Photovoltaic PHC pipe pile automatic statistical method and system based on unmanned aerial vehicle and optimized YOLOv11

Abstract

The invention discloses a photovoltaic PHC pipe pile automatic statistical method and system based on an unmanned aerial vehicle and optimized YOLOv11. The method comprises the steps that aerial photography is conducted through the unmanned aerial vehicle based on a target density self-adaptive algorithm, and flight parameters are dynamically adjusted; transmitting the video stream to an identification server in real time through an RTMP (Real Time Messaging Protocol) and an NGINX server; the server carries out water surface reflection suppression preprocessing on the image, and then a multi-module optimized YOLOv11 model is used for recognizing the PHC pipe pile; automatic counting and unique numbering are realized through a cross-frame time sequence consistency counting algorithm; and finally, visually displaying and storing a result on a Web platform, and constructing a traceable digital file. According to the method, key technologies such as unmanned aerial vehicle dynamic route coverage, high-speed streaming media transmission, high-performance AI recognition and consistency check counting are integrated, high precision, high real-time performance and complete traceability of statistical work are ensured, a traditional low-efficiency manual counting mode is replaced, and rapid and automatic digital statistics of the photovoltaic PHC pipe pile is achieved.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, and in particular relates to an automatic statistical method and system for photovoltaic PHC pipe piles based on UAVs and optimized YOLOv11. Background Technology

[0002] With the accelerated global energy structure transformation, the new energy industry, represented by photovoltaic power generation, is experiencing a large-scale construction boom. Photovoltaic power plants occupy vast areas and operate in complex terrains. Their foundation support structure—prestressed high-strength concrete (PHC) pipe piles—is numerous and densely distributed. Accurately calculating the quantity of these pipe piles is crucial for project schedule management, quality control, and cost settlement. Currently, the industry generally faces challenges such as continuously expanding project scale, tight construction schedules, and increasingly sophisticated management requirements, creating an urgent need for efficient and accurate engineering material and schedule verification technologies.

[0003] To address these needs, existing technologies primarily rely on two main approaches. First, drones equipped with cameras are used for aerial inspections. Streaming media technologies, such as RTMP (Real-Time Messaging Protocol), transmit the captured video streams back to ground control in real time, enabling non-contact data collection and preliminary monitoring of vast and complex terrains. Second, in the data backend processing stage, some solutions introduce deep learning-based object detection algorithms, such as RCNN and YOLO series general models, to automatically analyze the collected images or video frames to identify specific objects. These algorithms, combined with web frontend technologies, then graphically and tabularly display the identification results, constructing a basic technical framework from image acquisition to information visualization.

[0004] However, when these technologies are applied to real-world photovoltaic (PV) pile statistics, several shortcomings remain, primarily in recognition accuracy, statistical reliability, and system real-time performance. First, while drones and video transmission solve the data collection coverage problem, the final identification and counting of piles still largely relies on manual observation of videos or images, or on unoptimized general detection models. This makes it difficult to accurately identify densely distributed small cylindrical PHC piles under high-altitude aerial photography, especially when pile boundaries are unclear or there are shadows or water surface reflections, easily leading to missed or false detections. Second, traditional statistical methods often count based on single-frame recognition, lacking cross-frame temporal consistency analysis. When the same pile is identified multiple times in consecutive frames, it easily leads to duplicate counting, lacking a dynamic numbering mechanism to ensure cross-frame uniqueness. Third, existing drone aerial photography typically uses a preset fixed altitude, unable to adaptively adjust flight altitude and speed according to pile density in different areas, resulting in insufficient image resolution in some areas and reducing the overall recognition effect. Finally, the existing statistical process has a long chain from data collection and manual review to result generation, which leads to a lag in information updates, makes it difficult to achieve real-time synchronization of construction progress, and cannot provide reliable digital traceability records. Summary of the Invention

[0005] Purpose of the Invention: The purpose of this invention is to provide an automatic statistical method and system for photovoltaic PHC pipe piles based on unmanned aerial vehicles (UAVs) and optimized YOLOv11. This method utilizes UAV dynamic flight parameter optimization for real-time video stream data acquisition, a multi-module optimized YOLOv11 model for automatic pipe pile identification, and a cross-frame temporal consistency counting algorithm to achieve accurate, real-time, and traceable digital statistics of photovoltaic pipe piles. This method aims to solve the problems of low efficiency, high error rate, information lag, and poor identification performance in complex environments associated with existing manual statistical methods, providing reliable digital and intelligent support for photovoltaic project construction management.

[0006] Technical Solution: The automatic statistical method for photovoltaic PHC pipe piles based on UAVs and optimized YOLOv11 described in this invention includes the following steps:

[0007] S1. Using drones, optimize flight parameters based on target density adaptive algorithm, and take aerial photos of PHC pipe piles in the photovoltaic project area to obtain video stream data;

[0008] S2. The video stream data is transmitted in real time to the NGINX relay server via the RTMP protocol, and after distribution, it is pushed to the image recognition server to ensure low-latency transmission;

[0009] S3. The image recognition server receives the video stream and parses it into images frame by frame. First, it optimizes the image quality through a water surface reflection suppression algorithm, and then uses the YOLOv11 model optimized by multiple modules to identify the PHC pipe piles in each frame of the image.

[0010] S4. Based on the cross-frame temporal consistency counting algorithm, motion vector matching and trajectory association are performed on the identification results of each frame to realize automatic counting and unique cross-frame numbering of PHC pipe piles, avoiding duplicate counting;

[0011] S5. The identification, counting, and construction progress prediction results of PHC pipe piles are graphically displayed and stored on a Web visualization platform.

[0012] The automatic statistical method for photovoltaic PHC pipe piles proposed in this invention achieves automation and immediacy of data acquisition and transmission through real-time acquisition of video stream data via dynamic flight parameter optimization of UAVs, automatic identification of pipe piles using a multi-module optimized YOLOv11 model, and a cross-frame temporal consistency counting algorithm. The UAV dynamically adjusts flight parameters based on pipe pile density to ensure data acquisition quality. The transmission end utilizes an RTMP+NGINX architecture to achieve low-latency and highly stable data flow. The server-side uses a water surface reflection suppression algorithm and a multi-module optimized YOLOv11 model to analyze video frame by frame, accurately identifying PHC pipe piles in each frame while ensuring uniqueness through a cross-frame temporal consistency counting algorithm. Finally, a web visualization platform graphically displays and stores the identification, counting results, and construction progress forecasts. This effectively overcomes the problems of low efficiency, error-proneness, delayed updates, and poor scenario adaptability of traditional manual statistical methods, constructing an intuitive, reliable, and traceable digital management system for photovoltaic pipe piles, comprehensively promoting the evolution of photovoltaic project construction and management towards intelligence and refinement.

[0013] Preferably, in step S1, the target density adaptive algorithm enables the UAV to automatically match the optimal flight parameters in different pipe pile density areas, ensuring that the image resolution in high-density areas meets the requirements for small target recognition, while taking into account the acquisition efficiency in low-density areas, avoiding insufficient local resolution or waste of resources caused by fixed flight paths, and providing a high-quality data foundation for subsequent recognition.

[0014] Preferably, the algorithm is implemented as follows:

[0015] S11. During aerial photography, every 10 seconds, based on the preliminary identification results of the YOLOv11 model, the density ρ of the pipe piles in the current sub-region is dynamically updated. The calculation formula is as follows:

[0016]

[0017]

[0018] Where N is the number of pipe piles identified in the current area, S is the aerial coverage area of ​​the current sub-area, H is the shooting height, and θ is the number of pipe piles identified in the current area. h Let θ be the horizontal field of view of the camera. vCamera vertical field of view.

[0019] S12. Based on the real-time density ρ and the preset density threshold range, adaptively adjust the flight altitude H and speed V.

[0020]

[0021] Where H0 is the preset height, V0 is the preset speed, ρ1 is the high density threshold, and ρ2 is the low density threshold.

[0022] Preferably, in step S2, the video stream data is relayed and distributed via an NGINX server before being pushed to the image recognition server.

[0023] By introducing an NGINX server in step S2 to relay, distribute, and push video stream data, this method constructs an efficient and stable data transmission channel. This architecture effectively alleviates the pressure of data concurrency, ensuring that the video stream can be smoothly and with low latency transmitted to the image recognition server, thereby guaranteeing the continuity and stability of subsequent real-time recognition and processing. This not only optimizes the overall performance of the system under high-load scenarios but also further enhances the reliability and robustness of the solution in actual deployment, providing crucial support for completing highly time-sensitive and accurate recognition tasks.

[0024] Preferably, in step S3, the water surface reflection suppression algorithm can effectively suppress strong water surface reflection, filter noise, correct color deviation, significantly improve the contrast between the pipe pile and the background under complex water surface background, provide high-quality input for subsequent recognition models, and reduce the risk of false detection and false negative detection caused by reflection.

[0025] Preferably, the algorithm is implemented as follows:

[0026] S31. Convert the RGB image to the HSV color space and separate the luminance channel (V channel), hue channel (H channel), and saturation channel (S channel).

[0027] S32. Channel-specific brightness suppression: Adaptive threshold suppression is applied to the V channel, calculated using the following formula:

[0028]

[0029] Where V' is the suppressed luminance channel value; V is the original luminance channel value; T is the adaptive luminance threshold (calculated by the Otsu algorithm); and k is the suppression coefficient (0.3).

[0030] S33. Median filtering is applied to the S channel, with the window size set to 3×3, to remove surface noise interference.

[0031] S34. Perform histogram equalization on the H channel to correct the color shift caused by water reflection.

[0032] S35. Convert the processed H, S, and V channels back to RGB images and output a clear image with no glare after optimization.

[0033] Preferably, in step S3, the optimization of the YOLOv11 model using multiple modules includes: adding shallow convolutional layers to enhance the capture of small target features; introducing an edge enhancement feature extraction module for cylindrical small targets to enhance the boundary response of the pipe pile; adjusting the size of the detection head anchor frame through K-means clustering to adapt to the pixel distribution of the aerial pipe pile; embedding a model self-calibration module to adapt to environmental changes, and achieving accurate recognition in all scenes through multi-module collaboration.

[0034] Preferably, the model is constructed through the following steps:

[0035] S31. Add one shallow convolutional layer (3×3 kernel size, 1 stride, 128 output channels) to the backbone network layer to enhance the capture of small target features;

[0036] S32. After the shallow feature extraction network, an edge enhancement feature extraction module is inserted to improve boundary response capability. The specific steps are as follows:

[0037] S321. Use a Gaussian kernel with σ=1.0 to smooth the shallow feature map and remove noise;

[0038] S322. The gradient map is calculated using the Sobel operator in four directions (0°, 45°, 90°, 135°) to capture the multi-directional features of the circular edge of the pipe pile. The gradient calculation formula is as follows:

[0039]

[0040]

[0041]

[0042]

[0043] Where F is the shallow feature map, and ∂F / ∂x and ∂F / ∂y are the partial derivatives in the x and y directions, respectively.

[0044] S333. The four directional gradient maps are fused with the original shallow feature map by element-wise addition to output the edge-enhanced feature map. The calculation formula is as follows:

[0045]

[0046] Where ω1~ω4 are the fusion weights, all set to 0.25.

[0047] S33. Based on the actual pixel size distribution of the pipe pile target in the aerial image, the K-means clustering algorithm is used to re-cluster and generate 3 sets of anchor boxes [15,20], [25,30], and [35,40] to improve the regression accuracy;

[0048] S34. To improve the robustness of the model under different lighting conditions, a model self-calibration module is introduced. This module dynamically adjusts the recognition parameters based on real-time lighting conditions, further enhancing the model's recognition capability. The calculation formula is as follows:

[0049]

[0050]

[0051]

[0052] Where L is the light intensity, θ is the confidence threshold, and IoU thresh θ0 is the non-maximum suppression (NMS) threshold, W and H are the image width and height, I(i,j) is the gray value of pixel (i,j), θ0 is the baseline confidence threshold, IoU0 is the baseline NMS threshold, L0 is the standard illumination gray mean, and λ1=0.15 and λ2=0.1 are adjustment coefficients.

[0053] S35. Construct a dedicated dataset containing PHC pipe pile images under different lighting, terrain and weather conditions, and label the PHC pipe piles in the images;

[0054] S36. Divide the dedicated dataset into a training set, a validation set, and a test set;

[0055] S37. Train the multi-module optimized YOLOv11 model using the training set and validation set to obtain the special model.

[0056] Preferably, in step S4, the cross-frame temporal consistency counting algorithm can achieve stable association and unique identification of the pipe pile target across multiple video frames, thereby avoiding duplicate counting caused by repeated detection of consecutive frames and ensuring the uniqueness and accuracy of the statistical results.

[0057] Preferably, the specific implementation steps of the algorithm are as follows:

[0058] S41. For each frame of image, extract the center coordinates (x, y), bounding box size (w, h), and feature vector f of the identified pipe pile target;

[0059] S42. The motion vector from the previous frame to the current frame is calculated using the Lucas-Kanade (LK) optical flow algorithm. The position of the target in the previous frame is predicted to obtain the predicted coordinates (x', y').

[0060] S43. Calculate the overall matching degree between the detected target and the predicted target in the current frame. The matching degree is obtained by weighting the positional similarity and feature similarity. The specific calculation formula is as follows:

[0061]

[0062]

[0063]

[0064] Wherein, IoU is the intersection-union ratio between the predicted bounding box and the current detected bounding box, used to measure positional similarity; Sim is the feature similarity; Score is the overall matching degree; BB1 is the current detected target bounding box; BB2 is the predicted target bounding box; F1 is the current detected target feature vector; F2 is the predicted target feature vector; and α=0.7 is the weighting coefficient. When Score>0.6, it is determined to be the same target and inherits the unique number from the previous frame; when Score≤0.6, it is determined to be a new target and is assigned a new unique number.

[0065] S44. When the same target is not matched for 3 consecutive frames, the trajectory is marked as terminated, and the count confirmation is completed.

[0066] Preferably, in step S5, the storage includes automatically saving the raw video stream data of the inspection, images with identification tags, and statistical results of the pipe piles in each region. All data is automatically archived by date and region to form a structured digital archive.

[0067] Preferably, in step S5, the construction progress prediction is based on the average daily completion volume over the past 3 days, automatically estimating the remaining construction period using the following formula:

[0068]

[0069] If the "average daily completion volume in the past 3 days" is zero, the system will automatically mark the progress status as "no construction at present", which will intuitively indicate that there is currently no construction progress.

[0070] By automatically saving the original video stream data of the inspection, images with identification tags, and statistical results of the pipe piles in each region, this method constructs a complete, structured, and multimodal traceable digital archive system. This system not only achieves full-chain digital traceability of the statistical process and results, providing authentic and reliable original evidence for subsequent queries, verification, and backtracking, but also greatly enhances the transparency and credibility of project management, making quality review, progress auditing, and accountability traceability efficient and rigorous, thereby comprehensively improving the digitalization and standardization level of pipe pile engineering management in photovoltaic projects.

[0071] The automatic statistical system for photovoltaic PHC pipe piles based on UAVs and optimized YOLOv11 described in this invention includes:

[0072] The data acquisition and transmission module, including the UAV and communication unit, optimizes the UAV flight parameters based on the target density adaptive algorithm, acquires video streams of the PHC pipe pile area in real time, and transmits the video streams to the NGINX server in real time with low latency via the RTMP protocol. After being distributed by the NGINX server, the streams are pushed to the image recognition server, ensuring low latency and high stability of data flow.

[0073] The image recognition and processing module includes an image recognition server, which receives video streams and parses them frame by frame. First, it optimizes image quality through a water surface reflection suppression algorithm. Then, it uses a YOLOv11 dedicated model optimized by multiple modules to accurately identify pipe piles. Finally, it uses a cross-frame temporal consistency counting algorithm to achieve automatic counting and unique cross-frame numbering to avoid duplicate counting.

[0074] The data display and storage module, including a web visualization platform and database, is used to graphically display and structurally store the identification and counting results of PHC pipe piles, as well as the construction progress prediction based on the average completion volume of the past 3 days.

[0075] The photovoltaic PHC pipe pile automatic statistics system constructed in this invention integrates three core modules: data acquisition and transmission, image recognition and processing, and data display and storage. This achieves a fully automated closed-loop process, from real-time acquisition of on-site video streams and intelligent recognition based on a multi-module optimized YOLOv11 model, to the visualization and structured storage of results. The system effectively integrates key technologies such as UAV mobile coverage, high-speed streaming media transmission, high-performance AI recognition, and an interactive digital platform. It addresses the pain points of traditional, inefficient, and error-prone manual statistics methods and general detection models, ensuring high accuracy, strong real-time performance, and complete traceability in PHC pipe pile statistics. This promotes the intelligent, refined, and digital transformation of photovoltaic project construction and management.

[0076] Preferably, the system further includes a video stream relay server located between the data acquisition and transmission module and the image recognition and processing module, for receiving video streams transmitted by the UAV and distributing them to the image recognition server.

[0077] By adding a video stream relay server located between the data acquisition and transmission module and the image recognition processing module, a stable and efficient video data transfer and distribution hub was constructed. This design not only effectively decouples the key processes of front-end data acquisition and back-end recognition processing, enhancing the flexibility and scalability of the system architecture, but also ensures the smoothness, stability, and load balancing of data transmission through unified reception and intelligent distribution of video streams. This significantly improves the overall robustness and processing efficiency of the entire system in the face of high concurrency or complex network environments, providing a solid guarantee for continuous and reliable real-time recognition and analysis at the back end.

[0078] Preferably, the YOLOv11-specific model used in this invention has undergone a series of targeted multi-module deep optimizations, with core innovations covering four key modules. First, by adding convolutional layers to the shallow layers of the backbone network, the feature capture capability for typical small targets (such as pipe piles) in aerial images is significantly enhanced. Second, an edge enhancement feature extraction module for cylindrical targets is innovatively introduced. This module calculates gradients based on multi-directional Sobel operators and adaptively fuses them with the original feature map, effectively enhancing the boundary response capability of pipe piles and solving the difficulty of identifying their blurred edges. Third, to improve positioning accuracy, the model redesigns the anchor frame size of the detection head based on the K-means clustering algorithm, making it accurately adaptable to the actual pixel size distribution of PHC pipe piles under high-altitude aerial photography. Finally, the model embeds an environmental self-calibration module, which can dynamically adjust the confidence threshold and NMS threshold according to the real-time illumination intensity of the input image, thereby significantly improving the robustness and adaptability of the model under different lighting conditions. Through the collaborative optimization of the above-mentioned multiple modules, the model achieves accurate and stable identification of small, varied, and densely distributed pipe pile targets from an aerial perspective. It effectively overcomes various challenges caused by long shooting distance, dense targets, and fluctuations in ambient lighting. Ultimately, it provides the system with powerful, reliable, and highly professional core identification capabilities, ensuring the accuracy and robustness of the statistical results across the entire scene.

[0079] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: 1. This invention solves the problem of low efficiency in manual statistics by automatically identifying and counting the pipe piles through real-time video stream acquisition by UAVs and a multi-module optimized YOLOv11 model, achieving rapid and automatic digital statistics of photovoltaic PHC pipe piles; 2. This invention proposes a target density adaptive algorithm, enabling the UAV to dynamically adjust flight parameters to ensure that the image resolution in high-density areas meets the requirements for small target identification, providing a high-quality data foundation for subsequent identification; 3. This invention introduces a water surface reflection suppression algorithm for image preprocessing and deeply optimizes the YOLOv11 model, including the introduction of an edge enhancement feature extraction module and an environmental self-calibration module. Through these optimizations, the system significantly improves the feature capture capability of small aerial targets and the contrast between the pipe piles and the background in complex water surface backgrounds, ensuring high-precision identification and positioning, and effectively reducing false detection and false negative rates; 4. This invention adopts a cross-frame temporal consistency counting algorithm, which achieves stable association and unique identification of pipe pile targets across multiple video frames through motion vector matching and trajectory association, avoiding duplicate counting caused by repeated detection in consecutive frames, and ensuring the uniqueness and accuracy of statistical results. 5. The system, based on an RTMP+NGINX architecture, achieves low-latency, real-time transmission of video streams and frame-by-frame analysis on the server side. It can instantly complete identification and counting, and visualize and predict construction progress through a web platform. This effectively solves the problem of delayed information updates, while automatically saving multimodal data, constructing a complete digital record, and achieving reliable digital traceability and auditability of the statistical process and results. Attached Figure Description

[0080] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0081] Figure 2 This is a schematic diagram of a dataset example from the present invention;

[0082] Figure 3 This is a schematic diagram of the water surface reflection preprocessing algorithm results of the present invention;

[0083] Figure 4 This is a schematic diagram of the web visualization results of the present invention. Detailed Implementation

[0084] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0085] This invention provides an automatic statistical method and system for photovoltaic PHC pipe piles based on unmanned aerial vehicles and multi-module optimized YOLOv11, such as... Figure 1 As shown, it includes the following steps:

[0086] 1. Construct a multi-module optimized YOLOv11 pipe pile detection model

[0087] (1) Dataset Construction: A total of 1000 images of photovoltaic pipe piles were collected under different lighting conditions (sunny day, cloudy day), terrain (fishpond, tea garden, muddy ground), and weather conditions (sunny day, foggy day). The location (boundary box coordinates) and number of pipe piles were labeled. The number of pipe pile samples was 28851. A dedicated dataset was constructed and divided into training set, validation set and test set in an 8:1:1 ratio. Some image samples are shown below. Figure 2 As shown.

[0088] (2) Multi-module optimization: First, add one shallow convolutional layer (3×3 kernel size, stride 1, 128 output channels) to enhance the capture of small target features; then, insert an edge enhancement feature extraction module after the shallow feature extraction network, with the fusion weights all set to 0.25 to enhance the response capability of the pipe pile boundary; then, perform cluster analysis on the above-constructed dataset using the K-means clustering algorithm, adjust the size of the detection head anchor frame to adapt to the pixel distribution of the aerial pipe pile, and generate 3 sets of anchor frames, namely [15,20], [25,30], and [35,40]; finally, embed a model parameter self-calibration module, dynamically adjust the recognition parameters through real-time illumination intensity L, set the baseline confidence threshold θ0 to 0.2, set the baseline NMS threshold IoU0 to 0.15, and set the adjustment coefficients to λ1=0.15 and λ2=0.1.

[0089] (3) Model training: The model parameters were optimized using the stochastic gradient descent (SGD) algorithm. The learning rate η=0.001, momentum γ=0.9, weight decay λ=0.0005, and training epochs=150. The final model achieved a recognition accuracy of 94.76% and a recall of 93.81% on the test set. The calculation formulas are as follows:

[0090]

[0091]

[0092] Among them, TP is a true positive (actually a pipe pile and correctly identified), FP is a false positive (actually not a pipe pile but misidentified as a pipe pile), and FN is a false negative (actually a pipe pile but not identified).

[0093] 2. Data Acquisition and Real-time Transmission

[0094] (1) Equipment deployment: Select drones (DJI Mavic4 Pro) adapted to complex terrain, and plan aerial photography routes for different working areas of photovoltaic projects such as fishery-solar complementary and tea-solar complementary projects to ensure that all pipe pile areas are covered without blind spots. Configure an image recognition server, deploy an NGINX relay server and a Web visualization platform.

[0095] (2) Video Stream Transmission and Preprocessing: During aerial photography, the UAV automatically matches the optimal flight parameters based on the target density adaptive algorithm. The preset altitude H0 is set to 80m, the preset speed V0 is set to 7m / s, the high density threshold ρ1 is set to 0.02 roots / m², and the low density threshold ρ2 is set to 0.018 roots / m². The real-time video stream from the UAV is transmitted to the NGINX server (a high-performance HTTP and reverse proxy server, which acts as a "relay and distribution" mechanism, receiving the video stream from the UAV and efficiently forwarding it to the image recognition server to ensure transmission stability and load balancing) via the RTMP protocol (Real-Time Messaging Protocol, a network protocol specifically designed for low-latency transmission of audio and video data, which can quickly synchronize data between the acquisition end and the receiving end, and is commonly used in live streaming, real-time monitoring, and other scenarios) through the RTMP protocol. The server pushes the video stream to the image recognition server in real time, with an end-to-end transmission delay of less than 1 second, solving the problem of information lag in traditional manual reporting. The image recognition server parses the video stream frame by frame (resolution rate 25 frames / second), initiates a water surface reflection preprocessing algorithm for images containing water surfaces, sets the suppression coefficient k to 0.3, applies median filtering to the S channel to remove noise, and finally converts it back to an RGB image. Some images are shown below. Figure 3 As shown.

[0096] 3. Automatic identification and consistency verification counting of pipe piles

[0097] (1) The PHC pipe piles in each frame of the image are identified by the trained YOLOv11 model.

[0098] (2) Automatic counting and unique cross-frame numbering of PHC pipe piles are achieved through a cross-frame temporal consistency counting algorithm, with the weight coefficient α set to 0.7. When the overall matching score > 0.6, it is determined to be the same target and inherits the unique number of the previous frame; when the score ≤ 0.6, it is determined to be a new target and a new unique number is assigned. When the same target is not matched for 3 consecutive frames, it is marked as trajectory termination, and the counting confirmation is completed.

[0099] 4. Data visualization and record management

[0100] (1) The identification results are synchronized to a Web visualization platform to present the statistics of the number of pipe piles in each area and the prediction of construction progress in a graphical manner. For example Figure 4 As shown.

[0101] (2) Storage includes automatically saving the original video stream data of the inspection, images with identification tags, and the statistical results of the pipe piles in each area. All data is automatically archived by date and area to form a structured digital archive.

[0102] 5. Model robustness verification

[0103] Taking a photovoltaic power generation project as an example, three typical test areas were selected:

[0104] Area A: Aquaculture-solar hybrid area, fishpond surface, 860 pipe piles;

[0105] Area B: Tea-solar complementary area, tea garden topography, 630 pipe piles;

[0106] Area C: Mixed area, including muddy ground and grassland, with 1020 pipe piles.

[0107] The test results are shown in the table below:

[0108] Test area Actual quantity (roots) Count of roots in the optimized YOLOv11 model Count of roots in the unoptimized YOLOv11 model Optimize model error rate Unoptimized model error rate Statistics on time taken (minutes) Manually calculate the time (in minutes). Area A 860 857 802 0.35% 6.74% 8 75 Area B 630 628 589 0.32% 6.51% 6 60 Area C 1020 1015 937 0.49% 8.14% 10 90

[0109] The results show that in diverse and complex field scenarios such as fishery-solar hybrid water surfaces, tea-solar hybrid terrain, and mixed areas, the optimized YOLOv11 model stably controls the statistical error rate to within 0.5%, a significant reduction compared to the unoptimized YOLOv11 model's average error rate of over 6%, fully demonstrating the model's strong robustness and high accuracy. Furthermore, the automatic statistical calculation time is only about 1 / 8 to 1 / 10 of that of traditional manual calculation, efficiently solving the inefficiency and time-consuming drawbacks of traditional manual counting. In summary, this invention successfully verifies its applicability to all scenarios, extremely high statistical reliability, and significant efficiency advantages.

Claims

1. An automatic statistical method for photovoltaic PHC pipe piles based on UAVs and optimized YOLOv11, characterized in that, Includes the following steps: S1. Using drones, optimize flight parameters based on target density adaptive algorithm, and take aerial photos of PHC pipe piles in the photovoltaic project area to obtain video stream data; S2. The video stream data is transmitted in real time to the NGINX relay server via the RTMP protocol, and then pushed to the image recognition server after distribution; S3. The image recognition server receives the video stream and parses it into images frame by frame. First, it preprocesses the images using a water surface reflection suppression algorithm, and then uses the multi-module optimized YOLOv11 model to identify the PHC pipe piles in each frame of the image. S4. Based on the cross-frame temporal consistency counting algorithm, motion vector matching and trajectory association are performed on the identification results of each frame to realize automatic counting and unique cross-frame numbering of PHC pipe piles. S5. The identification, counting, and construction progress prediction results of PHC pipe piles are graphically displayed and stored on a Web visualization platform.

2. The method according to claim 1, characterized in that, In step S1, the UAV plans an initial aerial photography path based on the terrain and work surface type of the photovoltaic project, and dynamically adjusts the UAV's flight altitude and speed based on the target density adaptive algorithm during the aerial photography process.

3. The method according to claim 1, characterized in that, In step S2, the NGINX relay server is used to receive the video stream transmitted by the drone and distribute it to the image recognition server to achieve load balancing.

4. The method according to claim 1, characterized in that, In step S3, the water surface reflection suppression algorithm includes the following steps: Convert the RGB image to the HSV color space and separate the luminance channel, hue channel, and saturation channel; Adaptive threshold suppression is applied to the brightness channel; The saturation channel is subjected to median filtering. Histogram equalization is performed on the tone channels; The processed luminance, saturation, and hue channels are converted back to RGB images, and the optimized image is output.

5. The method according to claim 1, characterized in that, In step S3, the construction steps of the multi-module optimized YOLOv11 model include: A dedicated dataset containing images of PHC pipe piles under different lighting, terrain, and weather conditions was constructed, and the PHC pipe piles in the images were labeled. The dedicated dataset is divided into a training set, a validation set, and a test set; A YOLOv11 model was used as the base model, and the following structural optimizations were performed on the base model: a shallow convolutional layer was added to the backbone network layer of the base model; an edge enhancement feature extraction module was inserted after the shallow convolutional layer; anchor boxes were regenerated based on the pixel size distribution of the pipe pile targets in the dedicated dataset using the K-means clustering algorithm; and a model self-calibration module was introduced into the base model to dynamically adjust the model recognition parameters according to real-time lighting conditions. The training and validation sets are used to train the structure-optimized model to obtain the multi-module optimized YOLOv11 model.

6. The method according to claim 5, characterized in that, The processing steps of the edge enhancement feature extraction module include: Use a Gaussian kernel to smooth the shallow feature map; The gradient map is calculated using Sobel operators in multiple directions; The gradient maps from multiple directions are fused with the original shallow feature map to output an edge-enhanced feature map.

7. The method according to claim 1, characterized in that, In step S4, the cross-frame temporal consistency counting algorithm calculates the intersection-union ratio and feature similarity between targets in adjacent frames to obtain the comprehensive matching degree, thereby realizing the trajectory association of targets across frames.

8. The method according to claim 1, characterized in that, In step S5, the storage includes raw video stream data, images with identification marks, and statistical results of pipe piles in each region, to construct a complete and structured traceable digital archive system.

9. An automatic statistical system for photovoltaic PHC pipe piles based on unmanned aerial vehicles and optimized YOLOv11, characterized in that, include: The data acquisition and transmission module, including the UAV and communication unit, is used to optimize the UAV flight parameters based on the target density adaptive algorithm, acquire video streams of the PHC pipe pile area, and transmit the video streams to the NGINX relay server via the RTMP protocol. The NGINX relay server is used to distribute and push the video stream to the image recognition and processing module; The image recognition and processing module includes an image recognition server, which receives video streams and parses them into images frame by frame. It preprocesses the images using a water surface reflection suppression algorithm, uses a multi-module optimized YOLOv11 model to identify PHC pipe piles in the images, and automatically counts and assigns unique cross-frame numbers to the identified PHC pipe piles based on a cross-frame temporal consistency counting algorithm. The data display and storage module, including a web visualization platform and database, is used to graphically display and structurally store the identification, counting results, and construction progress prediction of PHC pipe piles.

10. The system according to claim 9, characterized in that, The multi-module optimized YOLOv11 model adds shallow convolutional layers to the backbone network layers of the YOLOv11 model, inserts an edge enhancement feature extraction module and a model self-calibration module, and adopts a model with anchor boxes regenerated based on a dedicated dataset.

Images