Roller compacted concrete dam warehouse surface resource identification method, system and device based on YOLOv10
By deeply optimizing the YOLOv10 model, the problem of multi-scale target identification under complex working conditions of roller-compacted concrete dam surfaces was solved, enabling accurate real-time monitoring of construction site resources and improving the intelligence and safety of construction management.
Patent Information
- Application Number
- CN202610364549.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-24
- Publication Date
- 2026-07-03
AI Technical Summary
Existing technologies cannot achieve accurate identification of multi-scale targets under complex working conditions on roller-compacted concrete dam surfaces. In particular, the identification accuracy is insufficient in environments with dense obstruction and strong background interference, resulting in low levels of intelligence in construction management and difficulty in timely warning of safety hazards.
A resource identification system for roller-compacted concrete dam surfaces was constructed by employing a deeply optimized YOLOv10 model, combined with an adaptive feature enhancement module, a multi-scale collaborative prediction head structure, and a scale-aware dual label allocation optimization strategy, to achieve real-time image preprocessing and target recognition of video streams.
It has achieved full-element automated, precise, and real-time monitoring of roller-compacted concrete dam surfaces, improving the level of intelligent construction management and on-site operation safety, and reducing the rate of missed detections and false detections.
Smart Images
Figure CN122336665A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent construction and computer vision technology, specifically relating to a method, system and equipment for identifying resources on the surface of roller-compacted concrete dams based on YOLOv10, and particularly to an implementation method for automated real-time identification and management of multiple types and scales of resources on the surface of roller-compacted concrete dams (RCCD) under complex working conditions. Background Technology
[0002] Roller-compacted concrete dams, as the mainstream dam type in water conservancy and hydropower projects, are characterized by fast construction speed, good economy, and strong adaptability. The construction operation belongs to the category of large-scale, high-intensity, and multi-trade collaborative operation. The dam surface, as its core construction platform, gathers various key resources such as compaction machinery, transport vehicles, and workers. It is the core area of dynamic operation. The level of dam surface management directly affects the dam construction quality, progress, and operational safety, and is the key to achieving refined and intelligent construction of the project.
[0003] Currently, the management of roller-compacted concrete dam surfaces still relies on traditional methods such as manual observation, walkie-talkie dispatching, and manual recording. This extensive management model has many inherent defects: on the one hand, information transmission is delayed and highly subjective, and information such as on-site resource status and construction conflicts cannot be synchronized in real time. Dispatch decisions are based entirely on the experience of management personnel, which can easily lead to problems such as idle machinery, poor process coordination, and missed hazard identification, thus slowing down construction efficiency. On the other hand, data collection is fragmented, safety control is lacking, it is difficult to retain complete construction data, it is impossible to carry out refined efficiency analysis, and it is difficult to provide timely warnings of hazards such as personnel illegally entering the work area and equipment operating improperly, resulting in insufficient safety assurance and seriously restricting the improvement of project construction quality and efficiency.
[0004] To address the shortcomings of traditional management methods, the industry has attempted to apply YOLO series early target detection algorithms based on convolutional neural networks to resource identification and management at construction sites, leveraging the algorithm's advantage of rapid detection to optimize on-site management. However, the construction environment of roller-compacted concrete dam surfaces presents unique challenges, including open work areas, significant differences in target size, dense obstruction by personnel and equipment, strong light and dust interference, and complex backgrounds. Existing early YOLO algorithms lack adaptability, exhibit low sensitivity to small targets on the dam surface, are significantly affected by background interference, and have high rates of missed and false detections under dense obstruction conditions. Their recognition accuracy and stability are insufficient to meet the real-time management needs of the dam surface, making it impossible to achieve accurate and efficient identification of dam surface resources.
[0005] In summary, the traditional manual method for resource management on roller-compacted concrete dam surfaces suffers from significant drawbacks and low levels of automation. Existing target detection technologies are ill-suited to the complex working conditions of the surface, struggling to identify multi-scale targets in environments with strong interference. This has become a technical bottleneck hindering the intelligent upgrading and efficiency optimization of surface construction. Therefore, developing a high-precision resource identification method adapted to the working conditions of roller-compacted concrete dam surfaces and overcoming the shortcomings of existing algorithms is a pressing technical problem that needs to be solved by those skilled in the art. Summary of the Invention
[0006] To overcome the shortcomings of existing technologies, this invention proposes a resource identification method, system, and device for roller-compacted concrete dam surface based on YOLOv10. By deeply optimizing the YOLOv10 framework, it specifically solves the problem of resource identification under multi-scale, dense occlusion, and strong background interference in RCCD dam surface scenarios, so as to achieve automated, accurate, and real-time monitoring of all construction elements.
[0007] The objective of this invention can be achieved through the following technical solutions.
[0008] A method for resource identification of roller-compacted concrete dam surface based on YOLOv10 includes the following steps: S1 Data Acquisition: Real-time video stream of the roller-compacted concrete dam surface is acquired through image acquisition devices deployed at the construction site; S2 Image Preprocessing: After deframing the video stream described in step S1, each frame of the image is normalized and then enhanced with an environment to obtain the enhanced surface image. Construction of S3 Dedicated Dataset for Compacted Concrete Dam Surface Resources: Define target recognition categories for roller-compacted concrete dam surface resources, and perform fine annotation on the surface images obtained in step S2 according to the defined categories to obtain annotated images, thus completing the construction of the dedicated dataset for roller-compacted concrete dam surface resources; S4 Model Construction and Training: Based on the dedicated dataset of roller-compacted concrete dam surface resources constructed in step S3, a deeply optimized roller-compacted concrete dam surface resource identification model was constructed and trained using the YOLOv10 model. S5 Resource Identification: Using the resource identification model of the roller-compacted concrete dam surface trained in step S4, inference is performed on the real-time video stream of the roller-compacted concrete dam surface, and the inference results are post-processed by the spatiotemporal correlation target tracking algorithm to obtain structured resource information.
[0009] Further, the specific process of step S1 is as follows: deploy several industrial-grade network cameras on the cableway system above the roller-compacted concrete dam surface and on the high points of the mountains on both sides, and obtain real-time video streams of the roller-compacted concrete dam surface through the deployed cameras.
[0010] Further, the specific process of step S2 is as follows: after deframing the video stream described in step S1, each frame image is standardized, including size unification and pixel value normalization, and an adaptive environment enhancement algorithm is used to perform illumination compensation and noise suppression on each standardized frame image to obtain an enhanced surface image.
[0011] Furthermore, the target identification categories for the roller-compacted concrete dam surface resources mentioned in step S3 include at least bulldozers, vibratory rollers, pavers, transport trucks, and construction workers; during the construction of the dedicated dataset for surface resources mentioned in step S3, labeled images under different external environmental conditions are collected simultaneously to expand the content of the dedicated dataset for roller-compacted concrete dam surface resources.
[0012] Furthermore, the roller-compacted concrete dam surface resource identification model described in step S4 is constructed based on the existing YOLOv10 model. An adaptive feature enhancement module is embedded at the output of each PAN path of the neck network of the YOLOv10 model. The prediction head of the YOLOv10 model adopts a multi-scale collaborative prediction head structure. The training strategy of the roller-compacted concrete dam surface resource identification model is divided into two steps: pre-training and transfer learning. First, the weights of the YOLOv10 model pre-trained on a large public dataset are loaded. Then, the model is transferred and fine-tuned using the roller-compacted concrete dam surface resource-specific dataset constructed in step S3. In addition, a scale-aware dual-label allocation optimization strategy is also adopted during the training process of the roller-compacted concrete dam surface resource identification model.
[0013] Furthermore, the adaptive feature enhancement module employs a lightweight contextual attention module, which aggregates multi-scale contextual information, enabling the model to better distinguish between the target and the background with similar textures, and adaptively enhances effective target features and suppresses noise features. The multi-scale collaborative prediction head structure adds a P2-level detection head specifically for small targets to the original prediction head of the YOLOv10 model. This detection head fuses high-resolution feature maps from shallower layers of the backbone network and works in conjunction with detection heads of other sizes to form a full-scale detection system covering targets from small to very large.
[0014] Furthermore, the scale-aware dual-label allocation optimization strategy is an improvement on the original label allocation mechanism of the YOLOv10 model, which dynamically adjusts the number of positive sample candidate boxes assigned to the actual target box according to its size.
[0015] The objective of this invention can also be achieved through the following technical solutions.
[0016] A resource identification system for roller-compacted concrete dam surfaces based on YOLOv10 includes a data acquisition module, an image preprocessing module, a dedicated dataset construction module for surface resources, a model construction and training module, and a resource identification module; among which, Data acquisition module: Acquires real-time video streams of the roller-compacted concrete dam surface through image acquisition devices deployed at the construction site; wherein, the image acquisition devices consist of several industrial-grade network cameras, arranged on the cableway system above the roller-compacted concrete dam surface and at high points on both sides of the mountain. Image preprocessing module: After defraction of the video stream acquired by the data acquisition module, each frame of the image is standardized sequentially, and an adaptive environment enhancement algorithm is used to enhance the environment of each standardized frame of the image to obtain the enhanced warehouse surface image. The dedicated dataset construction module for roller-compacted concrete dam surface resources defines target recognition categories for roller-compacted concrete dam surface resources. These categories must include at least bulldozers, vibratory rollers, pavers, transport trucks, and construction workers. Based on these categories, the enhanced surface images from the image preprocessing module are finely annotated to obtain annotated images, thus completing the construction of the dedicated dataset for roller-compacted concrete dam surface resources. During the construction of the dedicated dataset, annotated images under different external environmental conditions are simultaneously collected to expand the content of the dedicated dataset for roller-compacted concrete dam surface resources. Model building and training module: Based on a dedicated dataset of roller-compacted concrete dam surface resources, a YOLOv10 model is used to build and train a deeply optimized roller-compacted concrete dam surface resource identification model; Resource identification module: The trained roller-compacted concrete dam surface resource identification model is used to infer the real-time video stream of the roller-compacted concrete dam surface, and the inference results are post-processed by the spatiotemporal correlation target tracking algorithm to obtain structured resource information.
[0017] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described method for identifying resources on the surface of a roller-compacted concrete dam based on YOLOv10.
[0018] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the YOLOv10-based method for identifying resources on the surface of a roller-compacted concrete dam as described above.
[0019] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: (1) The present invention uses the YOLOv10 model to construct and train a deep-optimized roller-compacted concrete dam surface resource identification model. The model is used to infer the real-time video stream of the roller-compacted concrete dam surface, identify the target category, confidence level, location coordinates, etc., and the above identification results are post-processed by the spatiotemporal correlation target tracking algorithm to obtain structured resource information, including: the quantity of various resources, real-time location, dwell time, movement trajectory, etc.
[0020] (2) At the output of each PAN path of the neck network of the present invention, a lightweight context attention module is specifically embedded as an adaptive feature enhancement module to aggregate multi-scale context information, enhance target features and suppress background noise, that is, enhance the model’s feature perception of key resources (such as personnel and machinery) while suppressing interference from complex backgrounds.
[0021] (3) The prediction head of the model of the present invention adopts a multi-scale collaborative prediction head structure. On the basis of the original prediction head of the YOLOv10 model, a P2 level detection head is added, which is dedicated to small targets. It integrates high-resolution feature maps from shallower layers of the backbone network and works in collaboration with other detection heads (P3, P4, P5) to form a full-scale detection system covering targets from small targets to super-large targets.
[0022] (4) The present invention introduces a scale-aware dual label allocation optimization strategy during the model training process, which improves the original label allocation mechanism of the YOLOv10 model. The allocation of positive samples is dynamically adjusted according to the target size to ensure that both large and small targets can be fully learned.
[0023] This invention enables automated, precise, and real-time monitoring of all elements of RCCD warehouse construction, effectively improving the level of intelligent construction management and on-site operational safety. Attached Figure Description
[0024] Figure 1 This is a schematic diagram of the resource identification method for roller-compacted concrete dam surface based on YOLOv10 of this invention. Detailed Implementation
[0025] The present invention will now be further described with reference to the accompanying drawings.
[0026] This invention discloses a resource identification method for roller-compacted concrete dam construction surfaces based on YOLOv10. This method aims to solve the problem that the complex construction surface environment of roller-compacted concrete dams (RCCD) leads to insufficient identification accuracy and susceptibility to interference of existing target detection algorithms under conditions of multi-scale targets (especially small targets), dense occlusion, and strong background interference.
[0027] like Figure 1As shown, the present invention provides a method for resource identification of roller-compacted concrete dam surface based on YOLOv10, which specifically includes the following steps: S1 Data Acquisition: Real-time video streams of the roller-compacted concrete dam surface are acquired through image acquisition devices deployed at the construction site.
[0028] S2 Image Preprocessing: After deframing the video stream described in step S1, each frame of the image is normalized and then enhanced with an environment enhancement to obtain the enhanced surface image.
[0029] Construction of the S3 Dedicated Dataset for Roller-Compacted Concrete Dam Surface Resources: Define target recognition categories for roller-compacted concrete dam surface resources, and perform fine-grained annotation on the surface images obtained in step S2 according to the defined categories to obtain annotated images, thus completing the construction of the dedicated dataset for roller-compacted concrete dam surface resources.
[0030] S4 Model Construction and Training: Based on the dedicated dataset of roller-compacted concrete dam surface resources constructed in step S3, a deeply optimized roller-compacted concrete dam surface resource identification model was constructed and trained using the YOLOv10 model. S5 Resource Identification: Using the resource identification model of the roller-compacted concrete dam surface trained in step S4, inference is performed on the real-time video stream of the roller-compacted concrete dam surface, and the inference results are post-processed by the spatiotemporal correlation target tracking algorithm to obtain structured resource information.
[0031] In one possible implementation, preferably, the specific process of step S1 is as follows: Several (e.g., 4) image acquisition devices using industrial-grade network cameras are deployed on the cableway system above the roller-compacted concrete dam surface and on the high points of the mountains on both sides. The real-time video stream of the roller-compacted concrete dam surface is obtained through the deployed cameras, and the resolution can be 3840x2160.
[0032] In one possible implementation, preferably, the specific process of step S2 is as follows: after deframing the video stream described in step S1, each frame image is standardized, including size unification and pixel value normalization, and an adaptive environment enhancement algorithm is used to perform illumination compensation and noise suppression on each standardized frame image to suppress the influence of external environmental factors (such as strong light, shadow, fog or dust) on image quality, and finally obtain the enhanced warehouse surface image.
[0033] In one possible implementation, preferably, step S3 involves defining target identification categories for roller-compacted concrete dam surface resources. These categories include at least five types: bulldozers, rollers, pavers, trucks, and workers. Based on these defined categories, the surface images obtained in step S2 are refined with bounding boxes. Targets such as bulldozers, rollers, pavers, trucks, and workers are outlined on the surface images and assigned corresponding category labels, resulting in labeled images. This ultimately completes the construction of a dedicated dataset for roller-compacted concrete dam surface resources. Furthermore, this invention simultaneously collects labeled images under different external environmental conditions (light, dust, occlusion, etc.) to expand the content of the dedicated dataset for roller-compacted concrete dam surface resources and improve the model's generalization ability.
[0034] In one possible implementation, preferably, the roller-compacted concrete dam surface resource identification model described in step S4 is constructed based on an existing YOLOv10 model (e.g., the YOLOv10-L model), mainly including a backbone network, a neck network, and prediction heads.
[0035] The backbone network is a feature extraction layer, and its basic structure is the efficient CSP architecture adopted by YOLOv10, which is used to extract multi-scale features from low-level texture to high-level semantics in images.
[0036] The neck network is a feature fusion layer, and its basic structure is the Path Aggregation Network (PAN) used in YOLOv10. At the output of each PAN path in the YOLOv10 neck network, an adaptive feature enhancement module is specifically embedded to enhance the model's discriminative ability. This adaptive feature enhancement module employs a lightweight Contextual Attention Module, which can aggregate multi-scale contextual information, enabling the model to better distinguish targets (such as construction workers) from backgrounds with similar textures (such as concrete surfaces), and adaptively enhance effective target features while suppressing noisy features.
[0037] The prediction head belongs to the prediction layer and adopts a multi-scale collaborative prediction head structure. Based on the original prediction head of the YOLOv10 model, a P2-level detection head is added specifically for small targets, responsible for predicting targets ranging in size from 8x8 to 32x32 pixels. This detection head fuses high-resolution feature maps from shallower layers of the backbone network and works in conjunction with other detection heads of different sizes (P3, P4, P5) to form a full-scale detection system covering targets from small to very large.
[0038] The prediction head is responsible for densely sampling and generating the original bounding box coordinates and class probability distributions on the multi-scale feature map after feature fusion, providing high-density candidate outputs for subsequent detection. To address the inference latency and accuracy loss caused by non-maximum suppression (NMS) in traditional object detection models, YOLOv10 introduces a consistency dual-process matching mechanism. This mechanism, belonging to the matching and loss layers, is a key strategy for achieving NMS-free performance. During the training phase, it intelligently assigns positive and negative sample labels to the massive number of candidate boxes output by the prediction head through dual-path consistency constraints, ensuring high confidence and localization accuracy of positive samples while suppressing redundant predictions. This allows for high-precision and high-efficiency detection performance without the need for post-processing NMS. Based on this, the roller-compacted concrete dam surface resource identification model introduces a scale-aware dual-label allocation optimization strategy during training to optimize the matching process in the roller-compacted concrete dam surface scenario.
[0039] The scale-aware dual-label allocation optimization strategy is an improvement on the original label allocation mechanism of the YOLOv10 model. It dynamically adjusts the number of positive candidate boxes assigned to the ground truth bounding box based on its size. Specifically, it allocates more positive samples to large-scale targets such as large machinery to ensure their core features are fully learned; simultaneously, it ensures high-quality positive samples are assigned to small-scale targets such as construction workers to prevent them from being ignored during training. For example, the label allocation logic in the training code is modified by adding a judgment logic: if the area of the ground truth bounding box is greater than 10% of the total image area, the number of its positive candidate boxes increases by 30%; if it is less than 0.1%, it is given a higher matching priority during matching.
[0040] Furthermore, the training strategy for the roller-compacted concrete dam surface resource identification model consists of two steps: pre-training and transfer learning. First, pre-trained weights from the YOLOv10 model on a large public dataset (such as COCO) are loaded to enable the model to have general feature extraction capabilities. Then, the dedicated roller-compacted concrete dam surface resource dataset constructed in step S3 is used for transfer learning and fine-tuning of the model. For example, the training parameters are: batch size set to 16, initial learning rate set to 1e-3, AdamW optimizer used, and training epochs set to 300 epochs, dynamically adjusted during training, and early stopping is used to prevent overfitting.
[0041] In one possible implementation, preferably, step S5 involves the following process: using the roller-compacted concrete dam surface resource identification model trained in step S4 (this model can be deployed to edge computing devices, such as the NVIDIA Orin series, to achieve real-time detection and result visualization at the construction site), inference is performed on the real-time video stream of the roller-compacted concrete dam surface to identify the target's category, confidence level, location coordinates, etc. The identification results are then post-processed using a spatiotemporally correlated target tracking algorithm to obtain structured resource information, including: the quantity of various resources, real-time location, dwell time, motion trajectory, etc., which can be displayed to users or upper-level management systems. The spatiotemporally correlated target tracking algorithm can employ algorithms such as Deepsort. During the inference phase, it is responsible for spatiotemporally correlated with the detection boxes output by the model, tracking each target, calculating its velocity, and simultaneously associating the same target in consecutive frames to generate its complete motion trajectory. Accordingly, this invention can also provide early warnings for engineering safety conditions and monitor potential safety hazards in real time (such as personnel getting too close to dangerous machinery). For example, when the present invention detects that a construction worker has entered within 5 meters behind a vibratory roller that is in operation, it triggers an audible and visual alarm and broadcasts a warning message to the site via a loudspeaker.
[0042] Based on the principles and processes of the above methods, this invention also proposes a resource identification system for roller-compacted concrete dam surfaces based on YOLOv10, which mainly includes a data acquisition module, an image preprocessing module, a dedicated dataset construction module for dam surface resources, a model construction and training module, and a resource identification module.
[0043] The data acquisition module acquires real-time video streams of the roller-compacted concrete dam surface through image acquisition devices deployed at the construction site. These image acquisition devices consist of several industrial-grade network cameras, arranged on the cableway system above the roller-compacted concrete dam surface and at high points on both sides of the mountain.
[0044] The image preprocessing module deframes the video stream acquired by the data acquisition module, performs standardization processing on each frame sequentially, and uses an adaptive environment enhancement algorithm to enhance the environment of each standardized frame to reduce the interference of strong light, shadows, fog or dust at the construction site on image quality, ultimately obtaining the enhanced warehouse surface image.
[0045] The module for constructing a dedicated dataset for roller-compacted concrete dam surface resources defines target recognition categories for roller-compacted concrete dam surface resources. These categories must include at least bulldozers, vibratory rollers, pavers, transport trucks, and construction workers. Based on these categories, the enhanced surface images from the image preprocessing module are finely annotated to obtain annotated images, thus completing the construction of the dedicated dataset for roller-compacted concrete dam surface resources. During the construction of the dedicated dataset, annotated images under different external environmental conditions are simultaneously collected to expand the content of the dedicated dataset for roller-compacted concrete dam surface resources.
[0046] The model building and training module uses a YOLOv10 model to build and train a deeply optimized model for identifying roller-compacted concrete dam surface resources, based on a dedicated dataset of roller-compacted concrete dam surface resources.
[0047] The resource identification module employs a trained resource identification model for the roller-compacted concrete dam surface to infer information from the real-time video stream of the roller-compacted concrete dam surface. It then combines this model with a spatiotemporal correlation target tracking algorithm to post-process the inference results, thereby obtaining structured resource information.
[0048] Based on the above principles and processes, the present invention can also propose an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-mentioned method for identifying resources on the surface of roller-compacted concrete dams based on YOLOv10.
[0049] Based on the above principles and processes, the present invention can also propose a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-mentioned method for identifying resources on the surface of roller-compacted concrete dams based on YOLOv10.
[0050] Although the functions and working processes of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific functions and working processes described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these are within the protection scope of the present invention.
Claims
1. A method for resource identification of roller-compacted concrete dam surface based on YOLOv10, characterized in that, Includes the following steps: S1 Data Acquisition: Real-time video stream of the roller-compacted concrete dam surface is acquired through image acquisition devices deployed at the construction site; S2 Image Preprocessing: After deframing the video stream described in step S1, each frame of the image is normalized and then enhanced with an environment to obtain the enhanced surface image. Construction of S3 Dedicated Dataset for Compacted Concrete Dam Surface Resources: Define target recognition categories for roller-compacted concrete dam surface resources, and perform fine annotation on the surface images obtained in step S2 according to the defined categories to obtain annotated images, thus completing the construction of the dedicated dataset for roller-compacted concrete dam surface resources; S4 Model Construction and Training: Based on the dedicated dataset of roller-compacted concrete dam surface resources constructed in step S3, a deeply optimized roller-compacted concrete dam surface resource identification model was constructed and trained using the YOLOv10 model. S5 Resource Identification: Using the resource identification model of the roller-compacted concrete dam surface trained in step S4, inference is performed on the real-time video stream of the roller-compacted concrete dam surface, and the inference results are post-processed by the spatiotemporal correlation target tracking algorithm to obtain structured resource information.
2. The method for resource identification of roller-compacted concrete dam surface based on YOLOv10 according to claim 1, characterized in that, The specific process of step S1 is as follows: Several industrial-grade network cameras are deployed on the cableway system above the roller-compacted concrete dam surface and on the high points of the mountains on both sides to obtain real-time video streams of the roller-compacted concrete dam surface.
3. The method for resource identification of roller-compacted concrete dam surface based on YOLOv10 according to claim 1, characterized in that, The specific process of step S2 is as follows: After deframing the video stream described in step S1, each frame image is standardized, including size unification and pixel value normalization. An adaptive environment enhancement algorithm is then used to perform illumination compensation and noise suppression on each standardized frame image to obtain an enhanced surface image.
4. The method for resource identification of roller-compacted concrete dam surface based on YOLOv10 according to claim 1, characterized in that, The target identification categories for the roller-compacted concrete dam surface resources mentioned in step S3 include at least bulldozers, vibratory rollers, pavers, transport trucks, and construction workers; during the construction of the dedicated dataset for surface resources mentioned in step S3, labeled images under different external environmental conditions are collected simultaneously to expand the content of the dedicated dataset for roller-compacted concrete dam surface resources.
5. The method for resource identification of roller-compacted concrete dam surface based on YOLOv10 according to claim 1, characterized in that, The roller-compacted concrete dam surface resource identification model described in step S4 is built based on the existing YOLOv10 model. An adaptive feature enhancement module is embedded at the output of each PAN path of the neck network of the YOLOv10 model. The prediction head of the YOLOv10 model adopts a multi-scale collaborative prediction head structure. The training strategy of the roller-compacted concrete dam surface resource identification model is divided into two steps: pre-training and transfer learning. First, the weights of the YOLOv10 model pre-trained on a large public dataset are loaded. Then, the model is transferred and fine-tuned using the roller-compacted concrete dam surface resource-specific dataset constructed in step S3. In addition, a scale-aware dual-label allocation optimization strategy is also adopted in the training process of the roller-compacted concrete dam surface resource identification model.
6. The method for resource identification of roller-compacted concrete dam surface based on YOLOv10 according to claim 5, characterized in that, The adaptive feature enhancement module employs a lightweight contextual attention module, which aggregates multi-scale contextual information, enabling the model to better distinguish between the target and the background with similar textures, and adaptively enhances effective target features while suppressing noise features. The multi-scale collaborative prediction head structure adds a P2-level detection head specifically for small targets to the original prediction head of the YOLOv10 model. This detection head fuses high-resolution feature maps from shallower layers of the backbone network and works in conjunction with detection heads of other sizes to form a full-scale detection system covering targets from small to very large.
7. The method for resource identification of roller-compacted concrete dam surface based on YOLOv10 according to claim 5, characterized in that, The scale-aware dual-label allocation optimization strategy is an improvement on the original label allocation mechanism of the YOLOv10 model. It dynamically adjusts the number of positive candidate boxes assigned to the actual target box based on the size of the target box.
8. A resource identification system for roller-compacted concrete dam surface based on YOLOv10, characterized in that, It includes a data acquisition module, an image preprocessing module, a dedicated dataset construction module for warehouse resources, a model building and training module, and a resource identification module; among which, Data acquisition module: Acquires real-time video streams of the roller-compacted concrete dam surface through image acquisition devices deployed at the construction site; wherein, the image acquisition devices consist of several industrial-grade network cameras, arranged on the cableway system above the roller-compacted concrete dam surface and at high points on both sides of the mountain. Image preprocessing module: After defraction of the video stream acquired by the data acquisition module, each frame of the image is standardized sequentially, and an adaptive environment enhancement algorithm is used to enhance the environment of each standardized frame of the image to obtain the enhanced warehouse surface image. The dedicated dataset construction module for roller-compacted concrete dam surface resources defines target recognition categories for roller-compacted concrete dam surface resources. These categories must include at least bulldozers, vibratory rollers, pavers, transport trucks, and construction workers. Based on these categories, the enhanced surface images from the image preprocessing module are finely annotated to obtain annotated images, thus completing the construction of the dedicated dataset for roller-compacted concrete dam surface resources. During the construction of the dedicated dataset, annotated images under different external environmental conditions are simultaneously collected to expand the content of the dedicated dataset for roller-compacted concrete dam surface resources. Model building and training module: Based on a dedicated dataset of roller-compacted concrete dam surface resources, a YOLOv10 model is used to build and train a deeply optimized roller-compacted concrete dam surface resource identification model; Resource identification module: The trained roller-compacted concrete dam surface resource identification model is used to infer the real-time video stream of the roller-compacted concrete dam surface, and the inference results are post-processed by the spatiotemporal correlation target tracking algorithm to obtain structured resource information.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for identifying resources on the surface of a roller-compacted concrete dam based on YOLOv10 as described in any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for identifying resources on the surface of a roller-compacted concrete dam based on YOLOv10 as described in any one of claims 1 to 7.