction on the image using the obtained extrinsic and distortion parameters of the binocular camera. Combined with the depth map of the road image obtained from the binocular camera, it performs binarization, erosion, dilation, and finds the largest color patch image operations to obtain the coordinates of the road edge center point. Specifically, the machine vision-based road edge ranging method is as follows: First, the captured road image is binarized. Then, a global thresholding method is used to set pixels larger than the threshold to white and pixels smaller than the threshold to black. The formula is as follows: ； In the formula, f(x,y) is the pixel value after the operation, g(x,y) is the original pixel value, and T is the threshold. Image preprocessing is performed on the binarized image, including opening operations. The convolution kernel slides along the image. If all pixel values ​​of the original image corresponding to the convolution kernel are 1, then the central element retains its original pixel value; otherwise, it is 0, which is equivalent to performing an erosion operation on the image. Then, a dilation operation is performed. If any pixel value in the original image corresponding to the convolution kernel is 1, the pixel value of the center element is 1. The opening operation can eliminate noise, fill small holes in the image, and make the depth map smoother. After preprocessing, the largest color block in the image is found to obtain the road edge information. All pixel values ​​except those of the road edge are set to 0, and the coordinates of the road center point are calculated. Finally, calculate the distance between the water truck and the center point of the road edge: The formula for converting pixels to images is: ； In the formula, (u,v) are the pixel coordinates. , These represent the lengths of a single pixel in the x and y directions, respectively. The coordinates of the pixel in the camera coordinate system are: ； In the formula, For pixels Depth information, ( () represents the camera intrinsic parameters obtained from camera calibration; Then, convert the coordinates in the camera coordinate system to the world coordinate system: ； In the formula, The camera extrinsic parameters obtained from camera calibration, (X W ,Y W Z W ) represents its coordinates in the world coordinate system. coordinate; The final actual distance between the camera and the center point of the road edge is: ； If the camera height is h, then the lateral distance is: ； The image processing device calculates the target speed as follows: first, it reads video images frame by frame; then, it uses ByteTrack to track and obtain the coordinates of the target bounding box, calculates the height and distance of the target bounding box, and calculates the target speed based on the frame rate and distance. The principle of the pedestrian and non-motorized vehicle speed detection module is as follows: First, the video frame rate is read, and then the video images are read frame by frame; the coordinates of the target bounding box are obtained by ByteTrack tracking. ), and calculate the height of the target bounding box: H= - ； Estimate the distance of the target box by its height: ； In the formula, Here, f is the actual height of the object, and 'f' is the camera focal length. Since the bounding box must completely enclose the target object, the height of the bounding box must be higher than the object's height. Therefore, a method is used to correct the target bounding box height. ; Speed ​​is measured by comparing changes in the distance of an object between different frames; first, the distance of the object in each frame is recorded, and then the object's speed is calculated at frame intervals. ； In the formula, The distance to the target bounding box in the current frame. The distance to the target box in the previous frame is t, which is the time interval calculated from the frame rate. The distance determination formula is: ； In the formula, D is the distance between the front of the water truck and the pedestrian or non-motorized vehicle along the road direction, h is the distance between the camera and the target, and L is the lateral distance between the camera and the road edge. Based on the obtained target distance and speed, the controlled water spraying time is determined as follows: ； In the formula, t represents the time for controlling the water spraying, that is, reducing the amount of water sprayed after t seconds. The distance between pedestrians or non-motorized vehicles and the front of the water truck along the road. The relative speed between pedestrians or non-motorized vehicles and the water truck. The correction factor includes visual recognition time, system judgment time, and water spraying braking time.

2. The automatic water volume adjustment system for sprinkler truck nozzles based on machine vision according to claim 1, characterized in that, The image acquisition device is a Raspberry Pi.

3. The automatic water volume adjustment system for sprinkler truck nozzles based on machine vision according to claim 2, characterized in that, The water spraying control device includes a water pump and a three-bend front spray nozzle. The three-bend front spray nozzle is connected to the water pump, and the water pump is electrically connected to the Raspberry Pi. The Raspberry Pi adjusts the spraying distance by controlling the pressure of the water pump and controls the spraying time interval by controlling the start and stop of the water pump.

4. The automatic water volume adjustment system for sprinkler truck nozzles based on machine vision according to claim 3, characterized in that, The calibration method for a stereo camera during the acquisition of raw road images is as follows: the intrinsic parameter matrix A is solved using the Zhang Zhengyou calibration method, and the extrinsic parameter matrix is ​​obtained through the intrinsic parameter matrix and the homography matrix H. The radial and tangential distortions are corrected using a distortion correction model, thereby obtaining the radial distortion parameters k1, k2, k3 and the tangential distortion parameters p1, p2 of the binocular camera, and obtaining the forward projection mapping between the world coordinate system and the image coordinate system: ； In the formula, s is the scaling factor from the world coordinate system to the image coordinate system. These are the image coordinates.

5. The automatic water volume adjustment system for sprinkler truck nozzles based on machine vision according to claim 4, characterized in that, The method for converting the original image into a depth map using a binocular camera is as follows: the left and right cameras take the same picture of the same target, find the matching point of each pixel in the left camera image in the right camera image, and then calculate the parallax by comparing the positional differences of the corresponding pixels in the left and right camera images, thereby obtaining the depth information of the target.

6. The automatic water volume adjustment system for sprinkler truck nozzles based on machine vision according to claim 5, characterized in that, The image processing device uses a neural network to calculate the target category and distance. The neural network includes a backbone network, a fusion network, and a head network. The backbone network uses convolutional kernels of different sizes to extract feature layers of different scales from the input data. The fusion network fuses these feature layers of different scales in a diversified manner to generate a new fused feature map. The fused feature map is then input into the head network for target detection. The decision unit of the head network calculates the target category and distance based on these fused feature maps, predicts the target's coordinates, and generates a prediction box.

7. The automatic water volume adjustment system for sprinkler truck nozzles based on machine vision according to claim 6, characterized in that, The backbone network is a feature extraction network, which adopts the CSPNet structure. It reduces the amount of computation by dividing the feature map into two parts and processing them in different paths. The fusion network uses a feature pyramid structure to upsample low-resolution feature maps and fuse them with high-resolution feature maps in a bottom-up manner, thereby enriching the semantic and detailed information of the features to identify and locate targets. The head network uses several fully connected layers to convert the feature map into the category to which each pixel belongs and generates a prediction box.

8. The automatic water volume adjustment system for sprinkler truck nozzles based on machine vision according to claim 7, characterized in that, The target location coordinates predicted by the head network include the bounding box coordinates, class probability, and confidence level.

9. The method for automatically adjusting the water volume of a sprinkler truck nozzle using a machine vision-based automatic water volume adjustment system according to claim 8, characterized in that, First, the binocular camera is calibrated using Zhang Zhengyou's calibration method, and road images are acquired. Road edge detection is performed on the acquired images to obtain the center point of the road edge, thereby obtaining the road width. Then, the acquired images are fed into a neural network for training to obtain the target category, speed, and distance. Finally, the data is processed using a Raspberry Pi, and the water volume of the sprinkler truck is automatically adjusted by controlling the water pump pressure and the start and stop time of the water pump.

Images