Communication maintenance field operation big data management system based on visual analysis

Abstract

The invention discloses a communication maintenance field operation big data management system based on visual analysis, and belongs to the technical field of detection and analysis. By separating illumination and reflection components, the detail definition of the target in strong light, backlight and dark scenes can be effectively improved, and target missing detection caused by uneven illumination is avoided; the high-frequency texture is reconstructed through the generative adversarial network, so that the detail recovery rate of the compressed video can be effectively improved, the problem of information loss in a compression-analysis link is solved, and data support can be provided for equipment defect identification; the extracted illumination feature sequence provides an environment context for subsequent risk early warning, and the false alarm rate is reduced; after YOLOv5s is combined with illumination features, the target detection precision in strong light and backlight scenes can be effectively improved, and the omission factor is reduced; the multi-modal feature matrix can be migrated to different maintenance scenes, the model weight is dynamically adjusted through the illumination features, and the robustness of the system is enhanced.

Description

Technical Field

[0001] This invention relates to the field of detection and analysis technology, and more specifically to a big data management system for on-site communication maintenance operations based on visual analysis. Background Technology

[0002] As power companies face increasingly stringent operational requirements, power grids need to improve their communication maintenance capabilities to adapt to economic development. Currently, communication maintenance sites are mostly outdoors or underground equipment rooms, where lighting conditions vary greatly, such as strong light, backlight, and dim lighting, affecting the accuracy of target detection and recognition. Furthermore, existing technical solutions often compress video to save bandwidth and storage, leading to loss of detail and impacting subsequent monitoring and analysis of communication maintenance operations. Summary of the Invention

[0003] The purpose of this invention is to provide a big data management system for on-site communication maintenance operations based on visual analysis, which solves the technical problem of poor active optimization effect of target detection and recognition accuracy in existing solutions.

[0004] The objective of this invention can be achieved through the following technical solutions: A big data management system for on-site communication maintenance operations based on visual analytics includes: Multi-source data acquisition and processing module: Acquires raw video streams and light intensity data from the communication maintenance site, performs adaptive light compensation processing and reverse compression distortion repair on the raw video stream, generates enhanced video sequences, and extracts light feature sequences; Feature detection and processing fusion module: Performs target detection on enhanced video sequences using a pre-built detection model, constructs a multimodal feature matrix by combining illumination feature sequences, and outputs target detection results including single video frame index, optimized bounding box, corrected confidence score and label category; The detection result optimization evaluation prompt module processes and analyzes the single-frame optimization effect of the target detection results to obtain the first optimization label or the second optimization label. It uses the first optimization label and the second optimization label to calculate the optimization reliability rate and optimization impact rate of the existing scheme, performs joint analysis on the optimization reliability rate and optimization impact rate, and provides dynamic optimization management prompts.

[0005] Preferably, when performing illumination adaptive compensation processing on the original video stream, the original video frames in the original video stream are obtained, and the pixel values ​​of the original video frames are normalized to [0,1]. Then, they are converted into logarithmic brightness using the formula L(x,y)=log(I(x,y)+1), where I(x,y) is the pixel value of the original video frame, representing the brightness value of the pixel with coordinates (x,y) in the image; and L(x,y) is the logarithmic brightness value.

[0006] Preferably, the illumination component is calculated. ;in, is the Gaussian kernel at the i-th scale; i is the scale index, i=1,2,3; * indicates the convolution operation.

[0007] Preferably, the reflection components at the three scales are weighted and fused to obtain the final reflection component R(x,y); the final reflection component is linearly stretched to output the video frame after illumination compensation.

[0008] Preferably, when generating the enhanced video sequence, the video frames after illumination compensation are fused with the high-definition frames repaired by the generative adversarial network at the pixel level, and the fusion weights are dynamically adjusted.

[0009] Preferably, when extracting the illumination feature sequence, the illumination feature vector corresponding to each frame of video is extracted from the synchronous illumination intensity data; the illumination feature vector includes basic statistical features, illumination component features and temporal features; the extracted illumination feature vectors are sorted and combined to output the illumination feature sequence.

[0010] Preferably, each frame of the enhanced video sequence is normalized and input into the detection model for inference analysis, and a target list for each frame is output. When constructing a multimodal feature matrix by combining illumination feature sequences, the visual feature vector of each target in the target list of each frame image is obtained; the visual feature vector is concatenated with the illumination feature vector to form a 261-dimensional multimodal feature vector, and the multimodal feature matrix is ​​constructed with the timestamp as the row index and the target ID as the column index.

[0011] Preferably, the confidence level c is corrected using the illumination feature vector, and the relevant expression is: ;in, The corrected confidence level; This represents the illumination contrast in the illumination feature sequence.

[0012] Preferably, the optimized bounding box in the target detection result is obtained, and the corresponding single-frame optimized value YD is calculated using a formula; if YD > M, the first optimized label is generated; M is a natural number; Conversely, a second optimized label is generated.

[0013] Preferably, the total number of consecutive first optimized labels N1 and the total number of second optimized labels N2 are counted, and the optimization reliability rate is obtained by calculation. and optimization impact rate ; These are respectively optimizing the reliability threshold and optimizing the impact threshold; like and If the optimization and improvement effect is normal, the existing plan should be maintained. Conversely, if the optimization effect is abnormal, the operation and maintenance personnel will be notified and optimization management prompts will be sent.

[0014] Compared to existing solutions, the beneficial effects achieved by this invention are: This invention effectively improves the detail clarity of targets in strong light, backlight, and dim light scenes by separating the illumination and reflection components, avoiding missed detection of targets due to uneven illumination. By reconstructing high-frequency textures through generative adversarial networks, it effectively improves the detail recovery rate of compressed videos, solves the information loss problem in the compression-analysis link, and provides data support for equipment defect identification. The extracted illumination feature sequences provide environmental context for subsequent risk warnings. For example, by identifying scenes where backlighting leads to misjudgment through illumination change features, it reduces the false alarm rate.

[0015] This invention, based on YOLOv5s and combined with illumination features, can effectively improve target detection accuracy in strong light and backlight scenarios and reduce the false negative rate; the multimodal feature matrix can be transferred to different maintenance scenarios, and the model weights can be dynamically adjusted through illumination features, thereby enhancing the robustness of the system.

[0016] This invention utilizes the target detection results obtained from preprocessing, performs single-frame optimization calculations and analysis on single-frame videos, and integrates and analyzes the single-frame optimization results of different single-frame videos from different dimensions. This allows for proactive determination of the optimization and improvement effects of existing solutions, enabling multi-level and multi-dimensional proactive management of existing communication maintenance field operation big data. This improves the diversity and reliability of the processing and analysis of the implementation effects of existing optimization solutions. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart illustrating the operation of the big data management system for on-site communication maintenance based on visual analysis, as described in this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, this invention is a big data management system for on-site communication maintenance operations based on visual analytics, comprising: Multi-source data acquisition and processing module: Acquires raw video streams and light intensity data from the communication maintenance site, performs adaptive light compensation processing and reverse compression distortion repair on the raw video streams, generates enhanced video sequences, and extracts light feature sequences; specific steps include: When collecting raw video streams from communication maintenance sites, industrial-grade high-definition cameras are used, and the data is transmitted to the edge computing unit via a PoE interface. When collecting light intensity data, four light sensors are arranged at equal intervals around the camera lens to output the raw light intensity value. The video frame timestamps are aligned with the light sensor data timestamps using a GNSS module, with the error controlled within ±10ms, ensuring that each video frame corresponds to the average data from four light sensors; the GNSS module includes both BeiDou and GPS dual-mode; and the video frame timestamps are accurate to the millisecond level. In addition, when preprocessing the light intensity data, the raw values ​​of the four sensors are subjected to median filtering with a sliding window size of 5 to remove instantaneous strong light interference, such as flash lamps and direct sunlight spots, to obtain a smoothed light intensity sequence. The preprocessing implemented is a conventional technical solution, and the specific implementation steps are not described here. When performing illumination adaptive compensation processing on the original video stream, the original video frames in the original video stream are obtained. After normalizing the pixel values ​​of the original video frames to [0,1], they are converted into logarithmic brightness using the formula L(x,y)=log(I(x,y)+1), where I(x,y) is the pixel value of the original video frame, representing the brightness value of the pixel with coordinates (x,y) in the image; L(x,y) is the logarithmic brightness value. Gaussian kernels of three scales are used, specifically small-scale... Mesoscale Large scale Calculate the illumination component ;in, For the Gaussian kernel of the i-th scale, ; i is the scale index, i = 1, 2, 3; * indicates the convolution operation; It should be noted that small scale focuses on local details, such as device edges and textures, and captures high-frequency lighting changes, such as local shadows; medium scale balances the local and global, reflecting medium-range lighting unevenness, such as regional brightness differences caused by backlighting; large scale covers the global lighting trend, such as the overall brightness distribution of the image, and suppresses large-area background interference.

[0021] The reflection components at the three scales are weighted and fused to obtain the final reflection component, and the relevant expression is as follows: ;in, correspond , The default values ​​are 0.3, 0.5, and 0.2, and can also be customized according to the application requirements and specifications of the actual application scenario; Linear stretching of the final reflection component R(x,y) and outputting the illumination-compensated video frame can eliminate illumination unevenness and preserve target details. The relevant expression is: ;in, These are the maximum and minimum values ​​of the reflection component in the current frame, respectively; When performing reverse repair of compression distortion, a generative adversarial network is constructed. The network structure includes a generator, a discriminator, and a loss function. The generator adopts the U-Net architecture, with the input being the video frames after illumination compensation and the output being the reconstructed high-definition frames; the encoder contains 5 convolutional blocks, with the number of channels in each layer being 64, 128, 256, 512, and 1024 respectively, and the decoder fuses the high-frequency features of the encoder through skip connections; The discriminator uses a PatchGAN structure, with input being real high-definition frames and reconstructed high-definition frames, and output being a pixel-level probability map of authenticity. The loss function is ;in To combat the losses, In order to perceive loss, The default value is 10 for calculating the weight; The model training set collected 100,000 pairs of original high-definition frames and compressed low-definition frames, with a compression ratio of 10:1 to 20:1, covering different lighting scenarios, such as strong light, backlight, and dim light. The training parameters include the Adam optimizer, a learning rate of 0.0002, a batch size of 8, and 50,000 iterations. For each frame of video after illumination compensation, a reconstructed high-definition frame is output through a pre-trained generator as the result of compression distortion repair. It should be noted that the training and optimization of the model are based on existing conventional technical solutions, and the specific implementation steps will not be elaborated here. When generating enhanced video sequences, the video frames after illumination compensation are fused with the high-definition frames repaired by the generative adversarial network at the pixel level, and the fusion weight is dynamically adjusted; specifically, when the original illumination intensity is <500, the weight of the repaired frame is increased to 0.7, prioritizing the preservation of details; When the light intensity is >50000, the weight of the compensation frame is increased to 0.6, prioritizing the suppression of overexposure; When extracting the illumination feature sequence, the illumination feature vector corresponding to each frame of video is extracted from the synchronous illumination intensity data and labeled as follows. t is the frame index, t=1, 2, ..., N, N is the total number of frames, including basic statistical features, illumination component features and temporal features; Among them, the basic statistical characteristics include the mean light intensity. Standard deviation Contrast ; The illumination component features include the histogram entropy of the illumination components L(x,y) separated by the Retinex algorithm. ; Temporal features include the rate of change of illumination intensity between the current frame and the previous 5 frames. ; The extracted illumination feature vectors are sorted and combined to output the illumination feature sequence. , where N is the number of video frames.

[0022] It should be noted that the original video stream and illumination data are acquired synchronously to ensure that the illumination compensation and restoration process can dynamically adjust parameters based on real-time ambient illumination characteristics. For example, the weights of the generative adversarial network (GAN) are enhanced in low illumination. By eliminating illumination interference, the details lost in compression are restored through the GAN, avoiding the impact of illumination noise on the GAN's learning of high-frequency textures. The enhanced video sequence is used for subsequent target detection, and the illumination feature sequence serves as the environmental parameters for multimodal analysis.

[0023] In this embodiment of the invention, by separating the illumination and reflection components, the detail clarity of the target in strong light, backlight, and dim light scenes can be effectively improved, avoiding target omissions caused by uneven illumination; by reconstructing high-frequency textures through generative adversarial networks, the detail recovery rate of compressed videos can be effectively improved, solving the information loss problem in the compression-analysis link and providing data support for device defect identification; the extracted illumination feature sequence provides environmental context for subsequent risk warnings, such as identifying scenes where backlighting leads to misjudgment through illumination change features, reducing the false alarm rate.

[0024] Feature detection and processing fusion module: Performs object detection on enhanced video sequences using a pre-built detection model, constructs a multimodal feature matrix by combining illumination feature sequences, and outputs object detection results including single video frame indices, optimized bounding boxes, corrected confidence scores, and label categories; specific steps include: The pre-built detection model is based on YOLOv5. Channel attention (CA) and spatial attention (SA) sub-modules are added after the C3 module of YOLOv5 to enhance the feature extraction capability of key target regions, such as device labels and tool edges. It should be explained that the C3 module is a core component of YOLOv5, which is an improvement on the CSP (Cross Stage Partial Network) structure. It consists of three convolutional layers and multiple Bottleneck modules, with the following specific structure: Input splitting: The input feature map is divided into two parts, usually in a 1:1 ratio; Main branch: A portion of the feature map is reduced in dimensionality by 1×1 convolution, and then passed through multiple cascaded Bottleneck modules. Each Bottleneck contains 1×1 convolution, 3×3 convolution, and residual connections. Skip connections: The remaining feature maps are directly convolved with 1×1 convolutions to preserve their dimensions; Feature fusion: The feature maps of the main branch and the jump connection are concatenated and finally integrated and output through a 1×1 convolution.

[0025] The C3 module includes efficient feature extraction and gradient protection; the efficient feature extraction enhances feature representation through Bottleneck stacking, while 1×1 convolution reduces computational cost. Gradient protection utilizes the CSP structure to prevent gradient information from being repeatedly calculated during propagation, thus mitigating the gradient vanishing problem in deep networks.

[0026] Channel attention learns channel weights through global average pooling (GAP) and a fully connected layer (FC), and the relevant expression is: Where F is the input feature map; σ is the sigmoid activation function, the output channel weight vector, with the same dimension as the number of input channels; Spatial attention learns spatial weights through 7×7 convolutional layers, and the relevant expression is: ;in, The output is a single-channel weight matrix, with dimensions consistent with the spatial dimensions of the input feature map F; It is a 7×7 convolutional layer; This is for global max pooling; This means concatenating the results of global average pooling and global max pooling along the channel dimension to form a 2-channel intermediate feature map; When performing multi-scale feature fusion optimization, the 8×8, 16×16, and 32×32 feature maps output by the C3 module are upsampled and downsampled and then stitched together to enhance the feature representation of small targets, such as screwdrivers and connectors. Upsampling and downsampling are existing conventional technical solutions, and the specific implementation steps are not described here. Based on the light contrast in the light feature sequence Assign weights, A value less than 0.2 corresponds to low lighting conditions. The weight of the high-resolution feature map is increased to 0.6 to enhance detail capture. During the sample dataset collection and annotation, 500 video clips of communication maintenance scenarios were collected, including scenes with strong light, backlight, and nighttime conditions. 100,000 frames of images were extracted and labeled with three types of targets: Equipment, including power distribution cabinets, fiber optic junction boxes, etc., is categorized under "equipment". Personnel, including maintenance personnel, are categorized under the label "person". Tools, including wrenches, welding machines, etc., are categorized under the tag "tool". The image is randomly flipped, subjected to lighting perturbation, and Gaussian blur to simulate a complex lighting environment. The flipping includes horizontal and vertical rotation; the lighting perturbation includes brightness ±30% and contrast ±20%; and the Gaussian blur value ranges from 0.5 to 2. The model is trained and optimized using the sample dataset to obtain the detection model; the training and optimization of the model are existing conventional technical solutions, and the specific implementation steps will not be elaborated here. Each frame of the enhanced video sequence is normalized and input into the detection model for inference analysis, outputting a list of targets for each frame. ;in, represents the bounding box coordinates of the target; c is the confidence score, with a value range of [0,1]; cls is the class label; k represents the kth detected target instance in a single frame image, used to distinguish different target objects within the same frame; Among them, the target bounding box coordinates are numerical descriptions used to accurately locate the target region in an image or video frame, usually in pixels, representing the position and size of the target's bounding rectangle; When constructing a multimodal feature matrix by combining illumination feature sequences, the target list for each frame of the image is... Each target in the algorithm outputs a 256-dimensional visual feature vector through the YOLOv5 Head layer. It includes bounding box, class probability, and location information; It should be explained that the Head layer is the detection output module of YOLOv5, which adopts the Anchor-Based prediction mechanism. It designs detection heads for the three scale feature maps output by the Neck layer, and each detection head includes a convolutional prediction layer and post-processing. The convolutional prediction layer converts the feature map into a prediction tensor using a 3×3 convolution followed by a 1×1 convolution. Post-processing filters out redundant bounding boxes using non-maximum suppression, retaining the prediction results with the highest confidence. Visual feature vectors With illumination feature vector The features are concatenated to form a 261-dimensional multimodal feature vector. : Construct a multimodal feature matrix (MM) using timestamps as row indices and target IDs as column indices: The matrix elements are 261-dimensional feature vectors, with the number of rows and columns being the number of video frames N and the maximum number of targets per frame K, respectively, and K is preset to a value of 20. Using illumination feature vectors The expression involved in correcting the confidence level c is: ;in, The corrected confidence level; When the contrast is low, the feature blurring caused by lighting interference can be compensated by adjusting and increasing the confidence weight.

[0027] Final output target detection results It includes single video frame index, optimized bounding box, corrected confidence and label category.

[0028] In this embodiment of the invention, by combining YOLOv5s with illumination features, the target detection accuracy in strong light and backlight scenarios can be effectively improved, and the false negative rate can be reduced. The multimodal feature matrix can be transferred to different maintenance scenarios, and the model weights can be dynamically adjusted through illumination features, thereby enhancing the robustness of the system.

[0029] The detection result optimization evaluation and prompt module processes and analyzes the single-frame optimization effect of the target detection results to obtain a first optimization label or a second optimization label. Using the first and second optimization labels, it calculates the optimization reliability rate and optimization impact rate of the existing scheme. It then performs a joint analysis of the optimization reliability rate and optimization impact rate and provides dynamic optimization management prompts. Specific steps include: Obtain the optimized bounding box from the object detection results, and calculate the corresponding single-frame optimized value YD using the formula: ;in, These represent the optimized bounding box, the ground truth bounding box, and the unoptimized bounding box, respectively; Q() is the matching function. ;in, for The area of ​​the intersection of the two bounding boxes. ; for Bounding box coordinates; for Bounding box coordinates; for The area of ​​the union of the two bounding boxes. , They are respectively The area of ​​the two bounding boxes; If YD > M, it is determined that the overlap between the optimized bounding box and the ground truth box has improved, and the first optimized label is generated; M is a natural number, with a default value of 0, which can be customized according to the application requirements and specifications of the actual application scenario; Conversely, if the optimization results in a decrease or no improvement in positioning accuracy, a second optimization label is generated. Additionally, the total number of consecutive first-optimized labels N1 and the total number of second-optimized labels N2 are counted, and the optimization reliability rate is obtained through calculation. and optimization impact rate ; These are the optimization reliability threshold and the optimization impact threshold, respectively. The specific values ​​can be determined through simulation testing based on the previous sample data. For example, the median of the total number of all first-optimized labels appearing in the simulation test, the median of the total number of all second-optimized labels appearing, or customized by professionals in this field according to the application requirements and specifications of the actual application scenario. When determining the optimization improvement effect of the existing scheme based on the optimization reliability and optimization impact rate, if and If the optimization and improvement effect is normal, the existing plan should be maintained. Conversely, if the optimization effect is abnormal, the operation and maintenance personnel will be notified of the optimization management prompts. The optimization management prompts may involve adding, deleting, or modifying the implementation rules and content of the existing solutions. The specific optimization management steps and content are not limited in this embodiment of the invention.

[0030] Unlike existing technical solutions that can only monitor the effectiveness of optimization schemes using a single indicator, resulting in poor reliability of the monitoring and processing of optimization scheme implementation effects, this invention utilizes the target detection results obtained from pre-processing, performs single-frame optimization calculation and analysis on single-frame videos, and integrates and analyzes the single-frame optimization results of different single-frame videos from different dimensions. This proactively determines the optimization and improvement effect of existing schemes, enabling multi-level and multi-dimensional proactive management of existing communication maintenance field operation big data, and improving the diversity and reliability of the analysis and processing of the implementation effects of existing optimization schemes.

[0031] In the several embodiments provided by this invention, it should be understood that the disclosed system can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative; for example, the division of modules is only a logical functional division, and there may be other division methods in actual implementation.

[0032] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0033] Furthermore, the functional modules in the various embodiments of this invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in a combination of hardware and software functional modules.

[0034] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the essential characteristics of the present invention.

[0035] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A big data management system for on-site communication maintenance operations based on visual analytics, characterized in that: include: The multi-source data acquisition and processing module is used to acquire raw video streams and light intensity data at the communication maintenance site, perform adaptive light compensation processing and reverse repair of compression distortion on the raw video stream, generate enhanced video sequences, and extract light feature sequences. The feature detection and processing fusion module is used to perform target detection on the enhanced video sequence through a pre-built detection model, construct a multimodal feature matrix by combining the illumination feature sequence, and output the target detection results including single video frame index, optimized bounding box, corrected confidence and label category; The detection result optimization evaluation prompt module is used to process and analyze the single-frame optimization effect of the target detection results to obtain the first optimization label or the second optimization label. The first optimization label and the second optimization label are used to calculate the optimization reliability rate and optimization impact rate of the existing scheme. The optimization reliability rate and optimization impact rate are analyzed together and dynamic optimization management prompts are provided.

2. The communication maintenance field operation big data management system based on visual analysis according to claim 1, characterized in that, When performing illumination adaptive compensation processing on the original video stream, the original video frames in the original video stream are obtained. After normalizing the pixel values ​​of the original video frames to [0,1], they are converted into logarithmic brightness using the formula L(x,y)=log(I(x,y)+1), where I(x,y) is the pixel value of the original video frame and L(x,y) is the logarithmic brightness value.

3. The communication maintenance field operation big data management system based on visual analysis according to claim 2, characterized in that, Calculate illumination components ;in, is the Gaussian kernel at the i-th scale; i is the scale index, i=1,2,3; * indicates the convolution operation.

4. The communication maintenance field operation big data management system based on visual analysis according to claim 3, characterized in that, The reflection components at the three scales are weighted and fused to obtain the final reflection component R(x,y); the final reflection component is linearly stretched to output the video frame after illumination compensation.

5. The communication maintenance field operation big data management system based on visual analysis according to claim 4, characterized in that, When generating enhanced video sequences, the video frames after illumination compensation are fused with the high-definition frames repaired by the generative adversarial network at the pixel level, and the fusion weights are dynamically adjusted.

6. The communication maintenance field operation big data management system based on visual analysis according to claim 5, characterized in that, When extracting the illumination feature sequence, the illumination feature vector corresponding to each frame of video is extracted from the synchronous illumination intensity data; Illumination feature vectors include basic statistical features, illumination component features, and temporal features; The extracted illumination feature vectors are sorted and combined to output the illumination feature sequence.

7. The communication maintenance field operation big data management system based on visual analysis according to claim 6, characterized in that, Each frame of the enhanced video sequence is normalized and input into the detection model for inference analysis, outputting a list of targets for each frame. When constructing a multimodal feature matrix by combining illumination feature sequences, the visual feature vector of each target in the target list of each frame image is obtained; the visual feature vector is concatenated with the illumination feature vector to form a 261-dimensional multimodal feature vector, and the multimodal feature matrix is ​​constructed with the timestamp as the row index and the target ID as the column index.

8. The communication maintenance field operation big data management system based on visual analysis according to claim 7, characterized in that, The confidence level c is corrected using the illumination feature vector, specifically as follows: ;in, The corrected confidence level; This represents the illumination contrast in the illumination feature sequence.

9. The communication maintenance field operation big data management system based on visual analysis according to claim 8, characterized in that, Obtain the optimized bounding box from the target detection results and calculate the corresponding single-frame optimized value YD; if YD > M, generate the first optimized label; M is a natural number; Conversely, a second optimized label is generated.

10. The communication maintenance field operation big data management system based on visual analysis according to claim 9, characterized in that, The total number of consecutive first-optimization labels N1 and the total number of second-optimization labels N2 are counted, and the optimization reliability rate is obtained through calculation. and optimization impact rate ; These are respectively optimizing the reliability threshold and optimizing the impact threshold; like and If the optimization and improvement effect is normal, the existing plan should be maintained. Conversely, if the optimization effect is abnormal, the operation and maintenance personnel will be notified and optimization management prompts will be sent.

Images