Field observation device and positioning method for rice leaf roller based on multi-view stereo vision

Abstract

The present application belongs to the technical field of image processing, and proposes a rice leaf roller field observation device and positioning method based on multi-view stereo vision. The main scheme is as follows: the rice leaf roller is trapped through a trapping unit, the image information of the trapped rice leaf roller at different flight moments is captured through an image acquisition unit, the difference between the background processing unit and the trapped rice leaf roller is formed, and the image interference when capturing the image information of the trapped rice leaf roller at different flight moments is removed, the shooting angle of the image acquisition unit is controlled through a processing calculation unit, and an improved image processing algorithm and a three-dimensional positioning algorithm are called to process the image information of the rice leaf roller at different flight moments, calculate the individual behavior parameters and population behavior parameters of the rice leaf roller, and save. The present application can improve the observation accuracy, recognition and positioning accuracy of the rice leaf roller.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a field observation device and positioning method for rice leaf roller based on multi-view stereo vision. Background Technology

[0002] Rice, a staple food for more than half the world's population, is the most widely cultivated of the world's three major staple crops. The rice leaf roller is a core pest of rice, easily causing significant yield reductions. Accurate identification, spatial location, and monitoring data collection of the leaf roller are crucial for the scientific control of rice pests. Current monitoring technologies are transitioning from manual to intelligent methods, but multiple bottlenecks limit their effectiveness. Traditional manual surveys and monitoring lights cannot simultaneously achieve species differentiation, accurate location, and data retention. While computer vision-based solutions are a trend, they suffer from significant shortcomings in terms of equipment, algorithms, data support, and research background.

[0003] At the device level, existing image acquisition equipment is ill-suited to the needs of field pest monitoring. Fixed-view devices cannot capture pest morphology from a 360-degree angle without blind spots, portable devices are cumbersome to adjust and lack structural designs adapted to multi-view positioning. There is currently no integrated device that uses a funnel-shaped transparent structure to achieve efficient pest collection while simultaneously using a circular track to move the camera for 360-degree image positioning. This results in low efficiency in acquiring multi-view images and difficulty in accurately calculating the three-dimensional coordinates of pests. The field environment is complex, with numerous background interferences under natural lenses. The subject and background features are highly similar, and the boundaries are blurred, making it difficult to effectively separate the subject from the background using traditional methods, severely impacting the accuracy of target detection and recognition.

[0004] At the algorithm level, existing deep learning models lack adaptability. Traditional CBAM modules have fixed parameters and lack adaptive adjustment capabilities. They cannot dynamically optimize attention weights for different body sizes and morphological features of rice leaf rollers and rice planthoppers, and are easily affected by background noise such as rice leaf texture. The AFPN structure is not optimized for small rice pests and lacks a low-resolution P2 layer design. The fusion of small-scale pest features is insufficient, resulting in low species recognition accuracy and difficulty in accurately distinguishing two types of pests with similar body sizes.

[0005] At the data support level, extracting high-quality pest datasets is difficult. Existing datasets suffer from limited scene representation, incomplete annotation information, a lack of multi-view and 3D coordinate data from the field, and require manual collection and annotation, resulting in low efficiency and poor consistency. Furthermore, the data collection devices are disconnected from the data extraction functions, failing to achieve a closed loop of automated pest image acquisition, feature extraction, and dataset generation. This restricts model training and generalization capabilities, further impacting recognition and localization accuracy.

[0006] From a research perspective, current studies on the individual movement of the rice leaf roller largely rely on basic parameters such as population density and birth rate obtained from field surveys. They lack precise quantitative data on the spatial behavior of individual pests, leading to significant biases in model predictions of population dispersal and activity patterns. Furthermore, existing monitoring technologies cannot efficiently acquire the pest's three-dimensional movement trajectories and behavioral characteristics, hindering the optimization and upgrading of population dynamics models. Therefore, an integrated technical solution that bridges trajectory monitoring and population dynamics research is urgently needed. Summary of the Invention

[0007] The purpose of this invention is to provide a field observation device and positioning method for rice leaf roller based on multi-view stereo vision, so as to solve the problems that existing monitoring devices cannot accurately capture the flight point of rice leaf roller, have low efficiency in acquiring multi-view images, and have insufficient accuracy in pest identification and poor three-dimensional positioning effect of algorithms.

[0008] The technical solution adopted by this invention to solve its technical problem is as follows: On one hand, the present invention provides a field observation device for rice leaf roller based on multi-view stereo vision, comprising: Trapping units are used to trap rice leaf rollers; The image acquisition unit is used to capture image information of the trapped rice leaf roller at different flight times; The background processing unit is used to differentiate the image from the trapped rice leaf roller and remove image interference when capturing image information of the trapped rice leaf roller at different flight times. The processing and computing unit is equipped with improved image processing and 3D positioning algorithms to control the shooting angle of the image acquisition unit. It also calls the improved image processing and 3D positioning algorithms to process the image information of the rice leaf roller at different flight times, calculates and saves the individual behavior parameters and population behavior parameters of the rice leaf roller.

[0009] In some embodiments, the trapping unit is a cylindrical-cone shaped "integrated cylinder and cone" rice leaf roller observation and trapping barrel. The barrel body is made of acrylic material with a light transmittance of ≥90%. The cylindrical barrel body is 80cm high, the inverted funnel barrel is 20cm high, the upper opening diameter is 58cm, and the lower opening diameter is 330cm. The upper opening of the trapping barrel is tightly connected to the roof. The food attractant is placed on the inner edge of the lower opening of the trapping barrel, the inner edge is 2cm wide, and the barrel body is 1cm thick.

[0010] In some embodiments, the image acquisition unit includes a ring slide, a moving drive, and a high-definition industrial camera; The annular slide is coaxially fitted onto the outside of the trapping barrel. The annular slide is 1cm wide and 2cm high, and is made of stainless steel. The moving drive component is a stepper motor drive component, which is installed on the annular slide. The high-definition industrial camera mounted on the stepper motor drive component is suspended under the annular slide. The stepper motor drive component is used to drive the high-definition industrial camera and the baffle to make a 360-degree circular motion along the annular slide at a speed of 5cm / s, so as to realize the full-view image acquisition of rice leaf roller in the trapping bucket. The high-definition industrial camera has a resolution of 3840×2748 and a frame rate of 30 FPS. The high-definition industrial camera and the baffle are symmetrical about the central axis of the trapping barrel.

[0011] In some embodiments, the high-definition industrial camera is triggered by a built-in timer to capture image information of the rice leaf roller at different flight times by taking 10 consecutive shots per second. The high-definition industrial camera stops taking pictures after 22 minutes of surround shooting, and then performs 8 minutes of image processing and saving by the processing and computing unit.

[0012] In some embodiments, the background processing unit is a matte leaf-green background panel with dimensions of 100cm × 100cm and a thickness of 2cm. The background panel is positioned symmetrically to the high-definition industrial camera about the axis of the trapping barrel, parallel to the camera lens, and the center of the background panel is on the same horizontal line as the center of the trapping barrel.

[0013] In some embodiments, the processing and computing unit includes a main controller, a data processing module, and a storage module; The main controller uses an STM32 microcontroller, which is electrically connected to the high-definition industrial camera and the stepper motor driver. The main controller controls the working state of the stepper motor to drive the high-definition industrial camera to move along the track and adjust the shooting angle. The stepper motor drives the high-definition industrial camera to make a 360-degree circular motion along the circular track. The main controller is used to control the start and stop cycle of the high-definition industrial camera: the single working time is set to 1320s, during which the camera can move at a constant speed along the circular slide for 30 full circles, and return to the initial position after the movement ends; the single stop time is set to 8 minutes, which is used by the storage module to store the image information captured in this working cycle. The data processing module has built-in improved image processing and 3D imaging algorithms to ensure their normal operation. The data processing module completes image analysis and 3D coordinate calculation within 8 minutes, and obtains and saves the individual behavior parameters and population behavior parameters of the rice leaf roller.

[0014] In some embodiments, the individual behavioral parameters of the rice leaf roller include flight speed, activity radius, and trajectory curvature, while the population behavioral parameters include multi-target trajectory overlap and trajectory direction consistency.

[0015] On the other hand, the present invention also provides a field positioning method for rice leaf roller based on multi-view stereo vision, applied to the aforementioned field observation device for rice leaf roller based on multi-view stereo vision, comprising the following steps: Construct the overall spatial coordinate system of the trapping barrel; Image information of trapped rice leaf rollers at different flight times was captured, and image preprocessing and cross-scale enhancement were performed. Locate the geometric center of the rice leaf roller; Based on the geometric center of the rice leaf roller located by positioning, multi-view stereo vision 3D coordinate calculation is performed.

[0016] The beneficial effects of this invention are: the improved image processing algorithm and three-dimensional positioning algorithm of this invention can process image information of rice leaf folder at different flight times, calculate and save individual behavior parameters and population behavior parameters of rice leaf folder, thus providing high-precision spatial behavior input data for rice leaf folder population dynamics model, and improving the scientific nature of population size prediction and precise control decision-making. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the field observation device for rice leaf roller based on multi-view stereo vision in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the "integrated cone-shaped" rice leaf roller observation and trapping bucket in Embodiment 1 of the present invention; Figure 3 This is a detailed view of the image acquisition unit and background processing unit in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram illustrating the specific execution process of the field positioning method for rice leaf roller based on multi-view stereo vision in Embodiment 3 of the present invention.

[0018] Among them, 1 represents the trapping barrel, 2 represents the circular slide, 3 represents the moving drive component, 4 represents the high-definition industrial camera, 5 represents the leaf green matte background plate, and 6 represents the controller box. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0020] Example 1

[0021] See Figure 1This embodiment provides a field observation device for rice leaf roller based on multi-view stereo vision, including: Trapping units are used to trap rice leaf rollers; The image acquisition unit is used to capture image information of the trapped rice leaf roller at different flight times; The background processing unit is used to differentiate the image from the trapped rice leaf roller and remove image interference when capturing image information of the trapped rice leaf roller at different flight times. The processing and computing unit is equipped with improved image processing and 3D positioning algorithms to control the shooting angle of the image acquisition unit. It also calls the improved image processing and 3D positioning algorithms to process the image information of the rice leaf roller at different flight times, calculates and saves the individual behavior parameters and population behavior parameters of the rice leaf roller.

[0022] See Figure 2 In this embodiment, the trapping unit is a "cylindrical-cone integrated" rice leaf roller observation and trapping barrel 1 with an upper cylindrical shape and an inverted funnel shape at the bottom. The barrel body of the trapping barrel 1 is made of acrylic material with a light transmittance of ≥90%. The height of the cylindrical barrel body is 80cm, the height of the inverted funnel barrel is 20cm, the upper opening diameter is 58cm, and the lower opening diameter is 330cm. The upper opening of the trapping barrel is tightly connected to the roof. The food attractant is placed on the inner edge of the lower opening of the trapping barrel 1, with an inner edge width of 2cm. The overall thickness of the barrel body is 1cm. Therefore, it can efficiently trap rice leaf rollers without obstructing the camera image acquisition, making it easy to clearly capture the flight state of the pests inside the barrel.

[0023] See Figure 3 In this embodiment, the image acquisition unit includes an annular slide 2, a moving drive component 3, and a high-definition industrial camera 4. The annular slide 2 is coaxially fitted onto the outside of the trapping barrel 1. The width of the slide portion of the cylindrical trapping barrel 1 is 1 cm, and its height is 2 cm. Stainless steel is used to ensure structural stability. The high-definition industrial camera 4 has a resolution of 3840×2748 (10.5 million pixels) and a frame rate of up to 30 FPS. A stepper motor drive component is mounted on the annular slide 2, and the drive assembly, carrying the high-definition camera, is suspended below the annular slide 2. The moving drive component 3 can drive the high-definition camera and the baffle to perform a 360-degree circular motion along the annular slide 2 at a speed of 5 cm / s, achieving full-view image acquisition of the rice leaf roller inside the trapping barrel 1. The camera and the baffle are symmetrical about the barrel's central axis, and their rotation speed and direction are consistent.

[0024] In practical applications, if a fixed camera is used to take pictures from a single position, only a two-dimensional image of that single position can be obtained. Multi-angle shooting provides image information under different camera coordinates, offering different camera coordinate calibrations and providing more data support for three-dimensional positioning. Because the target is a flying individual, shooting from a single position cannot achieve clear and comprehensive capture of the target individual from all directions. 360-degree camera rotation effectively solves problems such as unclear insect images, poor posture resulting in only capturing the edges of the insect, and poor lighting conditions at a single angle during a certain period. Therefore, in this embodiment, the camera has a built-in timer that triggers continuous shooting at 10 frames per second, accurately capturing image information of the rice leaf roller at different flight moments. After 22 minutes of continuous shooting, the camera stops shooting and performs 8 minutes of image recognition and storage by the processing and computing components. This combination of processing time and intermittent analysis time ensures both shooting and processing time are available.

[0025] See Figure 3 In this embodiment, the background processing unit is a matte green background plate 5, measuring 100cm x 100cm and 2cm thick. This background plate is positioned symmetrically to the high-definition camera along the barrel axis, parallel to the camera lens, and its center is on the same horizontal line as the center of the trapping barrel 1. Without the background processing component, the transparent background barrel reflects the complex field environment behind it, resulting in a lot of interference. The adult rice leaf roller is yellowish-brown, similar in color to the rice paddy environment, making it difficult to distinguish and identify the insect against a complex background. The matte green material avoids image interference caused by light reflection and has a significant color difference from the insect's body, creating a high-contrast background during image acquisition. This facilitates rapid removal of background impurities through threshold segmentation during subsequent image processing, improving target recognition efficiency.

[0026] In addition, the processing and computing unit in this embodiment is located in the controller box 6, including a main controller, a data processing module, and a storage module. In this embodiment, the main controller uses an STM32 microcontroller, which is electrically connected to the high-definition camera and stepper motors. The main controller controls the working state of the stepper motors to drive the camera to move along the track and adjust the shooting angle. The stepper motors drive the high-definition camera to perform a 360-degree circular motion along the circular track. The driving components adopt a dual-stepper motor synchronous drive structure, with two sets of stepper motor drivers respectively installed at the connection points between the high-definition camera connecting rod and the track, and the connection points between the background plate connecting rod and the track. Through synchronous pulse signals, the STM32 main controller outputs completely consistent control commands to the two sets of drivers, achieving strict synchronization of speed, direction, and start / stop actions. During operation, the STM32 main controller outputs unified drive pulses and direction level signals. After receiving the same timing commands, the two drivers respectively drive the corresponding connecting rods to perform coaxial rotation around the circular track, driving the high-definition industrial camera and its supporting structural components to perform a 360° coaxial synchronous circular motion along the circular track. The main controller only controls the camera's start-stop cycle: a single working cycle is set to 1320 seconds (22 minutes), during which the camera can move at a constant speed along the circular track 2 for 30 full laps, returning precisely to its initial position after the movement. A single stop cycle is set to 8 minutes, allowing the storage module to store image information captured during this working cycle and to debug the data processing module, ensuring the image processing and 3D imaging algorithms operate normally. Within this time, the data processing module completes image analysis and 3D coordinate calculation. The data processing module is also controlled by the main controller. The controller initiates the image acquisition program; this module incorporates various improved image processing and 3D imaging algorithms to analyze and process the acquired images.

[0027] In practical applications, the rice leaf roller is active during the peak summer season when sunlight is strong. The midday sun angle in summer is high, and the light intensity in the field can reach strong light levels, easily creating a coexistence of bright and shadowed areas, resulting in high light contrast and potentially causing localized overexposure or underexposure of images. Simultaneously, the uneven distribution of light and frequent changes in light spots and shadows are influenced by clouds, rice canopy obstruction, and wind-driven leaf movement, creating a typical dynamic and complex lighting environment. Therefore, the device in this embodiment needs to adapt to summer lighting conditions. This module is responsible for identifying the target rice leaf roller from numerous image images. After identifying the target, it determines the center point of the target and uses the world coordinates of the center point as the individual target coordinates. Therefore, this embodiment can include a trajectory feature parameter extraction submodule to calculate the following key parameters based on the generated three-dimensional flight trajectory: 1. Individual behavior parameters: flight speed, activity radius, trajectory curvature; Population behavior parameters: 2. Multi-target trajectory overlap, trajectory direction consistency.

[0028] Example 2

[0029] Based on Example 1, this example provides a field location method for rice leaf roller based on multi-view stereo vision, which may include the following steps: Construct the overall spatial coordinate system of the trapping barrel; Image information of trapped rice leaf rollers at different flight times was captured, and image preprocessing and cross-scale enhancement were performed. Locate the geometric center of the rice leaf roller; Based on the geometric center of the rice leaf roller located by positioning, multi-view stereo vision 3D coordinate calculation is performed.

[0030] In practical applications, this embodiment implements a field positioning method for rice leaf rollers based on multi-view stereo vision through the following specific execution steps: I. Construct the overall spatial coordinate system of the trapping bucket.

[0031] With the center of the circular slide as the origin O of the world coordinate system w Establish a right-handed Cartesian coordinate system O w -X w Y w Z w :X w O w Y w The plane coincides with the plane of the circular slide and is parallel to the ground. Z w The axis is perpendicular to the ground and points upwards, coinciding with the central axis of the trapping barrel. The initial position of the camera corresponds to X. w In the positive direction of the axis, the azimuth angle θ increases as the camera rotates counterclockwise along the annular slide.

[0032] 2. Capture image information of the trapped rice leaf roller at different flight times, and perform image preprocessing and cross-scale enhancement.

[0033] (a) Image distortion correction and grayscale conversion 1. Image Distortion Reduction: Due to the "cylinder-cone integrated" shape of the trapping barrel (cylinder on top, cone on the bottom), a height-dependent radial scaling factor is introduced during distortion reduction based on the barrel's diameter of 58cm, cylindrical barrel height of 80cm, and conical barrel height of 20cm. This method achieves segmented distortion correction for cylindrical and conical segments. For radial and tangential distortions generated during image acquisition by the camera, a distortion correction algorithm is used to optimize the image coordinates. Let (x0, y0) be the coordinates of the ideal distortion-free image, and (x, y) be the coordinates of the corrected image. Let R be the radial distance from the pixel to the optical center, k1, k2, and k3 be the radial distortion coefficients, p1 and p2 be the tangential distortion coefficients, R = 29cm be the radius of the trapping barrel, H1 = 80cm be the height of the cylindrical segment, H2 = 20cm be the height of the conical segment, and h be the actual height of the current pixel inside the barrel. Model the cylindrical and conical segments separately: a. Cylindrical segment (0 ≤ h ≤ H1) With a constant radius, a standard distortion model is used: b. Conical segment (H1) <h The radius shrinks linearly with height, so a taper scaling factor is introduced. Then perform distortion correction and substitute the values ​​into formula (1): 2. Grayscale Conversion: In summer, rice paddies experience high light intensity, strong blue and green light components, and significant leaf reflection. This leads to image overexposure in the blue channel in highlight areas, resulting in high noise levels. Additionally, the green channel in green leaves has an excessively high proportion, weakening the grayscale difference between the insect and the leaf. The red channel is relatively stable and provides better identification of the yellowish-brown and dark brown areas of the insect. Therefore, it is necessary to reduce the weight of the blue and green channels and increase the weight of the red channel. The improved formula is used: In the formula These represent the red, green, and blue channel values ​​of the image at pixel (x,y), respectively.

[0034] (II) Adaptive threshold segmentation and Gaussian kernel optimization 1. Initial threshold acquisition: In response to the complex imaging environment of strong sunlight, leaf reflection, uneven lighting, and differences in brightness inside the trapping bucket in midsummer rice paddies, the "improved maximum inter-class variance method (OTSU) with the introduction of a light compensation factor" is used to determine the initial segmentation threshold, so as to achieve robust separation of the rice leaf roller target from the green leaves and the background of the bucket.

[0035] The improved formula for calculating inter-class variance is: ​ω0 and ω1 represent the pixel ratios of the foreground and background, respectively; μ0 and μ1 represent the average gray values ​​of the foreground and background, respectively; μ represents the average gray value of the entire image; L(x,y) is the local brightness normalization value of the image, ranging from [0,1], specifically characterizing the intensity of reflections from rice paddy leaves, water surface glare, and the reflective area of ​​the inner wall of the trapping bucket in midsummer, calculated by the ratio of the local pixel gray value to the global maximum gray value of the image; α is the strong light suppression coefficient, ranging from 0.03 to 0.08 to meet the usage scenarios of this patent, dynamically adjusted according to the light intensity of the cylindrical section (0~80cm) and the conical section (80~100cm) at different heights inside the trapping bucket, with the relatively uniform illumination of the cylindrical section taken as 0.03~0.05, and the gradually changing illumination of the conical section taken as 0.05~0.08.

[0036] With σ 2 The gray value corresponding to the maximum (T) is the optimal initial segmentation threshold, which completes the binarization mapping and effectively reduces the influence of summer highlights, water surface and leaf specular reflection on the threshold shift, thus achieving accurate initial separation of the rice leaf roller target from the green background and barrel interference area.

[0037] 2. Gaussian kernel optimization segmentation: To address the issues of blurred edges and distorted depth features caused by strong summer sunlight, as well as the false deletion and missed detection of target bounding boxes due to reflections from the inner wall of the trap, a Gaussian scattering kernel optimization segmentation threshold was designed to suit the specific conditions of the trap structure (cylindrical + conical) and the characteristics of adult rice leaf rollers (8-10mm in length and 18-20mm in wingspan). The minimum Gaussian radius r was calculated through classification and discussion to ensure that the segmentation boundary fits the outline of the insect, thereby improving the accuracy and robustness of target segmentation.

[0038] h and w are the height and width of the target prediction bounding box for the rice leaf roller, respectively, set in conjunction with the insect's body size. Pixels Pixels; overlap is the intersection-union ratio of the predicted bounding box and the ground truth bounding box. Based on the imaging characteristics of insects in the trap, a threshold overlap ≥ 0.6 is set. If it is lower than this threshold, it is considered invalid interference, namely leaf fragments and reflective spots. r1, r2, and r3 are the Gaussian radii under three matching states. Considering the gradual change in light intensity inside the trapping bucket, the light intensity in the conical section is weaker than that in the cylindrical section. Therefore, a light correction coefficient is introduced to ensure that the Gaussian radius in different areas is adapted to the clarity of the insect imaging. An illumination adaptation correction coefficient β is introduced. For uniform illumination in the cylindrical section, β=1.0 is used, and for gradual illumination in the conical section, β=1.2 is used to correct the Gaussian radius and compensate for the blurring of the insect's edge in low-light areas.

[0039] Solving for the minimum Gaussian radius r through case-by-case analysis: a. When the corner of the prediction frame and the truth frame are externally tangent to the corresponding cylindrical section of the trapping bucket, the lighting is uniform, and the insect image is clear. Solve b. For the conical section of the trapping bucket where the predicted box is inscribed with the corner points of the ground truth box, the light is weak and the edges of the insect bodies are blurred. , Solve c. For the connection between the cylinder and the cone of the trapping bucket where the predicted box has both inscribed and circumscribed correspondences, the light changes suddenly and the imaging of the insect bodies is unstable. , Solve Take the minimum Gaussian radius in the three states to ensure that the Gaussian kernel fits the minimum contour of the insect body and avoid excessive blurring or missed detection: Combined with the light intensity in midsummer and the target scale of Cnaphalocrocis medinalis, set the adaptive variance of the target scale, and at the same time limit the minimum variance to avoid the failure of the kernel function under strong light: In the formula, 2.5 is the minimum variance threshold for fitting the minimum size of the Cnaphalocrocis medinalis insect body, which is different from the general Gaussian kernel variance and can effectively retain the edge features of small-sized insect bodies. Finally, optimize the segmentation threshold and introduce strong light suppression and barrel light correction: In the formula: T0 is the initial segmentation threshold obtained by the improved OTSU method mentioned above; d is the distance from the pixel to the initial segmentation boundary, which is used to correct the segmentation boundary to fit the contour of the insect body; is the adaptive variance of the target scale, obtained from formula (9); is the strong light suppression coefficient in midsummer, which is linked to α in the initial threshold, with a value range of 0.06 - 0.12, adapting to the light intensity in different regions of the trapping bucket; L(x, y) is the local brightness normalization value, which is the same as formula (4) and is used to suppress the interference of high-light and reflective regions on the threshold; is the fitting coefficient for the trapping bucket area, z is the height of the pixel in the barrel body, taking for the cylinder section where 0 ≤ z ≤ 80 cm, and for the conical section where 80 cm < z ≤ 100 cm, compensating for the threshold offset in the weak light area of the conical section to avoid misdeleting insect bodies.

[0040] (III) Motion Blur Elimination and Cross-Scale Enhancement 1. Motion Blur Removal: Addressing the dynamic blur and trailing issues caused by the flight of live rice leaf rollers (0.3~0.8 m / s) and airflow disturbances within the trapping bucket under strong light conditions in paddy fields, a global-local joint motion compensation strategy is employed, taking into account the imaging characteristics of the trapping bucket's cylindrical + conical structure. This is combined with an improved Wiener filter to achieve blur restoration. The global motion vector is calculated using Fourier cross-correlation, as shown in the formula: Local motion compensation is based on gradient similarity (GSIM) block matching, and the matching block size is set according to the body size of the rice leaf roller to avoid mismatches. For dynamic blurry images generated by the flight of live rice leaf rollers, a combined algorithm of inverse filtering and improved Wiener filtering is used for restoration. A dedicated blur kernel is constructed based on the characteristics of strong summer light noise; the calculation formula is as follows: In the formula, The frequency domain representation of the restored image; H(u,v) is the project-specific blur kernel function, which is constructed by the flight speed of the rice leaf roller, the camera frame rate of 30fps, and the light intensity inside the trapping bucket. Let H(u,v) be the conjugate complex number; The noise power spectrum under strong summer sunlight is adapted to the noise characteristics of leaf reflection and barrel glare. The power spectrum of the rice leaf roller image is determined by the yellow-brown grayscale features of the insect body; G(u,v) is the frequency domain representation of the blurred image; Equations 11 and 14 introduce correction coefficients of 0.015 and 0.02, respectively, to suppress strong light noise interference; The adaptation coefficient for the trapping barrel area is the same as Formula 10, to compensate for the differences in the degree of ambiguity in different areas; L(x,y) is the strong light suppression coefficient, and L(x,y) is the local brightness normalization value. The correlation coefficient between the matching block and the reference block. These are the grayscale standard deviations of the matching block and the reference block, respectively. To avoid extremely small values ​​where the denominator is zero, this project specifically sets the denominator to 10. -6 D represents the image pixel span corresponding to a trapping barrel diameter of 60cm.

[0041] 2. Cross-scale detail enhancement: Taking into account the tiny size of the insect, and addressing the issues of blurred details such as wing veins and body segments being weakened by strong light and having blurred edges after blurring restoration of the rice leaf roller, this device's algorithm incorporates 3D shearlet decomposition of the restored image to obtain low-frequency structural coefficients A and high-frequency detail coefficients D. j,k Formula for enhancing high-frequency coefficients: In the formula, k0 is the enhancement coefficient specific to this project, which is set to 0.8~1.5 based on the detailed characteristics of the insect body. In the strong light area, it is 1.2~1.5, and in the weak light area, it is 0.8~1.0. To avoid a minimum value where the denominator is zero, this project sets the value to 10. -5 ; Weights were assigned to enhance the details of the insect body, ranging from 0.9 to 1.1, to suit the detailed features of the wing veins and body segments of the rice leaf roller. j and k were the decomposition scale and direction coefficient, respectively, with the decomposition scale set to 3-4 levels based on the insect size. After reconstruction, the insect body edges were optimized using a fuzz enhancement algorithm to address edge blurring issues under strong summer sunlight. The formula is as follows: In the formula The pixel values ​​of the image after edge optimization; These are the pixel values ​​of the reconstructed image; The mean gray value of the 3×3 neighborhood of the pixel (x,y); The standard deviation of the gray level in the neighborhood. Used to suppress strong light interference and avoid excessive edge enhancement.

[0042] (iv) Illumination equalization and noise suppression 1. Adaptive Light Equalization: Addressing issues of uneven lighting, localized overexposure (leaf reflection, water glare), and gradual changes in light and dark within the trapping container in summer rice paddies, this method integrates multi-exposure feature weights and an improved CLAHE algorithm. First, it calculates the three-dimensional feature weights, phase consistency (PC), local contrast (G), and color saturation (S). Then, it optimizes the weight map using guided filtering. The final light equalization formula is as follows: In the formula, The image pixel values ​​are after illumination equalization; GuideFilter is the guiding filter function used to optimize the weight map and avoid blockiness. The feature weights for this project are set to 0.4, 0.3, and 0.3 respectively to accommodate the feature differences between the insect and the background; ClipHist is the contrast limiting function, with T set to 3-4 to prevent over-enhancement and loss of texture in the insect area; HistEq is the histogram equalization function, which reduces the weight of highlight areas after strong light correction. The region adaptation coefficient for the trapping barrel is used to compensate for the lighting differences between the cylindrical and conical sections; I(x,y) is the pixel value of the input image.

[0043] 2. Salt and Pepper Noise Removal: To address the salt and pepper noise generated by strong summer sunlight, camera noise, and dust interference inside the trapping container, an improved median filtering algorithm is employed. This algorithm, combined with the edge features of the rice leaf roller, dynamically selects a 3×3 or 5×5 filtering window to precisely remove salt and pepper noise from the image while preserving the insect's edge features. The formula is as follows: In the formula, is the pixel value of the filtered image; k is the adaptive window radius for field noise, which is adaptively set to 1 or 2 (1=3×3, 2=5×5) according to the noise intensity of the pixel to specifically filter out salt-and-pepper noise from the rice paddy image sensor, while completely preserving the insect's edge and wing veins; The edge protection coefficient for the insect body is 0.95 to 1.05, with 1.05 for edge regions and 0.95 to 1.0 for non-edge regions to avoid edge blurring. Median is the median calculation function.

[0044] This embodiment can also employ a combination of low-rank sparse decomposition and dynamic median filtering to first separate the signal from the noise: A represents the low-rank signal component of the rice leaf roller, and E represents the sparse noise component of instantaneous noise and point interference in the field; these provide a clean insect signal for subsequent dynamic median filtering.

[0045] Then perform dynamic windowed mid-range filtering on A: In the formula, L(x,y) Consistent with the previous text. L(x,y) is shown in the note of Equation 16. (See note 18) 3. Multi-dimensional Sharpness Screening: This step uses the Laplacian gradient variance method to quantitatively evaluate the sharpness of the preprocessed burst images, selecting the sharpest single frame as the benchmark image for target recognition. This method characterizes sharpness by calculating the variance value of the Laplacian operator in the image; the larger the Q value, the sharper the insect and the more complete its features. The calculation formula is: In the formula, For the Laplace operator, a 3×3 operator is used; Here, Q is the variance calculation function, and Q is the sharpness evaluation value. The larger the Q value, the sharper the image. The weighting coefficient for the insect body region is set to 1.2, while that for the background region is set to 0.8. This emphasizes the importance of insect body sharpness and avoids misselection of images with sharp backgrounds but blurry insect bodies. - The image with the highest Q value is selected for subsequent recognition processes to address the issue of unstable insect imaging under strong summer sunlight.

[0046] (v) Morphological optimization and feature-level fusion matching 1. Morphological optimization processing: Dilation-erosion morphological operations are sequentially performed on the binarized image to fill holes in the target area and eliminate residual small noise points in the background. The calculation formula is as follows:

[0047] In the formula, For expansion operations, For the erosion operation, S is a 3×3 structuring element used to fill the holes in the insect's outline, eliminate minor noise in the rice paddy background, and maintain the integrity of the insect's shape. For binarized images, The image after dilation. The final morphologically optimized image.

[0048] 2. Two-stage feature fusion matching: Extracting Hu invariant moment features from the target region after morphological processing. These features are invariant to translation, rotation, and scaling, effectively matching the extraction of rice leaf rollers in different flight postures. They are adaptable to any flight posture and shooting angle of the insect, ensuring stable matching. In the formula, For the central moment, For normalized central moments; The coordinates of the center point of the target region are calculated from the morphologically optimized region. The weighting coefficient for the insect body region is consistent with Equation 21.

[0049] The spatiotemporal feature similarity formula is as follows, used to calculate the similarity between the flight trajectory and attitude of the rice leaf roller between consecutive frames, to distinguish interfering targets such as rice planthoppers and rice leafhoppers, and to reduce false detections: The spatiotemporal similarity correction coefficient is set to 0.9~1.0 to match the flight speed of the insect.

[0050] The feature weighted fusion formula is as follows: it fuses morphological features and spatiotemporal features to construct a patent-specific feature vector for the rice leaf roller, thereby improving the robustness of recognition in complex field backgrounds.

[0051] The spatiotemporal feature fusion coefficient is set to 0.3.

[0052] Two-stage feature refinement: Features are extracted from the center points of each facet of the target bounding box and the initial localized center points using bilinear interpolation. These features are then input into a multilayer perceptron (MLP) for attribute refinement and confidence prediction. The final confidence level is: This represents a first-stage confidence level. This represents the two-stage confidence level.

[0053] The extracted Hu invariant moment features are input into a pre-defined feature library for rice leaf folder matching. When the MatchScore ≥ 90% and If the pest is identified as a target pest, non-target pests such as rice planthoppers are removed, including: These are the standard feature vectors from the rice leaf folder feature library; L(x,y) is consistent with the previous text and is used to suppress matching errors caused by strong light interference.

[0054] 3. Locate the geometric center of the rice leaf roller.

[0055] (I) Heatmap generation and peak optimization 1. Target Center Heatmap Construction: Based on the morphologically optimized target region, a two-dimensional Gaussian heatmap is constructed to highlight the characteristics of the target center. In the formula, Let these be the initial centroid coordinates. The target scale is adapted to the variance; The weighting coefficient for the insect body region is set to 1.2, and for the background region it is set to 0.8 to avoid interference from background pseudo-centers. The adaptation coefficient for the trapping barrel area is used to compensate for the thermal map response deviation caused by the difference in light intensity in different areas. The value is 1.0 for the cylindrical section and 1.1 for the conical section.

[0056] 2. Heatmap Peak Extraction: Circular region max-pooling is used to screen local peaks. Combined with the small body size characteristics of the rice leaf roller, pseudo-center interference caused by leaf reflection and noise is suppressed. In the formula, r is the pooling radius, which is 2.5 to 3.5 pixels to adapt to the adult body size of the rice leaf roller; This is the peak correction factor, ranging from 0.98 to 1.02, used to calibrate the peak attenuation problem in thermal maps under strong light; peak values ​​are retained. Candidate center points.

[0057] (II) Weighted centroid and offset compensation positioning 1. Weighted Centroid Calculation: By introducing pixel grayscale weights and heatmap response values, the accuracy of centroid calculation is optimized. The formula is as follows: In the formula, N is the total number of pixels in the target region, (x i ,y i I(x) represents the pixel coordinates within the target region. i ,y i () represents the grayscale value of the corresponding pixel; The heatmap response value of the corresponding pixel obtained by Equation 31; The grayscale adaptation weight for the insect body is set to a value of 1.0~1.15. The grayscale value is 45~75. The typical grayscale range of the insect body is set to 1.15, and the rest are set to 1.0 to highlight the pixel weight of the insect body.

[0058] 2. Center Point Offset Compensation: To compensate for downsampling quantization errors, the center point positioning offset is calculated and compensated based on the image pixel span D corresponding to a diameter of 60cm within the trapping bucket. An exclusive penalty coefficient is introduced into the offset loss function, which is as follows: In the formula, To predict the offset, R=4 is the network output step size, and p is the true center coordinate of the target. The offset compensation penalty coefficient is set to a value of 0.85~0.95, balancing the accuracy of offset calculation with anti-interference capability. This represents the weighting coefficient for the insect body region, avoiding background interference in the offset calculation. The final center point coordinates are: In the formula, These are the offset compensation values ​​in the x and y directions. Consistent with Equation 10.

[0059] (III) Optimization and Verification of Positioning Results The Focal loss function is used to supervise and optimize the center point localization results to ensure localization stability. In the formula, These are predicted values ​​from the heatmap. The values ​​of α=2 and β=4.3 are used as the true values ​​of the heat map and are substituted into the formula as hyperparameters to balance the positive and negative samples and the weights of the positive and negative samples under strong light. The positioning accuracy is optimized through backpropagation iteration, and the positioning error of the center point of the rice leaf roller is finally less than 1 pixel.

[0060] IV. Multi-view stereo vision 3D coordinate calculation of the geometric center of the rice leaf roller based on localization.

[0061] (I) Viewpoint orientation modeling and camera pose calculation 1. Establish a global coordinate system: with the center of the circular orbit as the origin O of the world coordinate system. w The orbital plane is Xw O w Y w The plane, perpendicular to the orbital plane, is Z. w Axis; Camera coordinate system O c -X c O c Y c With the camera's optical center as the origin, Z c The axis points along the camera's optical axis toward the inside of the trapping barrel.

[0062] 2. Azimuth and Pose Mapping: Based on the real-time azimuth angle θ of the camera i Solve for the camera extrinsic parameter matrix [R] corresponding to the i-th image. i |t i (Rotation matrix R) i Translation matrix t i ): In the formula, H = 120cm is the vertical distance from the camera's optical center to the track plane, ensuring that the camera can completely capture the insects inside the bucket, and θ i The angle between the camera and the initial position when the i-th image is captured is acquired in real time by the stepper motor encoder with an accuracy of 0.1°. The circular track radius is adapted to a trapping bucket diameter of 58cm, ensuring multi-angle coverage of the entire area inside the bucket; This is the translation correction factor, with a value ranging from 1.02 to 1.05, used to compensate for translation matrix deviations caused by camera installation errors.

[0063] (II) Multi-view target point feature matching and filtering 1. Initial target point localization: Incorporate illumination suppression terms and bucket region coefficients for each image I. i The "heatmap peak extraction + weighted centroid calculation" method was used to locate the pixel coordinates of the center point of the rice leaf roller target (u). i ,v i And calculate the location reliability conf. i : In the formula, The target region's gray-level variance. denoted as the grayscale variance of the background region, MatchScore(i) is the matching score between the Hu invariant moment feature and the preset feature library, and 0.62 is the weight coefficient.

[0064] 2. Cross-view constraint matching: ORB feature descriptors are used for target point cross-view matching, combined with azimuth constraints to filter valid matching pairs: Set azimuth difference threshold Confidence product threshold Retain matching pairs that meet the conditions ((u i ,v i ),(u j ,v j The RANSAC algorithm is used to remove mismatches and retain matching pairs that satisfy the epipolar constraints. , where F ij Let be the fundamental matrix for viewpoints i and j.

[0065] (III) Initial solution of three-dimensional coordinates based on triangulation 1. Projection matrix construction: combining camera intrinsic parameters K and extrinsic parameters Construct the projection matrix of the i-th image. The projection matrix is ​​in the form of: 2. Multi-view triangulation solution: For the M groups of valid matching pairs after screening, a nonlinear triangulation method is used to minimize the multi-view projection error and solve for the three-dimensional coordinates of the target point. : The formula introduces the attempt weight. A value of 1.1 is used if there is a target view, and 0.2 is used if there is no target; an orientation consistency check is introduced. ;s i Let be the scale factor of the i-th image, which is solved through iterative optimization.

[0066] Orientation consistency check: If the three-dimensional coordinates of the target point satisfy If the coordinates are correct, the solution is retained; otherwise, it is considered an outlier and discarded.

[0067] (iv) Global optimization and accuracy improvement 1. Gradient Concentration Optimization: Introducing the PVSO (Per-View Target-Free Suppression) strategy amplifies the gradient contribution of effective views, suppresses interference from target-free views, and optimizes the objective function: Introducing gradient amplification factor , For effective view weights.

[0068] 2. Bundle Adjustment Joint Optimization: Using camera pose and target point coordinates as optimization variables, the Levenberg-Marquardt (LM) algorithm is used to iteratively solve for the optimal solution. The camera intrinsic parameters are fixed, and only minor perturbations to the extrinsic parameters and the target point coordinates are adjusted. (v) Results Output and Verification 1. Output the optimized world coordinates of the target point. Calculate the relative positioning error: For the height of the trapping bucket, if σ≤1.8%, output the coordinate results, and connect the flight trajectory coordinates according to the timeline. Based on the image capture time sequence, output the calculated 3D coordinate points P1, P2, ... P... n The data is sorted, and cubic spline interpolation is used to connect discrete coordinate points to generate a smooth three-dimensional flight trajectory curve.

[0069] 2. Calculate individual and population behavioral parameters based on trajectory curves: a. Flight speed: in A speed smoothing coefficient is introduced to represent the shooting time interval between adjacent coordinate points. .

[0070] b. Activity radius: Introducing radius correction factor .

[0071] c. Trajectory curvature: Introducing a curvature calibration coefficient Where L is the total trajectory length and D is the straight-line distance between the starting and ending points. Clustering: For multi-target trajectories, calculate the volume ratio of the spatial overlap area between any two target trajectories; the higher the ratio, the stronger the clustering. The trajectory curves are overlaid onto the spatial coordinate system of the trap bucket, and key point information is marked.

[0072] If the accuracy requirement is not met, expand the matching view window K and repeat steps (ii) to (iv) to recalculate. Example

[0073] Based on Examples 1 and 2, this example provides a specific field location method for rice leaf rollers based on multi-view stereo vision. See [link to example]. Figure 4 In its specific implementation, it can be achieved through the following steps: I. Assembly and Debugging of the Device.

[0074] 1. Select acrylic material with a light transmittance of ≥90% to manufacture a cylindrical-cone shaped trapping and collection bucket with an upper cylinder and an inverted funnel shape at the bottom. The cylindrical bucket is 80cm high, the inverted funnel bucket is 20cm high, the upper diameter is 60cm, and the lower diameter is 37cm.

[0075] 2. Install a 2cm wide stainless steel annular slide coaxially on the outside of the trapping tank. Mount a 3840×2748 high-definition industrial camera onto the slide using a stepper motor drive. Adjust the stepper motor to drive the camera in a 360-degree circular motion along the slide at a speed of 5cm / s.

[0076] 3. Fix a 100cm×100cm×2cm matte green background board on the opposite side of the high-definition camera, ensuring that the center of the background board is on the same horizontal line as the center of the trapping bucket and parallel to the camera lens.

[0077] 4. Electrically connect the STM32 microcontroller to the high-definition camera and stepper motor. Set the camera's continuous shooting parameters to 10 frames per second, a single working time of 1320 seconds, and a single stop time of 480 seconds. Debug the data processing module to ensure that the image processing algorithm and the 3D imaging algorithm are running normally.

[0078] 5. Place the assembled device in an area of ​​rice paddy with high incidence of pests, fix the bottom of the device, and ensure that the trapping bucket is vertical.

[0079] II. Field observation and data analysis.

[0080] s1: The insect entering the device is the rice leaf roller.

[0081] Adult rice leaf rollers are attracted by the bait in the trap and enter the trap through the opening at the bottom of the "tube-cone integrated" trap. The main controller triggers the image acquisition program, and the stepper motor drives the high-definition camera to move along the circular slide. The camera simultaneously takes high-speed continuous shots.

[0082] Assume that when the camera moves to an azimuth angle θ1=30°, it captures the first image I1 containing the rice leaf roller; when it moves to an azimuth angle θ2=90°, it captures the second image I2; and when it moves to an azimuth angle θ3=150°, it captures the third image I3.

[0083] After the images are transmitted to the data processing module, they undergo preprocessing such as distortion correction, grayscale conversion, and adaptive threshold segmentation. Then, the optimal clear frame of the three images is selected by using the Laplacian gradient variance method.

[0084] Extract Hu invariant moment features from the best clear frame and match them with a preset feature library. If the MatchScore is ≥ 90%, it is determined to be the target pest.

[0085] A two-dimensional Gaussian heatmap of the target center is constructed. The peak value is extracted by max pooling in the circular region. The weighted centroid is calculated by combining the pixel gray-level weights to obtain the pixel coordinates (u1, v1) of the target center point in image I1, (u2, v2) in image I2, and (u3, v3) in image I3.

[0086] The camera extrinsic matrix [R1|t1], [R2|t2], and [R3|t3] corresponding to each azimuth angle are calculated, and the projection matrix is ​​constructed by combining it with the camera intrinsic matrix K. The multi-view projection error is minimized by using a nonlinear triangulation method, and the world coordinates of the target point P1(X1,Y1,Z1), P2(X2,Y2,Z2), and P3(X3,Y3,Z3) at the three azimuth angles are calculated in sequence.

[0087] After joint optimization by Bundle Adjustment, the relative positioning error σ≤2% was verified. Points P1, P2, and P3 were connected in chronological order to generate the preliminary flight trajectory of the rice leaf roller in the trapping bucket.

[0088] s2: The insect entering the device is not the rice leaf roller.

[0089] Once a non-target pest enters the trap, the camera captures images according to a preset program and transmits them to the data processing module. After preprocessing and feature extraction, the data processing module inputs the Hu invariant moment features into a feature library for matching. If the matching score (MatchScore) is less than 90%, the system determines the insect is a non-target disturbance and immediately terminates the subsequent center point localization and 3D coordinate calculation process. The storage module only records the image acquisition information and the determination result of "not rice leaf roller, analysis terminated," without generating any location data or trajectory map.

[0090] s3: Two insects enter the device simultaneously.

[0091] When two insects enter the trap simultaneously, the same image captured by the camera contains two target regions, and the main controller maintains a normal shooting cycle. After image preprocessing, morphological dilation-erosion operations are used to separate the two independent target regions, which are labeled as target A and target B, respectively. Hu invariant moment features are extracted from both target regions and feature library matching is performed: if the matching score for target A is ≥90% and the matching score for target B is <90%, then only target A is localized for its center point and its 3D coordinates are calculated, generating its flight trajectory, while target B is labeled as an interfering target. If the matching scores for both target A and target B are ≥90%, then the center point pixel coordinates of the two targets in images at different azimuth angles are calculated, and their respective world coordinates P are determined. A1 P A2 P A3 and P B1 P B2 P B3Two independent flight paths are generated. Simultaneously, the overlap and directional consistency of the two paths are calculated. If the overlap is ≥30%, it is determined to be population aggregation behavior; if the angle between the directions of the two paths is ≤15°, it is determined to be a population diffusion trend in the same direction. The relevant parameters are stored in a feature parameter data file for calibration of the "aggregation-diffusion coefficient" in the population dynamics model. If the matching scores of both target A and target B are <90%, they are determined to be dual-interference targets, all analysis processes are terminated, and the determination results are recorded.

[0092] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A field observation device for rice leaf roller based on multi-view stereo vision, characterized in that, include: Trapping units are used to trap rice leaf rollers; The image acquisition unit is used to capture image information of the trapped rice leaf roller at different flight times; The background processing unit is used to differentiate the image from the trapped rice leaf roller and remove image interference when capturing image information of the trapped rice leaf roller at different flight times. The processing and computing unit is equipped with improved image processing and 3D positioning algorithms to control the shooting angle of the image acquisition unit. It also calls the improved image processing and 3D positioning algorithms to process the image information of the rice leaf roller at different flight times, calculates and saves the individual behavior parameters and population behavior parameters of the rice leaf roller.

2. The field observation device for rice leaf roller based on multi-view stereo vision according to claim 1, characterized in that, The trapping unit is a "cylindrical-cone integrated" rice leaf roller observation and trapping barrel with an upper cylindrical shape and an inverted funnel shape at the bottom. The barrel body is made of acrylic material with a light transmittance of ≥90%. The cylindrical barrel body is 80cm high, the inverted funnel barrel is 20cm high, the upper opening diameter is 58cm, and the lower opening diameter is 330cm. The upper opening of the trapping barrel is tightly connected to the roof. The food attractant is placed on the inner edge of the lower opening of the trapping barrel, which is 2cm wide. The barrel body is 1cm thick overall.

3. The field observation device for rice leaf roller based on multi-view stereo vision according to claim 2, characterized in that, The image acquisition unit includes a ring slide, a moving drive component, and a high-definition industrial camera. The annular slide is coaxially fitted onto the outside of the trapping barrel. The annular slide is 1cm wide and 2cm high, and is made of stainless steel. The moving drive component is a stepper motor drive component, which is installed on the annular slide. The high-definition industrial camera mounted on the stepper motor drive component is suspended under the annular slide. The stepper motor drive component is used to drive the high-definition industrial camera and the baffle to make a 360-degree circular motion along the annular slide at a speed of 5cm / s, so as to realize the full-view image acquisition of rice leaf roller in the trapping bucket. The high-definition industrial camera has a resolution of 3840×2748 and a frame rate of 30 FPS. The high-definition industrial camera and the baffle are symmetrical about the central axis of the trapping barrel.

4. The field observation device for rice leaf roller based on multi-view stereo vision according to claim 3, characterized in that, The high-definition industrial camera is triggered by a built-in timer to capture image information of the rice leaf roller at different flight times by taking 10 consecutive shots per second. The high-definition industrial camera stops taking pictures after 22 minutes of surround shooting, and then performs 8 minutes of image processing and saving by the processing and computing unit.

5. The field observation device for rice leaf roller based on multi-view stereo vision according to claim 2, characterized in that, The background processing unit is a matte leaf-green background panel, measuring 100cm x 100cm and 2cm thick. This background panel is positioned symmetrically to the high-definition industrial camera along the axis of the trapping barrel, parallel to the camera lens, with the center of the background panel and the center of the trapping barrel on the same horizontal line.

6. The field observation device for rice leaf roller based on multi-view stereo vision according to any one of claims 1-5, characterized in that, The processing and computing unit includes a main controller, a data processing module, and a storage module; The main controller uses an STM32 microcontroller, which is electrically connected to the high-definition industrial camera and the stepper motor driver. The main controller controls the working state of the stepper motor to drive the high-definition industrial camera to move along the track and adjust the shooting angle. The stepper motor drives the high-definition industrial camera to make a 360-degree circular motion along the circular track. The main controller is used to control the start and stop cycle of the high-definition industrial camera: the single working time is set to 1320s, during which the camera can move at a constant speed along the circular slide for 30 full circles, and return to the initial position after the movement ends; the single stop time is set to 8 minutes, which is used by the storage module to store the image information captured in this working cycle. The data processing module has built-in improved image processing and 3D imaging algorithms to ensure their normal operation. The data processing module completes image analysis and 3D coordinate calculation within 8 minutes, and obtains and saves the individual behavior parameters and population behavior parameters of the rice leaf roller.

7. The field observation device for rice leaf roller based on multi-view stereo vision according to claim 6, characterized in that, The individual behavioral parameters of the rice leaf roller include flight speed, activity radius, and trajectory curvature, while the population behavioral parameters include multi-target trajectory overlap and trajectory direction consistency.

8. A field positioning method for rice leaf roller based on multi-view stereo vision, applied to the rice leaf roller field observation device based on multi-view stereo vision as described in any one of claims 1-7, characterized in that, Includes the following steps: Construct the overall spatial coordinate system of the trapping barrel; Image information of trapped rice leaf rollers at different flight times was captured, and image preprocessing and cross-scale enhancement were performed. Locate the geometric center of the rice leaf roller; Based on the geometric center of the rice leaf roller located by positioning, multi-view stereo vision 3D coordinate calculation is performed.

Images