A control method for three-dimensional tracking and dynamic priority scheduling of plant diseases and insect pests

Abstract

The application discloses a control method for three-dimensional tracking and dynamic priority scheduling of plant diseases and insect pests, which comprises the following steps: acquiring a scene image in real time through an image acquisition device; screening a high-quality target image frame through a memory selection mechanism, and performing image recognition through a target recognition algorithm; calculating a dynamic priority score of the target through a priority arbitration algorithm; sending a control command to a scheduling module based on the arbitration result, switching to a target locking tracking mode, and driving a lower computer to execute a corresponding processing task by a control module; and constructing a technical system of "intelligent memory screening-multi-target priority arbitration-three-dimensional dynamic tracking-mode adaptive switching", solving the problems of low monitoring efficiency, insufficient tracking accuracy, multi-target processing conflict and poor data effectiveness of existing equipment, and realizing accurate positioning, dynamic tracking and intelligent prevention and control decision support of plant diseases and insect pests.

Description

Technical Field

[0001] This invention relates to the field of agricultural pest and disease monitoring and intelligent control, and in particular to a method for three-dimensional tracking and dynamic priority scheduling control of pests and diseases. Technical Background

[0002] Traditional pest and disease monitoring mainly relies on manual inspections, fixed-point sensor monitoring, or single-dimensional device scanning, which has many limitations. Manual inspections are inefficient and have limited coverage, and are constrained by terrain and crop density, making real-time dynamic monitoring difficult. Fixed sensor networks can only cover local areas and cannot actively track mobile pests and diseases. Existing monitoring equipment lacks a multi-target priority arbitration mechanism, easily becoming "stuck" to a single target and ignoring pests and diseases with higher levels of damage. Tracking methods are mostly two-dimensional planar positioning, lacking the ability to adapt to three-dimensional "close-up observation - wide-range overview," affecting monitoring accuracy. There is a lack of efficient search strategies after a target is lost, which can easily lead to monitoring interruptions. At the same time, the data processing does not perform intelligent filtering of target frames, and invalid data in blurred, deformed, or occluded states can easily interfere with tracking stability, making it difficult to achieve intelligent management of the entire process from positioning and tracking to trend prediction.

[0003] With the development of computer vision and intelligent control technology, target recognition technology based on visual language models has been applied to pest and disease monitoring. However, existing systems have not solved core problems such as multi-target conflict arbitration, three-dimensional dynamic tracking, target loss recovery, and effective data screening, making it difficult to meet the needs of precise monitoring in complex farmland environments. Summary of the Invention

[0004] Based on the above problems, this invention proposes a three-dimensional tracking and dynamic priority scheduling control method for pests and diseases. The proposed method addresses the problems of low monitoring efficiency, insufficient tracking accuracy, multi-target processing conflicts, and poor data validity of existing equipment by constructing a technical system of "intelligent memory screening, multi-target priority arbitration, three-dimensional dynamic tracking, and adaptive mode switching." This enables precise location, dynamic tracking, and intelligent prevention and control decision support for pests and diseases.

[0005] The technical solution adopted is a control method for three-dimensional tracking and dynamic priority scheduling of pests and diseases, the steps of which are as follows: S1, enter global scanning mode, and acquire scene images in real time through the image acquisition device; S2, the acquired scene images are input into the memory storage station, the memory selection mechanism is used to filter high-quality target image frames, and the target recognition algorithm recognizes the target image frames and outputs structured recognition content information; S3, based on the identified content information, obtain the target category and implement a differentiated tracking and identification strategy; S4, the identified content information is sent to the priority arbitration algorithm to calculate the dynamic priority score of the target and realize the arbitration of multi-target conflicts; The priority arbitration algorithm, deployed on edge hardware devices, calculates the dynamic priority score of the target based on dynamic weights and the real-time status of the target, thereby realizing multi-target conflict arbitration. S5, the tracking and control system sends control commands to the scheduling module based on arbitration information, switches from global scanning mode to target locking and tracking mode, and the control module drives the lower-level machine to execute corresponding processing tasks; S6. In target lock and tracking mode, if the target is lost, enter lost target search mode. The control module drives the lower device to perform a small-range search in the direction where the target last appeared. If the target is detected again during the search, the target lock and tracking mode is restored. If the target is not detected after the timeout, enter global scan mode.

[0006] Furthermore, the image acquisition device includes a camera, a motion platform for carrying the camera, an attitude sensor, a distance sensor, and an optical flow sensor; the camera is mounted on the motion platform, which includes, but is not limited to, a three-axis gimbal and an electric lifting rod; the attitude sensor acquires the angle information of the motion platform; the distance sensor and the optical flow sensor acquire target distance and image motion-related parameters.

[0007] Furthermore, the identified content information includes, but is not limited to, a structured data frame containing target ID, target type, hazard level, confidence level, and identification duration; the hazard level is obtained by querying based on preset pest and disease types; the confidence level is the confidence level output by the target identification algorithm; the identification duration is the duration for which the target is stably tracked; the target identification algorithm includes, but is not limited to, the Yolov8n model.

[0008] Furthermore, the memory selection mechanism is deployed on edge hardware devices and is equipped with a static weight model and an adaptive weight model; the static weight model dynamically allocates weights for relatively stationary objects such as plants; the adaptive weight model dynamically allocates weights for non-relatively stationary organisms such as insects. The memory selection mechanism selects scene images acquired by the image acquisition device and assigns them to an affinity score A(f). t ) and athletic score M(f t Based on the target category, different weighting models are used to calculate the comprehensive quality score of the target image frame. Filter and store high-quality target image frames; Affinity score A(f) t The calculation formula for ) (1) Affinity score A(f) tThe scoring criteria are based on the sharpness of the target image and its similarity to the target images in the memory module. In Formula 1, C(f) t S(f) represents the image sharpness score. t M) is the similarity score, α is the clarity weight, and β are the similarity weights, and α+β=1; Image sharpness score C(f) t The gradient magnitude is measured by normalizing it using the Frobenius norm; Image sharpness score C(f) t The calculation formula for ) (2) In formula 2, Indicates the current image gradient, This represents the maximum gradient value of image I in the memory bank; Similarity score S(f) t The feature vector of the current frame and the feature vector of the target image in the memory storage station are extracted, and the maximum cosine similarity is calculated. (3) In Formula 3, f t A feature or image patch representing the target image in frame t; This represents a memory bank, which stores a collection of historical plant or insect samples m; This represents the maximum value obtained from set M; Represents the feature extraction function; Motion scoring uses the motion speed and motion blur degree of the target image as scoring criteria. (4) In Formula 4, M(f) t ) represents the motion score for frame t. Represents the target image velocity in frame t. This represents the theoretical maximum speed of motion of the current target image. This represents the degree of motion blur in the target image of frame t. This indicates the maximum blur level set for the target image; The target image speed By analyzing the displacement response between two frames, and calculating the two-dimensional pixel displacement of the target image in the previous and current frames, the impact of the current motion on image quality can be reflected. (5) In formula 5, Let be the position vector of the target image in frame t. Let be the position vector of the target image in frame t-1; Motion image blur By calculating the average absolute difference between each pixel in the target region in the current frame and the pixel value of the corresponding position of the target in the previous frame, the larger the value, the greater the motion blur. (6) In formula 6, This represents the target image in the current frame. This represents the target image from the previous frame. Represents the set of pixels in the target region. This indicates the number of pixels in the target region. Represents pixel coordinates, The motion displacement of the target image from frame t-1 to frame t can be obtained from the optical flow sensor; When the target image frame is identified as a plant, a static weight model is executed. Weights are assigned to process the static target image, and a comprehensive quality score is calculated. (7) In formula 7, The overall quality score for frame t. Indicates the weight of the affinity score. Indicates the weight of the sports score; When the target image frame is identified as an insect, an adaptive weight model is executed. This model uses the sigmoid activation function to dynamically adjust the weights based on the deviation between the current affinity score and the historical average affinity score. (8) In formula 8, Affinity weight for frame t, The historical average of affinity. The historical average of athletic scores. , , For activation functions; (9) Motion score weights for frame t. Historical average affinity The calculation uses the average method to calculate the average affinity of historical frames in the memory storage station. (10) Historical average of athletic scores The calculation is similar. (11) In formulas 10 and 11, The length of historical memory data; Furthermore, the memory selection mechanism for filtering high-quality target image frames also includes storage decisions, memory station update strategies, memory station replacement mechanisms, and time decay storage rules. Storage decisions are made by comparing the overall quality score of the current plant frame with a pest and disease threshold, which makes the stringency of the storage threshold change over time. Storage decision (12) In formula 12, This indicates the overall quality score of the current plant frame. Indicates the threshold for pests and diseases, which changes over time; If the overall quality score of the current plant frame is ≥Pest and disease threshold If so, store the frame and proceed to the next filtering step; (13) In formula 13, The average quality score for the history of pests and diseases. Standard deviation of pest and disease quality scores over a period of time. This is an adjustable parameter used to control the leniency of the pest and disease threshold. The larger the size, the higher the threshold for pests and diseases, and the more stringent the storage conditions. Average quality score of pest and disease history , (14) Standard deviation of pest and disease quality scores over a period of time , (15); The memory station update strategy classifies the current frame into high-quality, medium-quality, and low-quality frames based on its quality score. For medium-quality frames, they can only be stored if their maximum similarity with the maximum number of frames in the memory storage station is less than a preset diversity threshold. Diverse decision functions , (16) The maximum similarity between the target image and the memory storage station; This is the diversity coefficient. The larger the coefficient, the higher the allowable similarity threshold, and the lower the requirement for diversity. Determining the conditions under which storage is permissible. (17) In formula 17, For high-quality thresholds, Low quality threshold; Ensure the storage of high-quality samples. For medium-quality samples, they can only be stored if they provide new information. Low-quality samples should be discarded directly. The memory station replacement mechanism first selects the worst sample in the current memory station when the memory storage station reaches its capacity and new samples need to be stored. Then, it compares the worst sample with the new sample based on replacement conditions to decide whether to replace the sample. The selection of the worst sample takes into account both the sample quality and the sample storage time. (18) In formula 18, As the worst sample, For time decay weight, Given the current mass of sample m, Current system time Sample m storage time, Here, is the decay constant and the time normalization factor. The smaller the value, the faster the time penalty increases; The method for determining the conditions for sample replacement is to replace the worst sample in the memory storage with the new sample when the quality gain of the new sample compared to the worst sample is greater than the replacement cost. (19) In formula 19, The overall quality score for the current frame. The overall quality score for the worst sample. For mass gain, To replace the tolerance factor, The similarity between the new sample and the worst sample; The time-decay storage rule states that the quality of sample m gradually decays over time t. (20) In formula 20, Let m be the mass of sample m after decay at time t. The quality when storing sample m. The attenuation coefficient is... Let m be the storage time for sample m. The time point is stored for sample m.

[0009] Furthermore, the dynamic priority score p is calculated as follows: Let there be N targets to be arbitrated, forming a set. (21), Formula 21 is the set formula for the targets to be arbitrated. Set weighting coefficients, weighting coefficient formula , (twenty two) Weighting is assigned to the severity level. Weights are assigned to confidence levels. Weighting is applied based on duration; Current goal (i=1,2,...,N), its dynamic priority score , (twenty three) In formula 23, To retrieve the hazard level score from the preset pest and disease type score table based on the target type, To identify the confidence score output by the algorithm, The duration for which the target is stably tracked; Hazard level classification , (twenty four) In formula 24, Let i be the type of the i-th target pest or disease. Use the preset mapping function to query the predefined pest and disease type score table; Confidence score , (25) In formula 25, The confidence level output by the target recognition algorithm; Duration , (26) The duration decreases linearly, as shown in Equation 26. For the goal Duration of stable tracking The attenuation coefficient; Calculate the target score for the current overall score, and make the final decision. As the current top priority target for tracking .

[0010] Furthermore, the control module drives the gimbal and lifting mast to perform corresponding processing tasks; specifically, the control module calculates the deviation between the target bounding box coordinates and the center of the image, calculates the PWM drive signal using a PID algorithm, and adjusts the gimbal's horizontal and vertical angles to ensure that the angular position deviation is less than a preset threshold; the control module compares the target's current pixel area with the preset ideal observation area and outputs lifting commands; specifically, If the current pixel area is less than the ideal observation area, control the lifting rod to descend at the rated speed and approach the target; If the current pixel area is greater than the ideal observation area, the lifting rod is controlled to rise at the rated speed to widen the viewing angle; if the deviation between the current pixel area and the ideal observation area is less than the rated value, the lifting rod remains stationary.

[0011] Furthermore, the switching logic and execution method of the working mode, specifically, the global scan mode, is entered by default after the system starts, and performs a region-wide scan. In target locking and tracking mode, after the memory storage module has completed the screening and the priority arbitration algorithm determines the optimal target, the system will start the tracking and control system. In the target loss search mode, when the target recognition module fails to detect the currently tracked target for several consecutive frames, the switch is triggered. The control module drives the gimbal to perform a short-term "Z"-shaped oscillation search within a small area where the target last appeared. If the target is detected again during the search, the target lock tracking mode is restored. If the target is not detected within the timeout period, the mode automatically returns to the global scan mode.

[0012] Furthermore, differentiated tracking and identification strategies, specifically... For tracking insect pests and diseases, based on target coordinate data from multiple consecutive frames during the pest's movement, the Kalman filter algorithm predicts the position of the next frame, allowing for pre-adjustment of the gimbal and lifting mast to achieve predictive tracking. Tracking; combining data from gimbal angle, lifting rod height, and distance sensors, periodically record the target's three-dimensional coordinates and analyze its activity trajectory; when the pest is stationary, accurately calculate the target's three-dimensional coordinates, and record its harmful behaviors such as feeding and egg-laying through image recognition to generate a behavior log; Track plant diseases and pests, record the location, size, and duration of plant lesions, combine environmental parameters to conduct source tracing analysis, determine the causes of disease, and predict the spread range in the next few days through a diffusion model; When there is no current target being tracked, the target with the highest dynamic priority score is selected directly, triggering a switch from global scanning mode to target lock tracking mode. When there is a current target being tracked, the difference between the dynamic priority score of the new target and the dynamic priority score of the current target image is calculated. If the difference is greater than or equal to the preset switching threshold γ, the current target is abandoned, and a switch to the lock tracking mode of the new target is triggered.

[0013] The image acquisition device in this invention can be placed on a wheeled agricultural robot; this invention is based on tracking, control and monitoring task execution. When a new device is connected, its onboard control module can automatically call the new device and complete the corresponding function application, thereby realizing the expansion of system functions.

[0014] This invention employs a memory selection mechanism to filter high-quality data, ensuring robust tracking; a priority arbitration algorithm enables intelligent sorting of multiple targets, preventing low-value targets from consuming monitoring resources; a differentiated tracking and identification strategy, combined with PTZ PID control and boom linkage, improves target monitoring accuracy; three working modes adaptively switch to ensure continuous monitoring; and the differentiated tracking strategy adapts to different types of pests and diseases, expanding application scenarios. The entire system achieves intelligent operation throughout the entire process, from target identification and dynamic tracking to data fusion and early warning decision-making, significantly improving the efficiency and accuracy of pest and disease monitoring, reducing manual intervention, and providing technical support for green pest control in agriculture.

[0015] This invention addresses agricultural pest and disease monitoring scenarios by constructing a comprehensive technical system encompassing "multi-dimensional perception—intelligent decision-making—three-dimensional tracking—memory optimization—closed-loop management." It achieves pest and disease target detection and localization through image acquisition and multi-sensor fusion, and utilizes a dynamic memory filtering mechanism based on affinity and motion scores to ensure the effectiveness and stability of tracking data. A priority arbitration algorithm enables intelligent sorting and scheduling of multi-target conflicts, and a three-dimensional tracking strategy linking gimbal leveling and boom height adjustment is combined. Ultimately, multi-source data fusion forms a complete closed loop of "monitoring-identification-tracking-decision-making."

[0016] The beneficial effects of this invention are as follows: By introducing a memory bank module and a dynamic memory selection mechanism, this invention significantly improves the image quality received by the target recognition algorithm, thereby greatly enhancing the stability and accuracy of pest and disease target identification and detection. Based on a dynamic priority intelligent arbitration mechanism of hazard level, confidence level, and tracking duration, combined with adaptive mode switching between global search, lock-on tracking, and lost target search, it achieves efficient scheduling and optimized resource allocation for multiple targets. Simultaneously, the system possesses robust lost target recovery capabilities and differentiated tracking strategies for insect and plant pests and diseases, fundamentally solving the bottlenecks of rigid tracking modes and lack of dynamic multi-target scheduling capabilities in traditional crop monitoring. The efficient utilization of the target recognition algorithm and optimized resource allocation ultimately achieve autonomous, precise, continuous, and efficient intelligent monitoring and control of agricultural pests and diseases. Attached Figure Description

[0017] Fig. 1 This is a control architecture diagram of the present invention; Fig. 2 This is a diagram illustrating the mode switching logic in this invention. Fig. 3This is a flowchart of the memory selection mechanism in this invention; Fig. 4 This is a diagram of the differential tracking and identification strategy in this invention. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The following embodiments are explanations of the present invention, but the present invention is not limited to the following embodiments.

[0019] See Figs. 1 to 4 As shown, a control method for three-dimensional tracking and dynamic priority scheduling of pests and diseases includes the following steps: S1, enter global scanning mode, and acquire scene images in real time through the image acquisition device; the image acquisition device includes a camera, a motion platform for carrying the camera, an attitude sensor, a distance sensor and an optical flow sensor; the camera is mounted on the motion platform, which includes, but is not limited to, a three-axis gimbal and an electric lifting rod; the attitude sensor acquires the angle information of the motion platform; the distance sensor and the optical flow sensor acquire target distance and image motion-related parameters.

[0020] S2, the acquired scene images are input into the memory storage station, and the memory selection mechanism therein is used to filter high-quality target image frames. The target image frames are then identified by the target recognition algorithm, and structured recognition content information is output. The recognition content information includes, but is not limited to, a structured data frame containing target ID, target type, hazard level, confidence level, and recognition duration. The hazard level is obtained by querying based on preset pest types. The confidence level is the confidence level output by the target recognition algorithm. The recognition duration is the duration for which the target is stably tracked. The target recognition algorithm includes, but is not limited to, the Yolov8n model.

[0021] The memory selection mechanism is deployed on edge hardware devices and is equipped with a static weight model and an adaptive weight model. The static weight model dynamically allocates weights for relatively stationary objects such as plants, while the adaptive weight model dynamically allocates weights for non-relatively stationary organisms such as insects.

[0022] S3, based on the identified content information, obtains the target category and implements a differentiated tracking and identification strategy; for insect pest tracking, based on the target coordinate data of multiple consecutive frames during the pest's movement, it predicts the position of the next frame using a Kalman filter algorithm, and adjusts the gimbal and lifting arm in advance to achieve predictive tracking. Tracking; combining data from gimbal angle, lifting rod height, and distance sensors, periodically record the target's three-dimensional coordinates and analyze its activity trajectory; when the pest is stationary, accurately calculate the target's three-dimensional coordinates, and record its harmful behaviors such as feeding and egg-laying through image recognition to generate a behavior log; Track plant diseases and pests, record the location, size, and duration of plant lesions, combine environmental parameters to conduct source tracing analysis, determine the causes of disease, and predict the spread range in the next few days through diffusion models.

[0023] S4, the identified content information is sent to the priority arbitration algorithm to calculate the dynamic priority score of the target and realize the arbitration of multi-target conflicts; The priority arbitration algorithm, deployed on edge hardware devices, calculates the dynamic priority score of the target based on dynamic weights and the real-time status of the target, thereby realizing multi-target conflict arbitration. The dynamic priority score p is calculated in the following steps: Let there be N targets to be arbitrated, forming a set. (21), Formula 21 is the set formula for the targets to be arbitrated. Set weighting coefficients, weighting coefficient formula , (twenty two) Weighting is assigned to the severity level. Weights are assigned to confidence levels. Weighting is applied based on duration; Current goal (i=1,2,...,N), its dynamic priority score , (twenty three) In formula 23, To retrieve the hazard level score from the preset pest and disease type score table based on the target type, To identify the confidence score output by the algorithm, The duration for which the target is stably tracked; Hazard level classification , (twenty four) In formula 24, Let i be the type of the i-th target pest or disease. Use the preset mapping function to query the predefined pest and disease type score table; Confidence score , (25) In formula 25, The confidence level output by the target recognition algorithm; Duration , (26) The duration decreases linearly, as shown in Equation 26. For the goal Duration of stable tracking The attenuation coefficient; Calculate the target score for the current overall score, and make the final decision. As the current top priority target for tracking .

[0024] S5, the tracking and control system sends control commands to the scheduling module based on arbitration information, switches from global scanning mode to target locking and tracking mode, and the control module drives the lower-level machine to execute corresponding processing tasks; S6. In target lock and tracking mode, if the target is lost, enter lost target search mode. The control module drives the lower device to perform a small-range search in the direction where the target last appeared. If the target is detected again during the search, the target lock and tracking mode is restored. If the target is not detected after the timeout, enter global scan mode.

[0025] The memory selection mechanism selects scene images acquired by the image acquisition device and assigns them to an affinity score A(f). t ) and athletic score M(f t Based on the target category, different weighting models are used to calculate the comprehensive quality score of the target image frame. Filter and store high-quality target image frames; Affinity score A(f) t The calculation formula for ) (1) Affinity score A(f) t The scoring criteria are based on the sharpness of the target image and its similarity to the target images in the memory module. In Formula 1, C(f) t S(f) represents the image sharpness score. t M) is the similarity score, α is the clarity weight, and β are the similarity weights, and α+β=1; Image sharpness score C(f) t The gradient magnitude is measured by normalizing it using the Frobenius norm; Image sharpness score C(f) t The calculation formula for ) (2) In formula 2, Indicates the current image gradient, This represents the maximum gradient value of image I in the memory bank; Similarity score S(f) tThe feature vector of the current frame and the feature vector of the target image in the memory storage station are extracted, and the maximum cosine similarity is calculated. (3) In Formula 3, f t A feature or image patch representing the target image in frame t; This represents a memory bank, which stores a collection of historical plant or insect samples m; This represents the maximum value obtained from set M; Represents the feature extraction function; Motion scoring uses the motion speed and motion blur degree of the target image as scoring criteria. (4) In Formula 4, M(f) t ) represents the motion score for frame t. Represents the target image velocity in frame t. This represents the theoretical maximum speed of motion of the current target image. This represents the degree of motion blur in the target image of frame t. This indicates the maximum blur level set for the target image; The target image speed By analyzing the displacement response between two frames, and calculating the two-dimensional pixel displacement of the target image in the previous and current frames, the impact of the current motion on image quality can be reflected. (5) In formula 5, Let be the position vector of the target image in frame t. Let be the position vector of the target image in frame t-1; Motion image blur By calculating the average absolute difference between each pixel in the target region in the current frame and the pixel value of the corresponding position of the target in the previous frame, the larger the value, the greater the motion blur. (6) In formula 6, This represents the target image in the current frame. This represents the target image from the previous frame. Represents the set of pixels in the target region. This indicates the number of pixels in the target region. Represents pixel coordinates, The motion displacement of the target image from frame t-1 to frame t can be obtained from the optical flow sensor; When the target image frame is identified as a plant, a static weight model is executed. Weights are assigned to process the static target image, and a comprehensive quality score is calculated. (7) In formula 7, The overall quality score for frame t. Indicates the weight of the affinity score. Indicates the weight of the sports score; When the target image frame is identified as an insect, an adaptive weight model is executed. This model uses the sigmoid activation function to dynamically adjust the weights based on the deviation between the current affinity score and the historical average affinity score. (8) In formula 8, Affinity weight for frame t, The historical average of affinity. The historical average of athletic scores. , , For activation functions; (9) Motion score weights for frame t. Historical average affinity The calculation uses the average method to calculate the average affinity of historical frames in the memory storage station. (10) Historical average of athletic scores The calculation is similar. (11) In formulas 10 and 11, The length of historical memory data; The memory selection mechanism filters high-quality target image frames, and also includes storage decisions, memory station update strategies, memory station replacement mechanisms, and time decay storage rules. Storage decisions are made by comparing the overall quality score of the current plant frame with a pest and disease threshold, which makes the stringency of the storage threshold change over time. Storage decision (12) In formula 12, This indicates the overall quality score of the current plant frame. Indicates the threshold for pests and diseases, which changes over time; If the overall quality score of the current plant frame is ≥Pest and disease threshold If so, store the frame and proceed to the next filtering step; (13) In formula 13, The average quality score for the history of pests and diseases. Standard deviation of pest and disease quality scores over a period of time. This is an adjustable parameter used to control the leniency of the pest and disease threshold. The larger the size, the higher the threshold for pests and diseases, and the more stringent the storage conditions. Average quality score of pest and disease history , (14) Standard deviation of pest and disease quality scores over a period of time , (15); The memory station update strategy classifies the current frame into high-quality, medium-quality, and low-quality frames based on its quality score. For medium-quality frames, they can only be stored if their maximum similarity with the maximum number of frames in the memory storage station is less than a preset diversity threshold. Diverse decision functions , (16) The maximum similarity between the target image and the memory storage station; This is the diversity coefficient. The larger the coefficient, the higher the allowable similarity threshold, and the lower the requirement for diversity. Determining the conditions under which storage is permissible. (17) In formula 17, For high-quality thresholds, Low quality threshold; Ensure the storage of high-quality samples. For medium-quality samples, they can only be stored if they provide new information. Low-quality samples should be discarded directly. The memory station replacement mechanism first selects the worst sample in the current memory station when the memory storage station reaches its capacity and new samples need to be stored. Then, it compares the worst sample with the new sample based on replacement conditions to decide whether to replace the sample. The selection of the worst sample takes into account both the sample quality and the sample storage time. (18) In formula 18, As the worst sample, For time decay weight, Given the current mass of sample m, Current system time Sample m storage time, Here, is the decay constant and the time normalization factor. The smaller the value, the faster the time penalty increases; The method for determining the conditions for sample replacement is to replace the worst sample in the memory storage with the new sample when the quality gain of the new sample compared to the worst sample is greater than the replacement cost. (19) In formula 19, The overall quality score for the current frame. The overall quality score for the worst sample. For mass gain, To replace the tolerance factor, The similarity between the new sample and the worst sample; The time-decay storage rule states that the quality of sample m gradually decays over time t. (20) In formula 20, Let m be the mass of sample m after decay at time t. The quality when storing sample m. The attenuation coefficient is... Let m be the storage time for sample m. The time point is stored for sample m.

[0026] The control module drives the gimbal and lifting mast to perform corresponding processing tasks; specifically, the control module calculates the deviation between the target bounding box coordinates and the center of the image, calculates the PWM drive signal using a PID algorithm, and adjusts the gimbal's horizontal and vertical angles to ensure that the angular position deviation is less than a preset threshold; the control module compares the target's current pixel area with the preset ideal observation area and outputs lifting commands; specifically... If the current pixel area is less than the ideal observation area, control the lifting rod to descend at the rated speed and approach the target; If the current pixel area is greater than the ideal observation area, the lifting rod is controlled to rise at the rated speed to widen the viewing angle; if the deviation between the current pixel area and the ideal observation area is less than the rated value, the lifting rod remains stationary.

[0027] The switching logic and execution method of the working mode are as follows: Specifically, the global scan mode is entered by default after the system starts, and it performs a region-wide scan. In target locking and tracking mode, after the memory storage module has completed the screening and the priority arbitration algorithm determines the optimal target, the system will start the tracking and control system. In the target loss search mode, when the target recognition module fails to detect the currently tracked target for several consecutive frames, the switch is triggered. The control module drives the gimbal to perform a short-term "Z"-shaped oscillation search within a small area where the target last appeared. If the target is detected again during the search, the target lock tracking mode is restored. If the target is not detected within the timeout period, the mode automatically returns to the global scan mode.

[0028] If there is no currently tracked target, the target with the highest priority score is selected directly, triggering a switch from global scanning mode to target lock tracking mode. If there is already a currently tracked target, the difference between the new target's priority score and the current target's score is calculated. If the difference is greater than or equal to the preset switching threshold γ, the current target is abandoned, and the switch is initiated. Switch between new target locking and tracking modes to avoid frequent switching.

Claims

1. A control method for three-dimensional tracking and dynamic priority scheduling of pests and diseases, characterized in that, The steps are as follows: S1, enter global scanning mode, and acquire scene images in real time through the image acquisition device; S2, the acquired scene images are input into the memory storage station, the memory selection mechanism therein is used to filter high-quality target image frames, and the target recognition algorithm recognizes the target image frames and outputs structured recognition content information. S3, based on the identified content information in S2, obtains the target category and executes a differentiated tracking and identification strategy; S4, the identified content information is sent to the priority arbitration algorithm to calculate the dynamic priority score of the target and realize the arbitration of multi-target conflicts; S5, the tracking and control system sends control commands to the scheduling module based on arbitration information to switch the working mode from global scanning mode to target locking and tracking mode, and the control module drives the lower computer to execute the corresponding processing tasks. S6, in target locking and tracking mode, if the target is lost, it enters lost target search mode, and the control module drives the lower device to perform a small-range search at the last location of the target; If the target is detected again during the search, the target lock tracking mode is restored; If no target is detected within the timeout period, the system will return to global scan mode.

2. The control method for three-dimensional tracking and dynamic priority scheduling of pests and diseases according to claim 1, characterized in that, The image acquisition device includes a camera, a motion platform for carrying the camera, an attitude sensor, a distance sensor, and an optical flow sensor; the camera is mounted on the motion platform, which includes, but is not limited to, a three-axis gimbal and an electric lifting rod; the attitude sensor acquires the angle information of the motion platform; the distance sensor and the optical flow sensor acquire target distance and image motion-related parameters.

3. The control method for three-dimensional tracking and dynamic priority scheduling of pests and diseases according to claim 1, characterized in that, The memory selection mechanism in S2 is deployed on edge hardware devices. The memory selection mechanism is equipped with a static weight model and an adaptive weight model. The static weight model is for dynamic weight allocation for relatively stationary objects such as plants. The adaptive weight model is for dynamic weight allocation for non-relatively stationary organisms such as insects. The memory selection mechanism described above calculates the affinity score A(f) of the target image frame for the scene image acquired in S1. t ) and athletic score M(f t Based on the target category, different weighting models are used to calculate the comprehensive quality score of the target image frame. Filter and store high-quality target image frames.

4. The control method for three-dimensional tracking and dynamic priority scheduling of pests and diseases according to claim 1, characterized in that, The identification content information in S2 includes, but is not limited to, structured data frames containing target ID, target type, hazard level, confidence level, and identification duration. The hazard level is obtained by querying the preset pest and disease types; the confidence level is the confidence level output by the target recognition algorithm; the recognition duration is the duration for which the target is stably tracked; the target recognition algorithm includes, but is not limited to, the Yolov8n model.

5. The control method for three-dimensional tracking and dynamic priority scheduling of pests and diseases according to claim 3, characterized in that, The affinity score A(f) mentioned above t The calculation formula for ) (1) Affinity score A(f) t The scoring criteria are based on the sharpness of the target image and its similarity to the target images in the memory module. In Formula 1, C(f) t S(f) represents the image sharpness score. t M) is the similarity score, α is the clarity weight, and β are the similarity weights, and α+β=1; The image sharpness score C(f) t The gradient magnitude is measured by normalizing it using the Frobenius norm. Image sharpness score C(f) t The calculation formula for ) (2) In formula 2, Indicates the current image gradient, This represents the maximum gradient value of image I in memory M; Similarity score S(f) t The feature vector of the current frame and the feature vector of the target image in the memory storage station are extracted, and the maximum cosine similarity is calculated. (3) In Formula 3, f t A feature or image patch representing the target image in frame t; This represents a memory bank, which stores a collection of historical plant or insect samples m; Represents the feature extraction function; The motion score uses the motion speed and motion blur degree of the target image as scoring criteria. (4) In Formula 4, M(f) t The motion score of the target image in frame t. Represents the target image velocity in frame t. This represents the theoretical maximum speed of motion of the current target image. This represents the degree of motion blur in the target image of frame t. This indicates the maximum blur level set for the target image; The target image speed It is obtained by calculating the two-dimensional pixel displacement of the target image between the previous and current frames, based on the displacement response between two frames. (5) In formula 5, Let be the position vector of the target image in frame t. Let be the position vector of the target image in frame t-1; The degree of motion blur By calculating the average absolute difference between the pixel value of each pixel in the target region in the current frame and the pixel value of the corresponding position in the target image in the previous frame, the larger the value, the greater the motion blur in the target image. (6) In formula 6, This represents the target image in the current frame. This represents the target image from the previous frame. Represents the set of pixels in the target region. This indicates the number of pixels in the target region. Represents pixel coordinates, The motion displacement of the target image from frame t-1 to frame t can be obtained from the optical flow sensor; When the target image frame is identified as a plant, a static weighted model is executed to calculate a comprehensive quality score by assigning weights. （7） In formula 7, The overall quality score for frame t. Indicates the weight of the affinity score. Indicates the weight of the sports score; When the target image frame is identified as an insect, an adaptive weight model is executed. This model uses the sigmoid activation function to dynamically adjust the weights based on the deviation between the current affinity score and the historical average affinity score. (8) In formula 8, Affinity weight for frame t, The historical average of affinity. The historical average of athletic scores. , , For activation functions; (9) Motion score weights for frame t. Historical average affinity The calculation uses the average method to calculate the average historical affinity in the memory storage station. (10) Historical average of athletic scores The calculation is similar. (11) In formulas 10 and 11, The length of historical memory data.

6. The control method for three-dimensional tracking and dynamic priority scheduling of pests and diseases according to claim 1, characterized in that, The memory selection mechanism in S2 also includes storage decision, memory station update strategy, memory station replacement mechanism, and time decay storage rules. The storage decision is made by comparing the current plant frame's overall quality score with a dynamically changing pest and disease threshold. Storage decision (12) In formula 12, This indicates the overall quality score of the current plant frame. Indicates the threshold for pests and diseases, which changes over time; If the overall quality score of the current plant frame is ≥Pest and disease threshold If so, store the frame and proceed to the next filtering step; (13) In formula 13, The average quality score for the history of pests and diseases. Standard deviation of pest and disease quality scores over a period of time. This is an adjustable parameter used to control the leniency of the pest and disease threshold. The larger the size, the higher the threshold for pests and diseases, and the more stringent the storage conditions; among them, Average quality score of pest and disease history , (14) Standard deviation of pest and disease quality scores over a period of time , (15); The memory station update strategy categorizes the current frame into high-quality, medium-quality, and low-quality frames based on its quality score. For medium-quality frames, they can be stored only if their maximum similarity to the frame in the memory storage station is less than a preset diversity threshold. Diverse decision functions , (16) The maximum similarity between the target image and the memory storage station; This is the diversity coefficient. The larger the coefficient, the higher the allowable similarity threshold, and the lower the requirement for diversity. Determining the conditions under which storage is permissible. (17) In formula 17, For high-quality thresholds, Low quality threshold; The aforementioned memory station replacement mechanism, when the memory storage station reaches its capacity limit and new samples need to be stored, selects the worst sample based on a comprehensive evaluation of sample quality and storage time. If the quality gain of the new sample outweighs the replacement cost, then replacement is performed. Specifically, the worst sample in the current memory storage station is first selected, and then the worst sample and the new sample are compared according to replacement conditions to decide whether to replace the sample. The selection of the worst sample takes into account both the sample quality and the sample storage time. (18) In formula 18, As the worst sample, For time decay weight, Given the current mass of sample m, Current system time Sample m storage time, Here, is the decay constant and the time normalization factor. The smaller the value, the faster the time penalty increases; The aforementioned method for determining the conditions for sample replacement stipulates that when the quality gain of the new sample relative to the worst sample is greater than the replacement cost, the new sample replaces the worst sample in the memory storage station. (19) In formula 19, The overall quality score for the current frame. The overall quality score for the worst sample. For mass gain, To replace the tolerance factor, The similarity between the new sample and the worst sample; The time-decay storage rule states that the quality of sample m decays exponentially with time t. (20) In formula 20, Let m be the mass of sample m after decay at time t. The quality when storing sample m. The attenuation coefficient is... Let m be the storage time for sample m. The time point is stored for sample m.

7. The control method for three-dimensional tracking and dynamic priority scheduling of pests and diseases according to claim 1, characterized in that, In S3, the dynamic priority score p is calculated using the following steps: Let there be N targets to be arbitrated, forming a set. (21)， Formula 21 is the set formula for the targets to be arbitrated. Set weighting coefficients, Weighting coefficient formula ， (22) Weighting is assigned to the severity level. Weights are assigned to confidence levels. Weighting is applied based on duration; Current goal (i=1,2,...,N), its dynamic priority score , (23) In formula 23, To retrieve the hazard level score from the preset pest and disease type score table based on the target type, To identify the confidence score output by the algorithm, The duration for which the target is stably tracked; Hazard level classification , (twenty four) In formula 24, Let i be the type of the i-th target pest or disease. Use the preset mapping function to query the predefined pest and disease type score table; Confidence score , (25) In formula 25, The confidence level output by the target recognition algorithm; Duration , (26) The duration decreases linearly, as shown in Equation 26. For the goal Duration of stable tracking The attenuation coefficient; Calculate the target score for the current overall score, and make the final decision. As the current top priority target for tracking 。 8. The control system for three-dimensional tracking and dynamic priority scheduling of pests and diseases according to claim 1. The method, characterized in that, The control module drives the gimbal and lifting arm to perform corresponding processing tasks. Specifically, the control module calculates the deviation between the target bounding box coordinates and the center of the image, calculates the PWM drive signal through the PID algorithm, and adjusts the horizontal and vertical angles of the gimbal to make the angular position deviation less than a preset threshold. The control module compares the current pixel area of ​​the target with the preset ideal observation area and outputs lifting commands. Specifically, If the current pixel area is less than the ideal observation area, control the lifting rod to descend at the rated speed and approach the target; If the current pixel area is greater than the ideal observation area, the lifting rod is controlled to rise at the rated speed to widen the viewing angle; if the deviation between the current pixel area and the ideal observation area is less than the rated value, the lifting rod remains stationary.

9. The control method for three-dimensional tracking and dynamic priority scheduling of pests and diseases according to claim 1, characterized in that, The switching logic and execution method of the working mode are as follows: Specifically, the global scan mode is entered by default after the system starts, and it performs a region-wide scan. In target locking and tracking mode, after the memory storage module has completed the screening and the priority arbitration algorithm determines the optimal target, the system will start the tracking and control system. In the target search mode, when the target recognition algorithm fails to detect the currently tracked target for several consecutive frames, the switch is triggered. The control module drives the gimbal to perform a short-term "Z"-shaped oscillation search within a small area where the target last appeared. If the target is detected again during the search, the target lock tracking mode is restored. If the target is not detected within the timeout period, the mode automatically returns to the global scan mode.

10. The control method for three-dimensional tracking and dynamic priority scheduling of pests and diseases according to claim 1, characterized in that, The aforementioned differential tracking and identification strategy, specifically, For tracking insect pests and diseases, based on the target coordinate data of multiple consecutive frames, the Kalman filter algorithm is used to predict the position of the next frame while the pest is in motion, and the gimbal and lifting rod are adjusted in advance to achieve predictive tracking. When pests are at rest, their harmful behaviors are recorded through image recognition, generating a behavior log. We track plant diseases and pests, record the characteristics of plant lesions, combine them with environmental parameters to conduct source tracing analysis, determine the causes of disease, and predict the spread range in the next few days using a diffusion model.

Images