Inspection identification method and system for dynamic environment

Abstract

The invention discloses an inspection identification method and system for a dynamic environment, and relates to the technical field of inspection identification. Preprocessing the image data, then determining an effective object, and determining an identification point location corresponding to the effective object; acquiring position change data of the identification point location in a preset time period, and determining a travel path track of the identification point location according to the position change data; performing safety judgment on the travel path track of the identification point location through a preset rule; if the safety judgment result is unsafe, marking the identification point as an unsafe point; and carrying out real-time dynamic tracking on the unsafe point location according to the position change data of the unsafe point location. According to the invention, inspection images are collected in real time, effective objects are accurately positioned, point locations are identified, unsafe point locations can be quickly locked and dynamically tracked in combination with travel path track analysis and safety judgment, the efficient and accurate inspection requirements in a dynamic complex environment are met, and the inspection identification efficiency in the dynamic environment is improved.

Description

Technical Field

[0001] This invention belongs to the field of inspection and identification technology, specifically relating to inspection and identification methods and systems for dynamic environments. Background Technology

[0002] Currently, inspection work relies heavily on manual patrols or semi-automated equipment, which not only consumes a lot of manpower but is also prone to omissions due to staff fatigue and subjective judgment bias, failing to fully cover potential risks within the inspected area. Furthermore, manual inspections struggle to capture real-time changes in the location and movement of inspected objects, resulting in a significant lag in safety risk identification. Risks are often only discovered after they have escalated or even led to accidents, hindering early warning and intervention. In addition, existing technologies lack unified and quantifiable standards for assessing the safety of inspected objects, relying heavily on subjective judgment based on human experience. This leads to insufficient accuracy and significant errors in risk assessment, making it difficult to accurately distinguish between safe and unsafe conditions. Even when unsafe locations are identified, the lack of a real-time dynamic tracking mechanism for their location changes easily leads to gaps in subsequent control, further exacerbating the risk.

[0003] Quadruped robots (robot dogs) have become core equipment for inspection tasks in complex environments due to their excellent ability to traverse unstructured terrain. To achieve precise observation and recognition, robot dogs are usually equipped with gimbals and vision sensors. However, during movement, the robot dog's body will generate high-frequency vibrations and sudden posture changes, resulting in severe motion blur, noise, and uneven lighting in the acquired images. This directly causes the input quality of subsequent target detection models to be unstable, and the foundation for object feature extraction is weak, ultimately leading to a decrease in inspection and recognition accuracy. This makes it difficult to meet the needs of efficient and accurate inspection in dynamic and complex environments, resulting in low inspection and recognition efficiency in dynamic environments. Summary of the Invention

[0004] The purpose of this invention is to solve the problem of low efficiency in inspection and identification under dynamic environments, and to propose an inspection and identification method and system for dynamic environments.

[0005] In a first aspect of this invention, a patrol identification method for dynamic environments is proposed, the method comprising: Real-time acquisition of image data of the inspection area, and preprocessing of the image data to obtain effective images; Identify valid objects in the valid images and determine the recognition points corresponding to the valid objects; Collect location change data of the identification points within a preset time period, and determine the trajectory of the identification points based on the location change data; The safety of the movement trajectory at the identified points is determined by preset rules; the safety determination is divided into safe and unsafe. If the security judgment result is unsafe, then the identified point is recorded as an unsafe point. The unsafe points are dynamically tracked in real time based on the location change data of the unsafe points.

[0006] Optionally, preprocessing the image data to obtain a valid image includes: A denoised image is obtained by processing the image data using a bilateral filtering algorithm; The first illumination component is obtained by smoothing the denoised image using a preset Gaussian kernel. The second illumination component is obtained by taking the logarithm of the denoised image and the first illumination component, subtracting them, and then performing an exponential operation. The grayscale value of the second illumination component is normalized to obtain a standard grayscale image; The standard grayscale image and the denoised image are fused to obtain an effective image.

[0007] Optionally, determining the valid objects in the valid image and determining the recognition points corresponding to the valid objects includes: The effective image is downsampled at multiple scales to obtain a high-scale image, a medium-scale image, and a low-scale image; Substitute the high-scale image, the medium-scale image, and the low-scale image into a preset convolutional attention module to obtain high-scale attention features, medium-scale attention features, and low-scale attention features, respectively; Substituting the high-scale attention features and the mesoscale attention features into the fusion model yields the first fusion feature; Substituting the first fusion feature and the low-scale attention feature into the fusion model yields the second fusion feature; After fusing the second fusion feature and the effective image, a 1×1 convolution operation is performed to obtain the third fusion feature; Substituting the first fusion feature, the second fusion feature, and the third fusion feature as the output of the neck network in the YOLOv8s model into the detection head yields an effective object recognition result. For each valid identification object, the valid identification object is substituted into the key point identification model to obtain the identification point.

[0008] Optionally, the working principle of the fusion model includes: Obtain the first input scale feature and the second input scale feature; the first input scale feature is greater than the second input scale feature. The first input scale feature is substituted into the depth convolution module to obtain the depth convolution feature, and the second input scale feature is substituted into the ghost module to obtain the ghost feature; The initial fused feature is obtained by normalizing the deep convolutional feature and the ghost feature respectively and then multiplying them. After upsampling the initial fusion features, a 1×1 convolution operation is performed, and the results are normalized after being substituted into the ghost module to obtain the output features of the fusion model.

[0009] Optionally, the safety assessment of the movement trajectory of the identified points using preset rules includes: If the length of the trajectory of the identification point is zero, and the location of the identification point is within a preset safe area, then it is determined to be safe; If the length of the trajectory of the identified point is greater than zero, then it is determined whether the trajectory of the identified point conforms to the preset safe trajectory range: If the travel trajectory is within the preset safe trajectory range, it is determined to be safe; If the travel trajectory is not within the preset safe trajectory range, it is determined to be unsafe.

[0010] In a second aspect of this invention, an inspection and identification system for dynamic environments is proposed, comprising: The preprocessing module is used to collect image data of the inspection area in real time and preprocess the image data to obtain effective images; The valid object recognition module is used to determine the valid objects in the valid image and to determine the recognition points corresponding to the valid objects; The trajectory determination module is used to collect position change data of the identification point within a preset time period, and determine the trajectory of the identification point based on the position change data; The movement trajectory safety judgment module is used to judge the safety of the movement trajectory of the identified points according to preset rules; the safety judgment is divided into safe and unsafe. An unsafe location marking module is used to mark the identified location as an unsafe location if the safety judgment result is unsafe. The dynamic tracking module is used to perform real-time dynamic tracking of the unsafe points based on the location change data of the unsafe points.

[0011] Optionally, the preprocessing module includes: A denoised image generation module is used to process the image data using a bilateral filtering algorithm to obtain a denoised image; The first illumination component generation module is used to smooth the denoised image using a preset Gaussian kernel to obtain the first illumination component. The second illumination component generation module is used to take the logarithm of the denoised image and the first illumination component, then perform the difference and exponential operation to obtain the second illumination component. A standard grayscale image generation module is used to normalize the grayscale value of the second illumination component to obtain a standard grayscale image; An effective image generation module is used to fuse the standard grayscale image and the denoised image to obtain an effective image.

[0012] Optionally, the effective object recognition module includes: A multi-scale sampling module is used to perform multi-scale downsampling on the effective image to obtain a high-scale image, a medium-scale image, and a low-scale image; The attention feature extraction module is used to substitute the high-scale image, the medium-scale image, and the low-scale image into a preset convolutional attention module to obtain high-scale attention features, medium-scale attention features, and low-scale attention features, respectively. The first fusion feature generation module is used to substitute the high-scale attention feature and the mesoscale attention feature into the fusion model to obtain the first fusion feature; The second fusion feature generation module is used to substitute the first fusion feature and the low-scale attention feature into the fusion model to obtain the second fusion feature; The third fusion feature generation module is used to fuse the second fusion feature and the effective image, and then perform a 1×1 convolution operation to obtain the third fusion feature; The effective object recognition generation module is used to input the first fusion feature, the second fusion feature and the third fusion feature as the output of the neck network in the YOLOv8s model into the detection head to obtain the effective object recognition result; The identification point determination module is used to input the identification object into the key point identification model for each valid identification object to obtain the identification point.

[0013] Optionally, the working principle of the fusion model includes: Obtain the first input scale feature and the second input scale feature; the first input scale feature is greater than the second input scale feature. The first input scale feature is substituted into the depth convolution module to obtain the depth convolution feature, and the second input scale feature is substituted into the ghost module to obtain the ghost feature; The initial fused feature is obtained by normalizing the deep convolutional feature and the ghost feature respectively and then multiplying them. After upsampling the initial fusion features, a 1×1 convolution operation is performed, and the results are normalized after being substituted into the ghost module to obtain the output features of the fusion model.

[0014] Optionally, the trajectory safety determination module includes: The first safety judgment module is used to determine that the location is safe if the length of the trajectory of the identification point is zero and the location of the identification point is within a preset safety area. The trajectory qualification judgment module is used to determine whether the trajectory of the identified point conforms to the preset safe trajectory range if the length of the trajectory of the identified point is greater than zero. The second safety judgment module is used to determine that the travel trajectory is safe if it is within a preset safe trajectory range; The unsafety judgment module is used to determine that the travel trajectory is unsafe if it is not within the preset safe trajectory range.

[0015] The beneficial effects of this invention are: This invention proposes an inspection and identification method for dynamic environments. By acquiring inspection images in real time, accurately locating valid objects and identification points, and combining trajectory analysis and safety judgment, it can quickly lock unsafe locations and dynamically track them. This not only improves the automation and accuracy of inspections, but also enables early detection and early warning of safety risks. It effectively reduces the omission rate and lag of manual inspections, ensures the efficiency and reliability of safety management in the inspection area, meets the needs of efficient and accurate inspections in dynamic and complex environments, and improves the efficiency of inspection and identification in dynamic environments. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 A flowchart illustrating an inspection and identification method for dynamic environments, provided as an embodiment of the present invention; Figure 2 This is a framework diagram of an inspection and identification system for dynamic environments provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] All other embodiments obtained by those skilled in the art without inventive effort in relation to the embodiments of this invention are within the scope of protection of this invention.

[0020] This invention provides a method for inspection and identification in dynamic environments. See also... Figure 1 , Figure 1 This is a flowchart illustrating an inspection and identification method for dynamic environments, provided as an embodiment of the present invention. The method includes the following steps: S101 collects image data of the inspection area in real time and preprocesses the image data to obtain effective images; S102, determine the valid objects in the valid images and determine the recognition points corresponding to the valid objects; S103, collect location change data of the identification points within a preset time period, and determine the trajectory of the identification points based on the location change data; S104, uses preset rules to make safety judgments on the movement trajectories of the identified points; S105, if the safety judgment result is unsafe, then the identified point is recorded as an unsafe point; S106, Real-time dynamic tracking of unsafe points based on location change data; Among them, safety judgment is divided into safe and unsafe.

[0021] The present invention provides an inspection and identification method for dynamic environments. By acquiring inspection images in real time, accurately locating valid objects and identification points, and combining trajectory analysis and safety judgment, it can quickly lock unsafe points and dynamically track them. This not only improves the automation and accuracy of inspections, but also enables early detection and early warning of safety risks. It effectively reduces the omission rate and lag of manual inspections, ensures the efficiency and reliability of safety management in the inspection area, meets the needs of efficient and accurate inspections in dynamic and complex environments, and improves the efficiency of inspection and identification in dynamic environments.

[0022] In one implementation, automatic image preprocessing and intelligent determination of objects and recognition points replace a large number of manual inspection operations, reducing manpower input and avoiding inspection omissions caused by human fatigue and subjective bias. The system captures the positional changes of recognition points in real time and generates movement trajectories, and quickly completes safety judgments in combination with preset rules, solving the lag of traditional inspections and enabling early detection and early warning of risks.

[0023] In one implementation, the preset time period is determined by technical personnel.

[0024] In one implementation, quantitative preset rules replace manual experience judgment to avoid subjective errors, ensure the consistency and accuracy of safety judgments, and accurately distinguish between safe and unsafe states; unsafe locations are tracked dynamically in real time to avoid subsequent control gaps, effectively curb the spread of risks, and improve the reliability of safety control in the inspection area.

[0025] In one implementation, unsafe locations are tracked dynamically in real time to avoid gaps in subsequent management and control, effectively curb the spread of risks, and improve the reliability of safety management and control in the inspection area.

[0026] In one implementation, after identifying unsafe locations, their core features (such as appearance features, location coordinate baseline values, and initial values ​​of motion attributes) are immediately associated to generate a unique identifier ID, avoiding confusion with other locations. For time-series location data collected within a preset time period, the location coordinates at each moment are concatenated in chronological order, and a continuous path is generated through polyline connections or curve fitting, visually presenting the location's movement trajectory in space. For the location data of two adjacent moments in the path, the ratio of spatial distance to time difference is calculated. The location data of the last two moments within the preset time period are taken, and a direction vector is calculated. After normalization, a unit direction vector is obtained, representing the current direction of movement.

[0027] In one embodiment, preprocessing image data to obtain a valid image includes: A denoised image is obtained by processing image data using a bilateral filtering algorithm; The first illumination component is obtained by smoothing the denoised image using a preset Gaussian kernel. The second illumination component is obtained by taking the logarithm of the denoised image and the first illumination component, subtracting them, and then performing an exponential operation. The grayscale value of the second illumination component is normalized to obtain a standard grayscale image; The standard grayscale image and the denoised image are fused to obtain the effective image.

[0028] In one implementation, bilateral filtering is used to remove image noise while preserving key details such as the edges and textures of valid objects, avoiding the blurring of details caused by traditional denoising algorithms. The illumination component is extracted by Gaussian kernel smoothing, and the illumination difference is separated and corrected by logarithmic transformation and exponential operation. Then, grayscale normalization is performed to effectively eliminate the influence of uneven illumination in the image and improve the overall brightness consistency of the image.

[0029] In one implementation, the Gaussian kernel is preset to 15×15; the second illumination component is obtained by taking the logarithm of the denoised image and the first illumination component, subtracting them, and then performing an exponential operation; the second illumination component is obtained by subtracting the logarithm of the first illumination component from the logarithm of the denoised image and then performing an exponential operation.

[0030] In one implementation, the advantages of illumination equalization of standard grayscale images and the detail integrity of denoised images are combined to make the features of effective objects clearer and more recognizable, reducing the misjudgment rate of subsequent target detection and point recognition. In response to the pain points of motion blur and illumination fluctuations caused by robot dog inspection, this process can specifically optimize image quality to ensure that subsequent algorithms can still work stably in complex environments.

[0031] In one embodiment, determining valid objects in a valid image and determining the corresponding recognition points for the valid objects includes: Multi-scale downsampling of the effective image yields high-scale, medium-scale, and low-scale images; Substitute the high-scale image, medium-scale image, and low-scale image into the preset convolutional attention module to obtain high-scale attention features, medium-scale attention features, and low-scale attention features, respectively. Substituting the high-scale attention features and the mid-scale attention features into the fusion model yields the first fusion feature; Substituting the first fusion feature and the low-scale attention feature into the fusion model yields the second fusion feature; After fusing the second fusion feature and the effective image, a 1×1 convolution operation is performed to obtain the third fusion feature; Substituting the first, second, and third fusion features as the outputs of the neck network in the YOLOv8s model into the detection head yields effective object recognition results. For each valid identification object, the valid identification object is substituted into the key point identification model to obtain the identification point.

[0032] In one implementation, downsampling at high, medium, and low scales can capture the overall features of large-sized effective objects as well as identify the detailed features of small-sized objects, avoiding the problem of missed detection of objects at a single scale; the convolutional attention module can automatically focus on the core region of the effective object, suppress background interference, and make feature extraction more targeted.

[0033] In one implementation, effective features at different scales are integrated through two-step layered fusion and 1×1 convolution dimension adjustment to compensate for the information deficiencies of single-scale features, making the feature expression richer and more accurate. Combining the efficient detection capability of the YOLOv8s model with the localization advantage of the key point recognition model, effective object recognition and recognition point extraction are realized simultaneously, balancing efficiency and accuracy, and providing reliable point data for subsequent trajectory analysis.

[0034] In one implementation, the preset convolutional attention module includes a channel attention module and a spatial attention module; the input features are first substituted into the channel attention module and then averaged with the input features to obtain the first attention feature; the first attention feature is then substituted into the spatial attention module and then averaged with the first attention feature to obtain the output feature.

[0035] In one implementation, the key point recognition model can be any one of the following: top-down paradigm, bottom-up paradigm, MTCNN, lightweight top-down model, etc.

[0036] In one embodiment, the fusion model works by: Obtain the first input scale feature and the second input scale feature; the first input scale feature is greater than the second input scale feature. Substitute the first input scale feature into the depth convolution module to obtain the depth convolution feature, and substitute the second input scale feature into the ghost module to obtain the ghost feature; The initial fused feature is obtained by normalizing the deep convolutional features and the ghost features respectively and then multiplying them. After upsampling the initial fused features, a 1×1 convolution operation is performed, and then the features are normalized after being substituted into the ghost module to obtain the output features of the fused model.

[0037] In one implementation, a deep convolution module is used to extract rich semantic information for large-scale features (first input), and a ghost module is used to efficiently generate feature maps for small-scale features (second input). The differentiated design makes the extraction of features at different scales more targeted. By multiplying the features after normalization, the correlation of effective information is strengthened, redundant interference is suppressed, and the initial fused features are more focused on core information.

[0038] In one implementation, the ghost module generates rich features at low cost, avoiding the high computational cost of deep convolution. At the same time, the 1×1 convolution simplifies the feature dimension, reducing the overall computational complexity of the model and adapting to the computational limitations of mobile inspection equipment such as robot dogs. The upsampling operation completes the spatial information of small-scale features, and the features are processed again by the ghost module and normalization, so that the final output features have both spatial integrity and semantic consistency, providing a high-quality foundation for subsequent fusion and detection.

[0039] In one implementation, the deep convolutional module and the ghost module are existing modules; when the first input scale feature is a medium-scale attention feature, the second input scale feature is a high-scale attention feature, and the output feature is the first fusion feature; when the first input scale feature is a low-scale attention feature, the second input scale feature is the first fusion feature, and the output feature is the second fusion feature.

[0040] In one embodiment, determining the safety of the movement trajectory of the identified points using preset rules includes: If the length of the trajectory of the identified point is zero and the location of the identified point is within the preset safe area, it is considered safe. If the length of the identified point's trajectory is greater than zero, then determine whether the trajectory of the identified point conforms to the preset safe trajectory range: If the travel trajectory is within the preset safe trajectory range, it is considered safe; If the travel trajectory is outside the preset safe trajectory range, it is deemed unsafe.

[0041] One implementation considers both static scenarios where the identification point is stationary (trajectory length is zero) and dynamic scenarios where the point moves (trajectory length is greater than zero), avoiding scenario omissions caused by a single judgment logic; the preset safe trajectory range is determined in advance by technical personnel; in dynamic scenarios, the focus is on whether the trajectory is compliant, directly locking the risk points that deviate from the safe trajectory; in static scenarios, it is confirmed whether the point is in the safe area, accurately meeting the core security requirements of preventing intrusion and unauthorized movement during inspections.

[0042] Based on the same inventive concept, this invention also provides an inspection and identification system for dynamic environments. See [link to related document]. Figure 2 , Figure 2 A framework diagram of an inspection and identification system for dynamic environments provided in this embodiment of the invention includes: The preprocessing module is used to collect image data of the inspection area in real time and preprocess the image data to obtain effective images; The valid object recognition module is used to identify valid objects in valid images and determine the recognition points corresponding to the valid objects; The trajectory determination module is used to collect location change data of the identification points within a preset time period and determine the trajectory of the identification points based on the location change data. The movement trajectory safety assessment module is used to assess the safety of movement trajectories at identified locations based on preset rules; the safety assessment is divided into safe and unsafe. The unsafe point marking module is used to mark the identified point as an unsafe point if the safety judgment result is unsafe. The dynamic tracking module is used to perform real-time dynamic tracking of unsafe locations based on the location change data of unsafe locations.

[0043] Based on the embodiments of the present invention, an inspection and identification system for dynamic environments is provided. By acquiring inspection images in real time, accurately locating valid objects and identification points, and combining movement trajectory analysis and safety judgment, it can quickly lock unsafe points and dynamically track them. This not only improves the automation and accuracy of inspections, but also enables early detection and early warning of safety risks. It effectively reduces the omission rate and lag of manual inspections, ensures the efficiency and reliability of safety management in the inspection area, meets the needs of efficient and accurate inspections in dynamic and complex environments, and improves the efficiency of inspection and identification in dynamic environments.

[0044] In one embodiment, the preprocessing module includes: The denoised image generation module is used to process image data using a bilateral filtering algorithm to obtain a denoised image; The first illumination component generation module is used to smooth the denoised image using a preset Gaussian kernel to obtain the first illumination component. The second illumination component generation module is used to take the logarithm of the denoised image and the first illumination component, then perform the difference and exponential operation to obtain the second illumination component. The standard grayscale image generation module is used to normalize the grayscale value of the second illumination component to obtain a standard grayscale image; The effective image generation module is used to fuse a standard grayscale image and a denoised image to obtain an effective image.

[0045] In one embodiment, the effective object recognition module includes: The multi-scale sampling module is used to perform multi-scale downsampling on the effective image to obtain high-scale, medium-scale, and low-scale images; The attention feature extraction module is used to substitute high-scale, medium-scale, and low-scale images into a preset convolutional attention module to obtain high-scale attention features, medium-scale attention features, and low-scale attention features, respectively. The first fusion feature generation module is used to substitute high-scale attention features and mid-scale attention features into the fusion model to obtain the first fusion feature; The second fusion feature generation module is used to substitute the first fusion feature and the low-scale attention feature into the fusion model to obtain the second fusion feature; The third fusion feature generation module is used to fuse the second fusion feature and the effective image, and then perform a 1×1 convolution operation to obtain the third fusion feature; The effective object recognition generation module is used to input the first fusion feature, the second fusion feature and the third fusion feature as the output of the neck network in the YOLOv8s model into the detection head to obtain the effective object recognition result; The identification point determination module is used to input the identification object into the key point identification model for each valid identification object to obtain the identification point.

[0046] In one embodiment, the fusion model works by: Obtain the first input scale feature and the second input scale feature; the first input scale feature is greater than the second input scale feature. Substitute the first input scale feature into the depth convolution module to obtain the depth convolution feature, and substitute the second input scale feature into the ghost module to obtain the ghost feature; The initial fused feature is obtained by normalizing the deep convolutional features and the ghost features respectively and then multiplying them. After upsampling the initial fused features, a 1×1 convolution operation is performed, and then the features are normalized after being substituted into the ghost module to obtain the output features of the fused model.

[0047] In one embodiment, the movement trajectory safety determination module includes: The first safety judgment module is used to determine that the location is safe if the length of the trajectory of the identified point is zero and the location of the identified point is within a preset safety area. The trajectory qualification judgment module is used to determine whether the trajectory of the identified point conforms to the preset safe trajectory range if the length of the trajectory of the identified point is greater than zero. The second safety judgment module is used to determine that the path is safe if the trajectory is within the preset safety trajectory range; The unsafety judgment module is used to determine that a movement trajectory is unsafe if it is outside the preset safe trajectory range.

[0048] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A patrol identification method for dynamic environments, characterized in that, The method includes: Real-time acquisition of image data of the inspection area, and preprocessing of the image data to obtain effective images; Identify valid objects in the valid images and determine the recognition points corresponding to the valid objects; Collect location change data of the identification points within a preset time period, and determine the trajectory of the identification points based on the location change data; The safety of the movement trajectory at the identified points is determined by preset rules; the safety determination is divided into safe and unsafe. If the security judgment result is unsafe, then the identified point is recorded as an unsafe point. The unsafe points are dynamically tracked in real time based on the location change data of the unsafe points.

2. The inspection and identification method for dynamic environments according to claim 1, characterized in that, Preprocessing the image data to obtain a valid image includes: A denoised image is obtained by processing the image data using a bilateral filtering algorithm; The first illumination component is obtained by smoothing the denoised image using a preset Gaussian kernel. The second illumination component is obtained by taking the logarithm of the denoised image and the first illumination component, subtracting them, and then performing an exponential operation. The grayscale value of the second illumination component is normalized to obtain a standard grayscale image; The standard grayscale image and the denoised image are fused to obtain an effective image.

3. The inspection and identification method for dynamic environments according to claim 1, characterized in that, Determining the valid objects in the valid images and determining the corresponding recognition points for the valid objects includes: The effective image is downsampled at multiple scales to obtain a high-scale image, a medium-scale image, and a low-scale image; Substitute the high-scale image, the medium-scale image, and the low-scale image into a preset convolutional attention module to obtain high-scale attention features, medium-scale attention features, and low-scale attention features, respectively; Substituting the high-scale attention features and the mesoscale attention features into the fusion model yields the first fusion feature; Substituting the first fusion feature and the low-scale attention feature into the fusion model yields the second fusion feature; After fusing the second fusion feature and the effective image, a 1×1 convolution operation is performed to obtain the third fusion feature; Substituting the first fusion feature, the second fusion feature, and the third fusion feature as the output of the neck network in the YOLOv8s model into the detection head yields an effective object recognition result. For each valid identification object, the valid identification object is substituted into the key point identification model to obtain the identification point.

4. The inspection and identification method for dynamic environments according to claim 3, characterized in that, The working principle of the fusion model includes: Obtain the first input scale feature and the second input scale feature; the first input scale feature is greater than the second input scale feature. The first input scale feature is substituted into the depth convolution module to obtain the depth convolution feature, and the second input scale feature is substituted into the ghost module to obtain the ghost feature; The initial fused feature is obtained by normalizing the deep convolutional feature and the ghost feature respectively and then multiplying them. After upsampling the initial fusion features, a 1×1 convolution operation is performed, and the results are normalized after being substituted into the ghost module to obtain the output features of the fusion model.

5. The inspection and identification method for dynamic environments according to claim 1, characterized in that, The safety assessment of the movement trajectory of the identified points using preset rules includes: If the length of the trajectory of the identification point is zero, and the location of the identification point is within a preset safe area, then it is determined to be safe; If the length of the trajectory of the identified point is greater than zero, then it is determined whether the trajectory of the identified point conforms to the preset safe trajectory range: If the travel trajectory is within the preset safe trajectory range, it is determined to be safe; If the travel trajectory is not within the preset safe trajectory range, it is determined to be unsafe.

6. A patrol and identification system for dynamic environments, characterized in that, The system includes: The preprocessing module is used to collect image data of the inspection area in real time and preprocess the image data to obtain effective images; The valid object recognition module is used to determine the valid objects in the valid image and to determine the recognition points corresponding to the valid objects; The trajectory determination module is used to collect position change data of the identification point within a preset time period, and determine the trajectory of the identification point based on the position change data; The movement trajectory safety judgment module is used to judge the safety of the movement trajectory of the identified points according to preset rules; the safety judgment is divided into safe and unsafe. An unsafe location marking module is used to mark the identified location as an unsafe location if the safety judgment result is unsafe. The dynamic tracking module is used to perform real-time dynamic tracking of the unsafe points based on the location change data of the unsafe points.

7. The inspection and identification system for dynamic environments according to claim 6, characterized in that, The preprocessing module includes: A denoised image generation module is used to process the image data using a bilateral filtering algorithm to obtain a denoised image; The first illumination component generation module is used to smooth the denoised image using a preset Gaussian kernel to obtain the first illumination component. The second illumination component generation module is used to take the logarithm of the denoised image and the first illumination component, then perform the difference and exponential operation to obtain the second illumination component. A standard grayscale image generation module is used to normalize the grayscale value of the second illumination component to obtain a standard grayscale image; An effective image generation module is used to fuse the standard grayscale image and the denoised image to obtain an effective image.

8. The inspection and identification system for dynamic environments according to claim 6, characterized in that, The effective object identification module includes: A multi-scale sampling module is used to perform multi-scale downsampling on the effective image to obtain a high-scale image, a medium-scale image, and a low-scale image; The attention feature extraction module is used to substitute the high-scale image, the medium-scale image, and the low-scale image into a preset convolutional attention module to obtain high-scale attention features, medium-scale attention features, and low-scale attention features, respectively. The first fusion feature generation module is used to substitute the high-scale attention feature and the mesoscale attention feature into the fusion model to obtain the first fusion feature; The second fusion feature generation module is used to substitute the first fusion feature and the low-scale attention feature into the fusion model to obtain the second fusion feature; The third fusion feature generation module is used to fuse the second fusion feature and the effective image, and then perform a 1×1 convolution operation to obtain the third fusion feature; The effective object recognition generation module is used to input the first fusion feature, the second fusion feature and the third fusion feature as the output of the neck network in the YOLOv8s model into the detection head to obtain the effective object recognition result; The identification point determination module is used to input the identification object into the key point identification model for each valid identification object to obtain the identification point.

9. The inspection and identification system for dynamic environments according to claim 8, characterized in that, The working principle of the fusion model includes: Obtain the first input scale feature and the second input scale feature; the first input scale feature is greater than the second input scale feature. The first input scale feature is substituted into the depth convolution module to obtain the depth convolution feature, and the second input scale feature is substituted into the ghost module to obtain the ghost feature; The initial fused feature is obtained by normalizing the deep convolutional feature and the ghost feature respectively and then multiplying them. After upsampling the initial fusion features, a 1×1 convolution operation is performed, and the results are normalized after being substituted into the ghost module to obtain the output features of the fusion model.

10. The inspection and identification system for dynamic environments according to claim 6, characterized in that, The trajectory safety assessment module includes: The first safety judgment module is used to determine that the location is safe if the length of the trajectory of the identification point is zero and the location of the identification point is within a preset safety area. The trajectory qualification judgment module is used to determine whether the trajectory of the identified point conforms to a preset safe trajectory range if the length of the trajectory of the identified point is greater than zero. The second safety judgment module is used to determine that the travel trajectory is safe if it is within a preset safe trajectory range; The unsafety judgment module is used to determine that the travel trajectory is unsafe if it is not within the preset safe trajectory range.

Images