CN-121999329-A - Live working range prediction method based on improved YOLOv algorithm

CN121999329ACN 121999329 ACN121999329 ACN 121999329ACN-121999329-A

Abstract

The invention discloses a live working range prediction method based on an improved YOLOv8 algorithm, and relates to the technical field of live working, wherein the method comprises the steps of obtaining multi-mode data of a live working site, and dividing the multi-mode data into a visible light data partition and an infrared data partition according to physical wave band attributes of the data; the method comprises the steps of establishing a characteristic index channel based on improvement YOLOv for visible light data partition, outputting a target characteristic index table by utilizing a self-adaptive deformable convolution and attention mechanism model, determining a key characteristic region of the visible light data partition based on the target characteristic index table, introducing a space-time accompanying prediction model to predict real temperature field distribution of an infrared data partition, performing collision detection on the real temperature field distribution and the key characteristic region by utilizing a space collision detection algorithm, judging whether a device space bounding box intersects with a thermal energy field contour or not to obtain a boundary judgment result, and complementarily improving recognition accuracy and reliability of a live working boundary by multi-mode data.

Inventors

SUN XUEBIN
WU HAOZHOU
WANG JIAFENG
LIU SHIJIAN
LIU YONGHUI
SUN SHAOYU
LIU JUNFA
GUO DACHUAN
Wang Nuodi
ZHAO JINHUI

Assignees

国网辽宁省电力有限公司鞍山供电公司

Dates

Publication Date: 20260508
Application Date: 20251224

Claims (9)

1. A live working range prediction method based on a modified YOLOv algorithm, the method comprising: acquiring multi-mode data of a live working site, and dividing the multi-mode data into a visible light data partition and an infrared data partition according to physical wave band attributes of the data; Establishing a feature index channel based on improvement YOLOv for visible light data partition, and outputting a target feature index table by using a self-adaptive deformable convolution and attention mechanism model; determining a key feature area of the visible light data partition based on a target feature index table, and introducing a space-time companion prediction model to predict the real temperature field distribution of the infrared data partition; And performing collision detection on the real temperature field distribution and the key feature area by using a space collision detection algorithm, and judging whether the equipment space bounding box intersects with the thermal energy field outline or not so as to obtain a boundary judgment result.
2. The live working range prediction method based on the modified YOLOv algorithm as claimed in claim 1, wherein the dividing the multi-modal data into visible light data partitions and infrared data partitions includes: acquiring multi-mode data comprising all time periods of a live working site from binocular acquisition equipment, wherein the multi-mode data comprises visible light images, infrared thermal images and infrared thermal image data and time stamps; According to the physical characteristics of an imaging wave band, the multi-mode data is divided into two parallel sub-data partitions, including a visible light data partition and an infrared data partition.
3. The live working range prediction method based on the modified YOLOv algorithm as claimed in claim 1, wherein the outputting the target feature index table using the adaptive deformable convolution and attention mechanism model includes: Embedding an adaptive deformable convolution in a backbone network of YOLOv to construct a base framework for adaptive feature extraction; wherein, construct the basic framework of self-adaptation characteristic extraction, include: Inputting visible light data into a trunk extraction network of YOLOv model in a partitioning way, and generating an input feature map containing multi-scale semantic information through downsampling and convolutional coding; meanwhile, extracting gradient characteristics of all pixel points from visible light data partitions, wherein the gradient characteristics comprise gradient amplitude and gradient direction, the gradient amplitude is characterized as the texture change rate of a local area, and the gradient direction comprises horizontal gradient and vertical gradient; Dividing an input feature map into a plurality of local areas, taking each local area as a feature extraction node, and taking gradient amplitude values in the areas as attributes of the nodes; Based on the local area, selecting a corresponding convolution kernel and a sampling path through a built-in regularity scoring strategy, wherein the convolution kernel comprises a conventional convolution sum and self-adaptive deformable convolution, and the sampling path comprises a static grid sampling path and a dynamic grid sampling path; performing coordinate mapping based on the sampling path, and forming a target feature index table through position index analysis; at the same time, the attention mechanism is cascaded SimAM at the output of the adaptive deformable convolution.
4. A live working range prediction method based on a modified YOLOv algorithm as claimed in claim 3, wherein the selecting of the corresponding convolution kernel by a built-in regularity scoring strategy includes: Traversing all local areas, obtaining gradient amplitude values of each feature extraction node, accumulating and summing the gradient amplitude values in a certain local area, taking an average value, and generating a regularity score: Comparing the regularity score with a preset score threshold: If the regularity score is smaller than a preset score threshold, selecting to use conventional convolution to establish a static grid sampling path; And if the regularity score is greater than or equal to a preset score threshold, selecting to use self-adaptive deformable convolution to establish a dynamic grid sampling path.
5. A live working range prediction method based on a modified YOLOv algorithm as claimed in claim 3, wherein the forming of a target feature index table by position index analysis includes: determining a current feature extraction node, and identifying a sampling path to determine a corresponding sampling coordinate, wherein the feature extraction node has a unique node mark; If the sampling path is a static grid sampling path, mapping the sampling path into a static sampling coordinate; If the sampling path is a static grid sampling path, determining a dynamic sampling coordinate through regression analysis; inputting the feature map processed by the sampling path into a SimAM attention mechanism, and executing an energy evaluation strategy, wherein the method comprises the following steps of: determining a local neighborhood range based on each neuron of the feature map; judging the linear difference degree of the neuron and the local neighborhood range, and obtaining the minimum energy value through Euclidean distance analysis; Adopting a Sigmoid function to perform normalized conversion on the minimum energy value to generate neuron energy weight; A target feature index table is determined based on the node markers, the sampling coordinates, and the neuron energy weights.
6. The live working range prediction method based on the modified YOLOv algorithm of claim 5, wherein the determining key feature areas of the visible light data partition based on the target feature index table includes: Traversing a neuron energy weight set in a target feature index table, and identifying each neuron energy weight; Screening feature extraction nodes with neuron energy weights larger than a standard energy threshold, identifying all sampling coordinates of the feature extraction nodes by combining a target feature index table, and constructing a key feature region by spatial clustering.
7. The live working range prediction method based on the modified YOLOv algorithm of claim 1, wherein the standard energy threshold is a dynamic value, determined based on historical neuron energy weights.
8. The live working range prediction method based on the modified YOLOv algorithm of claim 1, wherein the spatio-temporal companion prediction model includes a temporal companion prediction branch and a spatial companion prediction branch, and introducing the spatio-temporal companion prediction model predicts a true temperature field distribution of the infrared data partition, comprising: Importing the infrared data partitions into a space-time companion prediction model, and analyzing the infrared data partitions to obtain a three-dimensional infrared data space; Traversing each pixel unit in a three-dimensional infrared data space, backtracking M0 infrared temperature data in front of the pixel unit, and constructing a historical temperature value sequence; the airspace accompanying prediction branch is used for identifying a corresponding pixel unit in a three-dimensional infrared data space based on a current time frame and introducing a self-adaptive dynamic filtering window taking the pixel unit as a center; identifying infrared temperature data of all pixel units in a current filtering window, and extracting statistical indexes including extremum indexes, median indexes and difference indexes; presetting constraint conditions, including a first constraint and a second constraint; judging whether the statistical index accords with a preset constraint condition or not, comprising: Establishing a closed interval based on the extremum index, comparing the infrared temperature data of the current pixel unit with the closed interval, and judging whether the infrared temperature data of the current pixel unit meets a first constraint to obtain a first verification result, wherein the first constraint comprises that the infrared temperature data of the current pixel unit is larger than or equal to the lower limit of the closed interval and smaller than or equal to the upper limit of the closed interval; comparing the absolute value with a preset edge protection threshold value based on the absolute value of the difference index, and judging whether the absolute value meets a second constraint to obtain a second verification result, wherein the second constraint is that the absolute value of the difference index is smaller than the preset edge protection threshold value; Based on the time domain predicted temperature value and the space domain predicted temperature value, obtaining a real temperature field distribution value through complementary weighted summation, and obtaining real temperature field distribution through space coordinate recombination.
9. The live working range prediction method based on the modified YOLOv algorithm according to claim 1, wherein the collision detection of the real temperature field distribution and the key feature area using the spatial collision detection algorithm includes: Determining a thermal energy field profile based on the real temperature field distribution; Determining a device space bounding box based on the key feature region; Calculating the intersection area and the union area of the bounding box and the thermal energy field, and marking the ratio of the intersection area to the union area as the overlapping confidence; And screening the conditions that the area of the intersection area is nonzero and the overlapping confidence coefficient is greater than or equal to a preset overlapping threshold value, wherein the conditions indicate that the equipment space bounding box collides with the thermal energy field outline, and the equipment space bounding box is judged to be a charged boundary, otherwise, the equipment space bounding box is judged to be a non-charged area.

Description

Live working range prediction method based on improved YOLOv algorithm Technical Field The invention relates to the technical field of live working, in particular to a live working range prediction method based on an improved YOLOv algorithm. Background In the shielding operation of live working, it is important to accurately identify the boundary between a live region and a non-live region, and the safety and the operation efficiency of operators are directly related; the existing live working range prediction method has the following limitations: On one hand, the traditional method mainly relies on visible light image analysis or artificial experience judgment, but due to light conditions, shielding or complex background interference, the accuracy of a recognition algorithm based on visible light is limited, and misjudgment or missed detection is easy to cause; On the other hand, live working is analyzed by collecting visible light and infrared data, but as the fusion of the visible light and the infrared data mostly adopts simple superposition or threshold segmentation, the effective collaborative optimization of the two types of data features is lacking, so that the robustness of boundary identification is insufficient; Therefore, there is a need for an intelligent algorithm that combines visible light target detection with infrared temperature analysis. Disclosure of Invention (One) solving the technical problems Aiming at the defects of the prior art, the invention provides a live working range prediction method based on an improved YOLOv algorithm, adopts an improved YOLOv algorithm, adds an adaptive deformable convolution and attention-introducing mechanism, detects a shielding region in a visible light image with high precision, improves the recognition precision and reliability of a live working boundary through multi-mode data complementation, and solves the problems in the background art. (II) technical scheme In order to achieve the above purpose, the invention is realized by the following technical scheme: the application provides a live working range prediction method based on an improved YOLOv algorithm, which comprises the following steps: acquiring multi-mode data of a live working site, and dividing the multi-mode data into a visible light data partition and an infrared data partition according to physical wave band attributes of the data; Establishing a feature index channel based on improvement YOLOv for visible light data partition, and outputting a target feature index table by using a self-adaptive deformable convolution and attention mechanism model; determining a key feature area of the visible light data partition based on a target feature index table, and introducing a space-time companion prediction model to predict the real temperature field distribution of the infrared data partition; And performing collision detection on the real temperature field distribution and the key feature area by using a space collision detection algorithm, and judging whether the equipment space bounding box intersects with the thermal energy field outline or not so as to obtain a boundary judgment result. Further, dividing the multi-modal data into a visible light data partition and an infrared data partition includes: acquiring multi-mode data comprising all time periods of a live working site from binocular acquisition equipment, wherein the multi-mode data comprises visible light images, infrared thermal images and infrared thermal image data and time stamps; According to the physical characteristics of an imaging wave band, the multi-mode data is divided into two parallel sub-data partitions, including a visible light data partition and an infrared data partition. Further, outputting the target feature index table using the adaptive deformable convolution and the attention mechanism model, comprising: Embedding an adaptive deformable convolution in a backbone network of YOLOv to construct a base framework for adaptive feature extraction; wherein, construct the basic framework of self-adaptation characteristic extraction, include: Inputting visible light data into a trunk extraction network of YOLOv model in a partitioning way, and generating an input feature map containing multi-scale semantic information through downsampling and convolutional coding; meanwhile, extracting gradient characteristics of all pixel points from visible light data partitions, wherein the gradient characteristics comprise gradient amplitude and gradient direction, the gradient amplitude is characterized as the texture change rate of a local area, and the gradient direction comprises horizontal gradient and vertical gradient; Dividing an input feature map into a plurality of local areas, taking each local area as a feature extraction node, and taking gradient amplitude values in the areas as attributes of the nodes; Based on the local area, selecting a corresponding convolution kernel and a sampling path through a built-