CN-122024164-A - Roadside intrusion plant monitoring method based on visual servo holder

CN122024164ACN 122024164 ACN122024164 ACN 122024164ACN-122024164-A

Abstract

The invention relates to the technical field of computer vision and intelligent control, in particular to a method for monitoring a roadside intrusion plant based on a vision servo holder, which comprises the steps of collecting an image to be detected, inputting a target detection model and outputting a target boundary frame; the method comprises the steps of calculating pixel deviation by a target boundary box, judging a target aggregation state, generating a pitch angle control instruction by adopting piecewise proportion control with dead zone, amplitude limiting and hysteresis based on the pixel deviation and the target aggregation state, sending the pitch angle control instruction to a visual servo holder to adjust a visual angle, collecting new images and circularly executing to form a closed loop. The method realizes stable centering of the target and continuous available image acquisition, reduces omission and jitter, and improves the reliability and efficiency of monitoring the roadside invasive plant.

Inventors

QIAO XI
HUANG TAO
HUANG YIQI
LIU BO
WAN FANGHAO
QIAN WANQIANG

Assignees

中国农业科学院农业基因组研究所

Dates

Publication Date: 20260512
Application Date: 20260205

Claims (10)

1. A roadside intrusion plant monitoring method based on a visual servo holder is characterized by comprising the following steps: The method comprises the steps of obtaining an image to be detected, which is acquired by a camera device, inputting the image to be detected into an invasive plant target detection model to obtain a detection result containing a target boundary box position, wherein the invasive plant target detection model comprises a dynamic convolution feature extraction network and a multi-scale feature fusion network, and the dynamic convolution feature extraction network is used for adjusting convolution kernel weights according to input feature density. Determining pixel deviation based on the target boundary box position, determining a target aggregation state represented by the number and position distribution of the target boundary boxes, and generating a pitch angle control instruction by adopting segmentation proportion control of dead zone, amplitude limiting and hysteresis according to the pixel deviation and the target aggregation state; and outputting the pitch angle control instruction to a visual angle adjusting executing mechanism, wherein the visual angle adjusting executing mechanism is a visual servo holder so as to adjust the visual angle of the image pickup device and acquire a new image as the image to be detected, and the new image is repeatedly executed to form a closed loop.
2. The method of claim 1, wherein the detection result further comprises invasive plant class information and detection confidence.
3. The method of claim 1, wherein the dynamic convolution feature extraction network comprises a packet convolution unit and a dynamic weight adjustment unit that adjusts a convolution kernel weight of the packet convolution unit based on the input feature density; the multi-scale feature fusion network includes a top-down feature pyramid fusion path and a bottom-up feature aggregation path.
4. The method of claim 1, wherein the pixel deviation is determined by the target bounding box position, the pixel deviation is used to indicate an offset of the target bounding box position in a longitudinal direction relative to a longitudinal center of the image to be detected, and the pixel deviation includes both greater than zero and less than zero; And under the condition that the pixel deviation is smaller than zero, the pixel deviation is used for indicating that the target boundary frame position is located below the longitudinal center of the image to be detected.
5. The method of claim 1, wherein the target aggregate state comprises a cluster aggregate state and a sparse distribution state; the cluster aggregation state is that the number of target boundary boxes in the detection result is not smaller than a first number threshold, and the transverse distribution range and the longitudinal distribution range of the center point of the target boundary boxes in the image to be detected are smaller than a first span threshold; the sparse distribution state is the cluster aggregation state which is not satisfied, and the first quantity threshold and the first span threshold are both preset thresholds.
6. The method of claim 1, wherein the pitch control command generated by the piecewise proportional control is zero if the absolute value of the pixel deviation is less than 50 pixels.
7. The method of claim 1, wherein the segment scale control comprises a slow zone scale control and a fast zone scale control; generating the pitch angle control instruction by the segment ratio control by adopting the slow zone ratio control under the condition that the absolute value of the pixel deviation is not more than 180 pixels; and under the condition that the absolute value of the pixel deviation is larger than 180 pixels, the subsection ratio control adopts the fast zone ratio control to generate the pitch angle control instruction.
8. The method of claim 7, wherein the hysteresis comprises a first switching threshold for switching the slow zone proportional control to the fast zone proportional control and a second switching threshold for switching the fast zone proportional control to the slow zone proportional control, the first switching threshold being 200 pixels and the second switching threshold being 160 pixels.
9. The method of claim 7, wherein the first scaling factor of the slow zone scaling control is a pitch angle adjustment of 0.35 degrees per pixel offset, and the first maximum step size of the slow zone scaling control is 7.2 degrees; the second scaling factor of the fast zone proportional control is pitch angle adjustment quantity corresponding to 0.9 degree per pixel deviation, and the second maximum step length of the fast zone proportional control is 21.6 degrees.
10. The method of claim 5, wherein the piecewise proportional control generates the pitch control command in a plurality of loop iterations if the target aggregate state is the cluster aggregate state; And under the condition that the target aggregation state is the sparse distribution state, the segmentation proportion control generates the pitch angle control instruction in a single rapid adjustment mode, and under the condition that the pixel deviation meets the dead zone, the pitch angle control instruction which is not zero is stopped from being output.

Description

Roadside intrusion plant monitoring method based on visual servo holder Technical Field The invention relates to the technical field of computer vision and intelligent control, in particular to a road side intrusion plant monitoring method based on a vision servo holder. Background With the continuous aggravation of global trade and traffic activities, exotic invasive plants pose a continuous threat to the stability of the local ecosystem and agricultural production by virtue of the rapid propagation and diffusion capability. In the road along-line environment, the road is used as a typical linear corridor, and vehicle, human and livestock activities are overlapped, so that the spreading of invasive plant seeds is obviously accelerated, and the invasive plant presents the characteristics of spot aggregation, cluster expansion, multi-species hybrid coexistence and the like in a road side area. The invasive plant can occupy local plant habitat, reduce biological diversity, shade mark lines, invade a road shoulder drainage system, increase maintenance and cleaning cost, and bring higher fire risks and traffic safety risks in drought seasons. The existing invasive plant monitoring technology depends on an artificial sample plot investigation or an image recognition mode with a fixed visual angle, and although certain accuracy can be obtained in a local area, the problems of high labor intensity, low efficiency, limited space coverage range and the like generally exist, and the requirements of large-scale, continuous and high-frequency monitoring under a road side environment are difficult to meet. Along with the development of computer vision and deep learning technology, an automatic identification method based on a target detection model is gradually used for ecological monitoring, but the existing method is designed for a single target or independent plants, and generally assumes that the detection target boundary is clear and the dimension is relatively stable, so that the method is difficult to adapt to the actual conditions of obvious difference of the height of invasive plants in a field road side scene, complex cluster structure and frequent visual angle change. Meanwhile, under the condition of a mobile platform or visual angle adjustable equipment, a detection result is often directly used as a control input of a cloud deck or a steering engine, and when continuous frame detection has jitter, unstable scale or rapid change of multi-target distribution, frequent adjustment of visual angles is easily caused, so that image blurring, target loss and unstable effective image acquisition are caused, and continuous monitoring and positioning accuracy are affected. The existing closed loop monitoring schemes for researching multi-focus farmland or industrial scenes, aiming at roadside crowd-invading plants and carrying out collaborative design on target detection and visual angle control are still relatively lacking. Disclosure of Invention The invention provides a roadside intrusion plant monitoring method based on a visual servo holder, which is used for at least solving the problems that detection is unstable and executable visual angle closed-loop control is difficult to form due to the fact that a target is easy to deviate from a visual field in a mobile acquisition scene. The invention provides a road side intrusion plant monitoring method based on a visual servo holder, which comprises the following steps: Acquiring an image to be detected acquired by a camera device, inputting the image to be detected into an invasive plant target detection model to obtain a detection result containing the target boundary box position, wherein the invasive plant target detection model comprises a dynamic convolution feature extraction network and a multi-scale feature fusion network, and the dynamic convolution feature extraction network is used for adjusting convolution kernel weights according to input feature density; Determining pixel deviation based on the target boundary frame position, determining a target aggregation state represented by the number and position distribution of the target boundary frames, and generating a pitch angle control instruction by adopting segmentation proportion control of dead zone, amplitude limiting and hysteresis according to the pixel deviation and the target aggregation state; And outputting a pitch angle control instruction to a visual angle adjusting executing mechanism, wherein the visual angle adjusting executing mechanism is a visual servo holder so as to adjust the visual angle of the image pickup device and acquire a new image as an image to be detected, and repeatedly executing the new image to form a closed loop. In one possible implementation, the detection result further includes invasive plant class information and detection confidence. In one possible implementation, the dynamic convolution feature extraction network comprises a grouping convolution unit and a dynam